This module teaches FX from zero, but goes deep enough for a professional
It assumes no prior knowledge of trading and builds every idea from its foundations. Yet it goes deep enough to be useful to someone who already prices, hedges or builds systems for a currency desk.
Wherever a concept is technical, we stop and explain it in plain language before using it again. Wherever a term is jargon — pip, skew, last look, internalisation, PvP — we define it once, clearly, and then rely on it.
Two threads run through everything: the business and the technology
The business thread covers what FX trading is as a commercial activity. What a currency pair is, why corporates, funds and central banks trade, how a market maker quotes a two-way price and takes the other side, how thin spreads on enormous volume become a profit, and where the risk sits and how it is tamed.
The technology thread covers the machine that now runs the market. How prices stream by the microsecond, how a pricing engine skews and tiers a quote, how flow is netted internally before anything touches the market, how a trade is confirmed and settled through CLS, and how data quality at one step decides whether the next five work.
These two threads are not separate. FX is the market where the process and the platform became the same thing first and most completely. More than in any other asset class, the trading capability is software. You cannot understand the business without the technology, and the technology makes no sense without the business it serves.
How to read this module
Read it as a narrative, not a reference card
The module runs in five parts, following the ECM Module 1 structure. Part 1 sets the context: what FX is, where the desk sits, and the family of products. Part 2 walks the end-to-end trade flow and shows where the price actually comes from. Part 3 maps the actors and the people. Part 4 lays out the systems, data and technology front to back. Part 5 covers the risks, the real failures, and the transformation landscape. Each part assumes you have read the ones before it. If you are short on time, read the section titles and the "In a nutshell" boxes first, then come back for the detail. The titles are written to carry the argument on their own.
A suggested reading path across three sessions
This is a long document, and it is not meant to be read in one sitting. A workable path is three sessions: Part 1 and Part 2 first, to fix the market and the trade flow; then Part 3 and Part 4, for the actors and the systems; then Part 5, for the risks and where the business is heading. Each part closes with a practice box, an explain-it-back prompt and a myths box; do the practice before checking the Answer Key at the back. The cheat sheet, answer key and glossary at the end are reference material to return to, not to read straight through.
Each coloured box plays a fixed role
Knowing what each box signals lets you skim or dive as you choose.
Key insight distils the one idea from a section worth remembering even if you forget everything around it. Going deeper adds nuance for readers who already have the fundamentals. IT & data lens connects a business point to the systems and data that carry it. Watch out flags a trap, a conduct risk or a common misunderstanding. Worked example follows a concrete, illustrative trade. Europe / France in focus draws out how the European and French market, its venues and its rules differ. In a nutshell recaps a part. Check your understanding lets you test yourself before moving on. Practice gives you the calculations and judgements the desk actually makes; Explain it back asks you to teach an idea in your own words; and Myth vs reality dismantles a confident misconception. Worked solutions to the practice boxes are collected in the Answer Key at the back.
What you will master
By the end of this module you will be able to do five things. Each maps to one part of what follows.
You will understand the FX market and its products end to end
You will know why every FX trade is simultaneously a purchase of one currency and a sale of another, what spot, forwards, swaps, NDFs, cross-currency swaps and options are and how they relate, and how a market maker earns a thin spread on vast volume rather than by betting on direction.
You will understand where the price actually comes from
You will follow a trade from a streamed quote through execution, risk and settlement, and know how a pricing engine computes, skews and tiers thousands of quotes a second, why internalising flow before hedging externally is the single biggest driver of profit, and how algos and TCA make and measure execution.
You will be able to map every actor and role around a trade
You will know who provides liquidity and who takes it, how a fund reaches top-tier prices through a prime broker's credit, how the different client types shape the way a desk prices, and how the desk's own roles shifted from voice traders to quants, engineers and e-commerce owners as the market automated.
You will be able to trace an FX trade front to back through the systems
You will follow a trade through the pricing, aggregation, execution, risk, confirmation and settlement layers, into CLS and out to regulatory reporting. You will understand why FX has no single tape, why an error early in that chain reaches every system after it, and how FIX and SWIFT/ISO 20022 split the front and back office.
You will recognise the risk, conduct and transformation landscape
You will know the risks a currency desk runs — including the settlement risk that is uniquely acute in FX — the voluntary code and product-by-product rules that govern a largely unregulated market, the real cases where FX went badly wrong, and where better data and disciplined AI genuinely change the economics.
1Context and the Big Picture
Before any trade makes sense you need the map: what a currency pair is, where the desk sits inside the bank, and the family of products — from spot to options — that everything else builds on.
FX trading is buying one currency by selling another
Every foreign-exchange trade is a swap of one money for another, and that symmetry changes how you have to think. When you buy Apple stock, you give up dollars and receive something that is not dollars — a share. When you buy euros, you give up dollars and receive another currency that is itself a form of money. There is no "asset" on one side and "cash" on the other. Both sides are cash, just denominated differently. That single fact — that you are always long one thing and short another at the same instant — is the reason FX has its own vocabulary, its own quoting conventions, and its own peculiar risks. You never simply "own euros" in a vacuum; you own euros against something, most often dollars.
Because both legs are money, an FX position is a relative bet, not an absolute one. A trader who is "long EUR/USD" is betting the euro will strengthen relative to the dollar. If both currencies rise against everything else in the world but move together, the position makes nothing. If both fall together, it still makes nothing. Only the gap between them matters. This relative nature runs through everything that follows — the pricing, the products, the risk. Hold on to it, because it is the hinge that makes the rest of the market make sense.
An FX price is always a ratio, never an absolute value
An FX rate is the price of one currency measured in units of another, so it is a ratio and only a ratio. The number 1.0850 for EUR/USD does not mean euros are "worth 1.0850" in any absolute sense. It means one euro is worth 1.0850 US dollars right now. Flip the pair and you get the same information upside down: if one euro costs 1.0850 dollars, then one dollar costs 1 divided by 1.0850, roughly 0.9217 euros. The two statements are identical. There is no such thing as the price of a currency on its own, only its price against some other currency you have chosen as the yardstick.
This is why FX is quoted in pairs and why the choice of yardstick is a convention, not a law of nature. Markets could just as easily quote USD/EUR as EUR/USD, and the number would simply invert. The market has settled on conventions — which currency goes first, which goes second — largely by history and habit rather than logic. But once you know the convention for a pair, you know exactly what a rising or falling number means, and you can convert any amount in one currency into the other with a single multiplication or division. Everything downstream — spreads, forward points, profit and loss — is just arithmetic performed on this ratio.
The quote convention decides what "buy" even means
A currency pair is written as base/quote, and the base is the currency you are counting. In EUR/USD, the euro is the base currency and the dollar is the quote (also called the terms or counter) currency. The rate tells you how many units of the quote currency it takes to buy one unit of the base. So EUR/USD 1.0850 reads directly as "one euro costs 1.0850 dollars." When someone says they buy EUR/USD, they mean they buy the base — euros — and pay with the quote — dollars. Selling EUR/USD is the reverse: sell euros, receive dollars. The base currency is always the thing you are buying or selling; the quote currency is always the money you pay or receive.
Dealers never show a single price. They show two: a bid and an offer. The bid is the price at which the dealer will buy the base currency from you. The offer (also called the ask) is the price at which the dealer will sell the base currency to you. Because the dealer buys low and sells high, the offer is always above the bid. A dealer might quote EUR/USD as 1.0849 / 1.0851 — willing to buy euros at 1.0849 and sell euros at 1.0851. The gap between the two, here two-hundredths of a cent, is the spread. That spread is the dealer's raw compensation for standing ready to trade either way, and it is the first of the revenue sources we will unpack later. Notice the asymmetry from the client's point of view: you always cross the spread. You buy at the dealer's offer and sell at the dealer's bid, so the round trip costs you the full width of the spread even if the market never moves.
The pip is the standard unit for measuring how far a price has moved, and in the major pairs it is the fourth decimal place. A pip — the term comes from "percentage in point" — is the smallest conventional increment in a currency quote. For most majors quoted to four decimals, one pip is 0.0001. Moving from EUR/USD 1.0850 to 1.0851 is a one-pip move. Many dealers now quote a fifth decimal, a fractional pip or "pipette," to compete on price, so you will see 1.08505. The Japanese yen is the famous exception: because USD/JPY trades around 150 rather than around 1, it is quoted to two decimals, and a pip there is 0.01. The pip is not just jargon; it is how spreads, moves and profit-and-loss are all sized and communicated on the desk.
The "big figure" is the part of the price traders assume you already know and therefore drop. If EUR/USD is trading around 1.0850, the "08" is the big figure — the handle that changes slowly and is common knowledge on the floor. When a trader shouts a two-way price, they usually quote only the last two digits: "50 / 52" means 1.0850 / 1.0852, with the 1.08 understood. This shorthand exists because in fast markets nobody has time to repeat digits that are not moving. It also means that when the big figure does change, it is a genuine event — a break to a new handle — and everyone on the desk registers it. Understanding the big figure convention is the difference between reading a dealer's quote correctly and misreading it by a hundred pips.
Majors, crosses and emerging-market pairs differ in depth, not in kind
Every FX pair works the same way, but they are not equally deep, and depth is what separates the majors from everything else. The majors are the currency pairs that trade the US dollar against the other most heavily traded currencies: the euro, the yen, the British pound, the Swiss franc, the Canadian, Australian and New Zealand dollars. These pairs carry the most volume, the tightest spreads and the most continuous liquidity, because the dollar sits on one side of the overwhelming majority of the world's FX activity. When you can trade billions with a spread of a fraction of a pip, you are almost certainly in a major.
A cross is any pair that does not involve the US dollar, and historically it was literally built by crossing two dollar pairs. EUR/JPY, EUR/GBP and GBP/JPY are crosses. Because most liquidity historically pooled against the dollar, a bank pricing EUR/JPY would derive it from EUR/USD and USD/JPY — buy euros with dollars, sell those dollars for yen — and "cross" the two legs. Today the most active crosses trade directly in their own right with deep liquidity, but the term survives, and the mechanics still explain why a thinly traded cross can be wider than either of its dollar legs: you may be paying two spreads instead of one.
Emerging-market pairs trade on the same principles but are thinner, more volatile and often legally restricted. Pairs involving currencies such as the Brazilian real, Indian rupee, Chinese renminbi or Korean won carry wider spreads, less continuous liquidity and larger jumps. Some of these currencies are non-convertible or restricted: local rules limit who may hold or move them across borders, so you often cannot simply deliver the currency the way you would deliver euros. That restriction is exactly why a whole product — the non-deliverable forward — exists, and we will meet it below. The point to carry forward is that a real from São Paulo and a euro from Frankfurt are priced by the same ratio logic; they differ in how deep the pool is and whether you are even allowed to swim in it.
FX is an over-the-counter, decentralised, tiered market with no exchange and no tape
The single most important structural fact about FX is what it lacks: there is no exchange, no central order book and no consolidated record of trades. Understanding what fills that void — bilateral dealing, fragmentation and a tiered hierarchy of access — is the key that unlocks how the whole market behaves.
Over-the-counter means bilateral, not a public order book
FX is an over-the-counter market, which means trades happen bilaterally between two named parties rather than on a central exchange. Over-the-counter, or OTC, describes any market where a buyer and a seller agree a price directly with each other rather than by matching anonymous orders in a single public venue. When you buy a share, your order goes to an exchange, matches against a resting order from someone you will never know, and prints to a public tape that everyone can see. FX works nothing like that. When you trade EUR/USD, you trade against a specific counterparty — usually a dealing bank — on a price that counterparty has quoted specifically to you. There is no exchange sitting in the middle guaranteeing the trade, and there is no public record that the trade ever happened.
The market is quote-driven, not order-driven, and that distinction shapes everything. In an order-driven market, prices emerge from resting buy and sell orders that anyone can hit — the order book is the price. In a quote-driven market, dealers make prices and you trade on those prices; the dealer is obliged to quote but the price is theirs, tailored to who you are and what they know. FX is overwhelmingly quote-driven. You do not place an order into a book and wait; you ask a dealer for a two-way price, or you take a price the dealer is already streaming to you, and you deal on it or you do not. This is why relationships, credit and information matter so much in FX — you are always trading a specific dealer's specific price, not an anonymous public one.
Because the market is decentralised, there is no single price and no consolidated tape. Decentralised means the liquidity is spread across dozens of banks, electronic venues and platforms rather than concentrated in one place. There is no equivalent of the equity market's consolidated tape that stamps every trade with a time and a price for all to see. What you get instead are many overlapping streams of quotes from many providers, none of them official, none of them complete. The market runs roughly around the clock as the trading day rolls from Sydney to Tokyo to London to New York and back, and at any instant the "price" of EUR/USD is really a cloud of slightly different prices from slightly different sources. There is no referee. That absence is not a bug the market has failed to fix; it is the defining condition every participant has to engineer around.
Fragmentation is the market's defining feature and its central engineering problem
Fragmentation means the same pair can trade at genuinely different prices in different places at the very same instant. Because liquidity is scattered across many venues and many dealers, and because no rule forces those prices to line up, EUR/USD on one bank's platform can differ from EUR/USD on another venue at the same millisecond. In equities, regulation and a central tape push prices toward a single reference. In FX, nothing does. The market is a patchwork of pools, each with its own depth, its own participants and its own slightly different price. This is the structural reality that makes FX both frustrating and profitable to trade well.
Fragmentation creates the market's central engineering problem: how do you see the true price, and how do you get the best one? If prices differ across venues, then a participant who can look at many venues at once, compare them and route each order to the best available price holds an enormous advantage over one who sees only a single stream. That problem — collecting scattered quotes, cleaning them, and building a single coherent view of the market — is called aggregation, and solving it is one of the great technical achievements of modern FX. It is also the foundation of how dealers make money by internalising flow rather than paying away spread to hedge. We foreshadow it here and unpack it fully in Part 2, where electronic FX takes centre stage. For now, fix the idea: fragmentation is the problem; aggregation and internalisation are the answers the industry built.
The market is tiered by credit and size
Access to the best prices in FX is not open to everyone; the market is arranged in tiers, and where you sit depends on your credit and your volume. Picture a set of concentric rings. At the centre sit the largest global dealing banks — the Tier 1 interbank dealers — who trade enormous size with one another at the tightest prices anywhere. These banks deal directly and through primary electronic venues, sometimes called interbank ECNs (electronic communication networks), which are the wholesale marketplaces where the biggest players show and take liquidity. This inner core is where the deepest, tightest, most continuous pricing lives, and you cannot simply buy your way in; you need the credit standing and the trading relationships to be admitted.
Moving outward, the dealer-to-client tier is where banks price their customers, and prime brokerage is the bridge that lets smaller players borrow a bank's access. In the dealer-to-client tier, a bank quotes prices to its corporate, institutional and hedge-fund clients, typically wider than interbank prices because the bank is providing a service and taking on risk. Below the largest institutions, many funds and trading firms reach top-tier liquidity through prime brokerage: a large bank lends its name and credit standing so that a smaller firm can trade in the bank's name across many venues, with the bank guaranteeing settlement and managing the credit. Prime brokerage is the mechanism that lets a mid-sized hedge fund trade as if it were a global bank, for a fee.
At the outer edge sit retail traders and the aggregators that bundle them, whose small tickets are netted and pushed upstream. Retail participants — individuals trading through online platforms — deal in small sizes but enormous numbers. Retail brokers and aggregators collect these thousands of tiny trades, net them against each other, and pass the residual risk up to wholesale dealers. A retail trader never touches Tier 1 liquidity directly; their trade is wrapped, netted and relayed inward through several layers. The whole structure is a hierarchy of credit and size: the more creditworthy and higher-volume you are, the closer to the tight interbank core you can trade, and the better your prices. This is not a matter of fairness; it is a matter of who can be trusted to settle billions and who trades enough to be worth the tightest price.
The FX desk sits inside Markets and lives on flow, not directional bets
The FX desk is part of a bank's Global Markets business — the wholesale trading arm — and specifically its FICC division. FICC stands for Fixed Income, Currencies and Commodities, the part of an investment bank that trades bonds, rates, credit, currencies and commodities for clients. Within FICC, the currencies desk (often just "FX" or "the currency desk") sits between two worlds. On one side are the bank's clients — corporates, asset managers, funds — who reach the desk through sales teams and increasingly through electronic platforms. On the other side is the wholesale interbank and venue market, where the desk lays off, or offsets, the risk it takes on from clients. The desk is the pivot between these two worlds: it absorbs client demand on one side and recycles the resulting risk on the other.
The desk makes money on spread and volume, not on correctly guessing which way the market will go. This is the single most important thing to understand about the FX business and the reason it is different from a proprietary trading operation. A well-run FX desk is a flow business: it earns the spread between what it pays clients who sell and what it charges clients who buy, and it does this across huge volumes many times a day. It is not trying to hold a big directional view and profit from the market moving its way — that would be a bet, and bets are volatile and capital-hungry. The desk wants to be roughly flat on direction and long on flow. Getting the direction right is a bonus and, under modern risk limits, a small one; capturing spread on volume, over and over, is the actual business.
FX sits underneath every other asset class, because almost every cross-border trade in anything carries a currency leg. When a European pension fund buys US Treasuries, it must first buy dollars. When a Japanese investor buys German equities, a yen-into-euro trade rides along. When a corporate books revenue abroad, that revenue eventually has to come home in the home currency. This means FX is not a niche product sitting beside equities and bonds; it is the settlement layer beneath them. Foreign equity and bond trades almost always generate an FX trade, so the currency desk sees flow that originates in every other part of the firm and every other market in the world. That structural position — under everything, touching everything — is why FX is the largest and most continuous market on earth, a scale we will quantify at the end of this part.
The product family is one idea repeated: spot moved through time, then volatility added
The entire FX product range grows from a single seed: the spot trade. A spot transaction is the immediate exchange of one currency for another at today's rate, settling in the market convention of two business days. Everything else in the family is that same exchange, moved, stretched, or bent. Once you see spot as the atom, the rest of the products stop looking like a menu of unrelated instruments and start looking like variations on one theme. This is the most useful mental model a newcomer can carry onto a trading floor.
Move spot forward in time and you have a forward. A forward is an agreement to exchange the two currencies on a chosen future date at a rate agreed today. Nothing about the mechanics changes except the settlement date; the price differs from spot only by the forward points, which are the arithmetic of the two currencies' interest rates over the period, not a forecast. Chain two dated exchanges together — buy a currency now and agree to sell it back later — and you have an FX swap, the market's principal tool for short-term funding and for rolling positions from one date to another. An FX swap is not a new idea; it is two exchanges wearing one ticket.
Push the same logic into currencies that cannot be freely delivered and you get the NDF. A non-deliverable forward (NDF) is a forward on a restricted currency — one that capital controls prevent from being paid across borders freely — that never actually exchanges the two currencies. Instead, on the fixing date the two parties settle the difference between the agreed forward rate and an official reference rate, in a freely convertible currency such as the dollar. The economic exposure is identical to a forward; only the plumbing is cash-settled. Stretch the horizon from days to years, and let both sides pay interest to each other in the meantime, and the forward becomes a cross-currency swap, which we treat in full later in this part.
Every product named so far shares one crucial property: it is linear. A linear product moves one-for-one with spot. If the euro rises by one cent against the dollar, the value of a euro forward, an FX swap leg, or an NDF changes by a predictable, straight-line amount — double the spot move and you double the value change. This is why linear products can be risk-managed with a single number, the position size, and why they are the workhorses of hedging and funding.
An option breaks that straight line, and in doing so adds a genuinely new dimension. An option's payoff is non-linear: its value does not move one-for-one with spot, because the holder can walk away from a bad outcome and keep a good one. That asymmetry means an option's worth depends not only on where spot is, but on how much spot is expected to move — its volatility. Volatility is a second axis of risk that no linear product carries, and it demands an entirely separate toolkit to price and to hedge. The two big additions that follow — options and cross-currency swaps — are where the family stops being simple arithmetic on spot and starts to require its own machinery.
Spot is the core, and everything else is spot moved through time
There is really only one FX product, and it is spot — the immediate exchange of two currencies. Every other product is spot with a time dimension bolted on: a forward is spot pushed into the future, a swap is two spots at different dates, an NDF is a forward that cash-settles. Learn spot cold and the rest is arithmetic and plumbing.
Spot: the immediate exchange, settling T+2, that prices everything else
Spot is the exchange of two currencies at today's price for near-immediate settlement, and it is the reference from which every other FX product is built. The spot rate is the price for exchanging currencies "now." But "now" in FX has a precise meaning: settlement — the actual movement of the two currencies between accounts — normally happens two business days after the trade is agreed. This convention is written T+2, meaning trade date plus two business days. The two-day lag exists for a practical reason: currencies settle through banking systems in different countries and time zones, and moving real money between banks in, say, Frankfurt and New York takes time to instruct, confirm and clear. A handful of pairs settle faster; USD/CAD and a few others settle T+1, one business day, because the two banking systems are close enough to make it work.
Spot carries the deepest liquidity and the tightest spreads, and it is where electronic trading first took over. Because spot is the simplest product — no interest-rate maths, no future date to reason about — it is the most heavily traded, the most standardised and the easiest to automate. It was the natural place for electronic FX, or e-FX, to begin: streaming prices, click-to-trade platforms and algorithmic pricing engines all started with spot and spread outward from there. When we talk in Part 2 about pricing engines computing bids and offers in milliseconds, spot is the product they cut their teeth on. For now, hold the mental model: spot is the anchor. Every forward, swap and NDF price starts from a spot price and then adjusts it for time.
Outright forwards let a hedger lock a future rate
An outright forward is an agreement to exchange two currencies at a fixed rate on a specific future date beyond spot, and it is the classic corporate hedge. An outright forward — usually just "forward" — locks in today the exchange rate for a transaction that will settle weeks, months or years from now. Suppose a company knows it will need to convert dollars into euros in three months. Rather than wait and gamble on where the rate will be, it agrees today a fixed rate for that future exchange. When the date arrives, the two currencies change hands at the agreed rate no matter where spot has moved. The company has traded away uncertainty for certainty — it now knows exactly what it will get, which is the whole point.
The forward is the tool for hedgers, and it works because certainty is worth paying for. A corporate treasurer is not trying to profit from currency moves; they are trying to make the currency move irrelevant to the business. A forward does exactly that. It converts an unknown future rate into a known one, so the company can price its products, plan its cash flows and report its results without a currency surprise. The cost of that certainty is that the company also gives up any favourable move — if spot goes their way, they are still locked into the agreed rate. For a hedger, that trade-off is not a loss; it is the definition of a hedge. You give up the upside of the gamble to be rid of the downside, because you are in the business of making things, not betting on currencies.
Forward points are arithmetic, not a market view
The forward rate is not a forecast of where spot will be; it is spot mechanically adjusted for the difference in interest rates between the two currencies. This is one of the most misunderstood ideas in FX, so hold it firmly: the forward rate does not embody anyone's opinion about the future. It is fixed by arithmetic through a relationship called covered interest parity. The logic is a no-arbitrage argument. If you hold euros for three months, you earn the euro interest rate. If instead you convert to dollars, hold dollars for three months and earn the dollar interest rate, you must end up with the same value once you convert back — otherwise there would be a free lunch, and traders would arbitrage it away instantly. For those two paths to be equal, the future exchange rate must differ from spot by exactly the gap between the two interest rates. That gap, expressed in pips, is the forward points.
Forward points are simply added to or subtracted from spot, and their sign follows which currency pays the higher interest rate. If the dollar interest rate is higher than the euro rate, then holding dollars earns more, so the forward rate must compensate the euro holder — the forward EUR/USD comes out below spot. If the euro rate were higher, the forward would sit above spot. The size of the adjustment grows with the length of time and the width of the rate gap. None of this involves a view on the currency; it falls straight out of the two interest rates and the number of days. A dealer computing a forward is doing arithmetic, not forecasting.
The FX swap is the biggest product by turnover: two legs, a funding tool
An FX swap is two exchanges in one deal — a near-dated exchange one way and a far-dated exchange back the other way — and it is the single biggest FX product by turnover. In an FX swap you agree, in one transaction, to exchange two currencies today (or at spot) and to reverse that exchange at a fixed date in the future. For example: buy euros / sell dollars for spot value, and simultaneously agree to sell those euros / buy back the dollars three months later. The two legs are the same size in the base currency, and the only difference between the near price and the far price is the forward points — the same interest-rate arithmetic we just met. You are not taking a directional position at all, because whatever spot does, you are buying and selling the same amount and the exposures cancel.
The FX swap is a funding and liquidity tool, not a bet on where currencies are going. Because the two legs cancel any directional exposure, an FX swap is really a way to borrow one currency and lend another for a fixed period. A bank that is short euros today but will be long euros later can use an FX swap to bridge the gap without ever taking a currency view. Corporates, banks and asset managers use swaps to manage cash across currencies, to roll hedges forward, and to fund positions in a currency they do not permanently want to hold. This is why the swap, dull as it sounds, is the largest product by turnover in the entire FX market: it is the plumbing that moves liquidity around the system, used constantly and in vast size. The forward and the spot get the attention; the swap does the work.
Tom-next and rolling let a desk carry a position without taking delivery
Short-dated FX swaps let a desk keep a spot position alive from one day to the next without ever settling the currencies, and the workhorse is the tom-next swap. Recall that spot settles T+2. If a trader holds a spot position and does nothing, in two days real currencies must actually move between accounts — delivery. Often the trader does not want delivery; they simply want to keep the position open. The solution is a tom-next swap, short for "tomorrow-next": a one-day swap that pushes the settlement date forward by a day. "Tom" is tomorrow, the next business day; "next" is the day after. By trading tom-next each day, the desk rolls the position forward, sliding the settlement date along day by day so the currencies never actually have to be delivered.
Rolling turns a settlement obligation into a continuous carry, and the daily forward points are the cost or benefit of holding on. Each roll applies one day's worth of forward points, which is just one day of the interest-rate differential. So a trader who holds a position financed by rolling either pays or earns a little each day depending on the rate gap — the carry. This mechanism is invisible to the outside world but constant on the desk: it is how a book stays open overnight, over weekends and across weeks without a mountain of physical settlements. It is spot plus the swap arithmetic, applied one day at a time, and it is another reason the swap is the busiest instrument in the market.
NDFs cash-settle the currencies you cannot deliver
A non-deliverable forward is a forward that never delivers the currencies; instead only the profit or loss versus an agreed fixing is paid, usually in dollars. Some emerging-market currencies are restricted or non-convertible — local rules prevent free cross-border delivery, so you cannot simply hand over the currency the way you would hand over euros. Yet investors and corporates still need to hedge or express views on these currencies. The market's answer is the non-deliverable forward, or NDF. An NDF fixes a forward rate today just like an ordinary forward, but at maturity nobody delivers the restricted currency. Instead, the agreed forward rate is compared with an official fixing — a published reference rate on the maturity date — and the difference in value is settled in cash, almost always in US dollars.
The NDF lets you take on the economic exposure of a currency you are not allowed to hold or move, by paying only the net result. Suppose an investor wants exposure to the Korean won but cannot freely deliver won across borders. They trade a USD/KRW NDF. At maturity, if the won has moved in their favour relative to the agreed rate, they receive the difference in dollars; if it has moved against them, they pay the difference in dollars. The won itself never crosses a border. This cash-settled design is what makes emerging-market hedging possible at all where the currency is restricted. The economics mirror a normal forward — you are long or short the currency — but the settlement mechanism sidesteps the delivery problem entirely. The NDF is the clearest example of the market engineering a product around a legal constraint.
| Product | What it is | Typical use | Settlement | Relative turnover |
|---|---|---|---|---|
| Spot | Immediate exchange of two currencies at today's rate | The reference price; near-term conversion; the anchor for all other products | T+2 (T+1 for a few pairs such as USD/CAD) | Very large, but smaller than swaps |
| Outright forward | Exchange at a fixed rate on a single future date beyond spot | Corporate and investor hedging of a known future flow | Physical delivery on the forward value date | Modest relative to swaps and spot |
| FX swap | A near-dated exchange plus a reversing far-dated exchange, same amount | Funding, liquidity management, rolling positions, moving cash across currencies | Two physical settlements, one at each leg | The largest single product by turnover |
| NDF | A forward that cash-settles the P&L versus a fixing rather than delivering | Hedging or expressing views on restricted, non-convertible EM currencies | Cash settlement, usually in USD, against a published fixing | Growing; concentrated in EM currencies |
FX options price volatility, not just direction
What an option is, from zero
An FX option is a contract that gives its buyer the right, but not the obligation, to exchange two currencies at a fixed rate. That fixed rate is the strike. The right can be exercised on a set date (a European option) or at any point up to it (an American option). A call is the right to buy the base currency; a put is the right to sell it. Because every FX rate is a ratio of two currencies, a EUR call is automatically a USD put — the right to buy euros with dollars is the same instrument as the right to sell dollars for euros — which is why FX options are always named as a pair.
