Skip to content
Corporate & Investment Banking My progress Français

Home / Capital Markets & Securities Services / Capital Markets / Module 9

9Equity Trading (Cash & Algo)

Execution, Microstructure and the Machine · How a Bank Buys and Sells Shares for Clients

Buying and selling shares for clients, by hand and by algorithm.

Reading time
2 h 10
Sections
62
Recall questions
34
Quiz questions
20
Simulators
2

This module teaches equity trading from zero, but goes deep enough for a professional

It assumes no prior knowledge of trading floors and builds every idea from its foundations. Yet it goes deep enough to be useful to someone who already writes algos, runs a book, or builds the platforms underneath.

Wherever a concept is technical, we stop and explain it in plain language before using it again. Wherever a term is jargon, 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 equity trading is, why clients pay for it, how the bank makes money on razor-thin margins, how execution itself became a product, and where trading goes wrong.

The technology thread covers the front-to-back stack that carries an order, how one instruction travels from a click to custody, why liquidity scattered across a dozen venues can only be reassembled by software, and why data quality at one step decides whether the next five work.

These two threads are not separate. Equities is the most electronic corner of the bank, and here the process and the platform are the same thing seen from two angles.

How to read this module

Read it as a narrative, not a reference card

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.

The five parts build on each other in a single line

Figure 1 How the module fits together: each part answers the question the previous one raises, and one running example carries through all five.

Part 1 explains why the business exists and what it sells. Part 2 follows a single order from arrival to settlement, and is where the mechanics live. Part 3 asks who does all this and how they are kept apart. Part 4 opens the machine underneath. Part 5 asks how it fails and where it is going.

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 keeping even if you forget everything around it. Going deeper adds nuance for readers who already have the fundamentals. IT & data lens connects a business idea to the systems and data that carry it. Watch out flags a trap or a moving target. Worked example follows our running case. In a nutshell recaps a part. Check your understanding lets you test yourself.

Do the arithmetic yourself, and explain it out loud

Two habits will get you far more from this module than reading it carefully. First, redo every calculation. The worked examples deliberately show their arithmetic — the basis points, the share counts, the euros — so that you can reproduce them on paper. A cost you have computed yourself is a cost you understand; one you have only read is a number you will forget by the next page. Second, explain each section to someone else in three sentences before moving on. If you cannot, you have not understood it yet, and the gap in your explanation is precisely the thing to reread. Both habits come from the same principle: understanding is the ability to reconstruct an idea, not the feeling of having followed one.

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 equity trade end to end

You will walk through every step from the moment a client's order arrives to the moment shares and cash change hands and the position is safely held. You will understand how a high-touch block differs from a low-touch algo order and from a whole basket traded at once, and why the choice shapes everything downstream.

You will be able to map every actor

You will know who does what inside the bank and outside it — buy-side, venues, liquidity providers, clearing and settlement infrastructure, regulators — how these parties interact, and where the information barriers between teams become real engineering constraints.

You will understand where execution quality actually comes from

You will know how the order book sets a price, how an algorithm decides how much to trade and when, how a router decides where, and why the true cost of a trade is the impact you cause and the time you take — not the headline commission.

You will be able to trace an order front to back

You will follow a single order through the OMS, the EMS, the algo and smart order router, the venues, allocation, confirmation, clearing and settlement, and regulatory reporting. You will understand why an error early in that chain reaches every step after it.

You will recognise the risk and transformation landscape

You will know the regulations that shape every routing and pricing decision, the failure modes that have broken real firms and briefly broken the market itself, and the places where better technology — including AI — genuinely changes the economics, not just the appearance.

1Context and the Big Picture

Before any process makes sense, you need the map. This part defines what equity trading is, places it inside the bank, lays out the execution channels the desk sells, and explains the market forces — fragmentation, thin margins, and cyclicality — that everything else builds on.

Equity trading is the business of moving shares between owners

A share is a slice of a company, and trading is simply changing who owns the slice

A share of stock is a legal claim on a slice of a company: a slice of its profits, a slice of its assets if it is ever wound up, and usually a vote. That much you may already know. What matters here is narrower. Equity trading is the business of buying and selling those shares in the secondary market — for clients, or with the bank's own capital. It is not about deciding which shares are worth owning. That decision has already been made by someone else, usually a fund manager. The desk's job begins the moment that decision becomes an order.

The distinction between the primary and the secondary market is the first thing to fix. In the primary market, a company sells newly created shares and receives the cash — this is an initial public offering or a follow-on, and it is the world of Equity Capital Markets, the subject of Module 1. In the secondary market, investors trade existing shares among themselves, and not a cent reaches the company. The everyday image of the stock market, prices flickering on a screen, is the secondary market. That is where equity trading lives.

One more line to draw before we go further. Cash equities means trading the actual shares, and equity-like instruments such as exchange-traded funds and depositary receipts, for immediate delivery. It stands apart from equity derivatives — options, futures, swaps — which are contracts about shares rather than the shares themselves. This module is about the cash side and the electronic machinery around it. Derivatives are their own world, touched only where they hedge a cash position.

The desk's real job is to turn an intention into filled shares, provably well

A fund decides to buy two million shares of a company. That sentence hides the entire discipline. Two million shares may be several days of the stock's normal trading volume. Buy them all at once and the price runs away before the order is done; the very act of buying pushes the price up against the buyer. Trade too slowly and the market may drift away for reasons that have nothing to do with the order. The desk exists to solve exactly this: get the size done at the best achievable outcome — the right price, low market impact, across many fragmented venues, fast — and be able to prove it was done well.

That last clause is not decoration. In Europe especially, a bank must be able to evidence the quality of its execution to the client and the regulator. Proof is a deliverable, not an afterthought, and it is why measurement runs through this whole business.

Equities is the shop window of the trading floor

It is the most electronic, most transparent, most benchmarked market of all

Equities electronified earlier and further than any other asset class, for a simple reason: shares are standardised and exchange-traded. A share of a given company is identical to every other share of that company, and it trades on organised venues with published prices. That standardisation is what lets software take over. Compare a corporate bond, where thousands of separate instruments from one issuer trade by appointment over the phone, and you see why equities became the laboratory where electronic trading was invented and where it still runs furthest ahead.

Because prices are public and trades are benchmarked, equities is also the most measured market. Every fill can be compared to where the stock was when the order arrived, to the day's average price, to the closing auction. That measurability is a double-edged sword: it makes execution quality provable, and it makes it a commodity that clients shop for relentlessly.