The buyer pays for this right up front, and that payment caps the loss. The up-front payment is the premium. It is the most the buyer can ever lose: if the market moves against the option, the buyer simply lets it expire and forfeits the premium, nothing more. On the other side stands the seller, or writer, who collects the premium and accepts the mirror-image risk — an obligation to deliver at the strike if the buyer exercises, with a potential loss far larger than the premium received. The buyer's position is insurance; the seller's is underwriting.
A forward locks a rate; an option insures a rate. This is the cleanest way to feel the difference. A forward gives certainty: it costs nothing up front, but it removes both the downside and the upside — you will transact at the agreed rate whatever happens, even if the market has moved in your favour. An option keeps the upside: pay the premium, and you are protected against an adverse move while remaining free to abandon the contract and trade at a better market rate if one appears. The premium is the price of keeping that optionality. A treasurer choosing between the two is really choosing between a rate they must accept and a rate they may accept.
Who buys options and why
Corporates buy options when they want protection without surrendering a favourable move. Consider an illustrative European exporter expecting to receive USD 10 million in three months (figures illustrative). A forward would fix the conversion rate today and end the uncertainty. But the exporter believes the dollar might strengthen, which would let them convert at a better rate, and does not want to give that up. Buying a EUR call / USD put struck at, say, 1.10 protects the receivable if the euro rises above 1.10, while leaving the exporter free to convert in the open market if the dollar strengthens instead. The premium is the cost of that best-of-both position.
The premium can feel expensive, which is why the collar exists. Options do not come free, and treasurers often balk at the up-front cost. The popular answer is the zero-cost collar, also called a risk reversal: the exporter buys the protective EUR call / USD put and simultaneously sells a EUR put / USD call at a less favourable strike, using the premium earned from the sale to pay for the protection bought. The result costs nothing up front. The trade-off is that the sold option caps the exporter's benefit if the market moves in their favour — they have swapped unlimited upside for a free hedge, bounding their outcome inside a corridor between the two strikes. It is the most common structured hedge in corporate FX for exactly this reason.
The volatility surface: the market's price of uncertainty
The number that prices an option is not a rate but a volatility. Implied volatility is the options market's forecast of how much a currency pair will move over the option's life, expressed as an annualised percentage. It is called "implied" because it is backed out of the price the market is actually paying: a high premium implies the market expects large moves; a low premium implies calm. Volatility, not direction, is the raw material an options desk buys and sells. Two desks can agree exactly on where spot is heading and still disagree violently on what an option is worth, because they disagree on how bumpy the road will be.
But a single volatility number is a fiction — the market quotes a whole surface. The most-watched point is at-the-money (ATM) volatility, the implied vol for a strike near the current spot. Around it, two structures describe the shape. The risk reversal measures skew — whether the market is paying more for downside protection (puts) than upside (calls) or vice versa; a large risk reversal means fear is priced asymmetrically. The butterfly measures the smile — how much more the out-of-the-money "wings," the far strikes on either side, cost relative to the middle. Assemble ATM, risk reversal, and butterfly across every strike and every tenor (time to expiry) and you get a volatility surface: a three-dimensional map of the market's price of uncertainty. It is this surface, not a single figure, that an options desk quotes and risk-manages.
Delta hedging: why an options desk trades spot all day
An options desk wants to own volatility, not direction — so it continuously cancels its directional exposure. Delta measures how much an option's value moves for a small move in spot; a delta of 0.5 means the option gains roughly half a cent for every one-cent rise in spot. To strip out direction and be left holding pure volatility, the desk takes an offsetting position in spot equal to the option's delta. This is delta hedging. Because the option's payoff is curved, its delta changes as spot moves, so the hedge must be readjusted again and again throughout the day — buying spot as it rises, selling as it falls. This constant rebalancing is dynamic hedging, and it is why an options desk is one of the busiest traders of spot on the floor.
Two more Greeks explain the rhythm of the day. Gamma measures how fast delta itself changes — high gamma means the hedge must be adjusted often and sharply, which is profitable in a jumpy market and painful in a quiet one. Theta measures how much value the option loses simply from the passage of time, the daily "rent" an option holder pays as expiry approaches. A long-option desk is paid in gamma when markets move and pays in theta while they wait; a short-option desk collects theta and dreads gamma. The desk's real job is to manage this balance of Greeks — the sensitivities of the option's value — earning the spread between the volatility it sells and the volatility that actually materialises, while paying away time. Get the balance right and the desk earns the vol spread; get it wrong and time decay or a violent move eats the position.
Exotics extend the idea, and the risk
Beyond the plain "vanilla" option lies a large family of exotics that reshape the payoff. Barrier options switch on or off when spot touches a preset level: a knock-in comes to life only if the barrier is hit, while a knock-out is extinguished if it is — both are cheaper than a vanilla because the buyer accepts a chance of getting nothing. Digitals, or binaries, pay a fixed amount if a condition is met and zero otherwise, like a bet on whether spot finishes above a level. Bundle several options together and you can engineer structured products: target-redemption forwards (TARFs), which improve a hedging rate in exchange for the deal terminating once cumulative gains hit a target, and dual-currency deposits, which enhance a deposit's yield by embedding a sold option that may repay the investor in the weaker of two currencies.
Cross-currency swaps fund one currency with another over years
An FX swap funds for days; a cross-currency swap funds for years
The two instruments share a name and almost nothing else in scale. An FX swap is two near-dated exchanges of the same amount — buy euros today, sell them back in a week — with no interest changing hands beyond what is baked into the forward points. It is a short-term funding and liquidity tool, used to bridge a cash need or roll a position across dates. A cross-currency swap (XCCY) operates over years, not days, and it moves interest as well as principal. The family resemblance is real, but one is a weekend loan and the other is a mortgage.
A cross-currency swap exchanges principal at the start, interest throughout, and principal again at the end. At inception, the two parties exchange principal amounts in two currencies — say euros for dollars — at the prevailing spot rate. Over the life of the swap, each side pays interest to the other on the principal it received, in that currency: the party holding dollars pays dollar interest, the party holding euros pays euro interest, typically at floating rates. At maturity, the original principal amounts are re-exchanged at the same starting rate. This is precisely how an issuer that raised dollars in the bond market can turn them into dependable euro funding for years — or the reverse — without bearing the FX risk of doing it in the open market.
The cross-currency basis is the price of scarce dollar funding
In theory, funding one currency through another should be a wash. Covered interest parity is the principle that the cost of borrowing in one currency and swapping into another should exactly equal the cost of borrowing in the second currency directly — any gap would be a free profit that arbitrageurs would compete away. For decades this held closely. If it held perfectly, a cross-currency swap would carry no extra charge on either leg.
In practice a spread opens up, and it has a name: the basis. The cross-currency basis is an extra spread added to one leg of the swap when the demand to borrow a particular currency — most often the dollar — outstrips the supply of parties willing to lend it. A negative euro–dollar basis, for instance, means those seeking dollars must pay a premium to obtain them via the swap market. Three forces drive it. Post-2008 regulation made banks' balance sheets costlier to lend, shrinking the arbitrage that used to close the gap. Quarter-end and year-end reporting dates concentrate funding pressure into predictable spikes. And there is a persistent, structural global appetite for dollars from issuers, investors, and reserve managers. The basis is therefore not a pricing error waiting to be corrected — it is a live, closely watched market price of a real funding imbalance.
Who uses it and why it matters
The cross-currency swap sits at the junction of the world's funding needs. Bond issuers are its heaviest users: a company or bank that finds the cheapest borrowing in a foreign currency will issue there and swap the proceeds back to its home currency, locking in funding it could not get as cheaply at home. Banks and corporate treasuries use it to manage multi-currency funding books, matching assets in one currency against liabilities in another. Because the basis changes the true cost of every such transaction, it feeds directly into the pricing of long-dated forwards and into the cost of hedging any multi-year currency exposure.
The cross-currency basis is also the beating heart of a later Rates and Cross-currency module, where it is dissected alongside interest-rate swaps and curve construction. It appears here because this is the point where FX funding and interest rates first meet: the swap is an FX instrument in its principal exchange and a rates instrument in its interest streams. Understanding it as one object now makes the later, deeper treatment far easier.
The carry trade turns interest-rate differentials into a strategy
Borrow where rates are low, invest where they are high
The carry trade is the simplest speculative idea in FX, and one of the most powerful flows in the market. The carry trade means selling a low-yielding currency to buy a high-yielding one, and pocketing the interest-rate differential between them. Borrow in a currency paying near zero, convert, and hold a currency paying several percent, and — if the exchange rate stays put — you earn the gap. The differential does not appear as a coupon in the FX market; it appears as the forward points, the same arithmetic that prices every forward.
The forward "predicts" a move that often fails to arrive, and that gap is the profit. Because the high-yielding currency trades at a forward discount, the forward rate implies that it will weaken by exactly enough to cancel the interest advantage — that is what covered interest parity requires. But historically, high-yielding currencies have frequently not weakened as much as the forward implied, and sometimes strengthened instead. The carry trader is betting that this predicted depreciation will not fully materialise, and collects the differential when it does not. This single bet, repeated across many currencies and by many participants, is a major source of speculative flow in global FX.
The carry trade works until it violently does not
Carry is earned in a trickle and lost in a flood. The interest differential accrues slowly, day by day, which makes the strategy feel steady and safe for long stretches — returns that look almost like a bond coupon. The danger is that the trade is crowded: vast numbers of participants hold the same high-yield-long, low-yield-short positions at once. When stress hits and risk appetite reverses, everyone reaches for the exit simultaneously, selling the high-yielder and buying back the funding currency in a rush. The high-yielding currency collapses and the funding currency snaps higher, wiping out months of accumulated differential in days or hours. This is the liquidity and gap risk met earlier in this part, seen from the strategy side: the market that was deep and calm vanishes precisely when every carry trader needs it. The asymmetry — slow gains, sudden losses — is the defining signature of carry.
Currency overlay is FX risk management sold as a service
Asset managers hedge the currency risk of foreign assets
When an investor buys foreign assets, they buy two exposures, not one. A euro-based pension fund that holds US equities owns both the performance of those equities and the euro-dollar exchange rate, whether it wants the second exposure or not. Currency overlay is the practice of separating that FX risk from the underlying assets and managing it as its own dedicated programme, run independently of the stock or bond decisions beneath it. It lets the fund decide how much currency risk to keep quite apart from which assets to hold.
Overlay comes in two temperaments, passive and active. A passive overlay mechanically hedges a fixed proportion of the foreign exposure — commonly through hedged share classes, versions of a fund that roll forward contracts each period to strip out the currency move, leaving investors with close to the local-currency return of the assets. An unhedged share class does the opposite: it leaves the FX exposure in place, so investors feel every currency swing on top of the asset performance. An active overlay goes further, deliberately varying the hedge ratio — hedging more when the manager expects the foreign currency to weaken, less when it expects strength — in an attempt to add return rather than merely neutralise risk. Passive overlay is a discipline; active overlay is a view.
Why the bank provides it, and how it differs from the corporate hedge
The corporate and the asset manager hedge fundamentally different things. A corporate hedges a known commercial cash flow — an invoice, a receivable, a committed cost — a fixed amount on a fixed date. An asset manager hedges a fluctuating portfolio value measured against a benchmark, an exposure that drifts every day as markets move and that is judged not in absolute terms but by tracking error against an index. The corporate wants certainty on a number; the manager wants to match, or beat, a moving target. The hedging problem is continuous and rebalancing-heavy rather than one-and-done.
For the bank, overlay is a prized franchise, not a one-off trade. The desk supplies the execution, the rolling forwards and swaps, and sometimes designs and runs the entire overlay programme on the client's behalf. This is exactly the kind of client the taxonomy in the earlier client section flags as most valuable: recurring, benchmark-driven, high-volume flow that returns month after month as hedges roll, rather than a single transaction. Where the corporate relationship is episodic, tied to commercial events, the asset-manager overlay relationship is a standing mandate — which is why desks compete hard to win it.
The FX business is deeply cyclical, and the cycle drives the technology budget
Foreign-exchange revenue is not a steady annuity — it breathes with the market. A currency dealing desk earns money in two ways: it captures the bid-offer spread (the gap between the price at which it buys and the price at which it sells a currency) on every ticket, and it profits or loses on the inventory it holds while managing that flow. Both of those levers depend on the same underlying condition — volatility, the size and speed of price moves. When currencies move sharply, clients trade more to hedge and to reposition, turnover rises, spreads widen, and revenue swells. When currencies drift in narrow ranges, clients sit on their hands, turnover thins, spreads compress, and revenue shrinks. The business therefore moves in regimes — extended stretches of feast or famine — rather than around a stable mean.
The recent cycle makes the pattern concrete. Walk it in plain terms. In early 2020, the COVID shock detonated volatility across every asset class; FX volumes spiked, spreads gapped wider, and many banks booked record electronic-FX ("e-FX", meaning trading executed through automated platforms rather than by voice) revenue as clients scrambled to hedge and reprice risk. Through 2021, central banks held policy rates near zero and in lockstep; currencies traded in tight ranges, realised volatility fell to multi-year lows, and desk revenue thinned accordingly. In 2022 the picture flipped again: a synchronized global rate-hiking cycle, an energy shock, and a surging US dollar drove volatility and volumes back up, and trading revenue followed. Into 2023 and 2024, as inflation cooled and policy paths grew more predictable, conditions began to normalise — neither the panic of 2020 nor the torpor of 2021, but a middle regime. The specific magnitudes above are illustrative and indicative only; the shape of the cycle, not the exact figures, is the point.
The cycle has causes, and naming them explains the timing. Three forces set the tempo. The first is monetary-policy divergence: currencies are relative prices between two economies, so when central banks move in different directions — one hiking while another holds — the exchange rate between their currencies has to move, and it moves a lot. The second is macro shocks: wars, energy disruptions, pandemics and sudden shifts in growth reprice currencies violently and unpredictably. The third is risk sentiment: in calm periods capital flows smoothly and volatility stays low, while in stress episodes investors rush toward safe-haven currencies, amplifying moves. A desk earns more precisely when these forces are active — because clients trade more when they are uncertain, and spreads widen when the dealer's own risk of being caught on the wrong side rises.
Here is the point that matters for anyone building the platform: revenue is cyclical, but the platform is a fixed cost. An electronic-FX franchise is expensive infrastructure — pricing engines, connectivity to dozens of venues, risk and hedging systems, latency-sensitive networks, and the quantitative teams who tune them. That cost does not fall when volatility does. So a bank faces a structural temptation: build aggressively when revenue is fat, then cut engineers, freeze upgrades, and defer capacity when revenue thins. The problem is that platform quality compounds slowly and decays fast. A desk that cuts in the lean years arrives at the next volatility spike with stale pricing models, insufficient capacity, and a thinned team — and it loses share exactly when the market is paying the most for good execution. The rivals who kept investing capture that flow, and because flow is self-reinforcing (more flow means more data, tighter prices, and still more flow), the gap tends to become permanent rather than temporary.
The firms that win treat the e-FX platform as a through-cycle investment. They size the build to the average of the cycle rather than the peak, protect a core of engineering and quantitative capacity through the downturns, and use the quiet years — when the desk is not firefighting volume — to re-architect, test, and prepare for the next surge. This is a governance discipline as much as a technical one: it requires a management team willing to defend a stable technology budget against the pressure to flatter earnings in a slow year. The reward is asymmetric. Because volatility spikes are where the largest revenues are made, a platform that is ready when the spike arrives earns a return that a fair-weather platform structurally cannot.
For this audience, that reframes what a technology roadmap is for. A roadmap in a cyclical business is not a list of features to ship when times are good; it is a commitment to maintain a capability across conditions the business cannot control. When you are asked to justify platform spend in a soft revenue year, the argument is not "this will earn money now" — it often will not — but "this keeps us ready for the regime that pays, and readiness is the only thing that captures a volatility spike." Understanding the cycle is therefore not background colour. It is the frame that tells you when the platform earns its keep, why the temptation to cut is dangerous, and how to argue for the steady investment that separates the franchises that endure from the ones that are permanently rebuilding after every downturn.
Work these before reading the answer key at the back. They are the calculations and judgements a desk makes, not recall questions.
1.1 Forward
EUR/USD spot is 1.0850. The one-year EUR interest rate is 3.0% and the one-year USD rate is 5.0%. Compute the approximate one-year outright forward and the forward points. Which currency trades at a forward premium?
The forward is spot adjusted by the interest-rate differential: forward ≈ spot × (1 + USD rate) / (1 + EUR rate) = 1.0850 × 1.05 / 1.03 = 1.1061, so the forward points are +1.1061 − 1.0850 = +0.0211, or +211 pips. Because the USD rate is higher, the dollar trades at a forward discount and the euro trades at a forward premium — you receive more dollars per euro in a year than today. This is arithmetic (covered interest parity), not a forecast.
1.2 Cross rate
EUR/USD is 1.0850 and USD/JPY is 150.00. What is EUR/JPY?
EUR/JPY = EUR/USD × USD/JPY = 1.0850 × 150.00 = 162.75. The dollar cancels out; the cross is just the two dollar legs multiplied.
1.3 Pip value
You hold a long position of EUR 10,000,000 in EUR/USD. How much do you make, in USD, if the rate moves one pip in your favour?
One pip is 0.0001. EUR 10,000,000 × 0.0001 = USD 1,000 per pip. That is why a desk warehousing tens of millions watches every pip: the sums add up fast.
1.4 Product choice
A French exporter will receive USD 10m in six months, wants certainty of its euro proceeds, but would like to keep some benefit if the dollar strengthens. Name the two products that fit and state what the client pays for the optionality.
A forward locks the euro proceeds at zero up-front cost but surrenders any favourable move; an option (here a EUR put / USD call) keeps the upside for a premium paid up front — the premium is the price of the optionality, and the buyer's maximum loss. A zero-cost collar (buy the put, sell a call) removes the premium at the cost of capping the upside.
1.5 Spread cost
EUR/USD is quoted 1.0850 / 1.0852. You buy EUR 20,000,000 at the offer. How much more than the mid did you pay, in USD?
The mid is 1.0851. You buy at the offer, 1.0852 — one pip above mid. EUR 20,000,000 × 0.0001 = USD 2,000 above mid (half of the two-pip spread). That USD 2,000 is the desk's gross revenue on the trade before it hedges or internalises.
Explain to a friend who has only ever changed money at an airport bureau de change why a bank's FX desk shows two different prices for the same currency pair and earns its living on the gap between them — without using the words spread, bid, offer or market maker. Then push it: why can that gap be almost invisibly small for the desk and still make a fortune?
Why is every FX trade simultaneously long one currency and short another?
Because a price is a ratio of two currencies; buying the base is paying with, and therefore selling, the quote currency.
What is the spot rate, and why does it anchor every other FX product?
The immediate exchange rate settling T+2; forwards, swaps and options are all derived from it.
How does an option differ from a forward for a corporate hedger?
A forward locks a rate at no up-front cost but surrenders any favourable move; an option pays a premium to keep the upside while insuring the downside.
What is the cross-currency basis?
The extra spread on one leg of a cross-currency swap — a live deviation from covered interest parity that prices the scarcity of funding in one currency, usually the dollar.
What decides which tier of liquidity a participant can access?
Credit standing and trading size; top-tier interbank prices require the credit and volume to trade in the wholesale core, reached otherwise through D2C pricing, a prime broker or an aggregator.
Why is the FX desk called a flow business rather than a proprietary one?
It earns a thin spread by serving client flow and recycling the inherited risk, not by betting on direction.
1
Why is every FX trade long one currency and short another?
You always pay with another currency.
Spot trades are not swaps.
Right.
Hedging is a separate choice.
EUR/USD spot is 1.0850 and the forward points are +45. What is the forward?
The points are 0.0045, not 0.0450.
Right. With dollar rates above euro rates, the euro trades at a premium forward: spot plus the points.
That ignores the interest-rate differential.
That subtracts the points: the forward would be below spot, which contradicts the rate differential.
How does an option differ from a forward for a corporate hedger?
No up-front cost for a forward.
No.
That is a forward.
Right.
What does the cross-currency basis price?
No.
Right. A live deviation from covered interest parity.
No.
No.
2The End-to-End Trade Flow
The heart of the module. This part follows an FX trade from the moment a price is asked for to the moment it settles, and shows where the price actually comes from.
Every FX trade runs the same pipeline, much of it now in milliseconds
Every FX trade, from a retail click to a billion-dollar institutional order, runs through the same seven-step pipeline — and the first four steps can now happen faster than a human can blink. The steps are: request or stream, price, execute, risk and hedge, confirm and match, settle, and report and reconcile. What has changed over the past two decades is not the sequence but the speed of its front end. Pricing and execution, once a phone call and a scribbled ticket, now collapse into milliseconds of automated computation. But — and this is the point the picture is meant to burn in — no amount of speed at the front end changes the fact that the trade is not truly done until the currencies actually change hands at settlement, days later.
Step one is the request or the stream: the client either asks for a price or receives a continuous flow of prices. In a request-for-quote, or RFQ, the client asks the desk for a two-way price on a specific pair and size, and the desk responds. Alternatively, the desk streams live, continuously updating prices to the client's screen or system, and the client can deal on them at any moment. RFQ suits larger or less liquid trades where a tailored price makes sense; streaming suits liquid, standardised flow where speed and automation dominate. Either way, this is the moment the desk is asked to commit a price.
Step two is pricing: an engine computes a bid and an offer, skewed for the desk's position, the client and the state of the market. A pricing engine takes the current market, the desk's existing position, what it knows about this client's flow, and current volatility, and produces a two-way quote in a fraction of a second. The skew we met earlier lives here: if the desk wants to attract selling to flatten a long position, it shades the price accordingly. This step is pure computation, and for liquid spot it is fully automated.
Step three is execution: the client deals on the price and the trade is captured. When the client accepts a quote — clicks, or hits the stream, or confirms the RFQ — the trade is executed and immediately captured into the bank's systems as a live trade. For liquid spot, this whole front end, from request to execution, can be sub-second. The trade now exists and the desk owns the position.
Step four is risk and hedge: the new position hits the book and the desk decides to internalise it or hedge it. The moment the trade is captured, the desk is holding risk. It either internalises — nets the position against offsetting client flow already on the book or expected shortly — or hedges it in the wholesale market. Well-run desks internalise as much as they can, precisely because hedging externally costs spread. This step, too, is largely automated, with algorithms deciding in real time what to net and what to lay off.
Step five is confirm and match: both sides of the trade confirm the details and match them electronically. Confirmation is each party stating the terms it believes it agreed; matching is checking that the two statements agree — same pair, same amount, same rate, same date. Historically this was paper and phone; today it is electronic and near-instant for standardised trades. A trade that does not match is flagged for investigation before it can proceed, because an unmatched trade is an operational risk waiting to happen.
Step six is settlement: on the value date, the two currencies are actually exchanged, ideally through CLS on a payment-versus-payment basis. This is where the trade becomes real. On the value date — T+2 for most spot — the currencies move between accounts. Where the trade settles through CLS, both legs settle simultaneously, so neither party can lose principal to the other's failure. This is the step that discharges Herstatt risk, and it is the reason the trade is not truly complete until now, however fast the front end was.
Step seven is report and reconcile: the trade is reported to regulators where required and reconciled against the books. Reporting satisfies regulatory obligations — under EMIR in Europe, for instance, many FX derivatives must be reported to a trade repository. Reconciliation checks that the desk's records, the counterparty's records and the settlement records all agree. Only when everything reconciles is the trade genuinely closed. The industry's goal across this whole pipeline is straight-through processing, or STP: a trade that flows from execution to settlement without manual intervention at any step, because every manual touch is slow, costly and a source of error. The ideal is a trade that prices, executes, hedges, confirms, settles and reconciles itself. Steps one through four may take milliseconds; steps five through seven, and settlement above all, are what actually finish the job.
A worked example: a French exporter locks in a dollar receivable
A market maker quotes both sides and manages the risk it inherits
A market maker is a dealer that continuously quotes a price at which it will buy and a price at which it will sell, and then manages whatever position that quoting leaves it holding. To make a market is to commit to being a counterparty in both directions. The client chooses which way to trade; the market maker takes the other side, whichever it is. That commitment is a service — clients get instant, reliable liquidity — but it is also a risk, because every time a client trades, the market maker is left holding a position it did not choose. The art of the business is quoting prices that attract enough two-way flow to earn the spread while keeping the inherited risk manageable.
The two-way price: quote a bid and an offer, and stand ready to trade either side
The two-way price is the market maker's core obligation: show a bid and an offer simultaneously and be ready to deal on either. When a client asks for a price, the desk quotes both a bid (where it will buy the base currency) and an offer (where it will sell), without knowing in advance which side the client wants. This is what "making a market" concretely means — you commit to both sides before you learn the client's intention. The spread between the two is the desk's compensation for that commitment and for the risk of being picked off by someone who knows more than you do. Quote too wide and you lose the business to a competitor; quote too tight and you earn too little to cover the risk of the flow you take on. The whole skill sits in that tension.
Warehousing, not betting: the desk holds risk to serve clients and recycles it quickly
When the desk takes on a client trade, it warehouses the resulting risk temporarily to serve the client — and then works to recycle that risk quickly rather than to bet on it. To warehouse risk is to hold a position on the book for a while rather than immediately hedging every trade in the market. If a client sells the desk euros, the desk is now long euros. It does not have to rush out and sell those euros to another bank — which would cost spread. Instead it can skew its quotes: shade its price to make it slightly more attractive for the next client to buy euros from it, so that incoming flow naturally offsets the position. By adjusting where it shows its bid and offer, the desk pulls in the opposite flow it needs and flattens the risk without paying the market. Skewing to attract offsetting flow is one of the desk's most powerful and least visible tools.
Franchise flow is the asset the desk is really monetising, and the information it carries is a tightly governed edge. The steady stream of client business — the franchise — is valuable precisely because it is diverse and offsetting: buyers and sellers arriving through the day can be matched against each other inside the firm. A large, diverse franchise also gives the desk an information edge: seeing a broad cross-section of who is buying and selling helps the desk read the market's near-term pressure better than a participant who sees only their own trades. That edge is legitimate and central to market-making, but it is bounded by strict conduct rules. Regulators and industry codes govern how client information may and may not be used, how orders must be handled, and how the desk must treat clients fairly. The desk holds risk to serve clients and recycles it fast; it does not sit on a directional bet, and it does not misuse what it sees. Warehousing is a service function with an information benefit, not a licence to speculate.
FX electronified first and furthest: from voice to click to API
FX did not drift into electronic trading — it sprinted, and it got there before almost every other market. Today the large majority of foreign exchange by volume changes hands electronically, and spot trading in the major currency pairs is almost entirely machine-driven. To see why that happened, and why it matters for everything else in this module, it helps to walk through the three eras the market has passed through. Each era did not simply replace the last; it layered on top of it, and traces of all three survive today.
Three eras: voice, click, and machine-to-machine
The voice era ran on the telephone, and a price was a spoken thing that lived for seconds. A corporate treasurer or a fund manager who needed to trade would call a bank's dealing desk. A salesperson would shout the request across to a trader — "I need a hundred million euro-dollar, spot" — and the trader would quote a two-way price out loud: a bid at which the bank would buy and an offer at which it would sell. The client accepted or declined in the moment, and the price evaporated the instant the call ended. This was fast by the standards of its day and glacial by ours. A single trade took seconds of human attention. Relationships, memory and trust did the work that software does now. The voice era never fully died — large, sensitive or complex trades still get worked by humans on the phone or over chat — but it stopped being where the bulk of the volume lives decades ago.
The click era put a screen between the client and the price, and it changed the physics of the market. Beginning in the 1990s and accelerating through the 2000s, electronic platforms appeared that streamed live, continuously updating prices onto a client's monitor. Instead of calling for a quote, the client watched prices tick in real time and clicked to trade the one they wanted — "click-to-trade" — with the deal executing and booking automatically. Two flavours emerged. A single-dealer platform showed one bank's prices; a multi-dealer platform let several banks compete for the same trade on one screen, so the client could see who was tightest and click the best. The click era did something the voice era never could: it made price transparent and instantaneous, and it let a client compare dealers side by side. Once a client could see three banks' euro-dollar prices at once, the bank showing the widest spread simply stopped winning trades.
The API and algo era removed the human from the loop entirely, and this is where FX lives today. An API — an application programming interface — is a defined way for one machine to talk directly to another. In the modern market, a client's system connects straight to a bank's pricing system through an API and consumes a stream of prices programmatically, with no person watching a screen and no person clicking a button. Trades are requested, quoted, accepted and booked machine-to-machine in microseconds. Algorithms — pre-programmed execution strategies, which we will meet later in this part — slice large orders and work them automatically over time. In this era, the "trader" for a huge share of flow is code. Spot FX in the majors is now overwhelmingly transacted this way. The human has not disappeared; the human now builds, tunes, supervises and intervenes in the machine, which is a very different job from quoting a price out loud.
Why FX automated first: a standardised product, thin margins and a continuous market
FX automated before almost every other asset class because the product was practically built for a machine. Consider what a spot trade in euro-dollar actually is: an agreement to exchange one standardised, perfectly fungible thing for another standardised, perfectly fungible thing, at a single number, for delivery two days later. There is nothing to negotiate about the item itself. One euro is exactly like every other euro. Contrast that with a bespoke structured derivative, whose terms — the payoff, the underlying, the maturity, the credit terms — are unique to the deal and demand human judgement to price and document. A spot currency pair is a single number attached to a commodity so uniform that a computer can quote it, trade it and settle it without a person ever forming an opinion about it. When the product is that simple and that standardised, automation is not just possible; it is the obvious thing to do.
Thin margins made automation not a luxury but a survival requirement. The spread on a liquid pair — the gap between the price at which a dealer buys and the price at which it sells — is minute. In the majors it is often a fraction of a basis point on the notional. You cannot pay a human being to spend seconds quoting a trade whose entire profit is a sliver of a cent per unit; the economics collapse. The only way to make a fraction-of-a-basis-point spread into a real business is to earn it on enormous volume at near-zero marginal cost per trade — and that is exactly what a pricing engine delivers. Automation is what turns a spread too thin to notice into a franchise worth billions, by multiplying it across millions of trades that each cost almost nothing to make.
A market that never closes rewards a pricer that never sleeps. FX trades roughly around the clock, five and a half days a week, as liquidity passes from the Asian session to the European session to the North American session and back again. A human desk has to hand the book from one time zone to the next and cannot be equally sharp at every hour. A pricing engine is always on, equally awake at 3am Sydney time as at midday in London. Continuous trading rewards continuous, automated pricing, and it penalises anyone who has to switch the lights off.
And once one bank went electronic, everyone had to follow — a technology arms race with no exit. The moment a single dealer began streaming tighter prices, faster, at all hours, its rivals faced a stark choice: match it or watch the flow drain away. A client who could get a two-tenths-of-a-pip price from a machine was never going to accept a wider price from a human. Speed and tightness became competitive weapons, and each improvement forced a response. That dynamic — build a better engine, win more flow, invest the profit in a still-better engine — is why FX did not just automate but automated furthest, and it is why the rest of this part is really a tour of a technology stack.
The pricing engine is the FX franchise written in software
If you want to know what a modern FX business actually is, look at its pricing engine, because that piece of software makes every decision the old voice trader used to make — only thousands of times a second, for every client, in every pair, continuously. The engine is not a component of the franchise; it is the franchise, rendered in code. It decides how tight to quote, whom to show which price, how much risk to hold, and when to hedge. A better engine wins more flow at lower risk. Let us open it up and see what flows in and what flows out.
What goes in, what comes out
The engine consumes several streams of information at once, and its quality is largely a function of how well it fuses them. The first input is market data: the primary ECN feeds and the LP streams that, aggregated, tell the engine where the market is right now. The second is the desk's own state: its current position in each pair and the risk limits it is operating under — is it already long euros, and how much more can it hold? The third is the identity of the client asking: their tier and their flow profile, because, as we just saw, the same trade is worth a different spread depending on who brings it. The fourth is market conditions: how volatile things are, how thin the book is, whether a major data release is seconds away. The fifth, for anything beyond spot, is the forward-points and interest-rate curves that convert a spot price into a price for a future date.
From those inputs the engine computes an output, and the sequence of steps is worth naming precisely. It first calculates a fair mid — its best estimate of the true middle of the market. It then adds a spread around that mid, sized by the client's tier and the trade's size. It skews the two-way price to reflect its own position. It applies tiering so different clients see different spreads. It checks the whole thing against its risk limits. And it streams the resulting two-way price out to the client, refreshing or requoting in microseconds as the inputs change. The visible output is a live, streaming, two-way price. The invisible outputs matter just as much: per-client tiered pricing, automatic execution and booking when the client trades, a hedging signal telling the risk system what just happened to the position, and a record of every win and loss that feeds back to make the next quote smarter. Three of those mechanisms deserve their own explanation.
Skew nudges the next trade toward flattening the book
Skew is the engine deliberately making one side of its two-way price keener than the other in order to attract the trade that reduces its risk. Suppose the desk is already long euros — it has bought more than it has sold and is sitting on a pile of euros it would rather not hold. It does not simply want to quote a neutral, symmetric price; it wants the next client to be a buyer of euros, because a buyer takes euros off its hands. So the engine shifts its whole quote slightly: it shows a keener offer (a lower price to sell euros) to make buying from it more attractive, and a less generous bid (so it is less likely to be sold even more euros). The mid has not moved because the market changed; the engine has moved its mid to lean against its own inventory. Multiply this across every quote and skew becomes a continuous, automatic mechanism for steering the book back toward flat by making flattening trades cheaper for clients than position-building ones.
Tiering shows different clients different spreads
Tiering is the engine showing genuinely different prices to genuinely different clients, and it is both legitimate and central to how the economics work. A client whose flow is benign, who trades in size, and who does a lot of business across the franchise earns a tighter spread than a client whose flow is sharp, sporadic and toxic. This is the software version of the relationship judgement a voice trader used to make instinctively — "this is a good client, quote them well" — but applied consistently, at scale, and grounded in measured data about each client's flow quality rather than gut feel. Tiering is why there is no single "the price" in FX for a given moment; there is a price for you, and it reflects what your flow is worth to the dealer.
Auto-hedge offloads residual risk automatically when it breaches a threshold
Auto-hedging is the engine deciding, without a human, that its accumulated position has grown large enough to lay off, and then going to the market to do it. The desk sets a threshold — a level of position in a given pair beyond which it no longer wishes to warehouse the risk. As client trades push the position past that threshold, the system automatically sends a hedging order to an ECN or an LP to bring the position back within tolerance. The art, which we return to at the end of this part, is in where the threshold sits and how the hedge is worked: hedge too eagerly and you pay away the very spread you just earned; hedge too slowly and an adverse market move turns a warehoused position into a loss.
Put those pieces together and you can see why the engine is the franchise. It decides, thousands of times a second, how tight to quote, to whom, and how to manage the risk each trade creates — the exact bundle of decisions that used to define a great trader, now made continuously and consistently by software. A better engine quotes tighter without taking on more risk, wins more of the good flow, internalises more of it, and hedges the residual more cheaply. Building it and tuning it is not a project with an end date; it is a large, permanent effort by quantitative researchers and software engineers, and the banks that treat it that way are the banks that dominate. The engine is where the money is made and lost, one microsecond at a time.
Execution algorithms work a large order without moving the market
A big order has a problem that a small order does not: the act of executing it moves the price against you. If a fund needs to buy a billion euros and simply crosses the spread for the whole amount in one go, it will chew through every available offer, lifting the price as it goes and paying a worse and worse rate — and it will announce its intention to the whole market, inviting others to trade ahead of it. Execution algorithms exist to solve this. They are pre-programmed strategies that break a large parent order into many small child orders and work them into the market over time, so that the footprint is small and the price impact is minimised.
Slicing an order over time to reduce impact
The core idea of every execution algorithm is the same — do not show the whole order at once — but they differ in how they decide the timing and size of each slice. A TWAP algorithm, for time-weighted average price, spreads the order evenly across a chosen window: buy a steady trickle every few seconds for the next hour, so the average fill price tracks the market's average over that hour. A participation or POV algorithm, for percentage of volume, ties its pace to the market's activity: trade a set fraction — say ten percent — of whatever volume is printing, so you are busy when the market is busy and quiet when it is quiet, blending into the flow. A liquidity-seeking algorithm actively hunts for size across many venues, taking liquidity opportunistically wherever it appears rather than following a fixed schedule. A peg or limit algorithm rests passively at or just inside a reference price, waiting for the market to come to it rather than crossing the spread.
Every one of these is a choice on a single dial: speed versus patience. Execute fast and you are exposed to the market for less time, so you carry less timing risk — the risk that the price simply drifts away from you while you wait — but you pay more impact, because you are demanding liquidity urgently. Execute patiently and you minimise impact and often capture spread by posting passively, but you carry the timing risk of a market that may move against you over the hour it takes to finish. There is no universally right answer; there is only the right answer for this client, this order, this market. The algorithm is the tool that lets the client dial in exactly where on that spectrum they want to sit, and then executes that choice with a discipline no human clicking a mouse could match.
Transaction cost analysis proves the execution was good
Once you have worked an order through an algorithm, a natural and commercially loaded question arises: was that a good execution? Transaction cost analysis, universally shortened to TCA, is the discipline that answers it with data rather than assertion. It measures the quality of an execution against defined benchmarks and turns "I think we traded well" into a number you can defend, dispute, or improve upon.
What TCA measures and why both sides use it
TCA compares the price you actually achieved against one or more reference prices, and the gaps it exposes have names. The most common benchmark is the arrival price — the market price at the instant the order arrived — and the difference between that and your average fill is your slippage: how much the execution cost you relative to where the market was when you started. Another benchmark is the TWAP over the execution window, which tells you whether you beat or lagged the market's own average. A third, especially important in FX, is a fixing: an official benchmark rate struck at a set time, such as a widely used 4pm London benchmark, against which a great many orders are explicitly priced. TCA also tries to isolate market impact — how much your own trading moved the price — from the background drift that would have happened anyway.
Both sides of the market use TCA, for mirror-image reasons. The buy-side — the funds and corporates sending orders — uses it to grade the banks and the algorithms they trade through: which liquidity provider reliably delivers tight, low-impact execution, and which algorithm performed best for this kind of order? Poor TCA scores move business away. The sell-side — the banks — uses it to prove and improve the quality of its own execution and, critically, to defend its spreads: a bank that can show a client, with clean data, that its all-in execution cost was lower than a cheaper-looking competitor's has a powerful commercial argument. TCA is thus both a scorecard and a sales tool.
Doing TCA properly is a serious data problem, and this is where FX makes it hard. Meaningful analysis requires clean, precisely time-stamped records of prices and trades across a fragmented set of venues, all aligned to a common clock down to the millisecond or finer. In a market with no single official record of every trade — a point we return to forcefully later — assembling that dataset is genuine engineering work. Orders that reference a fixing need special care, because everyone trading to the same benchmark at the same moment creates its own crowd dynamics around the fix, and analysing execution into a fixing means understanding those dynamics rather than just reading a single number off a screen.
Last look and the FX Global Code govern a market with little law
FX has a peculiar governance problem: the largest financial market in the world contains a huge chunk — spot — that sits almost entirely outside formal financial regulation. Two things fill the gap where the law is thin. One is a specific and controversial market mechanism called last look. The other is a voluntary code of conduct that the industry wrote for itself after that mechanism, and others like it, went badly wrong. Understanding both is essential to understanding how the modern market polices itself.
What last look is, and where it turns from protection into abuse
Last look is a brief final check that a liquidity provider takes after a client accepts its streamed price, during which the LP may still reject the trade. Here is the sequence. The LP streams a price. The client, having seen it, clicks to accept it. But because the price was streamed a few milliseconds earlier, and because in a microsecond market the true market may have moved in those milliseconds, the LP takes one last look before confirming: has the market moved away from the quoted price? If it has not, the trade is confirmed. If it has moved beyond a small tolerance, the LP may reject the trade, and the client has to try again at a fresh price. The window is tiny — often single-digit milliseconds — but in a machine market that is an eternity.
The rationale for last look is real and defensible. In a world of streamed prices and networked latency, an LP is always quoting a price that is a few milliseconds stale by the time a client can act on it. Fast, sophisticated counterparties can detect when a streamed price has become stale and pick it off — deliberately trading against a price the LP would never have shown had it known the market had moved. Last look is, in its honest form, protection against being systematically picked off on stale quotes, and that protection is part of what lets an LP quote tightly in the first place. Remove all protection and the LP must widen its spreads to cover the picking-off risk, which makes everyone's prices worse.
But the same mechanism turns into abuse the moment it is applied asymmetrically or exploited. Asymmetric last look means rejecting the trade only when the price move favours the client, while happily filling when the move favours the LP — heads the LP wins, tails the client loses. That is not protection; it is a free option exercised against the client. Worse still is using the last-look window not merely to check the price but to trade ahead of the client's order — seeing that a large buy is incoming and buying first. These practices convert a legitimate risk control into a mechanism for extracting value from clients without their knowledge, and they are exactly what got the market into trouble.
The FX Global Code fills the gap where spot is unregulated
The FX Global Code is a voluntary, global set of principles of good practice for the wholesale FX market, and it exists precisely because so much of that market has no binding law. Spot FX, remember, is largely outside the perimeter of formal financial regulation in most jurisdictions — it is not treated as a regulated financial instrument the way a listed security or many derivatives are. That left a governance vacuum, and the vacuum was filled with scandal: the manipulation of benchmark fixings, in which traders colluded to move rates around the fix, and the last-look abuses just described. The industry's response, coordinated through the central-bank-sponsored Global Foreign Exchange Committee, was to write down a common standard of conduct that every participant could adhere to.
The Code covers the areas where the trouble actually happened, and it does so as principles rather than rules. It addresses ethics and professional conduct; execution, including how orders should be handled; information sharing, so that client information is not misused; last look, setting the expectation of transparent and non-abusive application; mark-ups, meaning the transparency of the margin a dealer adds; pre-hedging, the sensitive practice of a dealer trading in anticipation of a client order it expects to fill; and the mechanics of confirmation and settlement. Crucially, these are principles, not statutes. The Code does not carry the force of law; it carries the force of expectation.
Adherence works through a Statement of Commitment and, behind that, through commercial reality. A firm formally signs a public Statement of Commitment declaring that it aligns its practices with the Code. Nothing legally compels it to. But in practice, the major buy-side institutions increasingly expect their counterparties to have signed, prime brokers and venues expect it of their participants, and a firm that refuses to commit stands out in a way that costs it business. Adherence has therefore become a reputational and de facto commercial requirement — a voluntary standard that is not truly optional if you want to be in the game. That is how a market with little law polices itself: not with a rulebook and a regulator behind every trade, but with a shared standard whose real enforcement mechanism is the willingness of everyone else to trade with you.
Aggregation and smart routing turn fragmented liquidity into one executable book
FX liquidity is scattered across dozens of places, and turning that scatter into a single thing you can trade against — cleanly, without being fooled — is one of the hardest and most important engineering problems in the market. We met aggregation briefly earlier; now we take it apart properly, because the details are where the money and the danger both live.
The aggregation problem and the smart order router
The raw situation is fragmentation: many LPs, several ECNs, the primary pools, each streaming its own prices, its own available sizes, and its own quirks. No two feeds are identical. One LP might show a great price but in small size; another shows depth but a wider spread; an ECN shows a firm price but charges a fee to take it; another venue's price is subject to last look. An aggregator's first job is to combine all of these into a single, deduplicated composite book — a unified ladder of the best bids and offers across every source, with the real depth behind each level. That composite book is what the engine and the trader actually look at when they think "the market."
But building the book is only half the problem; deciding what to do with each order is the other half, and that is the job of the smart order router. A smart order router, or SOR, takes an order that needs to be executed and decides, in real time and per order, three things: where to send it among all the available venues, how to slice it if it is too big for any single venue's top-of-book, and when to take liquidity aggressively versus post it passively and wait. It makes those decisions weighing expected cost, the probability that a given venue will actually fill the order, and the risk of last-look rejection. The router is the tactical brain that turns the strategic composite book into a sequence of concrete actions. Three subtleties make this genuinely hard.
Phantom liquidity: the same price counted many times
Phantom liquidity is depth that appears in your composite book but is not really there, because the same underlying price has been counted more than once. Recall that a single liquidity provider often streams its price to many venues at once. If your aggregator naively sums the size available across all those venues, it will count that one LP's hundred million euros as though it were three hundred million — once on each venue the LP streams to. The book looks three times as deep as it is. Act on that phantom depth — fire an order sized for three hundred million into what is really one hundred — and you will get filled on a third of it and move the market chasing the rest. Deduplication, the work of recognising that the same LP price appearing on multiple venues is one piece of liquidity and not several, is therefore a hard and mission-critical problem. Getting it wrong means systematically overestimating how much you can trade, which is exactly the kind of error that produces bad automated quotes.
Fill quality beats the headline price
The best-looking price on the screen is worthless if you cannot reliably trade on it — so the router must weigh how often a venue actually fills you, not just how tight its level looks. A venue might consistently show the sharpest price in the composite book and then reject a large fraction of the orders sent to it under last look. Routing to that venue because it looks best is a trap: you get rejected, the market moves, and you end up worse off than if you had sent the order to a slightly wider but firmly reliable venue. Fill quality — the probability of actually completing the trade at the shown price — is a first-class input, sometimes more important than the headline level. And there is a second discipline here: spraying a large order across many venues at once leaks information, because observers see the order footprint appear everywhere simultaneously and infer that a big buyer or seller is active, then move the price against you. Discretion — being deliberate about how visibly you show your hand — is part of good routing.
Internalise first: the smartest first stop is the bank's own book
Before the router sends a single child order to any external venue, its smartest possible move is to check whether the trade can be filled inside the bank's own book. If the bank is already holding an opposite position, or has another client's offsetting flow available, the order can be matched internally at no external cost, with no market impact, no information leakage, and no last-look risk. The best venue is often no venue at all. This is why, in a well-built system, internalisation is not an afterthought bolted onto the routing logic; it is the first branch of the decision tree. Everything external is a fallback for the residual that could not be netted at home.
Anatomy of an FX order, and why deduplication is hard
To see why cleaning and deduplicating FX data is genuinely difficult, it helps to look at what a single FX trade or order actually is as a piece of structured data. A core FX order object carries a set of fields, and each field drives something specific downstream. There is the currency pair (euro against dollar, say), which determines which market and which risk book the trade belongs to. There is the side — whether the client is buying or selling the base currency, the first-named currency in the pair — which determines the direction of the resulting position. There is the notional, the size of the trade, which drives risk, margin and the size of any hedge. There is the price or rate agreed. There is the value date, the day settlement is due, which drives the entire settlement and funding process. There is the product type — spot, forward, swap — which determines how the trade is priced and processed. There is the client identity and tier, which drives pricing and relationship management. There is the venue on which it traded, which drives reporting. And there is a precise timestamp, which drives TCA, reconciliation and regulatory reporting alike.
Now put many of these objects together across many venues with no common referee, and you see why the data problem is real. FX has no single consolidated tape — no official, unified record of every trade and every price, the way listed equity markets have. The same quote from one LP is echoed across several venues, arriving as several separate records that describe one economic reality. To build an accurate picture of the market you must ingest all these partial, overlapping, slightly-inconsistent feeds, align their timestamps to a common clock, recognise which records are duplicates of the same underlying liquidity, and stitch the survivors into one clean book. That is an unglamorous but decisive engineering problem, because everything downstream — the fair mid, the skew, the aggregation, the auto-hedge — runs on this data. Feed the engine bad, double-counted, mistimed data and it will stream dangerous quotes automatically, at machine speed, before any human notices. Garbage in, dangerous quotes out.
When machines quote, machines must manage risk, with hard guardrails
The moment you let software stream prices and take trades without a human in the loop, you have handed a machine the power to lose money at machine speed — which means the risk management has to be automated too, and it has to be surrounded by guardrails that a human still controls. An automated desk that can price in microseconds can also accumulate a dangerous position in microseconds. The systems that keep it safe are as important as the systems that make it money.
Auto-hedging balances the cost of hedging against the risk of waiting
Every electronic trade updates the desk's position in real time, and the system has to decide, continuously, how much of that accumulating risk to keep and how much to lay off. The decision follows a clear priority order. First, internalise: offset each new client trade against opposite client flow already on the book, at no external cost — if one client sold euros and another is now buying them, the two cancel and nothing needs to be done in the market. Second, skew: for the residual imbalance that internalisation could not fully net, adjust the streamed prices to attract flow that flattens the book, as we described earlier, so the market itself does the hedging for free. Third, and only as a last resort, auto-hedge: if after internalising and skewing the residual position still breaches a limit, send an order to an external venue to bring it back within tolerance.
The tuning of that third step is where a great deal of P&L is quietly won or lost. Hedge too eagerly — lay off every small imbalance the instant it appears — and you pay away, in spread and impact, the very margin you just earned from clients; you become a machine for donating your profit to the external market. Hedge too slowly — let positions ride in the hope that internalising flow will net them out — and you expose the book to an adverse market move that turns a warehoused position into a real loss. The right threshold, and the right way to work the hedge when it triggers, sits between those two failure modes, and it depends on the pair, the volatility, the time of day and the character of the flow. There is no static answer; there is a continuously tuned policy, and the quality of that tuning is a real competitive difference between desks.
The guardrails that keep an automated desk safe
Around the automated pricing and hedging logic sits a set of hard controls whose entire purpose is to stop a fast machine from doing catastrophic damage fast. The first are position and loss limits: hard caps, set by currency pair, by desk and by book, beyond which the machine is not permitted to go. As the position or the loss approaches a limit, the engine responds automatically — widening its spreads to discourage further one-sided flow, and ultimately halting pricing altogether rather than breaching the cap. The limits are not advisory; they are enforced in the code.
The second guardrail is the kill switch, and it is the single most important safety feature on an automated desk. A kill switch is a mechanism — one-click for a human, or automatic on a trigger — that pulls every price the desk is streaming, instantly and everywhere. If something is going wrong — a data feed has gone haywire, the engine is streaming visibly bad prices, the market has entered a state the models were never built for — you do not want to reason about it while the machine keeps trading; you want to be able to stop everything in a single action and sort it out with the machine silent. The kill switch is the acknowledgement that the fastest safe response to a machine behaving badly is to turn it off. A desk without a reliable, well-tested kill switch is a desk with no brakes.
The third guardrail is pre-trade validation — the fat-finger and sanity checks that catch the absurd before it executes. Before any order goes out or any quote is streamed, the system checks it against reason: is this size ten times larger than anything this client has ever done? Is this price wildly away from where the market actually is? An order or a quote that fails these sanity checks is blocked automatically, catching both human input errors and software malfunctions before they reach the market. The fourth guardrail is the human: real-time monitoring by people who oversee the machine, watch its positions and its P&L and its behaviour, and stand ready to intervene — to widen, to halt, to hit the kill switch — when something looks wrong. Automation did not remove the human; it changed the human's job from making every price to supervising the thing that makes every price.
Operational resilience is not a nice-to-have in an automated FX market; it is a precondition for being allowed to run the business at all. An automated desk has moments that offer no second chance. A pricing engine that streams one badly wrong price can be picked off for a large loss before anyone blinks. A settlement that must complete on its value date must actually complete, or the failure cascades. Four capabilities keep a machine-run book safe. Redundancy means running duplicate systems, so no single component is a single point of failure. Automatic failover means a standby system takes over instantly and seamlessly when the primary fails, without waiting for someone to notice. Disaster recovery means a rehearsed, tested procedure for restoring service after a major outage — rehearsed, because a recovery plan no one has ever run is a hope, not a plan. And business continuity means the people and the processes, not just the servers, can keep functioning through a disruption. Downtime in FX does not merely cost you the flow you miss while you are dark; it tells every client that you are unreliable, and in a market built on always being there, that damages the franchise long after the systems come back up.
Internalisation is the profit engine of modern FX
We have circled internalisation throughout this part; now we put it at the centre, because it is the single most important economic fact about the modern FX market and the reason the whole industry looks the way it does. If you understand internalisation, you understand why a few banks dominate, why scale compounds, and why so much of the engineering we have described ultimately serves one purpose.
What internalisation is, and why it dominates the economics
Internalisation is a dealer matching one client's trade against another client's opposite trade on its own book, capturing spread on both sides, and never going out to the external market at all. Return to the simplest possible example. One client sells the dealer a hundred million euros. Shortly after, another client buys a hundred million euros from the dealer. The dealer takes the euros from the first client and hands them to the second. It has bought at its bid from the first and sold at its offer to the second, earning the full spread on both legs, and it has done so without touching an ECN, without paying an external spread, without moving the market, and without revealing anything to anyone. The trade lived and died entirely inside the dealer's book.
The economics of that are so favourable that internalisation, not the headline spread alone, is where the money is actually made. Consider what the dealer avoided: no external spread paid to another LP, no hedging cost, no market impact, no information leakage about what its clients are doing. And consider what it earned: the spread, captured twice, on flow that cost it almost nothing to match. The larger and more diverse a dealer's franchise flow, the more of it can be netted internally, because a big, varied client base is constantly generating offsetting trades — some clients buying while others sell, all day, in every pair. The top banks now internalise the large majority of their spot flow, sending only the residual imbalance — the net amount left over after all the internal matching — out to the external market to hedge. Their visible footprint in the market is far smaller than the volume they actually transact, because most of it never leaves the building.
Scale is self-reinforcing, and conduct must keep pace
Internalisation creates a flywheel: more flow means more internal matching, which means tighter prices, which wins more flow — and the flywheel concentrates the market among a few large banks. The logic is relentless. A bank with more client flow has more opportunities to net trades internally, so it incurs less external cost, so it can quote tighter, so it wins still more flow, so it can net still more internally. Scale begets scale. This is the reason FX liquidity has concentrated among a handful of top banks: not a conspiracy, but the mathematics of internalisation rewarding size. And because only the net unmatched residual reaches the ECNs and LPs, the largest banks have a market footprint dramatically smaller than their true client volume — they are far bigger than they look from the outside, which is itself part of the advantage.
But the same mechanism that makes internalisation so profitable makes conduct around it a live issue that the Code polices. When a dealer is matching client against client and warehousing flow on its own book, questions of fairness arise that were front and centre in the scandals we discussed. Is the mark-up the dealer adds transparent to the client? Is the dealer pre-hedging — trading in anticipation of a client order — in a way that serves the client's execution or in a way that front-runs it? Is client information being kept confidential rather than exploited? Internalisation, mark-ups and pre-hedging are exactly the areas the FX Global Code addresses, and a dealer that internalises at scale must do so under standards of fairness and transparency, or it converts a legitimate profit engine into a conduct scandal waiting to happen. Scale and integrity have to advance together.
FX makes money on spread and volume, not on getting the direction right
The FX desk is a volume business dressed up as a trading business. Its profit comes from doing an enormous number of trades at a small margin, from netting them cleverly against each other, and from adding services around them — not from a heroic call on the euro. This section is the profit-and-loss anatomy of a flow franchise.