It sits inside Global Markets, next to the businesses it feeds and is fed by

An investment bank is a collection of businesses under one roof. Equity trading sits inside the markets division — often called Global Markets or the trading floor — and specifically in the equities franchise. Around it sit the businesses it connects to. Equity Capital Markets creates the shares in the first place, in the primary market; those shares then trade on the desk in the secondary market forever after. Equity derivatives trade options and swaps that the cash desk often hedges. Prime brokerage lends shares and finances positions for hedge-fund clients, and shares plumbing with the cash desk. Research publishes the views that give clients reasons to trade — though, as we will see, regulation has forced a wall between paying for research and paying for execution.

The cash equities desk is the hub of this cluster. It is where an investment decision, wherever it originated, becomes an executed, settled trade.

The desk sells a spectrum of execution channels, from fully manual to fully automatic

Modern equities runs two connected businesses at once

The same desk runs a high-touch business and a low-touch business, and the difference is how much human judgement each order needs. High-touch is the world of the sales-trader: a person works large, illiquid or sensitive orders by hand, advises on strategy, sources hidden liquidity, and commits the bank's capital when the client wants certainty. Low-touch is the world of algorithms and direct market access: the client's orders flow straight into execution software with almost no human touching them. High-touch earns a higher commission and survives only where a human genuinely adds value. Low-touch earns a tiny commission and wins on scale.

Between and around these sit a handful of distinct products a client actually buys. Learn these and you can classify almost any piece of equity flow.

Five channels cover the field

Channel Touch Typical use Margin Key technology
Algo / DMA (low-touch) Machine Everyday institutional flow, sliced and worked automatically Very thin Algo suite, smart order router, EMS
High-touch cash / blocks Human Large, illiquid or sensitive orders; block liquidity; advice Higher Sales-trader tools, IOI network, block venues
Program / portfolio (basket) Blend Index rebalances, portfolio transitions, cash flows Thin–medium Basket engine, risk pricing, hedging
Capital commitment / risk Human + book Client wants the whole block done now, at a firm price Risk premium Risk book, pricing, balance sheet
ETF / electronic market-making Machine Continuous two-way quotes; ETF creation and arbitrage Spread Low-latency pricing, inventory management

Algo and direct market access is the everyday product. The client sends an order and chooses a strategy; software slices it and works it across venues while a person supervises. This is where most flow lives, and where margins are thinnest.

High-touch cash trading is the human premium. A sales-trader takes an order that is too big, too illiquid or too sensitive for a pure algo, advises on how to work it, quietly hunts for a natural counterparty, and — when asked — commits the bank's own money to take the block. This is judgement, discretion and a network: the things a machine cannot yet replicate.

Program trading straddles both. A whole list of stocks — tens or hundreds of names — is executed together as a single package, for an index rebalance or a portfolio transition. It is part electronic engine, part risk-priced capital commitment, and it is where single-stock execution meets portfolio risk.

Capital commitment is the balance-sheet product. For a big block the client may not want to wait: they want the bank to take the whole trade onto its book at once, at a firm price. The bank charges a risk premium and then has to unwind the position without losing more than it earned.

Electronic market-making is the always-on product. The desk quotes continuous two-way prices — in single stocks and in exchange-traded funds — earning the spread across millions of tiny trades and warehousing the resulting inventory. It is the bank competing directly with specialist high-frequency firms.

The reference case: working a large institutional order

Throughout this module we return to one reference case, because it exercises almost every mechanic in the business. A long-only fund wants to buy two million shares of a mid-cap company — roughly three days of that stock's average volume. It is too big to do at once without moving the price, too sensitive to reveal to the market, and subject to a duty to prove the execution was good. That single order will take us through strategy, liquidity sourcing, algorithms, routing, allocation, settlement and measurement. Whenever you see the Worked example box, it is this order, carried a step further.

The economics: thin margins, huge volumes, and a scoreboard that distorts behaviour

The money is made on volume, spread, capital and services — never on one fat trade

Equity trading is a scale-and-flow business. Per share, the desk earns almost nothing; in aggregate, across enormous volume, it earns a living. There are a few distinct revenue engines. Commissions are a fee per share or in basis points of value for executing client orders — the core of the agency business, tiny per trade but large in sum. Spread is the gap between the price the desk buys at and the price it sells at, captured on principal flow by standing in the middle. Capital commitment earns a risk premium for taking a block onto the book. Financing and securities lending — lending shares to short-sellers, financing long positions — is a steady, balance-sheet-driven income, closely tied to prime brokerage. And increasingly, execution services and technology — packaged algos, analytics, connectivity — are sold as a distinct, scalable line of their own.

A basis point is one hundredth of one percent. On a European cash-equity trade, the execution commission is often a small handful of basis points, and on the most competitive low-touch flow it is a fraction of that. When your margin is measured in fractions of a basis point, only two things keep you alive: volume and cost discipline. That single fact explains why the desk industrialised — why it automated everything it could and reserved humans for where they genuinely add value.

League tables and the broker scoreboard quietly shape every decision

Clients rank their brokers, and those rankings decide who gets the next order. Historically the ranking was soft — relationships, research, entertainment. Today it is increasingly hard and data-driven: the client measures each broker's execution quality with transaction-cost analysis and allocates future flow to whoever performs best, through a mechanism called the algo wheel that we meet in Part 2. Being top of a client's scoreboard is worth real money, because flow follows the ranking.

This scoreboard is mostly healthy — it rewards genuine execution quality — but it also distorts. It pushes the whole industry toward measurable, standardised, commoditised execution, and it intensifies the margin pressure that already defines the business. A broker that cannot prove its quality, cheaply and continuously, loses flow regardless of how good its traders are. The scoreboard rewards evidence as much as skill.

The commission is the smallest part of what a trade actually costs

Explicit costs are visible and small; implicit costs are invisible and large

Ask a client what a trade costs and many will quote the commission. That answer is wrong by an order of magnitude, and understanding why is the beginning of understanding execution. The costs of a trade divide into explicit costs, which appear on an invoice, and implicit costs, which never do.

The explicit costs are the ones you can point to. Broker commission — a few basis points, or a fraction of one on low-touch flow. Exchange and venue fees, charged per execution and varying by venue, which is one reason routers care where they send an order. Clearing and settlement fees, paid to the CCP and CSD per trade. And, in several European markets, transaction taxes, which are far from trivial: France levies a financial transaction tax on purchases of shares in large French-domiciled companies, Italy operates a similar regime, and the United Kingdom charges stamp duty reserve tax on share purchases. These taxes are typically levied in tens of basis points — an order of magnitude larger than the execution commission — and they fall on the buyer.

The implicit costs are larger still and never appear anywhere. Spread is what you pay to cross from bid to offer. Market impact is the price move your own order causes. Timing cost, or drift, is the market moving while you work. And opportunity cost is the part of the order you never managed to fill. None of these is invoiced; all of them are real money.