The revenue sources: spread, internalisation, skew, forward points, value-added services
The desk earns from five main sources, and direction is barely one of them. The first is the spread — the gap between the bid and the offer, captured every time the desk trades with a client. The second is internalisation: when the desk nets a client's buy against another client's sell inside the firm, it captures spread on both sides without ever paying the external market to hedge, which is the single most profitable thing it can do. The third is skew: by shading its quotes to attract the flow it wants, the desk both flattens risk cheaply and can earn a little extra on the position it is carrying. The fourth is forward points: on forwards and swaps, the desk earns a margin on the interest-rate-differential arithmetic in addition to the spot spread. The fifth is value-added services — hedging advice, structured solutions, execution algorithms, research and access — which clients pay for through wider spreads or explicit fees. Notice that "getting the direction right" is nowhere on this list. Directional profit exists, but it is small, risky and tightly limited; the durable money is in the other five.
The economics of a flow business: volume beats width
In a flow business, volume beats width: a thin spread earned millions of times is worth far more than a fat spread earned rarely, and the way to widen that gap is to internalise. The desk's economics reward scale. A tighter spread wins more business, and more business means more chances to net trades internally. Internalisation — matching offsetting client flow in-house rather than hedging externally — is the prize because external hedging costs money: you cross someone else's spread, you signal your position to the market, and you may move the price against yourself. Every trade you can net internally avoids all of that. So the desk with the largest, most diverse franchise can run the tightest spreads and keep the most margin, because it hedges least in the open market. This is a scale flywheel: more flow enables more internalisation, which enables tighter prices, which wins more flow.
Franchise flow is also defensive, and the desk must earn its spread net of real and rising costs. A large franchise is not just an offence; it is a defence. Steady, diverse client flow smooths the book and reduces reliance on any single trade or market view, which makes the whole business more stable and less capital-hungry. But running this business is not free. The desk pays for technology — pricing engines, aggregation, low-latency connectivity — for capital held against the risk it warehouses, and for settlement infrastructure that moves currencies safely. All of these are real, and several are growing as the market gets faster and more electronic. So the spread the desk earns must cover not just the risk of the flow but the full cost of the machine that prices, hedges and settles it. A desk that wins volume by pricing below its all-in cost is not running a business; it is subsidising its clients. The discipline is to earn the spread net of everything it takes to capture it.
The economics of the desk fit on one chart
Read the waterfall left to right and the economics of an electronic FX desk become obvious: the profit is a thin residual left after several large costs eat into a modest revenue. On the revenue side, two bars build the desk up. The first is gross spread captured — the aggregate difference between the prices at which the desk buys and sells, summed across huge volume. The second is the internalisation benefit: when the desk matches a client's buy against another client's sell inside its own book, it avoids paying the market to hedge, and that avoided cost is real money earned. Together these two bars are the desk's gross earnings. Then the costs subtract, bar by bar. External hedging cost is what the desk pays to offload the risk it could not internalise — crossing the spread on a venue, plus market impact. Technology and connectivity is the standing bill for co-located hardware, venue fees, market-data feeds and the engineering teams that keep the stack alive. Capital and settlement is the cost of the balance sheet the business consumes and the machinery of settling every trade. What survives all of that is a thin net margin — the small bar on the right. (All figures on the chart are illustrative and indexed to gross spread, not real numbers from any firm.)
The shape of this waterfall explains the entire strategy of modern e-FX: you win on volume, internalisation and cost discipline, not on wide spreads. Because the net margin per unit is so thin, the only way to a large profit is to run enormous volume across that thin margin — which requires the fast, reliable, low-latency stack described in Part 4. Because internalisation converts a cost into a saving, raising the internalisation ratio drops straight to the bottom line, which is why the largest banks — with the most flow to cross internally — enjoy a structural advantage. And because technology, capital and hedging are the bars that eat the revenue, disciplined control of those costs is as decisive as any trading skill. A desk that tried to widen spreads to fatten the margin would simply lose the volume to a cheaper competitor and shrink the very base its economics depend on. Thin margins at vast scale, ruthlessly internalised and tightly cost-controlled — that is the business the whole platform ecosystem is built to serve.
Fixings turn a moment in time into a benchmark — and a target
What a fixing is and why enormous flow references it
A fixing is an official reference rate struck at a set time, and it exists because thousands of institutions need one rate they can all agree on. A fixing (or benchmark fix) is a single exchange rate, published for a currency pair at a fixed moment each day, that market participants treat as the authoritative price for that instant. The most important is the 4pm London WM/Reuters fix — now published under the WM/Refinitiv name — which sets reference rates for hundreds of currency pairs. Central banks publish their own references too: the ECB euro reference rates are struck each afternoon and are widely used for accounting and contractual purposes across the euro area. These rates are not where most trading happens; they are where valuation and settlement anchor.
The reason vast order flow deliberately executes "at the fix" is that a benchmark removes an argument before it can start. Consider an index-tracking fund that must hold currencies in exactly the proportions of its benchmark index. When the index provider rebalances, the fund has to buy and sell currency to match — and it needs to do so at a rate that its custodian, its auditor and the index itself all recognise as fair. If the fund traded at some arbitrary intraday price, it would risk tracking error: a gap between its return and the index's return that it is contractually built to avoid. By executing at the published fix, the fund guarantees it transacts at the same rate the index uses to value the same position. The same logic pulls in passive mandates (portfolios run to replicate an index rather than beat it), custodians (banks that safe-keep and administer assets for asset owners), and month-end rebalancers (funds that realign to target weights at each month's close). All of them want a single, verifiable, non-negotiable number — so they all push their orders into the same short window, and the flow through that window becomes enormous.
How the fix is computed, and why that makes it fragile
The fix is not a poll of opinions; it is a measurement of actual trades and quotes captured inside a short window around the fixing time. For the WM/Refinitiv rates, a fixing window opens shortly before the stated time and closes shortly after. Historically this window was roughly one minute wide. After the benchmark reforms that followed the manipulation scandals — examined in Part 5 — the main window was widened to around five minutes for major currencies, precisely to make the rate harder to move. Inside the window, the mechanism captures live trades and firm quotes, then derives a representative rate, typically by taking a median (the middle value, which resists distortion by a few extreme prints better than an average would). The published fix is that derived number.
Because the rate is built from real trades in a narrow slice of time, anyone carrying large orders into that slice holds a lever. If a desk must buy a great deal of a currency at the fix and it concentrates that buying inside the window, the buying pressure can nudge the printed rate upward — and if the desk is being paid based on that printed rate, the nudge can transfer money in its favour. This is not a theoretical worry; it is the exact mechanism behind the fixing-manipulation cases that produced record fines and the reform of the windows, all covered in Part 5. The fragility is inherent: a rate that must be tied to genuine trading is, for that very reason, sensitive to whoever is trading in size at the decisive moment.
The bid-offer spread is not a markup on nothing — it is structural
A client who sees a dealer buy at 1.0850 and sell at 1.0851 often assumes the gap is pure profit taken for no work. It is not. The bid-offer spread — the difference between the price at which a dealer will buy a currency (the bid) and the price at which it will sell (the offer) — compensates the dealer for three real and separable costs, plus a thin margin on top. Each cost points at a different remedy, which is why understanding the decomposition matters: it tells you what a desk must get good at to price tighter than a rival without losing money.
The first cost is adverse selection — the winner's curse built into taking the other side of everyone's trade. Adverse selection means that some of the flow hitting your quoted price is better-informed than you are: it knows something you do not, or it is simply faster and is trading on information you have not yet processed. When such flow trades against you, the price tends to move against you immediately afterward, and you lose. Because you cannot tell in advance which client is informed, you must widen the spread enough to be compensated for the losers hidden inside the total flow. The remedy is to stop treating all flow as identical: profile flow toxicity (measure how often each client's trades are followed by adverse price moves) and tier prices so that benign, uninformed flow gets tight quotes and sharp, informed flow pays more or is declined.
The second cost is inventory risk — the cost of warehousing the position you inherit. When you fill a client who wants to sell euros, you now own euros you did not want; you must warehouse that position (hold it on your book) until you can offload it, and you may have to hedge it (take an offsetting trade) at a price worse than the one you just gave. Between the fill and the hedge, the market can move against you. The spread pays for carrying that risk. The remedy is to make risk cheap to recycle: internalise — match one client's buy against another client's sell inside your own book so you never touch the external market — and skew your quotes (shade your bid and offer to attract the flow that offsets your current position). A desk that internalises well warehouses less risk per ticket and can therefore quote a tighter spread for the same expected loss.
The third cost is immediacy and operational cost — the price of being able to say "done" instantly, always. The client does not want a price eventually; it wants a firm price (a quote it can trade on, not merely indicative) right now, at three in the morning, in size, without waiting. Providing instant, always-on liquidity is not free: it requires pricing technology, capital held against the positions, and connectivity to venues and clients that must run continuously. The spread funds that standing readiness. The remedy here is scale and automation — spread the fixed cost of the platform across a larger volume of tickets so the per-trade cost of immediacy falls, which is precisely the economics that reward the largest, most automated franchises.
Even at a single instant, size pays for immediacy — and that is the depth curve. A small ticket can be filled at the best available price, but a large ticket cannot: it consumes the liquidity available at the best price, then the next layer at a slightly worse price, then the next, walking down the book. Plotting price against ticket size traces the depth curve (also called the market-impact curve): the bigger the order, the worse the average price, because the order eats through successive layers of resting liquidity. This is exactly why execution algorithms — programs that slice a large parent order into many small child orders and work them over time — exist: by feeding the market in pieces rather than all at once, they let liquidity replenish between slices and reduce the impact the depth curve would otherwise impose.
So does full electronification remove the spread? No — it compresses it, but it cannot make it vanish. Electronic markets have driven spreads dramatically tighter by automating pricing, sharpening internalisation, and forcing competition into fractions of a pip. But adverse selection, inventory risk and the cost of immediacy are real costs that exist regardless of whether a human or a machine quotes the price — a faster informed trader still exists, a warehoused position still carries risk, and always-on liquidity still costs capital and technology. What digitalisation changes is who captures the spread and how thin it gets, not whether it must exist. The winners are the firms whose technology lets them bear those three costs more cheaply than anyone else — and then pass just enough of the saving to the client to win the flow.
Before a client can trade, it must be onboarded: the ISDA, the CSA and the credit line
A corporate or fund cannot simply call a bank and trade an FX forward — it must first be made *ready*. Client readiness is the FX-native gate that a counterparty must pass before its first derivative trade, the direct analogue of an issuer becoming ready to sell securities. Readiness is a bundle of legal agreements, credit arrangements, identity checks and internal systems, and until every piece is in place the client's trade cannot be legally documented, priced, credit-checked or settled. Define the pieces plainly.
The ISDA Master Agreement is the standing contract that governs everything two parties trade over the counter. Published by the International Swaps and Derivatives Association, the ISDA Master Agreement is the standardised master contract between two parties that sets the rules for all their OTC (over-the-counter, meaning privately negotiated rather than exchange-traded) derivatives: how obligations net against each other, what counts as an event of default, and how positions are closed out if one party fails. A single ISDA, signed once, covers every future forward, swap, option and NDF (non-deliverable forward — a cash-settled forward on a currency that cannot be freely delivered) between the pair, so the legal work is done up front rather than trade by trade.
The CSA is the collateral schedule bolted onto the ISDA. The Credit Support Annex specifies how the two parties collateralise their exposure to each other: who posts margin (cash or securities pledged to cover a position's value), in what currencies and instruments, how often it is exchanged, and what thresholds of unsecured exposure are tolerated before collateral must move. Post-crisis uncleared-margin rules — regulations requiring collateral on derivatives not settled through a clearing house — made the CSA central rather than optional, because it now determines the real economic cost of trading with a given client.
KYC, the legal-entity identifier and the credit line turn a willing counterparty into an approved one. KYC (know-your-customer) is the set of checks that verify who the client is and that it is permitted to trade; each entity is tagged with an LEI (Legal Entity Identifier — a global code that uniquely names a trading party). The bank then sets a credit line — a ceiling on how much exposure it will carry to that client — sized to the client's expected trading, and loads it into the pre-trade credit system so that every incoming order can be checked against remaining limit in real time before it is accepted.
For a corporate, readiness also means a hedging policy and the systems to run it. A board-approved hedging policy defines what exposures the treasury is allowed to hedge, how much of each, and with which instruments — the guardrails that keep hedging from becoming speculation. A treasury-management system (the software that records the company's exposures and books its trades) feeds those exposures into the process and captures the resulting trades. Without both, a corporate cannot demonstrate that a given trade is a sanctioned hedge rather than an unauthorised bet.
All of this is a technology and data programme, not a filing exercise. The ISDA and CSA terms, the credit line and the entity identifiers are useless as paper in a drawer — they only work if they are represented accurately in systems, because a trade physically cannot be priced without knowing the CSA collateral terms, cannot be accepted without a live credit check against the limit, and cannot be margined without the CSA thresholds loaded. Onboarding therefore produces reference data (the standing facts a trade is checked against) that the entire lifecycle depends upon.
Onboarding is where the trade lifecycle's foundation is laid as data. The ISDA terms, CSA thresholds, credit limits and LEIs captured during onboarding become the reference data that pricing, credit-checking, margining and settlement all read on every subsequent trade — a single client record consulted thousands of times. That data decays: credit lines are reviewed and change, CSA terms are renegotiated, LEIs lapse and must be renewed, and legal entities restructure. If the reference data is not actively maintained, trades get mispriced, credit-checked against stale limits, or margined on wrong terms — so onboarding is not a one-time gate but a data set the whole lifecycle silently depends on staying current.
Spot FX is the deliberate exception that proves the rule. A spot trade — an outright exchange of one currency for another, settled within a day or two — is a cash transaction with almost no forward credit exposure, so it needs far less of this machinery: much lighter documentation, minimal ongoing margin, and a small credit footprint. The heavy readiness apparatus exists because derivatives create exposure that lives for weeks, months or years, and that lifespan is exactly what the ISDA, the CSA and the credit line are built to govern.
2.1 Internalisation
In one minute your desk sells EUR 60m to client A and EUR 15m to client C, and buys EUR 45m from client B. What is the desk's net position, how much can it internalise, and how much must it hedge externally?
The desk sells 60m + 15m = 75m (short EUR 75m) and buys 45m (long EUR 45m), so the net position is short EUR 30m. It internalises the 45m that offsets, and must hedge the residual EUR 30m externally. The internalisation ratio on the sell side is 45 / 75 = 60% — and only the 30m residual carries external cost and market impact.
2.2 Skew
Your book is long EUR 40m and you would rather be flat. Which side of your two-way price do you make keener — the bid or the offer — and why?
You are long EUR 40m and want to sell, so you make the offer keener (lower it) to attract buyers who take euros off your hands and flatten the book. You shade the price to pull in the flow that helps you, without moving the mid.
2.3 Algo choice
A client must buy EUR 100m over the next five hours with minimum market impact and no urgency. Which algo style fits, and at what average hourly rate would a simple TWAP execute?
No urgency, minimise impact → a patient, liquidity-seeking or TWAP/POV style, not an aggressive sweep. A simple TWAP over five hours executes EUR 100m / 5 = EUR 20m per hour, spread evenly to blend into the market.
2.4 Last look
An LP rejects your trade only when the market has moved in your favour during the check window, and accepts when it has moved in the LP's favour. Is this consistent with the FX Global Code? State the principle at stake.
No — this is asymmetric last look and breaches the FX Global Code. Last look may be used only as a genuine, symmetric risk control against stale prices, not to reject solely when the move favours the client or to profit from the check window. The principle at stake is fair dealing and transparency around last look.
2.5 Fixing
You must buy EUR 500m at the 4pm fix, currently 1.0850. If concentrating your buying in the window pushes the printed fix up by 2 pips, how much extra do you pay — and who ultimately bears it?
Two pips on EUR 500,000,000 = 500,000,000 × 0.0002 = USD 100,000 extra. It is ultimately borne by whoever references that fix — typically the clients whose orders settle at the pushed rate — which is why concentrating flow to move a fix is a conduct breach, not clever trading.
Explain internalisation to someone who has run a busy theatre coat-check — coats coming in and going out all evening — and push the analogy until it breaks. Where does the coat-check stop capturing what an FX dealer actually does with the risk it takes in?
What are the seven stages every FX trade passes through?
Request or stream, price, execute, risk and hedge, confirm and match, settle, and report and reconcile.
What does a pricing engine do beyond quoting a mid?
It adds a spread by size, skews the price to attract flow that flattens the book, tiers by client, checks risk limits and streams in microseconds.
What is skew, and why does a desk use it?
Shading the two-way price to one side to attract the flow that reduces the desk's own position.
What is last look, and when does it become abusive?
A brief final check before confirming a trade; it turns abusive when applied asymmetrically — rejecting only when the move favours the client — or used to trade ahead.
Why is a benchmark fixing conduct-sensitive?
Because it is computed from trades in a short window, concentrating flow inside the window can move the printed rate — the mechanism behind the fixing-manipulation cases.
Why does internalisation dominate FX economics?
Matching one client's buy against another's sell in-house captures spread on both sides with no external hedging cost or market impact — the single biggest driver of profit.
2
What is skew?
Here it is a pricing tactic.
No.
Right.
That is a spread change.
When does last look become abusive?
It is permitted when fair.
It can be abused.
Right.
No.
What does a pricing engine do beyond quoting a mid?
Right.
No.
No.
Back office.
Which is NOT one of the seven stages of an FX trade?
A stage.
Right. That is a primary-market step.
A stage.
A stage.
3Roles, Actors and Interactions
Who is who around an FX trade — the liquidity providers, the clients, the prime brokers and the people inside the bank — and why the dealer sits on one side of almost every interaction.
FX is an ecosystem, not a venue: the web of players around a trade
Every FX trade is a meeting of several institutions, each playing a defined role, and the bank sits in the middle of almost all of them. A retail equity trade routes to an exchange and clears through a central counterparty; the venue is obvious and singular. FX has no such centre. When a French exporter sells dollars, the trade may be priced by a bank's engine, routed through a multi-dealer platform, hedged against liquidity sourced from an electronic communication network, credit-cleared through a prime broker, and finally settled in a utility owned by a consortium of banks. Six different kinds of institution can touch one ticket. Reading the FX market means knowing who plays which part in a given trade, and — just as important — whose credit and whose systems make the trade possible at all.
The six kinds of player, and why one firm can be several at once
The first player is the dealing bank, the sell-side market maker that quotes a two-way price and stands ready to trade. A market maker is a firm that continuously shows both a bid (the price at which it will buy) and an offer (the price at which it will sell), earning the spread between them and warehousing the resulting risk. The large dealing banks — a group often labelled the "top tier" of global FX — are the gravitational centre of the market. They price to clients, they price to each other, and they run the pricing engines that set the tone for everyone downstream. When this module says "the bank," it usually means a dealing bank of this kind.
The second player is the non-bank market maker, the principal trading firm that makes prices without a banking licence. A principal trading firm — often abbreviated PTF — is a technology-led company that trades its own capital, quoting prices algorithmically at very high speed. Over the last fifteen years, firms of this type moved from the equity and futures worlds into FX and now supply a large share of the liquidity on electronic venues. They hold no client deposits and take no client orders in the traditional sense; their edge is speed, models and inventory management. A dealing bank both competes with these firms for flow and, frequently, sources liquidity from them.
The third player is the client, and "client" spans a wider range of institutions than any other category. A corporate treasurer hedging invoice exposure is a client. So is a "real money" account — an asset manager, pension fund or insurer investing for the long term — rebalancing a currency-hedged bond portfolio. So is a hedge fund taking a directional view, a smaller bank that cannot make its own prices in every pair, a central bank managing reserves, and a retail individual clicking on a broker's app. What unites them is that they take prices rather than continuously make them. What separates them is size, sophistication, and — crucially for a market maker — the quality of the flow they bring, a point developed later in this part under the heading of flow toxicity.
The fourth player is the venue, the electronic platform where prices and orders meet. Venues come in several shapes. A multi-dealer platform, or MDP, lets a client request or stream prices from many banks at once and pick the best — examples include 360T, FXall and Bloomberg's FXGO. A single-dealer platform, or SDP, is one bank's own storefront, where its clients trade on that bank's prices alone. An electronic communication network, or ECN, is an anonymous order book where banks and PTFs meet each other — EBS and LSEG Matching are the classic interbank ECNs. An aggregator is a piece of software, sometimes run by the bank itself, that pulls prices from several venues and liquidity providers into one consolidated view. The same institution navigates all of these at once.
The fifth player is the prime broker, the large bank that rents out its credit and its name. Prime brokerage in FX lets a smaller fund or trading firm face the market under the prime broker's credit standing rather than its own. The client trades with an executing dealer, and the trade is then "given up" to the prime broker, who steps in as the counterparty to both sides, monitors credit, and handles margin. This is what lets a hedge fund with a modest balance sheet trade with dozens of banks it could never open direct credit lines with. The prime broker earns fees and financing, and takes on the job of watching that the client does not blow through its limits.
The sixth player is the infrastructure — the settlement utility, the overseers, and the vendors that make the market run. CLS, the settlement utility we examine in detail later, removes the risk that one side of a trade pays and the other does not. Central banks and the Global Foreign Exchange Committee — a body of public and private participants that publishes and maintains the FX Global Code of conduct — provide oversight and the standard of good behaviour. Around all of them sit data and technology vendors selling price feeds, connectivity, analytics and platforms. None of these trades directly, yet without them there is no market.
The reason the ecosystem is hard to read is that one firm is usually several of these players at the same time. A large European bank is a dealing bank when it quotes a corporate, a client when it hedges its own book on an ECN, a prime broker when it rents credit to a fund, and a venue user across all of the platforms above — sometimes within the same minute. There is no fixed org chart of the market. There is only, for a given trade, a specific set of roles that specific institutions happen to be playing, held together by specific lines of credit and specific systems.
Modern e-FX microstructure: who provides liquidity, and how it flows
"Microstructure" is the study of how a market is actually wired underneath the price — who supplies liquidity, through what venues, and on what terms. In FX the wiring is intricate, and getting a clear picture of it is the difference between thinking the market is a simple ocean of prices and understanding that it is a set of specific pools connected by specific pipes, some of which quietly loop back into each other. Let us trace the plumbing.
Liquidity providers, both bank and non-bank, stream the market's prices
A liquidity provider is any firm that continuously streams two-way prices that others can trade against — it stands ready to buy and to sell, and it makes its living on the spread and on managing the resulting position. For most of FX history the liquidity providers were banks. A handful of the very largest global banks — the top tier of the market — supply a disproportionate share of the depth in the major pairs; a small number of names account for a large slice of all spot volume. That concentration is not an accident, and by the end of this part you will see exactly why scale compounds in FX.
But the banks no longer have the pool to themselves. Over the past fifteen years a new species of liquidity provider has grown up: the non-bank market maker, also called a principal trading firm. These are technology-first trading houses that make markets with their own capital, using sophisticated quantitative models and very fast infrastructure, but without a traditional banking franchise of corporate clients behind them. In spot FX in the major pairs, the best of these non-bank firms now rival the banks as liquidity providers — in some venues they are among the tightest and deepest sources of price. Their rise matters because it changed the character of the market: liquidity increasingly comes from firms whose entire edge is speed and modelling, which raises the bar for everyone and sharpens the questions about flow quality we turn to next.
ECNs, aggregation and internalisation are the plumbing of modern flow
An ECN — an electronic communication network — is an anonymous electronic pool where many participants post and take prices without knowing who is on the other side. Think of it as a shared order book that no single bank owns. Liquidity providers stream prices into it, price-takers hit those prices, and the venue matches the two and reports the fill, with counterparties anonymised behind the venue's own name and credit arrangements. ECNs are where a lot of the wholesale, bank-to-bank and bank-to-non-bank price discovery happens. They are pools; the prices in them come from the liquidity providers who choose to stream there.
Aggregation is the act of combining many separate price streams into one composite view of the market. A serious FX participant does not consume one liquidity provider's stream; it consumes dozens — many individual LPs plus several ECNs — each arriving as its own feed with its own bid, offer and available size. An aggregator stitches these together into a single, unified book: the best bid across all sources, the best offer across all sources, and the depth stacked behind each. Done well, aggregation gives you one clean, executable picture of the whole market instead of dozens of partial ones. Done badly, it fools you — and we will see shortly why.
Internalisation is the quiet giant of modern FX: matching a client's buy against another client's sell inside your own book, so the trade never touches the outside market. When one client sells a hundred million euros and, minutes later, another client buys a hundred million euros, a dealer with enough flow can simply hand the second client's euros to satisfy the first, keep the spread on both sides, and never go out to an ECN or an LP at all. The external market never sees the trade. Internalisation is so central to how modern FX makes money that we devote the final section of this part entirely to it. For now, hold onto the idea that the plumbing is arranged so that as much flow as possible is matched in-house before any of it leaks out to the wider market.
Flow quality and toxicity decide who gets a good price
Not all order flow is equal, and a liquidity provider that ignores this goes broke — so it does not ignore it. "Flow quality" is the term for how a given client's trading tends to behave after the trade. Benign, or "franchise", flow comes from clients who trade for reasons unrelated to short-term price prediction: a corporate hedging an invoice, an asset manager rebalancing a portfolio, a traveller changing money. When you trade with them, the market is as likely to move in your favour afterward as against you. That flow is welcome; a dealer will quote it tightly, because on average it is profitable to warehouse.
Toxic flow is the opposite: sharp, informed trading that reliably runs against you. Some participants — certain fast quantitative funds, certain arbitrageurs — trade precisely because they have detected that the price is about to move. When you sell to them, the price tends to rise right after; when you buy from them, it tends to fall. Fill enough of that flow and you bleed. A liquidity provider watches the "mark-out" on every client — how the market moved in the seconds after each trade — and builds a profile. Clients whose flow reliably runs against the LP get quoted wider spreads, get shown less size, or in the extreme get declined altogether. This is not a moral judgement about the client; it is risk pricing. The tighter you want to quote the good flow, the more disciplined you have to be about the toxic flow, because the tight prices are subsidised by the profitability of the benign clients.
Prime brokerage rents out credit so the buy-side can reach top-tier liquidity
There is a structural problem in FX that has nothing to do with prices and everything to do with credit, and prime brokerage is the elegant solution the market built to fix it. To understand prime brokerage you first have to understand the problem it solves, because the solution looks strange until you see the problem clearly.
The problem PB solves: you need credit to trade with a top-tier LP
To trade with a top-tier liquidity provider, you need a bilateral credit relationship with it — and a mid-sized fund cannot possibly maintain one with every LP it wants to reach. When two parties agree an FX trade, each takes on the risk that the other fails to deliver its side; that risk is managed through credit lines, legal agreements and collateral arrangements that take time and balance sheet to establish. A large global bank will only face a counterparty it has vetted and extended credit to. A hedge fund that wants the tightest prices needs to reach a dozen or more top-tier LPs — but no fund can realistically negotiate and maintain full bilateral credit relationships with all of them, and the LPs may have no appetite to extend credit to a smaller, less-known fund at all. The fund is locked out of the best liquidity not because of its prices but because of its credit standing.
An FX prime broker — a large bank — solves this by lending the fund its own name and creditworthiness. The fund establishes a single relationship with the prime broker. It then trades in the market with many different LPs, but in the prime broker's name. The mechanism is called give-up: each trade the fund executes is immediately "given up" to the prime broker, who steps in as the counterparty to both sides — facing the LP on one side and the fund on the other. The LP is content because it is facing the prime broker's rock-solid credit, not the fund's. The fund is delighted because, through one relationship, it now reaches the entire top tier of liquidity. And the prime broker earns fees and financing spreads for the service, while taking on the job of managing the fund's risk, margin and limits. Everyone gets what they need; the fund gets liquidity, the LPs get credit certainty, the PB gets a business.
Credit is the product, and the risk is real
What the prime broker is actually selling is not access or technology but its balance sheet and its credit rating — and that means it is taking on genuine, sometimes dangerous, counterparty risk. When the PB faces both sides of every trade its client does, the PB is on the hook if the client fails. If a client blows up — say a fund that is heavily positioned when a currency makes a violent, gapping move — and cannot meet its obligations, the prime broker is left holding the trades and the loss. The product being rented is creditworthiness, and creditworthiness rented out is risk retained.
Prime brokers manage that risk with hard, real-time controls, and a breach can cut a client off instantly. The PB imposes margin requirements (collateral the client must post), position limits (caps on how large a position the client may build) and give-up limits (caps on how much the client may trade with any given LP under the PB's name). These are monitored continuously, and if a client breaches them the PB can stop accepting its trades immediately, severing the client's access to the market in real time. That is the PB's protection, and it is also a reminder of how tightly leashed a prime-brokered client actually is.
Because a few large banks dominate FX prime brokerage, the risk is concentrated, and a single blow-up can hurt. The FX PB business is not spread evenly across the industry; a handful of major banks intermediate the bulk of it. That concentration means the failure of a large, highly-leveraged client during a violent market move does not just hurt that client — it can inflict a serious loss on its prime broker, and history contains episodes where a sudden, extreme currency move left prime brokers nursing large losses from clients who could not pay. It is a genuine tail risk: rare, but capable of being severe when it arrives. The business of renting out credit is, in the end, the business of hoping the rent never comes due all at once.
The people who run FX are increasingly quants and engineers
The trading floor of the popular imagination — shouting voices, phones wedged under chins — has been quietly replaced by something closer to a software organisation. The people are still there, and the best of them are paid handsomely, but the centre of gravity has shifted. The person who used to quote spot prices by hand now supervises the engine that quotes them. The scarce skill is no longer a fast mouth on a phone; it is the ability to design, tune and control the systems that price and hedge thousands of times a second. To understand the modern FX business, you have to understand who sits around the desk and what each of them actually owns.
The roles on and around the desk
The spot trader manages the spot book, but increasingly oversees the pricing engine rather than quoting by hand. In liquid pairs, the trader no longer types a price in response to each request; the engine does that automatically thousands of times a second. The human's job becomes managing the residual risk the engine cannot fully hedge, setting the parameters the engine runs on, watching for abnormal conditions, and stepping in when something breaks. The trader is now part risk manager, part engine operator. The instinct for when a market is about to move still matters; the medium through which that instinct is expressed has changed entirely.
The forwards and swaps trader runs the curves, the points and the funding, and this work stays more manual and relationship-driven. As Part 1 explained, a forward price is built from the spot rate plus an interest-rate differential expressed as "points." Constructing and managing those curves — especially in longer-dated tenors and in emerging-market currencies where the market is thin — resists full automation. Pricing a five-year cross-currency swap for a corporate is a negotiation informed by funding costs, credit and relationship, not a click on a screen. This corner of the desk remains where markets knowledge and human judgment concentrate.
The e-FX trader, or quant, builds and tunes the pricing, the skew and the auto-hedging logic — this is the engine room. A quant here is not a distant researcher; the role sits close to live risk. This person decides how the engine sets spreads by client tier, how it skews prices to attract the flow it wants and repel the flow it does not, when it internalises a client trade against another rather than hedging it externally, and when it fires an automatic hedge into the market. When the engine's behaviour is right, the desk quietly makes money; when it drifts, the desk quietly bleeds. The e-FX quant owns that behaviour.
FX sales covers the clients — the human relationship laid over the electronic pipes. Even as execution moves onto screens and APIs, someone has to win the client, understand the client's business, structure the harder trades, and be the voice when a corporate treasurer wants to talk through a hedge. Sales does not typically make prices; it carries the franchise. The relationship a salesperson holds with a corporate treasury is often what determines whether that corporate's flow arrives on the bank's platform at all — and franchise flow is the good flow.
The e-commerce chief operating officer, or head of e-FX, runs the electronic business as a product. This is a role that barely existed a generation ago. It treats the bank's electronic FX offering the way a technology company treats a product: which venues to connect to, how to onboard clients quickly, how to build and maintain connectivity, what analytics to give clients, and how the platform itself makes or loses money. The head of e-FX owns venue strategy and platform profit-and-loss, and sits at the seam between the trading desk and the technology organisation.
Compliance polices conduct, and IT and data build and run the machinery. Compliance enforces the FX Global Code, runs surveillance over trader communications and behaviour, and evidences best execution — the obligation to obtain the best available result for a client. IT and data build and operate the pricing engines, the connectivity, the risk systems and the regulatory reporting. In a business this automated, these functions are not support; they are load-bearing. A pricing engine with bad data is a liability, and a desk that cannot evidence its conduct is a scandal waiting to happen.
The shift from voice traders to a software organisation
As spot execution automated, the classic spot-trader headcount shrank and the e-FX, e-commerce and technology roles grew in its place. This was not a one-off cut; it was a structural reweighting of the desk. A single well-designed pricing engine can do the price-making work that once took a row of traders, and it never sleeps, never fat-fingers a price under pressure, and prices consistently across thousands of clients. The value did not disappear — it migrated from the person quoting to the people building and governing the thing that quotes. The modern FX desk is a blend of two disciplines that used to sit far apart: deep markets expertise and serious software engineering. The firms that win are the ones that fuse the two rather than bolting a technology team onto a trading floor.
| Role | What they own | Primary technology |
|---|---|---|
| Spot trader | The spot book's residual risk; engine parameters and oversight | Pricing engine, risk/position system |
| Forwards & swaps trader | Curves, points and funding, especially long-dated and EM | Curve-building and pricing tools, risk system |
| e-FX trader / quant | Pricing, skew, internalisation and auto-hedging logic | Pricing & distribution engine, internalisation logic |
| FX sales | The client relationship and the franchise flow | SDP, CRM, sales-facing pricing tools |
| e-commerce COO / head of e-FX | Venue strategy, onboarding, connectivity, platform P&L | SDP/MDP connectivity, analytics, TCA |
| Compliance & IT/data | Conduct, the Code, surveillance, best execution; the systems themselves | Surveillance, reporting, pricing, connectivity, data |
The client is not one person: inside a corporate treasury and inside a fund
"The client" is a convenient fiction — behind the word sit several people whose incentives pull in different directions. A salesperson who treats the client as a single mind will price to the wrong person and lose the trade. The skill is to know which individual is asking, what that person is optimising for, and — separately — who has to sign off. Those are often not the same person, and the gap between them is where deals are won, stalled, or lost.
Inside a corporate treasury, four actors share the room and rarely agree. The group treasurer runs the day-to-day management of the company's cash and currency risk and wants certainty and a clean audit trail — a hedge that is defensible after the fact. The CFO (chief financial officer) sits above the treasurer and wants predictable earnings and low cost, which is not the same as certainty at any price. The board or audit committee wants policy compliance and no surprises — that the treasury stayed inside its board-approved mandate. And FP&A (financial planning and analysis) together with the operating business units want the hedge not to distort their reported numbers, because a hedge that swings the division's results creates explaining to do. Four people, four different definitions of a good outcome.
Inside a fund, a parallel set of actors divides the same way. The portfolio manager (PM), who owns the investment decision, wants the position — the exposure on, now, at the size the thesis calls for. The execution trader, who actually works the order into the market, wants best execution and low market impact, which usually means going slower than the PM would like. The risk function wants limits respected, regardless of how attractive the trade looks. And operations wants clean settlement — trades that book, confirm and settle without breaks. Each is doing their job well by pulling against the others.
These incentives collide in predictable, exploitable ways. The PM wants size immediately; the execution trader wants to work it slowly to reduce the market impact that the depth curve imposes on a large order — so the same fund fights itself over urgency. In the corporate, the treasurer wants a forward (a locked-in future rate, giving certainty) while the CFO eyes an option (the right but not the obligation to trade at a set rate, preserving upside for a premium) — so the same company fights itself over certainty versus flexibility. A good salesperson does not resolve these tensions from outside; they recognise which side is speaking and shape the pitch to it.
| Actor | What they optimise for | Where it can collide |
|---|---|---|
| Group treasurer (corporate) | Certainty, clean audit trail | Wants a forward for certainty; CFO wants an option for upside |
| CFO (corporate) | Predictable earnings, low cost | Resists paying option premium the treasurer would spend on flexibility |
| Board / audit committee (corporate) | Policy compliance, no surprises | Blocks trades outside mandate that treasury sees as commercially sensible |
| FP&A / business units (corporate) | Hedge not distorting their numbers | Resist hedges that swing divisional results even if group-optimal |
| Portfolio manager (fund) | Getting the position on, in size, now | Wants size immediately; execution trader wants to work it slowly |
| Execution trader (fund) | Best execution, low market impact | Slows the PM's order to cut impact, delaying the position |
| Risk / operations (fund) | Limits respected, clean settlement | Blocks or trims a trade the PM wants; enforces limits over conviction |
3.1 Client type and pricing
Two clients ask for the same EUR/USD price: a corporate treasurer hedging a quarterly invoice, and a fast systematic fund that reliably trades just before the market moves. Which gets the tighter spread, and why?
The corporate treasurer gets the tighter spread: benign, predictable franchise flow the desk can internalise. The systematic fund brings toxic, informed flow that reliably runs against the desk, so it is priced wider or declined.
3.2 Prime-brokerage give-up
A hedge fund executes with an LP, but the trade is "given up" to its prime broker. Who is the LP's credit counterparty on that trade, and what does the prime broker earn for standing in?
The prime broker is the LP's credit counterparty — the trade is given up to the PB, which faces both sides. The PB earns fees and financing for renting its credit and taking the counterparty risk if the fund fails.
3.3 PB margin
A fund runs a EUR 200m position through its prime broker, which requires 5% initial margin. How much collateral must the fund post?
5% × EUR 200m = EUR 10m of initial margin.
3.4 Actor conflict
Inside a fund, the portfolio manager wants to buy EUR 300m immediately while the execution trader wants to work it over the afternoon. State the trade-off each is optimising for.
The portfolio manager optimises for the position and return (get the exposure now); the execution trader optimises for best execution and low market impact (work it slowly). The tension is immediacy versus impact.
3.5 Franchise vs toxic flow
Classify each and say how the desk prices it: (a) a corporate's regular monthly hedge; (b) a latency-arbitrage fund picking off stale quotes.
(a) the corporate's monthly hedge is franchise flow — priced tight and internalised; (b) the latency-arbitrage fund is toxic flow — it picks off stale quotes and loses the desk money, so it is priced wide or declined.
Explain to a small-business owner who banks at a single local branch why a hedge fund needs a "prime broker" in order to trade with many different banks at once. Use the idea of a letter of introduction or a guarantor, and no trading jargon.
Why can one institution be several actors at once in FX?
The same bank can stream prices as an LP, take prices as a client, intermediate credit as a prime broker and use venues as a taker — roles defined per trade, not per firm.
How does the type of client change the price a desk shows?
Benign, steady franchise flow is quoted tighter and internalised; sharp, informed, toxic flow is priced wider or declined.
What problem does prime brokerage solve, and what does the prime broker actually rent?
It lets a fund reach many top-tier LPs through one relationship; the prime broker rents its balance sheet and credit standing, taking the counterparty risk.
How did the desk's roles change as FX automated?
The classic voice spot trader shrank while e-FX quants, e-commerce owners and technologists grew; the modern desk blends markets expertise and software engineering.
Why do desks prize long-only and real-money flow?
It is steady, lower-toxicity and easier to internalise against, unlike fast, price-sensitive professional flow.
Carrying forward from Part 1: a client without interbank credit still wants top-tier prices in the majors. From Part 1's tiering, which route reaches that liquidity, and which actor here makes it possible?
The prime-brokerage route: the fund trades in the prime broker's name, on the prime broker's credit.
Carrying forward from Part 2: why does a desk with a large, diverse franchise internalise more than one with narrow flow?
A diverse two-way flow lets client buys offset client sells in-house, so less residual has to be hedged externally.
3
What does an FX prime broker rent?
No.
Right.
No.
No.
How is toxic, informed flow treated?
No.
No.
Reversed.
Right. Benign flow is quoted tighter and internalised.
Why can one institution be several actors at once?
No.
It can.
Right.
No.
How did the desk's roles change with automation?
No.
No.
The opposite.
Right.
4Systems, Data and Technology
The technology that carries a trade front to back: the stack, the data, the messaging, the settlement plumbing, and the rules that shape every system in the chain.
The system landscape runs front to back, ideally with no re-keying
An FX trade passes through a stack of systems, and the ambition of the whole design is that a human never re-types the trade at any handover. Every time a person copies a trade from one system into another, they can transpose a digit, drop a leg, or delay a booking. In a market that trades in inches and settles in billions, those errors are expensive and occasionally fatal. So the architecture is built around a single idea: capture the trade once, correctly, and let it flow through pricing, risk, confirmation and settlement without human re-keying. That ambition has a name — straight-through processing — and the whole system landscape is organised to serve it.
The six layers a trade flows through
The first layer is pricing and distribution: the engines and streaming APIs that publish quotes. A pricing engine takes market data and the bank's own risk position and produces a stream of bids and offers, tiered and skewed by client. Distribution is how those quotes reach the outside world — pushed to single-dealer and multi-dealer platforms, and streamed over application programming interfaces, the documented connection points through which a client's own systems can consume prices directly. This layer is the shop window and the price factory at once.
The second layer is aggregation and connectivity: the software that gathers external liquidity and moves messages. An aggregator consolidates prices from many venues and liquidity providers into one view so the desk sees the best available market. A FIX gateway is the plumbing that speaks FIX — the Financial Information eXchange protocol, the standard messaging language of electronic trading front offices — to venues and counterparties. Venue links and smart order routers decide where an order should go to be filled at the best price with the least market impact. This layer is the nervous system connecting the bank to the wider market.
The third layer is e-trading and execution: the systems that capture the trade, apply controls, and book it. An order management system, or OMS, and an execution management system, or EMS, take the decision to trade and turn it into a live order and then a completed trade. This is where "last look" — the brief window in which a market maker may verify a price before accepting a trade, examined in Part 2 — is applied, where pre-trade risk checks and credit limits are enforced, and where the trade is finally written into the books. Nothing downstream is correct if this layer captures the trade wrong.
The fourth layer is risk and position management: the book that updates live. The moment a trade is booked, the bank's position changes, and the risk system must reflect it in real time. This is where internalisation is measured — the netting of one client's buy against another's sell — where skew is informed by the current inventory, and where auto-hedging fires when exposure breaches a threshold. A spot trader who cannot see the true position in real time is flying blind, so this layer is the instrument panel of the whole desk.
The fifth layer is confirmation and matching: the electronic agreement of what was traded. Confirmation is both sides formally agreeing the economics of the trade — the pair, the amount, the rate, the value date. Matching is the automated reconciliation of the two sides' records so that any disagreement surfaces immediately rather than on settlement day. In electronic FX this happens in seconds. In higher-touch products it can still involve manual steps, and every manual step is a place a break can hide.
The sixth layer is settlement and post-trade: the actual exchange of cash and the records that follow. Settlement is where the two currencies change hands, largely through CLS, on the value date. Around it sit regulatory reporting, reconciliation of positions and cash against the bank's own books and its nostro accounts (the accounts a bank holds in other currencies at other banks), and profit-and-loss attribution — the work of explaining exactly where the day's money came from. This layer is the least glamorous and the least forgiving; it is where the trade finally becomes real money or a real loss.
Cutting across all six layers is a band of capabilities that no single layer owns. Market data feeds every layer and must be clean, timely and correctly licensed. Reference data — the identifiers and static details of currencies, counterparties and instruments — has to be consistent everywhere. Credit and limits sit as checks at multiple points. Surveillance and audit watch the whole flow for misconduct and keep the record regulators will demand. These cross-cutting concerns are why an FX platform is never just a chain of boxes; it is a chain of boxes standing on a shared foundation of data and control.
Straight-through processing is the goal, and the fragmented edges are where it fails
In liquid spot, the whole chain from price to settlement runs end to end with no human hand on it, and that is where straight-through processing is essentially solved. A EUR/USD trade for a standard amount is priced, executed, risk-managed, confirmed, matched and settled by systems talking to systems. The human involvement is oversight, not operation. This is the industrialised heart of FX, and it works.
The effort now concentrates on the harder edges: the less liquid products and the fragmented data. Long-dated forwards, exotic options, restricted emerging-market currencies and bespoke structured trades still carry manual steps, because the products resist standardisation and the data is patchy. Fragmentation itself is an obstacle — with no single market data source and dozens of venues speaking slightly different dialects, gluing the picture together cleanly is genuine work. Straight-through processing in FX is not a finished project; it is a solved problem in the liquid core and an ongoing campaign at the periphery, and the periphery is where operational risk and cost still live.
Following one trade front to back: the data lineage
Trace a single trade through the stack and you see that each system creates a data object which the next system depends on, so an error at any point flows to every point after it. Data lineage is simply the documented path a piece of information travels — where it was born, what it became at each stage, and where it ended up. In FX this lineage is unusually consequential, because the objects are money and the final reader is a regulator. A wrong number near the front does not stay near the front. It rides the chain all the way to settlement and to the regulatory record, and the further it travels undetected, the more expensive it is to unwind.
The cascade principle: an error upstream reaches every system downstream
Walk the chain once and the cascade principle becomes obvious. The pricing engine creates the quote — a bid and an offer at a moment in time. The execution system, or OMS, turns an accepted quote into the trade record: this pair, this amount, this rate, this counterparty, this value date. The risk system reads that trade record and updates the position, changing what the desk believes it owns. The confirmation and matching layer turns the trade record into a matched trade, agreed by both sides. CLS turns the matched trade into a settlement, moving the two currencies. Reporting turns the trade into a regulatory record and an audit trail that the authorities can inspect.
Now inject one error and watch it travel. Suppose the trade is booked at the wrong rate, or one leg of a swap is mis-entered. The risk system now shows a position the desk does not truly hold, so hedging decisions are made against a phantom. Confirmation may catch it if the counterparty disagrees — or may not, if both sides captured the same wrong detail. If it survives to settlement, the wrong amount of cash moves, and clawing it back is a manual, cross-border, cross-timezone ordeal. And the regulatory report, built from the same tainted record, is now wrong too, which is its own problem with its own penalties. One bad object upstream becomes a bad position, a bad settlement and a bad report downstream. That is why controls cluster at the points where objects are created, and why the identifiers that pin each object down matter so much.
| Stage | System | Data object created |
|---|---|---|
| Pricing | Pricing & distribution engine | The quote (bid/offer at a timestamp) |
| Execution | OMS / EMS | The trade record (pair, amount, rate, value date, counterparty) |
| Risk | Risk & position system | The updated position |
| Confirmation | Confirmation & matching | The matched trade |
| Settlement | CLS | The settlement (the exchange of both legs) |
| Reporting | Reporting engine | The regulatory record and audit trail |
Identifiers pin down the currency, the entity and the trade
A trade record is only as trustworthy as the identifiers that name its parts, and FX relies on a small set of them. An identifier is a standardised code that means exactly one thing to every system that reads it, removing the ambiguity of free text. Get the identifiers right and systems reconcile automatically; get one wrong and the cascade above begins.
ISO 4217 three-letter codes name the currencies. EUR is the euro, USD the US dollar, JPY the Japanese yen. These codes are so universal that they are easy to take for granted, but they are the atoms of every FX record; a pair is just two of them. The Legal Entity Identifier, or LEI, names each legal counterparty for reporting. It is a global code that says precisely which legal entity — not which trading name or which desk — is on each side of the trade, which is exactly what a regulator needs to know. The Unique Trade Identifier, or UTI, names the individual trade so that both counterparties and the regulator reconcile to a single shared record rather than to two records that might not match. ISINs — International Securities Identification Numbers — appear for some listed or deliverable instruments and certain non-deliverable forwards, where the trade behaves enough like a security to warrant a security's identifier. Together these four codes turn a description of a trade into an unambiguous, machine-reconcilable, regulator-ready record.
| Identifier | What it names | Where it bites |
|---|---|---|
| ISO 4217 currency code | The currency (e.g. EUR, USD, JPY) | Every quote, trade and settlement |
| LEI (Legal Entity Identifier) | The legal counterparty | Regulatory reporting and reconciliation |
| UTI (Unique Trade Identifier) | The individual trade | Two-sided reconciliation and reporting |
| ISIN | Some listed/deliverable instruments and certain NDFs | Instrument-level identification and reporting |
No single tape: FX data is fragmented, costly, and a genuine edge
We have referred repeatedly to the fact that FX has no single record of its own prices — now we confront it directly, because this one structural fact shapes the entire market and turns data itself into a source of competitive advantage. Understanding why FX data is hard is understanding why the best-informed desks are also the best-pricing desks.
Why FX data is hard
FX has no consolidated tape, and that absence changes everything about how you know where the market is. In listed equity markets there is an official, consolidated record — a tape — of essentially every trade and every quote, so "the price" of a share is a known, published fact. FX has nothing equivalent. There is no authority that records every FX trade and price and publishes them as a single truth. As a result, "the market price" in FX is always an estimate — your best reconstruction of where the market is, assembled from whatever sources you can see. Two well-equipped desks looking at the same instant may disagree slightly about the mid, because they are looking at different, partial pictures.
The sources that do exist are fragmented, partial, and expensive. The primary ECNs and the major aggregators offer good real-time data, but each sees only its own slice of the market, and no single feed is complete. The best of these feeds — the primary-venue data that reveals where the core price is being discovered — is costly and tightly licensed; you pay real money for the right to consume it, and more to redistribute it. Filling the gaps at the daily level are the benchmark fixings and central-bank reference rates: agreed, published marks struck at set times that give the market a common reference point for valuation, accounting and settlement even though they say nothing about the price in any given millisecond. The data landscape is thus a patchwork of expensive partial real-time feeds and agreed periodic benchmarks, with no single complete truth anywhere.
Why better data is a differentiator
Because everyone is working from an estimate, the desk with the cleaner, faster, more complete estimate can quote both tighter and safer — and that is a durable edge. If your reconstruction of the true market is more accurate and less laggy than a rival's, you can afford to show tighter spreads without increasing your risk of being picked off, because you are more confident about where the real mid actually is. Conversely, every downstream system we have discussed inherits the quality of this data. The fair mid, the skew, the aggregated book, the auto-hedge — all of them are only as good as the data feeding them. Feed them a stale or distorted picture and they will quote and hedge confidently on false information. Garbage in, dangerous quotes out, at machine speed.
And one dataset no external feed can ever match is the bank's own client flow — which is why scale in franchise flow is also an informational advantage. When a large dealer watches its own clients trade, it sees real supply and demand: actual corporates buying dollars, actual funds selling euros, in real size, in real time. That is genuine information about where the market is heading, and no purchased data feed contains it, because it is proprietary to the franchise that generated it. A big bank effectively has a private window onto the market's true flow that its smaller rivals lack. Finally, all of this — the clean, unified, precisely time-stamped data — is the prerequisite for anything more sophisticated. You cannot build a predictive pricing model or a machine-learning system on messy, mistimed, double-counted data; the models will learn the noise. Clean data is the foundation on which the AI we discuss in Part 5 is built, and desks that neglected their data foundations will find they cannot build the next generation of tools at all.
Tick data and the quant pipeline are where the models actually live
Storing and querying billions of ticks needs specialised databases
Ordinary databases collapse under FX market data, because the data has a volume and a shape they were never built for. A tick is a single market event — a new quote or a trade — and a liquid currency pair generates thousands of ticks a second, tens of billions a day across a full universe of pairs and venues. A conventional relational database, optimised for transactional records and row-by-row updates, cannot ingest that firehose or answer the questions traders ask of it. Worse, almost every useful question about market data is a time-series question — "what was the volume-weighted price between 15:59:30 and 16:04:30?", "show me every tick within two milliseconds of this trade" — and a general-purpose database has no native notion of the time dimension that dominates these queries.
The industry answer is the specialised time-series, or tick, database, with the kdb+/q family as the de facto standard. kdb+ is a column-oriented database built for exactly this workload, and q is its terse, expressive query language; together they store and query billions of price and trade records with time treated as a first-class dimension rather than an afterthought. Column orientation means the engine reads only the fields a query needs, and the time-ordering of the data lets it answer window and as-of questions extremely fast. This store is the raw material for everything downstream: it feeds pricing (what should we quote?), skew (how should we lean our price given our inventory?), TCA (how good was our execution?) and quantitative research (what patterns can we find?). Without a database that can hold and interrogate the tick history, none of those functions has anything to work with.
From data to a live model: the MLOps pipeline
A quantitative model reaches the market through a disciplined pipeline that runs from raw ticks to a live quote and then loops back on itself. The pipeline has recognisable stages. First, feature engineering: from clean tick and flow data, the team computes the inputs a model will reason over — recent volatility, order-flow imbalance, the shape of the book. Second, training and validation: a model is fitted to historical data and then tested offline on data it has never seen, to check it generalises rather than memorises. Third, deployment: the validated model is pushed into production, where it informs skew, pricing or routing decisions in real time. Fourth, monitoring, which feeds a decision to retrain when the model's edge decays. This deploy-monitor-retrain loop, together with the tooling that makes it repeatable and auditable, is what the industry calls MLOps — machine-learning operations, the discipline of running models in production the way software engineering runs code in production.
The hard part is that a model informing a microsecond quote must be both fast to evaluate and tightly governed. A model that takes a millisecond to produce a number is useless on a path measured in microseconds, so models on the price path are engineered to evaluate in a handful of microseconds — which constrains how large and complex they can be. At the same time, a model that influences the prices a bank shows to clients is a source of risk, so it cannot be deployed on a quant's judgement alone; it must pass through model-risk governance — the controls, sign-offs and ongoing validation examined in Part 5. Most of this pipeline is proprietary and closely guarded, because it is the edge: two banks with the same tick data and different models will quote differently, and the difference is the whole competitive game.
The pipeline is only ever as good as its data and its guardrails, and both are easy to underestimate. A model trained on dirty data — mis-timestamped ticks, quotes that were never really tradeable, or records wrongly deduplicated so that real repeated trades were collapsed into one — will learn a distorted picture of the market and quote dangerously in production. A model with no monitoring is worse in a subtler way: it will drift silently as the market regime changes underneath it, still confidently producing numbers that were right last year and are wrong today. This is why the unglamorous work — tick storage, data cleaning, feature stores (managed, reusable repositories of engineered features), and continuous model monitoring — is where much of the real e-FX engineering effort actually goes. The headline model is a small part of a large plumbing problem.
Where FX actually trades: multi-dealer, single-dealer and interbank venues
FX does not trade in one place — it trades across a layered set of venues, and a serious bank participates in all of the layers at once, for different reasons. Understanding the venue landscape means understanding three distinct types of platform and why no major player can afford to ignore any of them.
MDPs, SDPs and ECNs, and why most banks use all of them
A multi-dealer platform, or MDP, is a client-facing venue where several banks compete to price the same trade, and it exists to give the client reach and best-execution evidence. On an MDP the client puts out a request and multiple banks respond with prices, or streams from multiple banks appear side by side, and the client trades the best one. The value to the client is competition made visible: several dealers fighting for the trade, and a clean record showing they took the best available price — which matters enormously for demonstrating best execution to their own stakeholders and regulators. Well-known MDPs in FX include 360T, FXall and Bloomberg's FXGO. For a bank, streaming to the MDPs is non-negotiable, because that is where a large share of client flow congregates and shops around.
A single-dealer platform, or SDP, is a bank's own portal or API showing only that bank's prices, and it exists to keep the relationship and differentiate on content. On its SDP the bank is not one of five competitors; it is the whole show. That would sound like a disadvantage — why would a client come to a venue with no competition? — except that the SDP is where the bank bundles everything that is not just a price: its execution algorithms, its analytics and TCA tools, its market research and commentary, its structuring capability, its post-trade services. The SDP is how a bank turns a commoditised price into a differentiated relationship. A client who values the bank's research, algos and service comes to the SDP; a client who only wants the tightest price on a vanilla spot trade shops the MDPs. Every bank wants to pull as much high-value flow as possible onto its own SDP, because flow that arrives there is stickier, more profitable, and richer in the franchise data we will discuss later.
An interbank ECN is the anonymous wholesale layer where dealers and non-bank market makers trade with each other, and historically two pools dominated price discovery in the majors. These are the venues where the market's core price is discovered — the primary pools. For decades, EBS Market and Refinitiv Matching, now under the LSEG umbrella, were the dominant anonymous primary venues for the major currency pairs, each tending to be the reference pool for particular currencies. Around and beyond those primary pools sits a wider ecosystem of non-primary streaming and aggregated venues, plus the retail and proprietary-trading platforms that serve smaller participants. The primary ECNs matter far beyond their own volume because the price discovered there feeds everyone else's fair-mid calculation.