Figure 2 The all-in cost of a European share purchase, in basis points. The broker commission is the smallest bar; impact, drift and transaction tax dominate (illustrative, based on a French FTT case).

Unbundling severed research from execution, and changed the business

Once, research was paid for with trading commissions — then the regulator forced them apart

For decades, the way a fund paid for a bank's research was indirect and opaque: it simply routed its trades to that bank, and the execution commission quietly bundled in the cost of the research, the analyst calls, the corporate-access meetings. Nobody priced research explicitly, and flow followed relationships and ideas as much as execution quality. MiFID II ended that. Since 2018, European rules require research to be paid for separately and explicitly — "unbundled" — so a fund can no longer settle its research bill by directing trades. Execution must be bought on its own merits, and research on its own.

The consequences ran deep. Execution became a more purely competitive, measurable product, judged on transaction-cost analysis rather than on the quality of a bank's research or its entertainment budget. Research itself came under pressure: with clients now paying explicitly and sparingly, research budgets shrank and coverage of smaller companies thinned. And the whole change reinforced the direction this module keeps returning to — execution as a standalone, commoditised, evidence-driven business, where flow follows the scoreboard rather than the relationship.

The market cycle: volumes and volatility rise and fall together

Trading revenue is cyclical because volume and volatility are cyclical

The desk makes money when clients trade, and clients trade most when markets move. Calm, trending markets produce thin volumes and thin revenue; volatile, fearful or euphoric markets produce heavy volumes and fatter revenue. The business is therefore cyclical, and the cycle is driven by volatility. This matters for more than the annual bonus. It shapes when a bank is willing to invest in technology: the case for a new low-latency stack or a better algo suite is easiest to make in a high-volume year and hardest to defend in a quiet one, even though the platform has to be there in both.

Figure 3 The equity-trading cycle: volumes and volatility rise and fall together, driving revenue and the appetite for technology investment (illustrative).

The recent cycle makes the point. The retail and volatility surge of 2021, when meme-stock episodes and a wave of new retail investors drove volumes to records, gave way to the rates shock and high volatility of 2022, and then to a lower-volatility grind in 2023–2024. Each phase changed how much the desk earned and how confidently it could fund the next platform upgrade. A business that must run the same expensive, resilient infrastructure through every phase, while revenue swings with the cycle, has a permanent tension at its heart: fixed cost meets variable income.

Scale is the defence against the cycle

The response to cyclicality, and to thin margins, is the same: scale. The largest players spread their fixed technology and connectivity costs across the most flow, so their cost per trade is lowest and their margin cushion is largest. That is why the equities business consolidates — why a handful of banks and a handful of specialist electronic firms dominate, and why the pressure on everyone else is relentless. Scale is not a luxury here; it is survival.

  1. What is the difference between the primary and the secondary market, and which one does equity trading operate in?

  2. What are the four things the desk actually sells?

  3. Why did equities electronify earlier and further than other asset classes?

  4. What is the difference between high-touch and low-touch, and why is low-touch margin so thin?

  5. Why is the business cyclical, and why does that matter for technology?

  6. How does the broker scoreboard distort behaviour?

  7. Why is negotiating the commission a poor way to reduce the cost of a large trade?

1

  1. In which market does equity trading operate?

  2. Why did equities electronify earlier than bonds?

  3. Why is low-touch margin so thin?

  4. Which is NOT one of the four things the desk sells?

2The Trade, End to End

This is the longest and most important part. It follows an order from the moment it arrives to the moment it settles, and it stops at every mechanic that matters: how the desk commits (or refuses) capital, how the fragmented market forms a price, how algorithms and routers do the work, how blocks are sourced in the dark, how fills are allocated, and how the trade finally clears and settles. The single hardest problem — market impact versus timing — sits at the centre, because every algorithm and every sales-trader is solving it.

The lifecycle in one view

An equity trade passes through seven stages, front to back

Every order, however it is worked, follows the same skeleton. The glamour is at the front; the operational risk and cost live at the back.

Stage What happens Typical timing Who leads
1. Order received Client sends name, side, size, limits, urgency via sales-trader, EMS or algo channel Seconds Client / sales-trader
2. Strategy & routing Choose the approach — algo, block, capital commitment — and how to slice and route Minutes Sales-trader / e-desk
3. Execution The order is worked in the market; fills accumulate across venues Minutes to days Algo + SOR / trader
4. Allocation Average-priced fills are split across the client's underlying accounts Trade day Client / middle office
5. Confirmation & affirmation Trade details matched between broker and client, ideally same-day T to T+0 Middle / back office
6. Clearing (CCP) Central counterparty novates and nets, manages default risk T to T+1 CCP / operations
7. Settlement (CSD, T+1) Shares and cash exchange at the depository; custody records the position T+1 CSD / custody

Effort and risk concentrate at execution, and again at the close

The pressure in an equity trade is not spread evenly. It peaks during execution — and within execution, it peaks at the closing auction, the single largest liquidity event of the day. The front two stages set the plan; the back three carry operational risk but are mostly automated. Execution is where the money is won or lost, where impact is caused, where information can leak, and where a human still watches the machine.

Figure 4 The trade lifecycle with an effort-and-risk curve: intensity peaks during execution, especially into the close.

Our running order enters the market

Meridian Asset Management, a long-only fund, wants to buy two million shares of Northwind — a mid-cap trading around twenty euros, with average daily volume of roughly seven hundred thousand shares. The order is therefore about three days of volume: far too large to do at once. Meridian's instruction to the desk is spare and typical — buy two million Northwind, minimise impact, do not reveal the size, and be able to show it was worked well. Everything that follows is this order, carried a step further. (Illustrative throughout.)

Step 1 — The order arrives, and the first choice is agency or principal

Who bears the price risk is the question that defines the trade

Before a single share moves, one decision shapes the economics: does the bank work for the client or trade with the client? This is the agency-versus-principal distinction, and it is the trading equivalent of who underwrites a deal.

Agency means the bank acts as broker. It finds the other side and executes on the client's behalf, earning a commission, with no firm capital at risk. Its interests are aligned with the client's: it is paid to get a good result, and the market risk stays with the client. Most modern equity flow is agency.

Principal, or risk, trading means the bank trades from its own book. It buys from or sells to the client directly, earning the spread and taking the price risk onto itself. Now the bank owns the position and its profit depends on where the market goes next.

Riskless principal is a middle form common in electronic markets: the bank buys and sells almost simultaneously, holding the position only for an instant, so it captures a spread without meaningfully warehousing risk.