Put the three layers together and the strategy is clear: reach *and* retention. A major bank streams to the MDPs and the ECNs to maximise reach — to be present wherever flow gathers and price is discovered — while simultaneously investing heavily in its own SDP and API to retain the high-value flow it most wants to keep. It is not a choice between the venues; it is a portfolio across all of them, each serving a different purpose in the same franchise.
In a microsecond market, the plumbing is a competitive weapon
When prices live for milliseconds and machines trade with machines, the physical infrastructure connecting you to the market stops being an IT cost and becomes a source of edge or of loss. This section is about latency — the time it takes information to travel — and about the connectivity engineering that exists to fight it. In a market this fast, the pipes matter.
Why latency matters, and the connectivity that answers it
A price you show is only as good as it is fresh, and a slow connection makes every price you show dangerously stale. If your quote reaches the client, or the client's acceptance reaches you, more slowly than your competitors' equivalent messages, two bad things happen. First, you get picked off: fast counterparties trade against your stale price before you can update it, exactly the risk that last look exists to blunt. Second, you miss trades: by the time your price arrives, someone faster has already won the flow. Latency directly affects P&L, and it directly affects your last-look rejection rate, because a slower pricer sees more stale-price situations and must reject more often, which annoys clients. Speed is not vanity; it is the difference between quoting safely and bleeding.
The first weapon against latency is physical proximity, achieved through co-location. Light and electrical signals travel fast but not infinitely fast, and over the distances between cities the delay is real and measurable. Co-location means placing your pricing and matching servers in the very same data centres that host the trading venues, so that the distance a signal must travel — and therefore the time it takes — is as small as physics allows. In FX the canonical locations are the data centres near the major financial hubs: LD4 in the London area and NY4 in the New York area are the classic examples, housing the matching engines of the primary venues and the servers of the participants who want to be right next to them. If your server is in the same building as the venue's, you are as close as anyone can be.
The second weapon is the way information travels once it leaves the server. Orders and prices move over messaging protocols; the industry standard is FIX (the Financial Information eXchange protocol), a common language that lets different firms' systems talk to each other. But FIX, being general and human-readable in origin, is not the fastest possible option, so the most latency-sensitive participants also use proprietary binary APIs — leaner, machine-optimised message formats that shave off precious microseconds. The choice of protocol is itself a performance decision.
The third weapon is network engineering, and it goes surprisingly deep. Firms lease dedicated fibre-optic routes and pay for the most direct paths between data centres, because a straighter cable is a faster cable. They use kernel-bypass networking, a technique that lets network traffic skip slow parts of the computer's operating system and reach the application faster. At the extreme, they push pricing logic into hardware — FPGAs, field-programmable gate arrays, are chips that can be configured to execute specific tasks far faster than general-purpose software — so that the most time-critical decisions happen in silicon rather than in code. And underpinning all of it, they run precise, synchronised clocks so that every event is timestamped accurately to a common reference — which is required both for meaningful TCA and to satisfy MiFID II's demands for accurate time-stamping of trades.
Speed lives on-premises; scale lives in the cloud
The latency-critical core cannot simply move to the cloud
The parts of the stack that must answer in microseconds stay on dedicated hardware placed physically next to the venues, and no amount of cloud marketing changes that. Pricing engines (the code that generates the streamed quote), matching engines (the code that pairs buy and sell orders), and smart order routers (the code that decides which venue to hit) all sit on the latency-critical path. For these, the enemy is not just how fast the software runs but how long a signal takes to travel. That is why banks and venues practise co-location: they rent rack space in the same data centres that host the trading venues themselves. In FX the canonical sites are LD4 in Slough, west of London, and NY4 in Secaucus, near New York — the two buildings where a large share of the world's electronic FX matching physically happens.
The binding constraint is the length of cable between your server and the venue's matching engine, and public cloud cannot shorten it. Light in fibre travels roughly two-thirds the speed of light in vacuum, so every extra metre of cable adds delay you can measure. A server in the same room as the matching engine is a few metres away; a public-cloud region is typically kilometres away and, worse, reached over a shared network whose delay varies trade to trade. That variability — jitter — is often more damaging than raw distance, because a pricing engine cannot plan around a delay it cannot predict. Public cloud also cannot, in general, place a rack of your machines inside the venue's own cage. For the microsecond core, then, bare-metal, co-located hardware is not a preference; it is a physical requirement.
Analytics, TCA and model training belong in the cloud
Everything that is not on the microsecond path benefits enormously from the cloud's elastic, pay-as-you-go compute. Transaction cost analysis (TCA — measuring, after the fact, how good execution was versus a benchmark), backtesting (replaying strategies against historical data), surveillance (scanning trading for abusive patterns), model training, and regulatory reporting all share a common shape: they are heavy but not time-critical. Nobody loses money because a TCA report finishes at 6:00 rather than 5:55. What these workloads want is the ability to summon vast compute briefly and then give it back. You can spin up hundreds of machines to train a pricing model overnight, use them for six hours, and release them at dawn — paying only for those hours rather than owning idle silicon year-round.
The result across the industry is a hybrid architecture, deliberately split by function rather than by accident of history. Bare-metal at the edge — co-located, tuned, expensive, always on — carries the price path where microseconds are money. Elastic cloud in the back carries the analytics, research and reporting where scale and cost flexibility matter and latency does not. The two halves are connected by pipelines that move data from the edge (where trades happen) to the cloud (where they are analysed), and move trained models back the other way. This is not a transitional state on the road to "everything in the cloud"; it is the stable, correct design for a business with both microsecond and overnight workloads.
The line between the two halves is drawn by two forces: latency and data gravity. Latency says the price path cannot tolerate the variable delay of a shared network. Data gravity says that the enormous volume of tick data is expensive and slow to move, so compute tends to migrate to wherever the data already sits, rather than the reverse. Push the price path into the cloud and you surrender microseconds you cannot afford to lose; keep the analytics on-premises and you pay for expensive hardware that sits idle most of the time. The real engineering skill is drawing that dividing line in the right place and building efficient ways to move data across it — a permanent architectural question the firm re-answers as workloads change, not a one-off migration project that finishes.
Time itself is a regulated quantity
Synchronised clocks are the backbone of measurement and surveillance
You cannot measure or reconstruct anything in an electronic market unless every system agrees, to a fine grain, on what time it is. Ask a simple question — did this order arrive before or after that price update? — and you find it is unanswerable unless the machine that logged the order and the machine that logged the price share the same clock. The whole apparatus of the electronic market depends on this agreement. Measuring latency requires it: latency is a difference of two timestamps, and if the two clocks disagree by more than the latency you are trying to measure, the number is noise. Running TCA requires it, because TCA compares your execution time to a benchmark time. Sequencing events across venues requires it. And reconstructing what happened during an incident — a flash move, a disputed trade, a suspected manipulation — is impossible if the logs from different systems cannot be laid on one shared timeline. This is why serious trading firms synchronise their clocks and make that synchronisation traceable: able to be proven, after the fact, to have been correct.
MiFID II made timestamps a legal obligation
Under MiFID II, accurate timestamping stopped being good practice and became law. The European Union's Markets in Financial Instruments Directive II (MiFID II) and its accompanying technical standards require trading systems to timestamp reportable events to a demanding granularity — for the fastest electronic activity, to the microsecond — and to keep those timestamps traceable to Coordinated Universal Time (UTC), the global time standard. Traceable means a firm must be able to demonstrate that its clocks were synchronised to UTC within a specified maximum divergence, and to evidence it if a regulator asks. This is not a formality; it is enforced, and firms invest real engineering to meet it.
Meeting that obligation pushed the industry from NTP to PTP. The older Network Time Protocol (NTP) synchronises clocks across a network to roughly the millisecond — fine for a laptop, far too coarse for a market that measures itself in microseconds. The Precision Time Protocol (PTP), by contrast, synchronises clocks to sub-microsecond accuracy, typically with help from specialised network hardware that timestamps packets as they pass. A firm running PTP can honestly claim its distributed systems agree on the time to well within a microsecond, and can prove that agreement is traceable to UTC. Accurate, traceable time is therefore the prerequisite for three things the market cannot do without: producing best-execution evidence (proving a client got a fair price at a fair moment), running surveillance, and reconstructing the exact sequence of a fast market when regulators, clients or risk managers need to know what happened. It is the most unglamorous layer of the entire stack, and the whole electronic market stands on it.
Checking credit in a microsecond is its own engineering problem
You cannot accept a price without first knowing the client can pay
Before a desk confirms a streamed trade, it has to be sure the counterparty can actually settle it — and it has to be sure in less time than the price will live. Every trade creates a settlement obligation: at value date, one side must deliver the currency it sold. If the counterparty cannot pay, the bank is left carrying the loss. So in principle, before agreeing any trade, the desk must confirm the client still has enough credit line — pre-approved capacity to owe the bank — to cover the new exposure. In the old voice world this was easy: trades were infrequent and a human had time to check.
In electronic FX that check has become an almost impossible timing problem, because the price the client is trading on may exist for only a few milliseconds. A streamed quote is a promise good for an instant; if the client clicks, the bank must decide now whether to honour it. A conventional credit check — open a database, look up the client's limits, sum their open positions, compare, respond — can take tens of milliseconds or more. By the time that answer comes back, the quote is stale, the market has moved, and the bank is exposed to being picked off. The tension is blunt: you cannot skip the credit check without risking a default, and you cannot run a slow credit check without destroying the price. Something has to give, and the answer is engineering.
Pre-trade credit hubs decide in microseconds
The solution is a real-time credit engine that pre-loads every client's limits and checks them inline, in microseconds, as each trade passes. Instead of querying a slow system of record at trade time, the venue, the prime broker and the bank each hold each client's authorised limits in fast in-memory state. When a trade arrives, the engine performs two operations in one atomic step: it checks that the remaining limit covers the trade, and it decrements the available limit by the amount consumed. The next trade sees the reduced limit. When a client exhausts its line, the very next order is blocked instantly — there is no human in the loop and no round-trip to a database. The whole decision fits inside the lifetime of the quote.
In the prime-brokerage model this engine is the backbone of the give-up credit chain introduced in Part 3. Recall that a fund without direct relationships at every venue trades in the name of its prime broker (PB): the fund executes against a liquidity provider, but the trade is "given up" to the PB, who steps in as the credit counterparty to both sides. For that to work safely, the fund is trading on the PB's limits at each venue, and those limits are loaded into the same real-time credit engine. The PB monitors consumption continuously and — this is the point — can cut or reduce the line in real time if the fund's risk grows or its collateral falls. The credit decision that used to take a phone call now happens automatically, per trade, faster than a human could perceive.
Real-time pre-trade credit is a genuinely hard distributed-systems problem, not a simple lookup. The same limit is consumed across many venues simultaneously — the fund may be trading on EBS, a bank's single-dealer platform and a multi-dealer venue at the same instant, all drawing down one shared PB line. The credit state therefore has to be shared and reconciled across those venues fast enough to prevent double-spending the line (two venues both approving trades that, together, breach the limit), yet without adding measurable latency to any single trade. Get it wrong in one direction and you block good trades and lose the client; get it wrong in the other and you let a client trade beyond its limit, which is a direct, uncollateralised credit loss.
The vendor and platform stack: what to build and what to buy
Across the whole stack, the bank faces the same recurring decision — build it or buy it — and the answer follows one consistent logic. Some parts of the system landscape are commodities: everyone needs them, they do the same job for everyone, and buying them from a specialist vendor is cheaper and better than building your own. Other parts are the bank's competitive edge: they are the reason a client picks this bank's price over another's, and building them in-house is the whole point. The pattern of what banks buy and what they build is not random. It traces the line between commodity and edge.
The build-versus-buy landscape across the stack
At the venue layer, banks almost always buy — they connect to platforms rather than owning them. The client-facing multi-dealer platforms — 360T, FXall, Bloomberg FXGO — are third-party venues a bank plugs into. The interbank ECNs where banks and PTFs meet — EBS, now owned by CME, and LSEG's Matching platform, formerly Refinitiv — are likewise external, as are newer venues such as Cboe FX and Euronext FX. A bank does not build a venue; it earns its place on the venues that exist.
At the pricing and distribution layer, the picture is mixed, and this is where "build" starts to matter. Some components — market-data handlers, streaming infrastructure — may be bought or assembled from vendor parts. But the core pricing logic tends to be built, because it encodes the bank's own view of risk and its own client strategy. Aggregation and execution management, one layer down, is similarly split: some banks buy an aggregator or EMS, others build proprietary ones, and many run a hybrid.
At the prime-brokerage layer, banks run give-up, credit-monitoring and margining platforms — the systems that let them stand behind clients' trades, watch those clients' credit in real time, and call margin when exposure grows. At the risk and booking layer sit trade capture, real-time risk, limits and profit-and-loss systems, some vendor-supplied, some built, usually deeply customised. At the post-trade and settlement layer, banks lean toward buying: confirmation and matching services, CLS connectivity, and the reporting engines that produce EMIR and MiFIR submissions are commoditised enough that building them offers no advantage.
The pattern that runs through all of it is simple to state and hard to argue with: banks buy the commoditised connectivity and venues, and build the decision logic that is their edge. Pricing, skew and internalisation logic — the machinery that decides what price to show, to whom, and whether to hold or hedge the resulting risk — is where one bank beats another. That machinery is almost always proprietary. Everything around it, from the wires that carry the messages to the utilities that settle the cash, can be rented. The decision engine cannot, because renting it would mean renting the very thing that makes the franchise worth anything.
Post-trade and settlement: CLS tames the risk that is uniquely FX
FX carries one risk that no other market has in the same form: because the two legs of a trade are two payments in two countries, one side can pay and then fail to receive. In a stock trade, cash and shares change hands together at the clearing house. In an FX trade, you pay away one currency in one country's payment system and expect to receive another currency in another country's payment system, and the two systems run on different clocks in different time zones. For a few hours, you have delivered and not yet received. That gap is the market's oldest and most dangerous exposure, and taming it is the central achievement of FX post-trade infrastructure.
CLS settles both legs at once, removing Herstatt risk
CLS — Continuous Linked Settlement — closes the gap by settling both legs of an FX trade simultaneously. The mechanism is called payment-versus-payment, or PvP: you only pay away one currency if, and at the same moment that, you receive the other. There is no interval in which you have delivered and not received. For the currencies CLS covers, this removes settlement risk — the specific danger, named after the 1974 Herstatt failure we examine later, that a counterparty collapses in the window between the two legs. CLS is owned by its member banks, which gives the market itself a stake in the utility that protects it.
Beyond removing the risk, CLS also nets the payments, which dramatically shrinks the cash that actually has to move. Netting means offsetting all of a member's obligations in a currency against each other so that only the difference is paid. A bank that owes ten billion dollars across many trades and is owed nine-and-a-half billion moves only the half-billion net, not the full gross amounts. This makes settlement not only safer but vastly more efficient in the cash and liquidity it consumes.
The critical limitation is that many emerging-market currencies sit outside CLS, so for them the old settlement risk is still live. CLS covers a set of major currencies; a great many others are not eligible. When you trade a currency outside the system, you are back in the Herstatt world — paying one leg and hoping the other arrives — and managing that exposure through bilateral limits and careful timing. This is not a historical curiosity. It is a present regulatory concern, and it is why supervisors keep pressing the industry to extend PvP coverage and why settlement risk remains on the agenda even in a market that has largely solved it for the majors.
The rest of post-trade: confirm, net, report, reconcile
Around settlement sit four routine but non-negotiable jobs, and each has its own systems. Confirmation and matching is both sides electronically agreeing the economics of the trade — near-instant for electronic FX, slower for higher-touch products. Netting and funding is the offsetting of obligations so that far less cash moves, with the treasury function funding the net amounts that remain. Reporting is the submission of the trade to regulators: FX derivatives are reported under EMIR and, when traded on-venue, under MiFIR, while spot FX is largely outside the reporting net. Reconciliation and controls is the continuous checking of positions, cash and nostro accounts against the bank's own records, with any break — any mismatch — investigated until it is explained. None of this makes money. All of it prevents the loss of money, which in a market of thin margins amounts to the same thing.
Two messaging worlds meet in post-trade: FIX at the front, SWIFT and ISO 20022 at the back
The front office and the back office speak different languages, and post-trade is where the two meet. In the front office, FIX and binary application programming interfaces carry quotes and orders between trading systems in microseconds; the priority is raw speed, and the messages describe prices and trades. In the back office, SWIFT MT messages — the long-standing standard format of the interbank payment network — and their modern ISO 20022 successors (often called MX messages) carry payment and settlement instructions; the priority is reliability and completeness, and the messages describe money moving. The split between these two messaging worlds is the front-office/back-office divide made concrete. A trade is born in the fast, terse language of FIX and dies, settled, in the rich, deliberate language of SWIFT and ISO 20022, and post-trade is the translation zone between them.
| Layer | Standard | What it carries | Timescale |
|---|---|---|---|
| Front office | FIX / binary API | Quotes and orders | Microseconds |
| Post-trade | Confirmation / matching | Matched economics | Minutes |
| Settlement | CLS | PvP exchange of both legs | Value date |
| Back office | SWIFT MT / ISO 20022 | Payment and reporting instructions | T+2 |
The regulatory picture is a patchwork, with spot at its centre uncaptured
FX is not governed by one comprehensive law; it is governed by a patchwork of a voluntary code plus a set of product-by-product rules, with the largest product — spot — sitting mostly outside all of them. Spot FX is generally not treated as a financial instrument in the way a share or a bond is, which means the heavy machinery of market regulation does not fully reach it. Derivatives on FX — forwards, swaps, options, non-deliverable forwards — are captured, because they are instruments. This leaves an odd shape: the market's centre of gravity is lightly regulated by hard law, and the gap is filled by a voluntary standard of conduct. To design systems for this market, you read each regime not as a legal summary but as a specification: what, concretely, does it demand that the systems be able to do?
Five regimes, each read as a system specification, not a legal summary
The FX Global Code is voluntary but central — the global standard of conduct for wholesale FX. Published and maintained by the Global Foreign Exchange Committee, the Code is not law; institutions adhere to it by public commitment. It sets expectations around fair dealing, the proper use of information, transparent handling of last look, and good market practice generally. What it demands of the systems is evidence: surveillance that can show conduct was fair, records of how last look was applied, and controls that keep information from being misused. In a market where hard law is thin, the Code carries much of the weight that statute carries elsewhere.
MiFID II and MiFIR govern FX derivatives in the EU — best execution, transparency, transaction reporting and venue rules — with spot out of scope as an instrument. Best execution is the obligation to obtain the best available result for the client; transparency rules govern what must be published about prices and trades; transaction reporting requires a detailed record of each in-scope trade to reach the regulator. What this demands of the systems is concrete: accurate timestamps, a transaction-reporting feed, the analytics to demonstrate best execution, and the venue classifications that determine which rules apply. Spot FX, not being an instrument, largely escapes — which is precisely why the Code matters for spot.
EMIR governs OTC FX derivatives — forwards, swaps and NDFs — through reporting, risk-mitigation and margin, with reliefs for deliverable commercial-hedging forwards. The European Market Infrastructure Regulation requires that these trades be reported to a repository, that counterparties follow risk-mitigation steps such as timely confirmation and portfolio reconciliation, and that margin be posted for many uncleared derivatives. There are reliefs recognising that a plain deliverable forward used by a company to hedge a real commercial exposure is not the same animal as a speculative derivative. What EMIR demands of the systems is a reporting feed with UTIs and LEIs, a reconciliation capability, and a margin calculation.
MAR — the Market Abuse Regulation — bites on manipulation and misconduct, especially around fixings, and underpins much of the post-scandal scrutiny. MAR prohibits market manipulation and the misuse of inside information. Its relevance to FX crystallised around benchmark fixings, where coordinated behaviour to move a widely-used reference rate is exactly the kind of abuse it targets. What MAR demands of the systems is surveillance: the ability to monitor trading and communications, detect anomalous patterns around sensitive moments, and produce the evidence trail an investigation would require.
Prudential and conduct rules — capital, leverage and counterparty requirements under Basel and CRR, plus conduct supervision by national regulators — sit underneath everything. These are not FX-specific; they govern how much capital a bank must hold against its risks, including counterparty and market risk from FX, and how national supervisors oversee the firm's behaviour. What they demand of the systems is accurate, real-time risk and exposure data feeding capital and leverage calculations, and the governance records supervisors expect to see. This is the foundation the rest of the patchwork stands on.
| Regime | What it governs | What it demands of the systems |
|---|---|---|
| FX Global Code | Wholesale FX conduct (voluntary, central) | Surveillance, last-look transparency, information controls, evidence of fair dealing |
| MiFID II / MiFIR | FX derivatives in the EU (spot out of scope) | Timestamps, transaction-reporting feed, best-execution analytics, venue classification |
| EMIR | OTC FX forwards, swaps, NDFs | Reporting feed (UTI/LEI), risk-mitigation, reconciliation, margin calculation |
| MAR | Manipulation and misconduct (e.g. fixings) | Trade and communications surveillance, anomaly detection, audit trail |
| Prudential & conduct (Basel/CRR + national) | Capital, leverage, counterparty risk; conduct supervision | Real-time risk/exposure data, capital and leverage calculations, governance records |
| Spot FX — special | Largely uncaptured as an instrument | Little hard-law reporting duty, which is exactly why the voluntary Code matters |
Operational resilience became law: DORA
Resilience is no longer just good practice — it is a regulation
We treated operational resilience earlier as an engineering virtue: build systems that keep pricing, connectivity and settlement running through failures. In the European Union that virtue has now been written into binding law. The instrument is the Digital Operational Resilience Act, known everywhere as DORA, which applies to financial firms from early 2025. DORA takes the loose expectation that a well-run trading business should not fall over and converts it into a set of specific, supervised obligations with legal force. The shift matters because it changes who judges whether a firm is resilient enough. It is no longer only the firm's own risk committee; it is a regulator with the power to inspect, demand evidence, and sanction.
What DORA demands of an FX business
DORA rests on a handful of concrete demands, and the first is formal ICT risk management. ICT stands for information and communication technology — the systems, networks, data stores and software on which the business runs. DORA requires firms to identify their critical ICT assets, map how a failure in one would propagate, set risk tolerances, and govern all of it through documented policies owned at board level. For an FX desk this means the pricing engine, the exchange connectivity, the market-data feeds and the settlement links must each be inventoried and understood as points of potential failure, not treated as invisible plumbing that simply works.
The second demand is mandatory reporting of major ICT incidents. When a significant operational failure occurs — an outage, a serious cyber breach, a corruption of critical data — DORA obliges the firm to classify it against defined criteria and report it to its supervisor within tight timeframes, with follow-up analysis. The effect is that a serious system failure is no longer a private embarrassment to be quietly fixed. It becomes a reportable event that the regulator sees, tracks, and can compare across the market, which raises the stakes on both preventing incidents and detecting them honestly.
The third demand reaches outside the firm to its critical third parties, and this is the part most consequential for a modern FX stack. A third party here is any outside provider the firm depends on — most importantly the cloud platforms and specialist data vendors on which pricing, analytics and connectivity increasingly run. DORA requires firms to oversee these providers actively: to map the dependency, assess concentration, secure contractual rights to audit and to exit, and demonstrate they could continue operating if a key provider failed. The most systemically important providers can themselves be brought under direct supervisory oversight. Resilience, in DORA's logic, is only as strong as the weakest external supplier in the chain.
The fourth demand is regular resilience and threat-led penetration testing. Firms must test their defences on a schedule, and the largest must periodically undergo threat-led penetration testing — controlled exercises in which skilled testers attack the live environment the way a real adversary would, to see what actually breaks. For an automated FX business whose pricing, connectivity and settlement are all software, this turns "the system must not go down" from an aspiration the team assures itself of into a claim it must prove, repeatedly, to a supervisor who checks the working.
Nobody designed the FX estate — it accumulated
The technology behind an FX business was never drawn on a whiteboard and then built. It grew one deal at a time, one venue at a time, one regulatory deadline at a time, and the shape it has today is the fossil record of every commercial decision the firm ever made. A new electronic venue appeared, so a connection was bolted on. A client wanted a currency pair the pricing engine did not cover, so a workaround was written. A regulator demanded a new report, so a system was wired to a system that was never meant to feed it. The result is an estate — the full collection of applications, databases, connections and feeds a business runs on — that no single person designed and few people fully understand. You cannot change what you do not understand, so reading the estate for what it is comes before any attempt to improve it.
Technical debt is undocumented coupling, not ugly code. The phrase technical debt — the accumulated cost of past shortcuts that makes future change slower — is usually pictured as messy, badly written code. That picture is misleading. Clean code that everyone can read is cheap to change; the real debt is coupling you cannot see, where one system silently depends on another in a way nobody wrote down. The pricing engine reads a field that a downstream settlement system happens to populate; nobody remembers this, because the person who wired it left years ago. When a team tries to retire the settlement system, prices go wrong in a way no one predicted, because the dependency was never documented — it lived only in the behaviour of the running software. The danger is not the code you can see and dislike; it is the connection you cannot see and do not know to test.
Build-versus-buy is a map of where the firm believes its edge is. Every estate is a series of decisions to build a capability in-house or buy it from a vendor, and the boundary between the two is not an accident of budget — it is a statement of strategy. Banks build the parts they believe make them money: the pricing logic, the skew (deliberately shading a quote to attract the flow you want and repel the flow you do not), the internalisation engine that matches client buys against client sells inside the firm rather than paying an external venue. They buy the parts that are the same for everyone: connectivity to venues, market-data feeds, post-trade utilities that confirm and settle. Read the build-versus-buy boundary and you are reading the firm's honest answer to the question "where do we think our advantage actually is?" A firm that has bought its pricing engine has told you, whether it meant to or not, that it does not compete on price.
When the same fact lives in five systems with five values, every calculation downstream is suspect. Reference data is the slow-moving factual backbone every trade leans on: currency codes, holiday calendars that say which days a currency cannot settle, LEIs (Legal Entity Identifiers, the global 20-character code that says exactly which legal entity you traded with), CSA terms (the collateral rules in the Credit Support Annex governing what each side must post), and credit limits. The problem is duplication. The same client's credit limit may be recorded in the pricing engine, the risk system, the limit-monitoring tool and two spreadsheets, and if those five copies disagree, then every check that reads one of them is unreliable — you can breach a limit in one system while sitting comfortably inside it in another. The answer is a golden source: one system designated as the single authoritative copy of a given fact, with all others required to read from it rather than keep their own. Reaching a golden source is mostly governance, not technology — the hard part is not building the database but getting every team to agree to stop keeping its own version and to accept someone else's as the truth.
The reference-data layer is increasingly built from open, vendor-neutral identifiers, and knowing the four that matter saves confusion. ISO 4217 gives every currency its three-letter code (USD, JPY, CHF). The LEI identifies the legal entity you face. The UTI (Unique Trade Identifier) tags each individual trade so two firms and their regulators can all point at the same transaction. And FIGI (the Financial Instrument Global Identifier) is an open, vendor-neutral code for financial instruments, which complements the currency codes rather than replacing them — it lets systems refer to an instrument without being locked to any one data vendor's proprietary numbering. Together these four are the vocabulary that lets separate systems agree on what they are talking about, which is the precondition for a golden source to mean anything.
Estates decay because they are funded as projects and then abandoned. A project is funded to a deadline and a budget, delivers a system, and dismantles the team; a product is owned by a standing team through the whole life of the thing it runs. Firms overwhelmingly fund technology as projects, which means every system is born healthy on delivery day and then left without an owner. Small changes accumulate with no one responsible for keeping the whole coherent, the documentation goes stale, the dependencies multiply unseen, and five years later the system is legacy — not because it was built badly but because nobody was funded to keep it well. A product team that owns a system for its life pays down debt continuously, keeps the map current, and refuses the shortcuts that a deadline-driven project accepts.