Capital commitment is the sharpest version of principal trading. For a large block, the client may want the bank to take the whole trade onto its book at once, at a firm price — a risk price, sometimes called a bought deal. The client offloads all execution risk and gets certainty; the bank charges a premium and then has to unwind the position without giving the premium back.

The internal gauntlet before the bank commits its balance sheet

A client cannot simply demand a risk price and have it appear. Committing the bank's capital, or even deploying an algorithm into the live market, runs through an internal gauntlet designed to stop the desk taking risks it should not. Onboarding and KYC come first: the client must be known, checked and approved, with credit and settlement lines in place, before any order is accepted. Pre-trade risk limits sit in the order path: every order is checked in real time against position, notional and "fat-finger" limits before it can leave the building, and a kill-switch can pull the desk out of the market instantly. Capital-commitment approval governs risk prices: a block above a threshold needs sign-off because it consumes balance sheet and risk-weighted capital, and a sharp move against the book turns the risk premium into a loss. And algo governance governs the machines: under European rules, every algorithm must be tested, version-controlled and owned by a named individual, with change control and detailed records, before it touches a client order.

Figure 5 The pre-trade gauntlet: every order clears client, size, limit, algo-governance and kill-switch checks before it reaches a venue — in microseconds, in the order path.

Pricing a risk trade means pricing the cost of getting out

When a client asks for a risk price on a block, the number the desk quotes is not a guess. It is built from the cost the desk expects to incur unwinding the position, plus a margin for the risk that the unwind goes badly. Four inputs drive it. The size relative to average daily volume sets how many days the unwind will take. The expected market impact of that unwind is the central cost. The volatility of the stock sets how far the price could move against the desk before it is out. And the cost of balance sheet — the capital the position consumes while held — is a real charge, because a bank must hold regulatory capital against risk-weighted assets and satisfy a leverage ratio on everything it owns.

That last input is worth pausing on, because it is invisible to clients and decisive internally. A position on the book is not free even if the price never moves. It consumes capital that could support another business, and the desk is charged internally for it. This is why capital commitment is rationed and why approval thresholds exist, and it is why a desk with a cheaper balance sheet can quote tighter risk prices than a competitor with identical trading skill. In a business where execution ability has largely commoditised, the cost of capital is one of the few remaining structural advantages.

Step 2 — Strategy and routing: the fragmented market the order must cross

After MiFID, one stock trades in many places at once

To choose a strategy, the desk has to understand the terrain. A single European stock does not trade in one place — it trades across a dozen competing venues at the same time, by design. This fragmentation is a direct product of regulation. Before 2007, national rules often forced trading onto the home-country exchange; the first Markets in Financial Instruments Directive abolished that "concentration rule," and competition flooded in. The result is a landscape of venue types, each with a different transparency and a different purpose.

Regulated exchanges are the primary listing venues — Euronext, the London Stock Exchange, Deutsche Börse, Nasdaq. They run deep, lit order books and set the reference price and the all-important close. Multilateral trading facilities, such as Cboe Europe and Turquoise, are exchange-like lit venues that broke the exchange monopoly by competing on fees and speed. Dark pools display no pre-trade quotes at all, matching hidden orders — often at the midpoint of the lit price — so large trades can execute with less market impact. Systematic internalisers are banks and market-makers executing client flow against their own book, off-exchange, under a regulated quoting regime. Periodic auctions run frequent, very short auctions that batch orders into a single price — a lit but low-impact alternative that grew after regulators capped continuous dark trading. And over-the-counter or off-book trades are bilaterally negotiated, typically large blocks, reported to the market but not executed on a continuous book.

Figure 6 One stock, many venues. Lit venues are pre-trade transparent; dark venues display nothing before the trade. The smart order router sweeps them into a single execution.

Europe and the United States solved fragmentation in opposite ways

Both markets fragmented, but their regulators chose different remedies, and anyone working across the two must hold the difference clearly. The United States imposed a rule; Europe imposed a duty.

Under the American Regulation NMS, the venues are linked into a single virtual book. A consolidated feed publishes the National Best Bid and Offer — the best price available anywhere — and the order protection rule forbids executing a trade at a worse price than the NBBO while that better price is displayed elsewhere. Trading through a better price is prohibited, so the market is stitched together by regulation into something that behaves like one venue.

Europe took the other path. MiFID created competition without a consolidated tape and without a trade-through prohibition. Instead it imposed best execution as a principle: a firm must take all sufficient steps to obtain the best possible result for its client, weighing price, cost, speed, likelihood of execution and settlement — and must be able to evidence that it did. There is no single official best price to point at, which is precisely why Europe has debated a consolidated tape for years, and precisely why transaction-cost analysis matters more in Europe than a compliance checklist would.

United States (Reg NMS) Europe (MiFID II)
Core mechanism Rule: order protection against the NBBO Duty: best execution, evidenced
Consolidated price Yes — the NBBO, published continuously Not yet — a consolidated tape is still being built
Trade-through Prohibited Not prohibited; judged against best-execution duty
Retail model Wholesaler internalisation, payment for order flow Systematic internalisers; PFOF restricted
Practical effect Compliance is largely mechanical Compliance is analytical and evidence-driven

The order book is where a price is actually made

Beneath the venue names is a single mechanism that forms the price: the central limit order book, or CLOB. An exchange matches orders through a live list of every bid (buy order) and every offer (sell order) at each price level. Two rules govern it. Price-time priority means the best price is matched first, and at the same price, whoever arrived earliest wins — fair and deterministic. The spread is the gap between the best bid and the best offer: the cost of an instant round-trip, and a direct measure of liquidity. Depth is how much size rests at each price level — how far the book can absorb a large order before the price moves.

Figure 7 A central limit order book: bids below, offers above, the spread between the best of each, and the depth of size resting at every level.

The order types are the vocabulary of the book, and every algorithm ultimately speaks it. A market order executes now at the best available price — certain to fill, but you accept whatever the book gives, which for size means walking up through the levels. A limit order executes only at a stated price or better — it controls the price but may never fill, and it is the building block of the book. An iceberg shows only a small "tip" on the book and refreshes as it fills, hiding true size to limit impact. A stop lies dormant until the price hits a trigger, then becomes a market or limit order. And pegged and conditional orders track the midpoint or rest quietly in the dark, firing only when a match appears. Everything an algo does is a clever sequence of placing, cancelling and re-placing these orders.

A useful image for market impact. Think of buying every seat in a concert hall. The first few tickets cost what the sign says. But as you keep buying, the cheap seats are gone and you work your way up the price list — and worse, the box office notices someone is buying everything and stops discounting. The price you eventually pay per seat is nothing like the price on the sign, and the act of buying is what changed it. That is walking the book and market impact, and it is why nobody sensible buys two million shares in one click.