Three integration patterns decide how fresh the data is: batch, API, and event streaming
How systems in the FX chain talk to each other determines how fresh anyone's data is, and there are three patterns worth telling apart. The choice among them looks like a technical detail buried far below the trading floor, yet it decides whether your risk system knows about a trade the instant it happens or six hours later, and whether two systems can quietly drift into disagreement. Each pattern is a different answer to one question: when a fact changes over here, how soon does over there find out?
Batch is the legacy default: systems exchange files on a schedule, and the data is stale between runs. In a batch pattern one system writes a file — all the trades from the day, all the positions as of midnight — and another system reads it on a timer, overnight or hourly. It is simple and robust, which is why it dominates older estates; a file is easy to produce, easy to check, and easy to rerun if something breaks. The cost is freshness. Between two runs the receiving system is working from a snapshot that is already out of date, and any error in the file surfaces only when the next batch is processed, often the following morning, by which point the wrong number has been used all day.
An API answers on demand, but it still hands back a point-in-time snapshot. An API (Application Programming Interface, a defined way for one system to ask another a question and get an answer immediately) replaces the scheduled file with a live request: the risk system asks the trade system "what is my position in EUR/USD right now?" and gets the answer now, not tomorrow. This is the standard modernisation step and it removes the overnight lag. Its limit is that it pulls — the receiving system only learns something if it thinks to ask, and the answer is true only at the instant it was asked. Nothing tells the risk system that the position just changed; it has to keep asking.
Event streaming inverts the flow: systems publish changes as they happen, and everyone who cares reacts in real time. In an event-streaming pattern (often built on Kafka-style publish/subscribe, where producers publish events to a stream and any number of subscribers consume them) a system announces every change the moment it occurs — a trade is done, a position moves, a limit is updated — and every interested system receives that same event and reacts immediately. Nobody has to ask. The trade that just printed reaches the risk engine, the limit monitor and the client-reporting feed in the same instant, from the same message.
Moving from batch to streaming is not merely a speed upgrade — it removes the need for reconciliation. Reconciliation is the daily work of comparing two systems' records and hunting down every place they disagree, and it exists overwhelmingly because batch systems drift apart between runs: each keeps its own copy, the copies are refreshed at different times, and by morning they no longer match. When every system is fed the same event as it happens, there are no two separate overnight snapshots to compare, because there is only one event and everyone consumed it. In an automated, real-time FX book — where trades print in microseconds and positions move continuously — this is decisive: reconciliation is not made faster, it is largely designed out of existence.
Resilience is measured, not asserted: change freeze, observability, RTO and RPO
Every firm calls its trading systems "mission-critical," and the word is almost always a slogan rather than a measurement. Resilience — the ability of a system to keep working, or recover fast, when something goes wrong — is either designed and quantified or it is merely hoped for, and the difference shows only on the bad day. What separates the two is not intention but specifics: named controls and actual numbers. Four of them do most of the work.
A change freeze is the cheapest resilience control there is: stop changing the system when the stakes are highest. Most incidents are not caused by hostile events; they are caused by changes the firm made to its own systems. A change freeze is the deliberate decision to make no changes during the highest-risk windows — a major fixing (the daily benchmark moment when large volumes trade at a published rate), a central-bank rate decision, quarter-end — because those are precisely the moments when a self-inflicted outage is least survivable. It costs nothing to implement; you simply do not touch the system. Its failure mode is the pile-up: every change deferred during the freeze lands in a rush the moment it lifts, and a large batch of simultaneous changes is itself a leading cause of incidents. A freeze buys safety during the window at the price of concentrated risk just after it, so the changes have to be released in a controlled trickle, not a flood.
Systems fail silently far more often than loudly, so monitor the work, not the machine. Observability is the ability to tell, from the outside, what a system is actually doing. The naive version watches the machine — is the server up, is the CPU busy — and answers the wrong question, because a server can be perfectly healthy while the thing it is supposed to do has quietly stopped. The version that matters watches the work: are prices streaming to clients, and are the prices sane? A pricing engine that is "up" but publishing a stale or absurd EUR/USD rate has failed completely, and only a business-level check catches it. Good observability asks the questions a trader would ask, not the questions a system administrator would ask.
RTO and RPO turn resilience into two honest numbers for the systems that actually matter. The Recovery Time Objective is how fast you must be back after a failure; the Recovery Point Objective is how much data you can afford to lose. Both are meaningless as adjectives and meaningful only as figures, stated for the two or three systems on the critical path — the pricing engine, the risk system, the settlement feed. Saying the pricing engine has an RTO of thirty seconds and an RPO of zero (illustrative) is a testable claim you can rehearse against; saying it is "highly resilient" is not. These numbers are the honest test of whether resilience was engineered or merely asserted, because you can fail the test.
| Concept | What it guarantees | Why it matters on the critical day |
|---|---|---|
| Redundancy | A spare component already exists for anything whose failure would stop trading | When one server, feed or site dies mid-session, another is already running and no single failure halts the book |
| Automatic failover | The switch to the spare happens by itself, in seconds, with no human in the loop | A human deciding to fail over during a central-bank surprise is too slow; the cutover has to be automatic to matter |
| Change freeze | No changes are made during the highest-risk windows | Removes the most common cause of incidents — self-inflicted change — exactly when an outage is least survivable |
| Observability | You can see whether the real business function is working, not just whether the machine is on | Catches the silent failure — an engine that is "up" but streaming stale or absurd prices — before clients trade on it |
| RTO | A stated maximum time to be back after a failure | Converts "we'll recover quickly" into a rehearsed, testable number you can prove or fail against |
| RPO | A stated maximum amount of data you can lose | Decides whether a failure costs you the last few seconds of trades or the last several hours of them |
The asset to protect is the stack, the flow data and the payments, not a deal secret
Security for an FX business protects a different set of crown jewels than security for a deal-making business. There is no single months-long secret to guard the way an equities advisory team guards a pending acquisition; there is a running, automated machine that prices and hedges continuously, a stream of proprietary data about who is trading what, and a pipe through which real currency payments leave the firm. Get the asset wrong and you defend the wrong thing. The three things worth protecting are the automated pricing stack (the connected set of systems that price and hedge), the proprietary flow data (the information about client trading that the firm accumulates), and the settlement payments (the actual movements of money at settlement).
The dangerous actor is rarely the outsider breaking in; it is the insider or the stolen credential already inside. A threat model is a clear-eyed statement of who you are actually defending against, and in FX the honest answer is uncomfortable. The authorised insider — a trader, an operations clerk, a departing employee — already has legitimate access to the pricing engine, the flow data and in some cases the payment rails, and needs to break nothing to do damage. The compromised credential — a real employee's login, phished or stolen — is the same problem wearing a trusted face: it walks in through the front door with valid permissions. The external hacker attacking the perimeter is real but is the case the whole industry is best defended against; the insider and the stolen credential are more dangerous precisely because they start on the inside.
Defence works in four layers, because no single wall holds. The first is network and perimeter: segmentation so a breach in one zone does not reach the whole estate, and encryption in transit so intercepted traffic is unreadable. The second is identity and access: per-desk entitlements so a person can touch only what their role requires, least privilege as the default, multi-factor authentication so a stolen password alone is not enough, and disciplined joiner-mover-leaver hygiene so access is granted, changed and — critically — revoked as people move and leave. The third is application controls: the kill switch that stops all automated pricing at once, fat-finger limits that reject an order too large to be a real intention, and dual approval on payment instructions so no one person can send money alone. The fourth is monitoring and surveillance: e-communications surveillance reading chat and email for misconduct, trade surveillance watching for manipulation, data-loss prevention (tooling that spots sensitive data leaving the firm), and anomaly detection that flags behaviour outside the normal pattern.
The attacks that actually succeed in FX are not the dramatic ones. By far the most common real loss is payment-instruction fraud and the business-email-compromise around settlement: a spoofed or altered instruction redirects a currency payment to an account the fraudster controls, and because the payment is genuine currency moving on genuine rails, it leaves and does not come back. Next is credential theft that hands an attacker access to the pricing or payment systems, letting them act with a legitimate user's rights. Third is leakage of proprietary client-flow data, which is a genuine informational asset — knowing what a firm's clients are about to do has real value to a competitor or a counterparty. Note what is absent from this list: the theft of a single guarded deal secret, which is an equities concept and not how FX firms actually lose.
| Control | The failure it is built for |
|---|---|
| Network segmentation and encryption in transit | An intruder or interceptor who reaches one zone reading everything and moving freely across the estate |
| Per-desk entitlements and least privilege | A user — insider or compromised account — touching systems and data far beyond what their role needs |
| Multi-factor authentication | A stolen or phished password on its own being enough to log in as a real employee |
| Joiner-mover-leaver access hygiene | A mover or leaver keeping access they no longer need, turning a stale account into an open door |
| Kill switch and fat-finger limits | A runaway algorithm or a mistyped order flooding the market before a human can react |
| Dual approval and surveillance on payments and comms | A single person redirecting a settlement payment, or misconduct and data leakage passing unwatched |
4.1 FIX vs SWIFT
For each, say whether it belongs to the front office or the back office and what it carries: (a) a streamed quote and an order; (b) a settlement and payment instruction.
(a) the streamed quote and order travel by FIX in the front office; (b) the settlement and payment instruction travels by SWIFT MT / ISO 20022 in the back office. Two messaging worlds, front and back.
4.2 Batch vs streaming
Two systems reconcile because each takes an overnight snapshot and the two are compared. Which integration pattern removes the need for that reconciliation, and why?
Event streaming (publish/subscribe) removes the need for reconciliation: every system reacts to the same event as it happens, so there are no two overnight snapshots that can drift apart. Reconciliation exists largely because batch systems diverge between runs.
4.3 RTO / RPO
The pricing engine has a Recovery Time Objective of 60 seconds and a Recovery Point Objective of zero. In plain words, what two promises is the desk making about a failure?
RTO 60 seconds = the desk promises to restore the pricing engine within 60 seconds of a failure. RPO zero = it promises to lose no committed data — no trades or price updates dropped — across that failure.
4.4 Latency
Co-locating your engine in the venue's data centre shortens the cable. London to New York is roughly 5,570 km and light in fibre travels around 200,000 km per second. Estimate the one-way and round-trip propagation time, and say why that number is the case for co-location.
One-way ≈ 5,570 / 200,000 s = 0.02785 s ≈ 27.9 ms; round-trip ≈ 55.7 ms. Every one of those milliseconds is time your quote is stale or your reaction is late, so shortening the cable by co-locating next to the venue directly reduces the risk of being picked off — the whole case for co-location.
4.5 Golden source
A client's CSA collateral threshold is stored as three different values in three systems. Name the problem and state the fix.
This is the golden-source problem: one fact (the CSA threshold) held as three conflicting copies makes every downstream margin calculation suspect. The fix is a single authoritative source that every system reads, enforced by data governance — not five independent copies kept "in sync" by hope.
Explain to someone who has only ever used a home Wi-Fi router why an FX bank puts its pricing computers physically inside the trading venue's building — and why "just move it all to the cloud" does not solve the speed problem.
What is the difference between FIX and SWIFT/ISO 20022 in the FX stack?
FIX carries quotes and orders in the front office in microseconds; SWIFT MT and ISO 20022 carry settlement and payment instructions in the back office over hours to days.
Why is phantom liquidity a hard problem for aggregation?
The same LP price is echoed across venues, so naive aggregation double-counts depth that is not really there; deduplication is mission-critical.
Why can the latency-critical price path not simply move to the cloud?
Public cloud adds variable latency and cannot be co-located inside a venue; speed-critical pricing and matching stay on co-located hardware while analytics and model training move to the cloud.
What does CLS do, and why do emerging-market currencies remain a concern?
CLS settles both legs simultaneously (payment-versus-payment), removing settlement risk for the currencies it covers; many EM currencies sit outside it, so the old Herstatt risk persists.
Why is redundancy alone insufficient for resilience?
Duplicate systems do not help if failover is untested, monitoring is silent, or recovery has no measured time and data objectives; resilience must be designed and rehearsed, not asserted.
What three identifiers pin down an FX trade, and what does each name?
ISO 4217 currency codes name the currencies, the LEI names each legal counterparty, and the UTI names the individual trade so both sides reconcile to one record.
Why does a bank build its pricing and internalisation logic but buy connectivity and venues?
The decision engine is its competitive edge and stays proprietary; commoditised connectivity and venue access are cheaper to buy.
Carrying forward from Part 2: give one example where a wrong reference-data value turns into a dangerous quote. — A stale holiday calendar or a wrong credit or CSA value can mis-price a forward or block a valid trade — bad data in, dangerous quote out.
Carrying forward from Part 3: why does the prime broker depend on the real-time credit system described here?
Give-up trades consume the prime broker's venue limits, which must be checked and decremented per trade so a breaching client can be cut off instantly.
4
Why is phantom liquidity a problem for aggregation?
It is.
No.
No.
Right. Deduplication is critical.
Why can't the latency-critical price path move to the public cloud?
No.
No.
Right. Analytics can move; pricing stays co-located.
No.
What does CLS do?
Right. Many EM currencies remain outside it.
No.
No.
No.
Which standard carries quotes and orders in microseconds?
Back office.
Back office.
Right. SWIFT and ISO 20022 carry settlement and payments.
No.
5Risks, Pain Points and Transformation
What can go wrong, how to tell a good desk from a lucky one, the real cases where FX broke, and where better data and disciplined AI change the economics — closing with the ideas that matter most.
Risk ends at settlement: what a currency desk actually runs
A currency desk runs a stack of distinct risks, and the trade is not truly finished until the last of them — settlement — is discharged. It is tempting to think the risk in FX is just "the price might move." That is one risk, but it is far from the only one, and it is not even the most dangerous. The desk runs market risk, intraday risk, settlement risk, counterparty and credit risk, liquidity and gap risk, and operational and model risk. Each has its own character and its own control. The table below lays them out; then we single out the one that is uniquely acute in FX and the reason a whole piece of market infrastructure exists.
| Risk type | What it is | How the desk manages it |
|---|---|---|
| Position (market) risk | The risk that the currency price moves against a position the desk is holding | Skew quotes to attract offsetting flow; internalise; hedge residual risk in the market; run position limits |
| Intraday risk | The risk building up within the day as positions accumulate before they are netted or hedged | Real-time position monitoring; intraday limits; continuous internalisation and hedging through the day |
| Settlement (Herstatt) risk | The risk that you pay away one currency but the counterparty fails before paying you the other | Settle through CLS on a payment-versus-payment basis; set settlement limits per counterparty; net where possible |
| Counterparty & credit risk | The risk that a counterparty defaults on its obligations before or at settlement | Credit lines, margin and collateral; prime-brokerage guarantees; netting agreements; ongoing credit monitoring |
| Liquidity & gap risk | The risk that the market becomes thin or jumps, so you cannot trade at or near the expected price | Diversify liquidity sources; size positions to market depth; widen quotes in stress; hold liquidity buffers |
| Operational & model risk | The risk of failures in systems, processes or the pricing models themselves | Straight-through processing; controls and reconciliation; model validation and governance; conduct oversight |
Settlement risk is the danger that is uniquely acute in FX
Settlement risk is the danger that you pay away the currency you owe and your counterparty fails before paying you the currency they owe — and it exists in FX in a way it does not in most markets. Because an FX trade exchanges two currencies that settle through two different banking systems, often in two different time zones, the two legs do not necessarily happen at the same instant. You might pay dollars in New York in the morning and expect to receive euros in Frankfurt — but between the two, in the hours of misalignment, the counterparty could go bankrupt. You would have paid away your dollars and received nothing. This is not a price risk; the price could be perfect. It is the risk of losing the entire principal — the whole amount — because the two sides of the exchange are not synchronised.
The classic case is the 1974 collapse of Herstatt Bank, and it is the reason the industry built a settlement utility. In June 1974, the German regulator shut down Bankhaus Herstatt during the day. Counterparties had already paid Deutsche Marks to Herstatt that morning, expecting dollars in New York later in the day. The bank was closed before those dollar payments were made. The Deutsche Mark legs had settled; the dollar legs never did. Counterparties lost the full principal on those trades. The episode was so instructive that settlement risk in FX is still called Herstatt risk to this day. It drove decades of work that culminated in CLS, Continuous Linked Settlement — a global utility that settles both legs of an FX trade simultaneously, so that one currency is only paid if and when the other is. That principle is payment-versus-payment, and it removes the window in which a Herstatt-style loss can happen. We cover CLS and the settlement machinery fully in Part 4; here, simply register that settlement risk is the risk that ends the trade, and taming it required building new infrastructure for the whole market.
The full risk picture fits on one map
FX runs six distinct risks, and the value of naming them is that each points to a specific defence. They are market risk (spot moves against the position), intraday risk (exposure that builds and unwinds within a single day), settlement risk (paying away one currency before receiving the other), counterparty or credit risk (the other side failing before the trade completes), liquidity or gap risk (the market thinning or jumping just when you need to trade), and operational or model risk (a failed process, a mis-set system, a wrong assumption). The list is really a map of controllability: some of these risks can be engineered almost entirely away, while others can only be sized, priced, and watched.
One of the six is uniquely acute in FX, and it is also the one most fully tamed. Settlement risk — the danger of delivering your currency and never receiving the other, named Herstatt risk after the 1974 bank failure that made it infamous — exists because the two legs of an FX trade historically settled in different countries and time zones, hours apart. It is the one risk on the map that is specifically an FX problem rather than a hazard shared with every other trading business. It is also the one most completely defended, through CLS (Continuous Linked Settlement), the utility that settles both legs simultaneously so that neither party can pay without receiving. The remaining five risks are shared with all trading and are managed, not eliminated.
What still hurts: the pain points that set the agenda
For all the industrialisation, a set of frictions remains unresolved, and they are the agenda the whole business is working against. These are not exotic risks; they are the daily grind of running an FX operation, and every transformation initiative is, in the end, an attempt to soften one of them. Name them plainly.
The unresolved frictions
Fragmentation and the absence of a single tape make it genuinely hard to see the true market or to prove best execution. With no consolidated record of all trades, a bank assembles its picture from dozens of partial feeds, and demonstrating that a given fill was the best available means reconstructing a market that was never recorded in one place. The latency race is a costly, never-ending treadmill — the infrastructure to shave microseconds off pricing and execution is expensive, and the advantage it buys erodes the moment a competitor matches it. Internalisation and skew management leak money when mis-tuned — an engine skewed wrong either gives away spread or quietly accumulates unwanted risk, and the models that govern it drift as market conditions change, so tuning is continuous rather than one-off.
Last-look conduct invites scrutiny and litigation when it is asymmetric or opaque — using the last-look window to reject trades that move against the maker while accepting those that move in its favour is exactly the kind of behaviour the Code and the courts examine. Settlement risk in emerging markets remains real — the currencies outside CLS still carry the Herstatt exposure that PvP removed for the majors. Data quality and cost sit under all of it — market data is fragmented, expensive and encumbered by licensing, and bad data does not merely inconvenience an automated engine; it makes the engine quote dangerous prices at machine speed.
Three real failures map the three ways FX breaks
There are exactly three ways the FX business goes seriously wrong, and history has furnished a landmark example of each. One is a failure of conduct — behaving unfairly when you hold an information advantage. One is a failure of settlement — paying one leg of a trade and not receiving the other. One is a failure of the market itself — a violent move with no liquidity to trade against. The three cases below are real, and each maps cleanly onto one of these modes. Learn the three cases and you have learned the risk map of the entire business.
Case 1 — the FX benchmark-fixing scandal: a conduct and manipulation failure
Between roughly 2013 and 2015, traders at several of the largest global banks were found to have coordinated around a widely-used benchmark fixing. The relevant benchmark was the 4pm London fix, a reference rate calculated from trading activity in a short window and used by countless funds and companies to value and execute currency transactions. In private chat rooms — with names that became notorious once disclosed — traders shared information about client orders and attempted to nudge the rate in their favour around the fix. Regulators across multiple jurisdictions imposed fines that were reported to total in the billions of dollars, and a number of individuals faced enforcement and, in some cases, criminal proceedings.
The episode directly drove the creation of the FX Global Code and sharpened the application of the Market Abuse Regulation to FX. The lesson is uncomfortable and precise: in a market with a genuine information edge and comparatively little hard law, conduct is the risk. There was nothing wrong with the plumbing. The people misused what they knew.
Case 2 — Herstatt Bank: the settlement failure that named the risk
On 26 June 1974, German authorities closed Bankhaus Herstatt during the trading day, and the timing exposed a risk the market had not properly named. Counterparties around the world had already paid Deutsche Marks to Herstatt that morning, expecting to receive US dollars in return later in the day, New York time. Because of the time-zone gap between Frankfurt and New York, the dollar leg had not yet been paid when the bank was shut. The counterparties had delivered one currency and never received the other. They lost the dollar leg outright.
The episode gave "Herstatt risk" its name, and it is why, decades later, the industry built CLS and payment-versus-payment settlement. The lesson is the foundation of everything in the post-trade section above: an FX trade is not done when it is agreed, or confirmed, or even when one leg pays. It is done only when both legs settle.
Case 3 — the Swiss franc de-peg: a market and liquidity gap failure
On 15 January 2015, the Swiss National Bank abruptly abandoned the floor it had held under EUR/CHF at 1.20, and the market convulsed. For over three years the central bank had promised to prevent the franc from strengthening past 1.20 per euro, and traders had built positions on the assumption the floor would hold. When it was removed without warning, the franc surged by around 30% in minutes. Liquidity vanished at exactly the moment everyone wanted to trade the same way, and prices gapped — jumping across levels with no trades in between rather than moving smoothly. The retail broker Alpari UK became insolvent. FXCM, another broker, required an emergency rescue loan reported at around USD 300 million. Banks, brokers and prime-brokerage clients took large losses, some because their clients could not cover the moves and the credit flowed back up the chain.
The lesson is a familiar one in the language of liquidity: liquidity that looks abundant in calm markets can disappear in an instant, and automated books and leverage amplify the gap rather than cushion it. A pricing engine that assumes it can always hedge is an engine that will one day be caught unable to.
| Case | Failure type | Where it maps | Lesson |
|---|---|---|---|
| Benchmark-fixing scandal | Conduct / manipulation | Fair use of an information edge | With little hard law, conduct is the risk |
| Herstatt Bank | Settlement | The value-date exchange of both legs | A trade is not done until both legs settle |
| Swiss franc de-peg | Market / liquidity gap | The ability to exit in a violent move | Liquidity can vanish; leverage amplifies the gap |
Two build stories show what "good" looks like
The abstractions become concrete when you follow two institutions actually building something. The first is a bank constructing an electronic FX capability from a weak starting point. The second is a corporate treasury imposing discipline on a sprawling hedging problem. Both are composites drawn to illustrate the shape of real programmes; the specifics are illustrative rather than reports of a particular firm. What matters is the sequence of decisions, because the sequence is what separates a coherent build from an expensive collection of parts.
Transformation, part one: deeper automation and more straight-through processing
The first arm of the transformation agenda is not a new idea but the extension of an old one: push automation deeper into the products and processes that have resisted it. Spot is solved. The frontier is everything spot's automation left behind — the longer-dated products, the larger tickets, the manual handovers, the screen-bound distribution. The direction of travel is consistent, and it is worth stating as a set of concrete movements rather than a vague promise of "more technology."
The direction of travel
Automation is spreading from spot into forwards, swaps and NDFs, and up the size curve into larger tickets. The pricing and hedging logic that mastered liquid spot is being extended, cautiously, into products that used to demand a human. Execution algorithms and transaction-cost analysis are becoming standard even for mid-sized clients, not just the largest accounts — a corporate treasury now expects the algo and the post-trade analytics that were once reserved for the biggest funds. Full straight-through processing — end to end with no manual re-keying — is the operational target, closing the fragmented edges where breaks still hide. Distribution is going API-first: the single-dealer platform becomes a set of documented endpoints a client's systems consume directly, not merely a screen a human clicks. Cloud and elastic compute give the models more horsepower at controlled cost, letting a desk run heavier calculations without owning a data centre's worth of idle hardware. And the venue and technology landscape is consolidating — CME absorbing EBS, LSEG absorbing Refinitiv — while banks rationalise their own sprawling stacks into fewer, better-integrated systems.
The effect on people is a migration, not an extinction. The human moves toward the exceptions the machine cannot handle, the large and complex trades, the client relationships, and — most importantly — the oversight of the machines themselves. The competitive edge shifts from people to platform, data and engineering. The best trader in the world cannot out-quote a good engine in liquid spot; the firm with the best engine, the cleanest data and the strongest engineering wins.
| Trend | Where it attaches | Why it matters |
|---|---|---|
| Deeper automation | Forwards, swaps, NDFs; larger tickets | Extends the profitable spot model into higher-touch products |
| Algo & TCA everywhere | Execution and post-trade, even mid-sized clients | Lowers impact and evidences best execution as standard |
| Full STP | Every handover front to back | Closes the fragmented edges where breaks and cost hide |
| API-first distribution | The SDP and client integration | Makes the platform a set of endpoints, not just a screen |
| Cloud & elastic compute | Model calculation and analytics | More horsepower for models at controlled, variable cost |
| Venue & tech consolidation | Venues and the bank's own stack | Fewer, better-integrated systems; rationalised cost |
Apply one filter to every "innovation" the market presents. Does it change the economics or the risk of the business, or does it merely change the appearance? Automating a manual forwards workflow changes economics — it removes cost and operational risk. A slicker screen that prices no better than the old one changes appearance and nothing else. This filter cuts through most of the noise. Judge each transformation claim by whether it actually moves cost or risk, and discard the ones that only move the pixels.
Transformation, part two: data and AI, with honest caveats
The second arm of the agenda is data and artificial intelligence, which is where the genuine promise and the genuine danger both live. FX is, in some ways, an ideal home for machine learning — enormous quantities of data, fast feedback, clear objectives. In other ways it is hostile territory — no single source of truth in the data, an adversarial market, and a regulatory expectation of explainability that black boxes cannot meet. The honest position is neither hype nor dismissal. It is a clear-eyed account of where the models help and where they must be held on a short leash.
Where AI and data genuinely help
Predictive pricing and skew is the most direct application. Models trained on flow and market data can set smarter spreads and skews than static rules — learning, for instance, which conditions precede a move and pricing accordingly. Flow and internalisation optimisation is close behind: models help decide what to internalise, what to skew, and what to hedge, and they help profile the toxicity of incoming flow, the very judgment central to modern FX profit. Smarter algo execution adapts the slicing and routing of a large order to live market conditions rather than following a fixed schedule. Surveillance and transaction-cost analysis use anomaly detection to find the needle — the suspicious pattern, the outlier fill — in datasets far too large for humans to read. And operations and data quality benefit from models that clean, reconcile and validate the fragmented data the whole business depends on.
The honest caveats
Data is the bottleneck, and it is the first caveat because it undermines all the others. With no single tape and with data that is fragmented, expensive and licensed, the models are only ever as good as the partial, costly picture they are fed. Markets are adversarial, which is the deepest problem. FX is non-stationary — its statistical behaviour changes over time — and reflexive, meaning the act of trading on a pattern can destroy the pattern. A model that worked yesterday can be gamed by competitors who reverse-engineer it, or can simply break in a regime shift like the Swiss franc de-peg. Explainability and control are non-negotiable. Automated pricing and hedging must run on models that can be explained and constrained, with hard limits and kill switches; a black box that no one can interrogate and no one can stop is not an asset but a hazard. And conduct and fairness bind everything. AI used in pricing must stay inside the FX Global Code — no unfair skewing, no misuse of information, no opaque last look dressed up as a model output. An algorithm that manipulates is still manipulation.
Generative AI earns its place only where it changes economics or risk
Generative AI — models that produce text and analysis rather than prices — has a real but bounded role, and it earns its place only where it changes economics or risk. The credible use cases share a shape: they take work that consumed skilled human time and do it faster, without touching the pricing and hedging decisions that must stay explainable and controlled. First, drafting and summarising trade documentation and client-facing materials — turning the raw facts of a trade or a relationship into clean prose a person then checks. Second, reviewing and querying large volumes of communications and policy for surveillance and compliance — reading years of chat and thousands of pages of policy far faster than a compliance officer could. Third, assembling research, market commentary and client-ready analytics — pulling scattered inputs into a first draft of the story a salesperson tells a client.