The auctions, especially the close, are the day's real liquidity

Continuous trading gets the attention, but the auctions batch a great deal of the volume. The opening auction sets the day's first price; the closing auction sets the official close and is the single most important liquidity event of the day. Index funds must trade at the close to match their benchmark, so enormous size concentrates there. This is why a whole family of algorithms exists just to target the close, and why the intensity curve peaks there.

Liquidity is not spread evenly through the day, and the shape it does take governs every scheduling decision an algorithm makes. Volume follows a U: heavy at the open, thinning through the morning to a lunchtime trough, building again through the afternoon, and spiking enormously into the closing auction.

Figure 8 The intraday volume profile. Trading against this shape is what a VWAP algorithm does; the closing spike is why a whole family of algorithms targets the auction. (Illustrative.)

This curve explains several things at once. It explains what a VWAP algorithm is actually doing — tracing this shape with its own child orders. It explains why the lunchtime trough is expensive: the same order size is a far larger share of volume at 12:30 than at 16:30, so it causes more impact. And it explains the gravitational pull of the close: if a fifth of the day's volume prints in one auction, that is where size can be done with least impact, which is why closing-auction share has grown year after year.

Step 3 — Execution: solving impact versus timing

The central problem: the act of trading changes the price you get

Here is the intellectual heart of the whole business. A stock's book only holds so much size near the current price. Try to buy more than that and you "walk the book" — each successive fill is at a worse price, because your own order is consuming the available liquidity. This is market impact: the price move your order causes, working against you, and it grows the bigger and faster you trade.

Impact has two flavours. Temporary impact bounces back after you stop trading — you pushed the price up while buying, and it eases once you are done. Permanent impact does not come back, because the market inferred information from your trade: a persistent buyer looks like someone who knows something. Alongside impact sits information leakage — if others detect a large buyer, they trade ahead of you, worsening your price before you finish. And underneath it all is the brute constraint of days to trade: an order that is several days of average volume simply cannot be done at once at a reasonable cost.

The trade-off is impact against timing, and there is an efficient frontier

The response to impact is to trade slowly. But slowness has its own cost. Trade fast and you get certainty but high impact; trade slow and you cut impact but expose yourself to the price drifting away from you while you wait — timing risk. Good execution is not "trade slowly" or "trade quickly"; it is finding the point that minimises the total of impact plus timing cost. Economists draw this as an efficient frontier: a curve on which every point trades one cost for the other, with a single optimum that minimises the sum.

Figure 9 The execution trade-off: trading faster raises market impact, trading slower raises timing risk. The optimum minimises their sum — the logic behind the Implementation Shortfall algorithm.

Two more choices ride on top of this frontier. Lit versus dark: lit venues show your hand and risk leakage; dark pools and blocks hide it, cutting impact at the cost of less certain fills. And liquidity is not constant: it is deepest at the open and close and in large-cap names, and thin in small caps and mid-session, so the strategy must adapt to the stock and the hour.

A worked example: measuring the cost of our order

Numbers make the trade-off concrete. Take Meridian's two-million-share Northwind buy. (Every figure here is illustrative.)

The algorithms: how the desk actually works the order

Each algo solves impact-versus-timing differently, against a different benchmark

An execution algorithm is software that slices a parent order into child orders and works them to a chosen strategy and benchmark. The suite is small and worth knowing by name, because the whole low-touch business is built on it.

Figure 10 The core execution algorithms, positioned by urgency and by how far they follow a fixed schedule versus opportunistically hunting liquidity.
VWAP — trade in line with the day's volume

A volume-weighted average price algo trades in proportion to the day's expected volume profile, aiming to match the VWAP benchmark. It is passive, predictable and benchmark-driven — the natural choice when the client is measured against VWAP.

TWAP — spread it evenly over time

A time-weighted average price algo spreads the order evenly across a set window, ignoring volume. It is simple and steady, and it suits illiquid names where the volume profile is unreliable.

POV — track a percentage of volume

A percentage-of-volume (or participation) algo trades a fixed share of live market volume — say 10% — speeding up when the market is active and slowing when it is quiet. It adapts to real liquidity but reveals a steady footprint that others can detect.

Implementation Shortfall — minimise slippage to the arrival price

The Implementation Shortfall algo minimises total slippage against the arrival price. It front-loads trading to cut timing risk, then eases off, balancing impact against drift — the "get it done" algo, and the one that directly implements the efficient frontier. This is what the desk chose for Meridian.

Liquidity-seeking and close algos — hunt blocks, or target the auction

A liquidity-seeking algo hunts hidden and displayed liquidity across lit and dark venues, firing when size appears rather than following a schedule — it prioritises finding a block. Close and auction algos target the closing auction, the day's biggest liquidity event and the benchmark for index funds, and they take a large and growing share of volume.

Smart order routing: the algo decides how much and when, the router decides where

Beneath every algo, a router chooses a venue for every slice

An algorithm decides how much to trade and when. It does not, by itself, decide where. That is the job of the smart order router, or SOR — the layer that takes each child order the algo generates and decides, in real time, which venue or venues to send it to for the best fill. It continuously weighs the price and displayed size at each lit venue, the likely hidden liquidity in dark pools and internalisers, the fees and rebates each venue charges or pays, the probability and latency of a fill at each destination, and anti-gaming logic to avoid leaking its intentions or being picked off. It sprays, pegs, cancels and re-routes in microseconds, reassembling fragmented liquidity into one coherent execution.

Routing quality is where real edge lives, and it is measurable: fill rates, price improvement, low leakage. It is also where conflicts hide. Routing to the bank's own dark pool first, or chasing a venue's rebate, must not come at the client's expense — that is a best-execution and conduct issue, and every routing decision is logged and analysed to prove it did not.

Sourcing size in the dark: blocks, IOIs and conditional orders

Trading an elephant without spooking the market is a discipline of its own

For a large order, the best outcome is often to find a single natural counterparty and cross the whole block at once, away from the lit market. This is the liquidity-sourcing craft, and it runs on discretion. Not all liquidity is visible. Lit liquidity is displayed on the public book, with full transparency — but showing it reveals your intention. Dark liquidity carries no pre-trade display and matches, often at the lit midpoint, so large trades avoid moving the market. Dark exists for one honest reason: a big institution wants to trade size without signalling it and being front-run. The trade-off is less impact and leakage, but uncertain fills and the risk of interacting with predatory flow — which is why dark venues run anti-gaming controls.

There are several ways to trade size discreetly. A block trade is a large parcel negotiated in one go, sourced through the sales-trader's network of indications of interest (IOIs) or on a dedicated block venue such as Cboe LIS or Liquidnet. Conditional orders rest indications in several dark venues at once and, when a match firms up, commit the order to just one — an efficient way to hunt blocks without over-committing. Periodic auctions batch orders into frequent lit micro-auctions at a single price, low-impact and fast-growing. And large-in-scale orders, above a regulatory size threshold, receive a transparency waiver — the sanctioned way to trade a genuine block in the dark.

For Meridian, the desk seeks a block first, then works the residual

Who is on the other side: participants and adverse selection

The desk must read the counterparty's likely motive

A stock's price is set by very different players with very different horizons, and good execution depends on reading who is on the other side of your trade. Real-money, long-only investors — pension funds, insurers, mutual funds, index trackers — are large, slow and benchmark-driven, the core institutional client of the desk. Hedge funds are active, opportunistic and higher-turnover — demanding on execution and often the most profitable flow. Market-makers quote continuous two-way prices and warehouse risk, supplying liquidity. High-frequency and proprietary firms are ultra-fast electronic traders — market-making, arbitrage and latency strategies — now a huge share of lit volume. Retail investors trade small tickets but large aggregate flow. And corporates and others — issuers doing buybacks, ETF authorised participants — feed the flow.

Why does the mix matter? Because of adverse selection: you want to trade against uninformed flow and avoid being picked off by someone who knows more than you. A slow long-only rebalance is very different from a fast, informed hedge-fund order. Knowing the likely motive of the counterparty shapes how you route, how much you show, and whether you trade lit or dark. This single idea — trade with the uninformed, avoid the informed — drives the whole design of lit-versus-dark and block trading.

High-frequency trading and electronic market-making

The fast firms that now provide much of the visible liquidity

The other side of a lit trade is increasingly a machine. High-frequency traders are ultra-fast, technology-driven firms trading in microseconds. Their strategies are electronic market-making (quoting two-way prices continuously, earning the spread across millions of tiny trades), arbitrage (capturing fleeting price differences between venues, or between a stock and its future or ETF), and latency strategies (reacting to order-book signals faster than anyone else to win queue priority or rebates). Firms like Citadel Securities, Jane Street, Optiver and Virtu are now central liquidity providers — not banks.

The debate about them is genuine and unresolved. The upside is real: tighter spreads, more continuous liquidity, and lower explicit costs for ordinary trades. The concern is equally real: speed advantages, fleeting quotes that vanish when hit, and the risk of destabilising events — the 2010 Flash Crash being the emblem. And a specific worry: HFT liquidity can evaporate in stress, because a market-maker with no obligation to quote may simply pull back exactly when the market needs it. The bank's response is to run its own electronic market-making with anti-gaming logic, while regulators impose registration, testing, kill-switches and circuit breakers.

ETFs and electronic market-making: the always-on side of the desk

An ETF trades like a share but is built from a basket, and that link is the business

Exchange-traded funds deserve their own treatment, because they are now a huge share of equity volume and a business the desk runs continuously. An ETF is a fund that trades like a single share, but it is a wrapper around a basket of underlying stocks, and the whole market-making opportunity lives in the link between the two. Because an ETF can be created and redeemed, its price is tethered to the value of its underlying basket by arbitrage. When the ETF trades richer than its basket, an authorised participant buys the underlying shares, delivers them to the fund and receives new ETF units to sell — creation. When the ETF trades cheaper than its basket, the reverse happens — redemption. This create-redeem mechanism is the arbitrage that keeps an ETF's price honest, and running it is a core electronic market-making activity.

The desk's role is to quote continuous two-way prices in ETFs and single stocks, earning the spread across millions of tiny trades while managing the inventory it accumulates. It is a low-margin, high-volume, technology-intensive business — the bank competing head-on with specialist firms like Jane Street and Optiver that built themselves around exactly this. The primary-versus-secondary distinction reappears here in miniature: most investors trade ETF units in the secondary market, but the creation and redemption of units is a primary process happening quietly in the background, and the desk sits on both sides of it.

Short selling and the borrow: how you sell what you do not own

Selling first and buying back later requires borrowing the shares

Not every sell order is a holder reducing a position. A short sale is selling shares you do not own, in the expectation of buying them back cheaper later — and mechanically, to deliver shares you do not have, you must first borrow them. This borrow comes from the securities-lending market: a holder (often a pension fund or index tracker) lends its shares for a fee, and the short-seller posts collateral and returns the shares later. The desk's job is to arrange the locate — confirming borrowable shares exist before the short is executed — and the financing around it.

Short selling is legitimate and useful: it improves price discovery and provides liquidity. But it is tightly regulated because it can amplify falls and, in the extreme, enable abuse. Europe's Short Selling Regulation requires disclosure of significant short positions, bans "naked" shorting where no borrow has been located, and lets regulators impose temporary bans in stressed conditions. For the desk, this means the systems must check the locate, monitor aggregate short positions against disclosure thresholds, and be able to enforce a ban instantly when one is imposed.

Program and portfolio trading: where execution meets portfolio risk

Trading a whole basket at once is a different animal

A program, portfolio or basket trade is the simultaneous execution of a whole list of stocks — often tens or hundreds of names — as a single package. It is used for index rebalances (trackers adjusting to index changes), portfolio transitions (moving between managers or mandates), cash flows (investing inflows or funding redemptions across a portfolio), and hedging or ETF creation. The choice, as ever, is agency or risk. An agency program executes the basket for a commission via algos, with the client keeping the market risk. A principal or risk program guarantees a price on the whole basket up front — often against the close — taking it onto the bank's book so the client offloads all risk.

Baskets behave differently from single stocks in one crucial way: diversification and natural offsets. A basket that is partly buys and partly sells has a net market exposure far smaller than the sum of its individual names, so its net risk — and the price the bank charges to warehouse it — is much lower than trading each name alone. The bank prices the basket's risk from its correlation, liquidity and net exposure, hedges the residual with an index future, and works out of the single names over time. Index rebalance days, when many funds must trade the same names at the same moment, are among the highest-volume and most contested events of the calendar.

Step 4 — Allocation: splitting average-priced fills across accounts

One order, many owners, one fair price

A fund manager rarely trades for a single fund. Our Meridian order might in truth be for a dozen underlying sub-accounts, and the fills arrived over three days at many different prices. Allocation is the mechanism that splits the executed shares fairly across those accounts, almost always at the average price. Every account that participated gets shares at the same volume-weighted average fill, in proportion to its share of the order, so no account is advantaged by the accident of which fill came first. This is the trading analog of deciding who gets what in an oversubscribed deal: the competing interests are the client's own sub-accounts, and the fairness rule is the average price. It is also where a subtle operational risk lives — an allocation error assigns the wrong shares or the wrong price to the wrong account, and it must be caught before settlement.