The specific risks are as concrete as the uses. Material non-public information and confidentiality: a model must never leak or misuse client-flow information or other material non-public information, a risk that is acute in a business built on knowing who is trading what. Hallucination and accuracy: a generative model that fabricates a number in a client document, or feeds a wrong input into a process, is dangerous precisely because its output looks confident and clean. Explainability and model-risk governance: an automated decision or output has to be traceable and governed like any other model in the bank, not waved through because it came from a fashionable tool. The gain is real, but it is only realised if the governance is as serious as the capability.
| Use of GenAI | The gain | The governing risk |
|---|---|---|
| Drafting and summarising trade docs and client materials | Speed and consistency on routine documentation | Hallucinated figures; leakage of client or MNPI |
| Reviewing and querying communications and policy for surveillance | Reads vast text far faster than humans | Missed context; false confidence; confidentiality |
| Assembling research, commentary and client analytics | Faster first drafts of the client story | Accuracy of inputs; explainability of conclusions |
Market-risk capital is now a data problem: FRTB
The capital a desk consumes decides which business is worth doing
A bank cannot run a trading book for free — it must hold regulatory capital against the risk that the book loses money, and that capital is a cost that decides which FX business is worth doing. Regulatory capital is the cushion of shareholders' own money a bank is required to set aside so that a bad run of losses is absorbed by owners rather than by depositors or taxpayers. Market risk is the risk that the value of the trading book moves against the bank because prices — exchange rates, interest rates, volatilities — change. The more capital a given desk is forced to hold, the more profit that desk must generate simply to earn an acceptable return on the money tied up behind it. Capital, in other words, is not an accounting afterthought; it is a price signal that tells the business which currencies, which tenors, and which products are economic to trade at all.
The Fundamental Review of the Trading Book, universally shortened to FRTB, is the Basel Committee's ground-up rebuild of how that market-risk capital is calculated — and it lands hardest exactly where price data is thin. The Basel Committee on Banking Supervision is the global standard-setter whose rules national regulators then write into local law. FRTB replaced a patchwork of older market-risk rules that the 2008 crisis had exposed as too easy to game and too blind to illiquid positions. The reform has been years in the making and is being phased in across major jurisdictions through the mid-2020s, with implementation dates that continue to move. For a technologist the headline is not the formula. It is that FRTB rewrites capital as a function of how much genuine, observable market data a bank can produce to justify its own risk estimates.
FRTB turns capital into an evidence problem
FRTB offers two routes to a capital number, and the choice between them is really a choice about how much a bank is willing to invest in data and modelling. The first route is the standardised approach, in which capital is computed from prescribed sensitivities of the book to a defined set of risk factors. A risk factor is any market variable whose movement changes the book's value — a specific exchange rate, a point on a yield curve, an implied volatility. A sensitivity measures how much the book gains or loses for a small move in one such factor. Under the standardised approach the regulator supplies the risk weights and the aggregation rules; the bank plugs in its sensitivities and arrives at a capital charge with little discretion. It is robust, auditable, and generally more expensive in capital terms, because prescribed weights are deliberately conservative.
The second route, the internal models approach, lets a bank use its own risk models to compute capital — but only if it can pass a set of hard, ongoing tests, and this is where FRTB becomes an evidence problem rather than a modelling one. The internal models approach can produce a lower, more risk-sensitive capital number, which is why sophisticated banks want it. The price of admission is proof. FRTB does not simply ask whether a model is clever; it asks whether the bank can demonstrate, factor by factor and day after day, that the model is anchored in real observed prices and that it explains the profit and loss the desk actually books.
The gatekeeper is the risk-factor eligibility test, and its sharpest tooth is the concept of the non-modellable risk factor, or NMRF. To treat a risk factor as modellable — and therefore eligible for the more favourable internal-models capital — a bank must show it has a sufficient number of real, verifiable price observations for that factor over a defined window. A common threshold discussed in the rules is on the order of a couple of dozen genuine observations a year with no excessive gaps, though the precise criteria are technical and jurisdiction-specific. Any risk factor that fails this test is declared non-modellable and is pushed into a separate, punitive capital calculation known as a stressed capital add-on. The practical consequence is blunt: a currency pair or a tenor for which the bank cannot evidence enough real trades and quotes attracts materially more capital, whatever the model says its risk actually is.
On top of eligibility sits the profit-and-loss attribution test, which checks that the model's story about risk matches the desk's real earnings. Profit and loss, always abbreviated P&L, is the daily gain or loss the book records. The attribution test compares two versions of that number: the P&L the risk model predicts from its risk factors, and the P&L the front-office systems actually produce. If the two drift too far apart, the model is judged to be missing something real about the book, and the desk can be forced off internal models and onto the more expensive standardised approach. The test turns a modelling question into a reconciliation question, and reconciliation is a data and systems discipline.
FRTB is, at bottom, a data and systems build wearing the costume of a capital rule. Proving that a risk factor is modellable means capturing and storing real transaction and quote data across every currency and every tenor the desk touches, then keeping that evidence clean and queryable for years. Running the framework means computing sensitivities across the entire book every day, aggregating them under the regulator's rules, and reconciling the risk model's P&L against the actual P&L booked by the front office. For an FX desk trading dozens of pairs across many tenors, that is not a project with an end date — it is a permanent data-engineering obligation. And because FX has no consolidated tape, no single official record of every trade and price the way an exchange-traded market does, assembling enough observed data to clear the modellability bar is harder here than almost anywhere else in the bank.
The move to T+1 settlement put FX operations under a new clock
A shorter securities cycle collided with an unchanged FX cycle
In 2024 the United States shortened the time it takes to settle a securities trade from two business days to one, and that single change quietly reset the clock for FX operations around the world. Settlement is the moment the trade actually completes: the buyer's cash and the seller's asset change hands for good. The convention is written as a "T plus" number, counting business days from the trade date, T. US equities and most US securities moved from T+2 — settlement two days after the trade — to T+1, settlement the next business day. The intent was sound: a shorter cycle means less time for a counterparty to fail and less risk sitting in the system between trade and completion.
The problem is that spot FX did not move — it still settles T+2 — so the securities leg and the currency leg that funds it now finish on different days. Consider a European investor buying US shares. The shares must be paid for in dollars, and those dollars are usually obtained through an FX trade selling euros for dollars. Under the old regime both legs settled T+2 and lined up neatly. Now the shares demand payment on T+1 while the standard FX to raise the dollars does not deliver them until T+2. The investor needs the dollars a full day before the ordinary currency trade produces them. A mismatch that simply did not exist before now sits at the heart of every cross-border securities purchase, and it is an operational and funding gap, not a market view.
The answers are same-day FX, pre-funding, and same-day PvP
There are three practical ways to close the one-day gap, and each buys alignment at a cost. The first is to trade the FX for same-day settlement — T+0, delivery on the trade date itself — so the dollars arrive in time for the T+1 securities leg. Same-day FX is available but typically prices worse than standard spot, because the market must scramble the liquidity into place on short notice, and the cut-off times to arrange it are early and unforgiving. The second is to pre-fund: hold a buffer of dollars ready in advance so the securities payment can be met without waiting for any FX to settle. Pre-funding removes the timing risk but ties up cash that would otherwise be earning a return elsewhere, and someone has to fund and manage that idle balance.
The third answer is to settle the FX itself faster, using same-day payment-versus-payment rails such as CLSNow. Payment versus payment, or PvP, is the mechanism that ensures both currencies in an FX trade move together or not at all, so that neither side can pay away its currency and fail to receive the other — the settlement risk we treated earlier as the one exposure in FX with no second chance. CLSNow is a same-day PvP service that lets two parties settle certain currency pairs within the day while preserving that all-or-nothing protection. It solves the timing problem without giving up safety, but it covers a limited set of currencies and still demands that both parties operate to a tighter, intraday schedule.
The lesson runs wider than one 2024 rule change: a settlement cycle is a shared clock, and when one market shortens its clock the shock travels straight into FX operations. Other jurisdictions are following the US toward T+1, and the direction of travel across markets is toward shorter cycles still, with T+0 already discussed. Every compression squeezes the window in which FX can be arranged, funded, and settled, and pushes more of the burden onto same-day processing and pre-positioned liquidity. For an FX operations team the strategic point is that its own tempo is set partly by decisions taken in adjacent markets it does not control.
New competitors are attacking the cross-border franchise
Fintechs are compressing margins at the retail and SME edge
The most profitable slice of cross-border FX has long been the retail and small-business end, where customers paid wide spreads for currency conversion because they had no easy way to see or shop the price — and that is precisely the slice payment fintechs have attacked. A spread here is the gap between the true wholesale exchange rate and the rate the customer is actually given; historically that gap was large and largely invisible, buried in an "all-in" rate with no explicit fee. Payment fintechs — the money-transfer specialists and multi-currency account providers that grew up over the past decade — built their proposition on tearing that opacity open: show the customer the mid-market rate, charge a small, explicit fee, and settle quickly. By competing on transparency and price rather than on branch networks, they pulled a profitable customer base away from banks and forced spreads down across the segment.
This is a genuine business threat to a lucrative edge of the franchise, even though it is a small share of wholesale volume. The wholesale market — banks and large institutions trading currency risk in size — dwarfs retail and small-business flow in notional terms, and the fintechs have made only limited inroads there. But margin, not volume, is what matters to profitability, and the retail and SME edge carried far higher margins than wholesale ever did. Losing share there, and being forced to match fintech transparency on the flow that remains, compresses a disproportionately profitable business. A bank can be dominant in wholesale FX and still watch a valuable revenue pool erode from the retail side.
Stablecoins and CBDCs could rewire the settlement layer
Two newer technologies aim not at the pricing of FX but at the rails beneath it — the plumbing that moves value across borders — and they deserve a measured look rather than either hype or dismissal. A stablecoin is a digital token designed to hold a fixed value against a reference currency, typically a major one like the US dollar, so that it can be sent quickly and cheaply across borders while behaving like the currency it tracks. A central bank digital currency, or CBDC, is a digital form of central-bank money itself — the same claim on the central bank that physical cash represents, issued in electronic form. Both are, in principle, new ways to hold and move currency value that do not necessarily pass through the existing chain of correspondent banks.
Correspondent banking — the network of banks holding accounts with one another to move money between countries — is the plumbing these technologies could partly bypass, and that is where their real potential lies. Today a cross-border payment often hops through several correspondent banks, each adding time, cost, and a point of failure. A stablecoin or a CBDC arrangement could, in time, let value move more directly, and could even provide an alternative rail for the settlement of FX trades themselves — the final exchange of one currency for another. The change they threaten is to the settlement and payment layer, not to the business of deciding what a currency is worth.
The right posture is measured: the wholesale risk-transfer business is not obviously threatened, but the layer beneath it might be. Pricing currency risk and warehousing it — standing ready to take the other side of a client's exposure and managing the resulting position — is a business built on judgement, balance sheet, and access to liquidity, none of which a new settlement rail replaces. What a stablecoin or CBDC could displace is the mechanics of how the agreed value actually moves once the trade is struck. Confusing the two leads to overreaction. The pricing and warehousing franchise sits on firmer ground than the plumbing does.
Apply the same filter to these new competitors that we applied to every other part of the transformation story: does the change alter the ECONOMICS of the business, alter its RISK, or merely alter its appearance? On that test the payment fintechs pass clearly — they genuinely changed the economics of retail and SME cross-border FX by collapsing spreads and exposing prices, and that shift is already banked, not hypothetical. Stablecoins and CBDCs are, for now, a potential change to the settlement rail rather than a realised change to economics or risk. They are worth watching precisely because settlement is the one part of FX with no second chance: a currency that leaves your account against a payment that never arrives is simply gone. Any new rail must first prove it is as safe as the PvP protection of CLS before it can earn the trust of the wholesale market — and until it does, it is a frontier to monitor, not a threat to price in.
Model governance is now a regulatory regime, not an internal policy
The models that run an FX desk have crossed the line from internal good practice into formal regulation. The pricing model that quotes a spread, the skew logic that shades quotes to attract the flow the desk wants, the auto-hedging engine that lays off risk without a human, and the surveillance models that watch for misconduct are no longer governed only by the firm's own sense of what is prudent. They fall under model-risk management as a supervised discipline, and increasingly under AI-specific law. Treating model governance as an internal policy the firm writes for itself, and can revise or waive at will, is now a misreading of where the rules come from.
Classic model governance assumes you can validate a model against the truth and reproduce its output — and the newest models break both assumptions. Model-risk management grew up around models you could pin down: feed the model known inputs, check its output against a ground truth, and confirm that running it again gives the same answer. Validation (independently testing that a model does what it claims) and reproducibility (getting the identical output from the identical input) were the load-bearing ideas. Machine-learning and generative models — models that produce new content rather than a single computed number — resist both. They can be hard to validate against any clean ground truth, they may not reproduce exactly from one run to the next, and they drift as the data around them shifts. Governance has to adapt rather than pretend the old tests still fit: continuous monitoring instead of one-off validation, challenger models (a second model run in parallel to test whether the primary one still looks right), hard limits the model cannot exceed, and a human override that can always stop it.
The EU AI Act turns all of this into a tiered legal regime with a hard floor. The Act classifies AI uses into risk tiers — unacceptable (banned outright), high (heavily regulated), limited (transparency obligations), and minimal (largely unregulated) — and scales the obligations to the tier, so a high-risk use carries far more duty than a minimal one. It also sets baseline obligations that bite regardless of tier: AI literacy for staff who use these systems, and transparency that a system is AI at all when a person interacts with it. Most firms hit the same wall first, and it is not legal interpretation — it is inventory. You cannot govern models you have never catalogued, and firm after firm has discovered late how many models and AI tools were quietly in use across desks, spreadsheets and vendor products. The register comes before the regime; without knowing what you run, the tiers and obligations have nothing to attach to.
| Dimension | Traditional model | Generative model |
|---|---|---|
| Validation | Tested against a ground truth, pass or fail | Often no clean ground truth to test against; judged on behaviour over time |
| Reproducibility | Same input gives the same output every run | May give different outputs from the same input; exact reproduction is not guaranteed |
| Explainability | The logic can be traced and explained step by step | The path from input to output is largely opaque, even to its builders |
| Ownership | A named owner and a documented specification | Often a vendor's model behind an interface, with the firm owning the use but not the internals |
| Uncontrolled change | Changes only when the firm deliberately re-releases it | Can change under the firm's feet as the vendor updates the model or as it drifts |
5.1 Reject-ratio KPI
Of 1,200 client attempts to trade on your streamed price, 90 are rejected under last look. What is the reject ratio, and why can a high number damage the franchise even when each rejection was individually defensible?
90 / 1,200 = 7.5%. Even individually defensible rejections, in aggregate, tell clients your price is unreliable; they route their flow elsewhere, so a high reject rate erodes the franchise regardless of each justification.
5.2 Internalisation KPI
A desk internalises EUR 800m of EUR 1,000m gross client flow. What is its internalisation ratio, and how much residual risk reaches the external market?
800 / 1,000 = 80%; the residual EUR 200m reaches the external market to be hedged. The higher the ratio, the less spread the desk pays away and the smaller its market footprint.
5.3 Failure-mode mapping
Match each real case to its failure mode: (a) the benchmark-fixing scandal; (b) Herstatt, 1974; (c) the 2015 Swiss-franc de-peg.
(a) fixing scandal → conduct / manipulation failure; (b) Herstatt → settlement failure; (c) Swiss-franc de-peg → market / liquidity-gap failure. The three cases are the three ways FX breaks.
5.4 GenAI governance
A quant wants to paste a week of live client-flow data into an external chatbot to summarise trading patterns. State the specific risk and the control that should stop it.
The risk is leaking material, price-sensitive client-flow data into an ungoverned external tool — it cannot be recalled and leaves no surveillance trace. The control is data-loss prevention plus a hard block on pasting client or price-sensitive data into external AI tools, with an approved internal alternative and clear policy.
5.5 FRTB judgement
A desk trades a thinly-traded emerging-market pair with few observable prices. Under FRTB, what happens to the capital on that risk, and why?
With few observable prices, the risk factor is deemed non-modellable (an NMRF) and attracts punitive capital. The thin emerging-market pair is expensive to hold precisely because the data needed to model it is missing — a capital cost driven by data scarcity.
Explain to someone who has never traded why the 1974 failure of a single German bank still shapes how every currency trade in the world is settled today — without using the words settlement risk, Herstatt or CLS.
What makes settlement risk uniquely acute in FX, and what tamed it?
Two currencies exchange across time zones, so one leg can be paid before the other fails; CLS payment-versus-payment settlement removed it for the covered currencies.
How do the three real failure cases map to three failure modes?
The fixing scandal is a conduct failure, Herstatt a settlement failure, and the Swiss-franc de-peg a market and liquidity failure — the three ways FX breaks.
What single filter separates real transformation from cosmetic change?
Whether it moves the economics or the risk of the business, rather than merely its appearance.
How should generative AI be governed in an FX business?
As tightly as it is useful: used where it changes economics or risk, with the human kept in the loop as a control, and the input — a paste of live deal material — watched as closely as the output.
Why is FRTB as much a data problem as a capital rule?
A risk factor only escapes punitive capital if the bank can prove it is modellable with enough real observed price data — hard in a fragmented market with no single tape.
Carrying forward from Part 2: how does the internalisation you learned to prize also reduce a risk in this part?
Netting flow in-house cuts external market impact and settlement exposure, lowering execution and settlement risk on the residual.
Carrying forward from Part 4: why does operational resilience belong in a chapter on risk?
Because the automated stack failing — a bad quote streamed at scale, or a missed settlement — is itself a top operational risk, and resilience is its main defence.
5
Why is settlement risk uniquely acute in FX?
No.
No.
No.
Right. CLS tamed it for covered currencies.
Which case is a market and liquidity failure?
Right. The fixing scandal is conduct; Herstatt is settlement.
Conduct.
Settlement.
The de-peg is.
How should generative AI be governed in FX?
Right.
Not the position.
No.
No.
Which filter separates real transformation from cosmetic change?
No.
No.
Right.
No.
The Six Ideas That Matter Most
If the reader forgets every table, every case and every acronym in this module, these six ideas are the ones that carry the whole thing. They are the load-bearing walls; everything else is decoration hung on them.
Idea 1 — In FX, the product is a pair, so price and risk are always relative
There is no absolute price of a currency, only its price against another, and this single fact shapes everything. Every FX trade is a simultaneous purchase of one currency and sale of another, so a "price" is a ratio and a "position" is always a relative exposure — long one currency is short another. There is no neutral cash to retreat to; even holding dollars is a position against everything else. This relativity is why the market has thousands of pairs rather than a list of prices, why risk is measured as exposure between currencies, and why the trade that looks simplest — swap one money for another — generates a book that must be managed with care.
Idea 2 — With no exchange and no tape, seeing the true market is the central problem
FX has no central exchange and no consolidated record of trades, so the hardest and most valuable thing is simply seeing the true market. Prices live scattered across single-dealer platforms, multi-dealer platforms and ECNs, none of them complete, and no single tape stitches them together. That is why banks spend so heavily on aggregation, data and connectivity: the consolidated view that a stock trader gets for free must be assembled, at cost, in FX. The firm that sees the real market a fraction more clearly prices better, hedges better and proves best execution more easily. Fragmentation is not a nuisance around the edge of the business; it is the central problem the whole technology stack exists to solve.
Idea 3 — Internalisation, not the spread alone, is where modern FX profit is made
The visible spread is thin and getting thinner; the real profit comes from internalising flow — matching one client's buy against another's sell without touching the market. A bank with enough flow can offset a large share of its trades internally, capturing spread on both sides while paying nothing to hedge and taking no market impact. This is why scale wins: more flow means more internalisation, which means better prices, which attracts still more flow. It is also why flow quality matters as much as flow quantity — "toxic" flow from counterparties who are systematically right leaves the bank holding losing positions, while benign franchise flow nets cleanly. Modern FX profitability is a game of internalising good flow at scale, not of quoting a wide spread.
Idea 4 — The pricing engine is the franchise, and compliance and resilience are core architecture
The pricing, skew and internalisation engine is the one thing a bank must build rather than buy, because it is the franchise itself — and the controls around it are part of that architecture, not an afterthought. Everything else in the stack can be rented: venues, connectivity, settlement, reporting. The decision engine cannot, because it encodes the bank's edge. And the things that keep it safe — adherence to the FX Global Code, kill switches that halt a misbehaving engine, payment-versus-payment settlement, and regulatory reporting — are designed in from the start, not bolted on afterwards. A pricing engine without built-in conduct and resilience is not a faster franchise; it is a scandal or a settlement failure waiting to happen.
Idea 5 — Settlement is where the market's oldest risk still lives, and CLS is why it is mostly tamed
An FX trade is not done until both legs settle, and the risk that one side pays and the other fails to receive is the market's oldest and most dangerous exposure. Herstatt named it in 1974; CLS and payment-versus-payment mostly tamed it decades later by settling both legs simultaneously for the major currencies. But "mostly" is the operative word. The many emerging-market currencies outside CLS still carry the full Herstatt risk, so settlement remains live architecture and a standing regulatory concern rather than a solved historical problem. The most sophisticated pricing engine in the world is worthless if, on value date, the cash does not arrive.
Idea 6 — Real transformation moves cost or risk, including disciplined, governed AI
The test of any innovation is whether it changes the economics or the risk of the business, and that test applies to AI as strictly as to anything else. Automating a manual forwards workflow moves cost; a slicker screen that prices no better moves nothing. Machine learning genuinely helps with pricing, internalisation, surveillance and data quality, and generative AI helps with documentation, compliance review and research — but only when governed as tightly as it is useful, with hard limits, explainability and protection of confidential information. In a market this fast, this adversarial and this lightly regulated by hard law, the human in the loop is a control, not a limitation. Real transformation moves cost or risk; everything else is a new coat of paint.
Process Cheat Sheet — The Whole FX Trade on One Page
A single reference for the entire FX flow, from a client's request for a price to settled cash and a filed report. Use it to place any conversation, system or control in the pipeline.
The trade lifecycle, stage by stage
Every FX trade runs the same seven-stage pipeline. In liquid spot the first four stages collapse into milliseconds; the trade is only truly done at settlement.
| Stage | Timing | Core activity | Key output | Lead and systems |
|---|---|---|---|---|
| 1. Request or stream | µs–seconds | Client sends an RFQ, or the bank streams live prices | A live two-way price on the client's screen | Sales, SDP / MDP, pricing engine |
| 2. Price | microseconds | Engine computes a fair mid, spread, skew and client tier | A streamed, skewed bid / offer | Pricing engine, market-data feeds |
| 3. Execute | sub-second | Client deals; the trade is captured and booked | A booked trade record | OMS / EMS, FIX / binary API |
| 4. Risk and hedge | real-time | Position hits the book; internalised or auto-hedged | A flatter book | Internalisation and auto-hedge engines, risk system |
| 5. Confirm and match | minutes | Both sides electronically agree the economics | A matched trade | Confirmation / matching platform |
| 6. Settle (CLS) | T+2 (T+1 some) | The two currencies are exchanged simultaneously | Settled cash, no leg risk | CLS, SWIFT MT / ISO 20022, nostro accounts |
| 7. Report and reconcile | end of day | Trade / transaction reporting, reconciliation, P&L | Regulatory reports, clean books | EMIR / MiFIR engines, reconciliation |
Key numbers and rules of thumb
The figures below are market rules of thumb, not precise quotes; illustrative prices are flagged as such in the body.
| Metric | Rule of thumb |
|---|---|
| Daily FX turnover | ~7–8 trillion US dollars, larger than all world equity markets combined |
| Trading hours | ~24 hours, across Sydney, Tokyo, London and New York |
| Spot settlement | T+2 (T+1 for USD/CAD and a few pairs) |
| Most-traded pair | EUR/USD |
| Biggest product by turnover | FX swaps |
| Typical spread, liquid major | a small fraction of a pip |
| A pip | the 4th decimal place in most pairs |
| Share of FX that is electronic | the large majority by volume; spot almost entirely |
| Internalisation, top banks | the large majority of spot flow netted in-house before hedging |
| Settlement mechanism | CLS payment-versus-payment, across the major currencies |
| Currencies outside CLS | many emerging-market pairs — settlement risk persists |
| Conduct standard | the FX Global Code (voluntary, but effectively expected) |
| Origin of settlement risk | the 1974 Herstatt Bank collapse |
| US securities settlement | T+1 since 2024, versus FX spot T+2 — a one-day funding gap |
| Option premium | the option buyer's maximum loss — the price of keeping the upside |
The entire trade in one sentence
A client needs to swap one currency for another; the desk streams a skewed, tiered two-way price from its engine, the client deals in milliseconds, the desk nets the risk against other client flow and hedges only the residual, and days later the two currencies are exchanged simultaneously through CLS — a thin spread on vast volume, almost all of it automated.
Sources and further reading
This is a teaching document, not a legal or regulatory reference. The running example and the worked exercises use invented figures, chosen for clarity and labelled illustrative wherever they appear. The tables below let a reader trace every named rule, standard and dated fact to its origin. As-at date: July 2026. Regulation in this field changes on a scale of months — verify anything time-sensitive against the current official text before relying on it.
Rules and standards named in the text
| Short name | What it is, and where to find it |
|---|---|
| FX Global Code | Voluntary global standard for wholesale FX conduct, maintained by the Global Foreign Exchange Committee (GFXC). |
| MiFID II / MiFIR | EU Directive 2014/65/EU and Regulation 600/2014; technical standards from ESMA. Governs FX derivatives, not spot. |
| EMIR | EU Regulation 648/2012 on OTC derivatives — reporting, risk-mitigation, margin. |
| MAR | EU Market Abuse Regulation 596/2014 — manipulation and misconduct, including around fixings. |
| DORA | EU Digital Operational Resilience Act, Regulation (EU) 2022/2554. |
| EU AI Act | Regulation (EU) 2024/1689 — risk-tiered rules for AI systems. |
| FRTB | Basel Committee "Minimum capital requirements for market risk" (the Fundamental Review of the Trading Book). |
| Basel III / CRR | Basel III standards, implemented in the EU through the Capital Requirements Regulation. |
| CLS | CLS Bank International — member-owned payment-versus-payment FX settlement utility. |
| ISDA / CSA | International Swaps and Derivatives Association Master Agreement and its Credit Support Annex. |
| BIS Triennial Survey | Bank for International Settlements survey of global FX turnover. |
| ISO 4217 / ISO 20022 | ISO currency codes / ISO financial-messaging standard. |
| LEI / UTI / FIGI | Legal Entity Identifier (GLEIF) / Unique Trade Identifier / Financial Instrument Global Identifier (an open, vendor-neutral instrument code). |
| AMF / ACPR / ESMA / ECB | French market and prudential regulators / the EU securities authority / the European Central Bank. |
Dated facts stated in the text
| Fact as stated | Origin to verify against |
|---|---|
| FX turnover of roughly 7–8 trillion US dollars a day | BIS Triennial Central Bank Survey (most recent edition). |
| US securities settlement moved to T+1 in 2024 | US SEC rule; effective 28 May 2024. |
| DORA applies from January 2025 | Regulation (EU) 2022/2554; application date 17 January 2025. |
| The EU AI Act's risk-tier obligations phase in from 2025 | Regulation (EU) 2024/1689; entry into force 2024, phased application thereafter. |
| FX benchmark-fixing enforcement and fines | Multi-regulator actions, 2013–2015. |
| The Herstatt collapse | 26 June 1974. |
| The Swiss-franc de-peg | Swiss National Bank, 15 January 2015. |
| The main FX fixing window widened to around five minutes | Benchmark reform following the 2013–2015 scandals (from February 2015). |
Charts and figures
The market-cycle chart in Part 1 is an illustrative, indicative shape, not plotted from a specific data series; for actual FX volatility and turnover series, consult BIS statistics and central-bank releases. The spread-decomposition, depth-curve, pricing-engine, P&L-waterfall and forward-points figures are schematic illustrations of the mechanics, not market data. The running example (the French exporter) and every worked exercise use invented numbers by design.