Retail flow and sub-distribution: how the smallest orders reach the market

Retail is small per ticket, vast in aggregate, and routed differently

Not all flow comes from institutions. Retail orders — individual investors, mostly via brokers — are tiny per ticket but enormous in aggregate, and they reach the market by a different path. In Europe, most retail flow is routed to venues or internalised by market-makers under the systematic-internaliser regime. In the United States, a distinctive model prevails: retail brokers often sell their order flow to wholesalers who internalise it, a practice known as payment for order flow, which is contentious and regulated differently on each side of the Atlantic. The mechanics differ by jurisdiction, but the principle is constant: retail flow is aggregated, routed or internalised, and it is prized precisely because it is uninformed — the opposite of the hedge-fund order the desk approaches warily.

Steps 5 to 7 — Post-trade: confirmation, clearing and settlement

A great execution that fails to settle is a loss

Execution is the front; the back is where a trade becomes a legally owned, safely held share. Confirmation and affirmation come first: the trade details are matched between broker and client, ideally same-day and automated. Then clearing, where the central counterparty steps between buyer and seller — becoming counterparty to both through a process called novation, netting exposures so only the difference settles, and managing the risk that one side defaults. Then settlement at the central securities depository, where shares and cash actually change hands through delivery-versus-payment — the guarantee that neither leg moves without the other. Euroclear and Clearstream are the European giants of this layer. Finally the position lands in custody, where ownership is recorded and from which corporate actions — dividends, splits, rights issues — flow.

Novation and netting are what make a modern market possible at all

The clearing step deserves more than a definition, because it solves two problems that would otherwise make high-volume trading unworkable. Novation is the legal substitution: the original contract between buyer and seller is replaced by two contracts, each with the CCP in the middle. Neither side needs to assess the other's creditworthiness, because both now face the same well-capitalised counterparty, which collects margin from everyone and maintains a default fund. Netting is the arithmetic that follows: once the CCP faces everyone, it can compress a day's many obligations in the same security into a single net position per participant.

Figure 11 Before clearing, every pair of counterparties has a separate obligation. After novation and netting, each participant faces only the CCP, with one net position to settle.
Figure 12 Two messaging worlds carry the trade: FIX in the front office (order and execution, milliseconds), SWIFT and ISO 20022 in post-trade (confirmation and settlement, seconds to minutes).

T+1 turns post-trade from an afterthought into a live front-office concern

The settlement cycle is compressing. Europe, together with the United Kingdom and Switzerland, targets a move to T+1 settlement in October 2027, following the United States, Canada and others who moved in 2024. (This date is a real, announced target.) Halving the time from two days to one leaves no room for overnight manual fixes: same-day affirmation, straight-through processing and exception management become critical, and cross-border investors must source cash and foreign exchange faster or risk a fail. Under the European Central Securities Depositories Regulation (CSDR), failing to settle on time already incurs cash penalties, so post-trade discipline is now a hard cost driver, not a back-office nicety. A great fill that fails to settle is a loss and a reputational hit — which is why the plumbing has become a front-office concern.

The lifecycle of a held position: corporate actions do not stop at settlement

Once shares are owned, the events that happen to them must be processed too

Settlement is not the end of a share's story — it is the start of the position's life. As long as the shares are held, the company keeps doing things to them, and each event must be captured, valued and applied to the holder's record. These are corporate actions, and they fall into two kinds. Mandatory actions happen to every holder automatically: a dividend is paid, a stock splits two-for-one, shares are consolidated, one company is absorbed by another. Voluntary actions require the holder to choose: a rights issue offers the right to buy more shares, a tender offer invites you to sell, a dividend may be taken in cash or in new stock. For voluntary actions the custodian must collect and act on an election before a deadline, or the holder loses the choice by default.

Why does a trading module care about post-settlement plumbing? Because corporate actions are a notorious source of operational risk and data error. A mishandled dividend, a split applied late, a missed election deadline — each corrupts the position record and the valuations that depend on it, and each is the cascade principle at work after the trade is long done. The systems that process corporate actions must ingesting a messy, non-standard stream of announcements from issuers and depositories, interpret them correctly, and apply them to thousands of positions without error. It is unglamorous, and it is exactly where quiet losses accumulate.

Post-execution: the feedback loop and the position that must be unwound

The trade is not over when the fills stop

Two things happen after the last fill. First, if the desk committed capital and took a block onto its book, it must unwind that position — trading out of the single names via algos while watching impact, and lifting any index-future hedge as the cash clears. This is the risk-trading equivalent of aftermarket support: the bank earned a premium for certainty and now has to exit without giving it back. Second, and for every order, comes measurement: transaction-cost analysis compares the average fill to the arrival price and to VWAP, quantifies the impact and the drift, and feeds a loop that makes the next order better. That loop — execute, measure, diagnose, adjust — is the subject of Part 2's close and the engine of the whole modern business.

  1. Why does market impact grow with the size and speed of an order?

  2. What is the difference between measuring a trade against arrival price and against VWAP, and why can they disagree?

  3. What does a smart order router decide that an algorithm does not?

  4. What is adverse selection, why does dark trading exist, and what is the regulator's concern about it?

  5. What internal controls must clear before the desk commits capital or deploys an algorithm, and where do they sit?

  6. If an order is 2.8× daily volume and the algorithm participates at 25% of volume, roughly how long does it take, and what does that imply?

  7. What do novation and netting achieve at the CCP, and why has T+1 made post-trade a front-office concern?

2

  1. Why does market impact grow with order size and speed?

  2. What does a smart order router decide that an algorithm does not?

  3. Arrival price versus VWAP: which captures total cost including drift?

  4. Why does dark trading exist?

3Actors, Roles and Interactions

An equity trade is a conversation between many specialised parties, and the bank sits in the middle of almost every exchange. This part maps the ecosystem, then goes deep on the two things that decide how a desk behaves: who its clients really are — the buy-side taxonomy that drives routing and allocation — and the information barriers that keep one institution from trading on what another part of it knows.

Four camps, and the bank in the middle of every interaction

The ecosystem has a client, an executor, a marketplace and a plumbing layer

Strip away the detail and equity trading has four camps. The buy-side generates the order: asset managers, pension funds, hedge funds, running their own dealing desks, execution systems and analytics. The sell-side broker or bank executes it — via sales-traders, algos and direct market access — and may commit capital and provide research and analytics. The marketplace is where orders actually match: exchanges, MTFs, dark pools, systematic internalisers and periodic auctions, populated by liquidity providers — bank and non-bank market-makers and high-frequency firms that quote continuously. And the plumbing finalises ownership: central counterparties for clearing, central securities depositories for settlement, and custodians for safekeeping, all wired together by data, connectivity and the FIX protocol, and watched by regulators.

Figure 13 The bank sits at the centre: the buy-side sends orders, venues and liquidity providers supply the market, post-trade infrastructure finalises ownership, and data and regulators surround it all.

The shape of it is simple to state and important to hold: the desk is a party to nearly every interaction in the chain. The buy-side generates the order; the broker executes it across competing venues where liquidity providers stand; the trade flows through CCPs, CSDs and custodians; and regulators oversee the whole thing. No single actor sees the whole picture, which is exactly why data, connectivity and post-trade recur as themes — the system is only as good as the seams between its parts.

Every major flow has a direction and a stage

From To What flows Stage
Buy-side Broker desk Order (name, side, size, limits, urgency) Order received
Broker Venues Routed child orders Execution
Venues Broker Fills (price, time, venue) Execution
Liquidity providers Venues / broker Quotes, block indications Execution
Broker Buy-side Fills, average price, TCA report Execution / post-trade
Broker CCP Cleared trades for novation and netting Clearing
CCP CSD Net settlement instructions Settlement
CSD / custodian Buy-side Recorded holding, corporate actions Settlement / lifecycle
Broker Regulator Transaction and best-execution reporting Post-trade

The buy-side taxonomy: who the clients really are

Reading the client's type is the first act of good execution

The desk's clients are not interchangeable. They differ in size, in how long they hold, in how much they care about price versus certainty, and in how informed their flow is — and every one of those differences changes how the desk should route, show and work the order. This is not a soft classification; it is the operational heart of adverse selection, which we met in Part 2. Trade with the uninformed, be wary of the informed.

Long-only, real-money funds — the core client

Long-only funds — mutual funds, insurers, index trackers — are large, slow and benchmark-driven. They rebalance rather than speculate, hold for months or years, and are relatively price-insensitive on any single trade because their horizon is long. The desk prizes them because their flow is mostly uninformed in the short term: a rebalance is not a bet on tomorrow's price. They are the bread-and-butter client, and their orders are the ones a desk is happiest to take the other side of.

Pension and sovereign wealth funds — the patient giants

Pension and sovereign wealth funds are the most patient capital of all. Their horizons run to decades, their tickets are enormous, and they are the natural counterparty for a block: a fund quietly building or trimming a long-term position is precisely the "natural liquidity" a sales-trader hunts for. Their patience makes them low-impact partners, and their size makes them the anchor of many block trades.

Hedge funds — active, opportunistic, and often informed

Hedge funds — long/short, event-driven, quantitative — are active, opportunistic and high-turnover. They are demanding on execution and often the most profitable flow a desk handles, but they carry the sharpest adverse-selection risk: a hedge-fund order may reflect a genuine informational edge, so trading against it can mean being picked off. The desk welcomes the commissions and watches the flow carefully.

Retail — small tickets, vast aggregate, politically charged

Retail investors trade small per ticket but represent large aggregate flow, and they are prized precisely because they are uninformed. Their orders are routed, internalised or, in some markets, sold to wholesalers. Retail also carries a political dimension — payment for order flow, access and fairness are recurring public controversies — that makes it a sensitive as well as a valuable segment.

Investor type Holding horizon Price sensitivity Why the desk wants or fears them
Long-only / real-money Months to years Low per trade Wants: large, uninformed, benchmark-driven core flow
Pension / sovereign wealth Years to decades Very low Wants: the natural block counterparty, patient and low-impact
Hedge fund Days to months High Mixed: profitable and active, but often informed — adverse selection
Retail Varies Varies Wants: uninformed aggregate flow, but politically sensitive
Figure 14 The buy-side mapped on the two axes that matter: how long they hold, and how much information their flow carries. The desk's economics improve as you move down and to the right.

Two axes explain almost everything about how the desk treats a client. The horizontal axis is holding horizon, from the microseconds of a quant market-maker to the decades of a sovereign wealth fund. The vertical axis is how much information the flow carries — whether trading against it is likely to leave you on the wrong side. The bottom-right quadrant, patient and uninformed, is where the desk most wants to be: this is natural block liquidity, and it is why a pension fund quietly trimming a position is the counterparty a sales-trader dreams of finding for our Meridian order. The top-left quadrant, fast and informed, is where the desk must be most careful, because standing opposite it repeatedly is a slow way to lose money.

The diagonal matters too. Corporate buybacks sit awkwardly high: an issuer buying its own shares is patient, but it plainly knows things about itself, which is precisely why buyback execution is hedged about with regulatory safe-harbour conditions. And retail sits low and slightly left — uninformed, which is why its flow is so commercially prized and so politically contested.

The client's own trading desk is who you actually talk to

A fund manager rarely calls the broker — the dealing desk does

One structural detail changes how the whole relationship works, and newcomers routinely miss it. At most institutional asset managers, the person who decides *what* to buy and the person who decides *how* to buy it are different people in different teams. The portfolio manager forms the investment view and generates the order. It then passes to the firm's own dealing desk — a team of buy-side traders whose entire job is execution — and it is they who choose brokers, select algorithms, monitor progress and challenge the results. The sell-side sales-trader's counterpart is therefore another trading professional, not a stock-picker, and the conversation is technical rather than thematic.

This matters because it explains what the buy-side desk is optimising for. They are measured on execution quality against benchmarks, not on whether the stock went up. They run their own execution management systems, their own transaction-cost analysis, and increasingly their own algo wheels. When they push back on a fill, they are doing so with data. The relationship is peer-to-peer between two execution specialists, and a sales-trader who talks to them as if they were the portfolio manager loses credibility immediately.

Not every institution has the scale for this. Smaller managers increasingly use outsourced trading — a third-party firm that runs the dealing function on their behalf, providing access to brokers, venues and analytics without the fixed cost of an in-house desk. It is a fast-growing segment and it adds an intermediary layer between the asset owner and the executing broker, with its own conflicts to manage.

Non-bank firms are now competitors, not just counterparties

The most formidable players in electronic equities are not banks

Part 2 introduced high-frequency and electronic market-making firms as liquidity providers. Seen from the desk's commercial position, they are something sharper: direct competitors for the most profitable parts of the business. Firms such as Citadel Securities, Jane Street, Optiver and Virtu compete with banks in market-making, in ETF pricing, and — in the United States especially — in the retail wholesaling business. They carry no legacy platform, no branch network and a far lighter regulatory perimeter than a systemically important bank, and they invest in technology with a single-mindedness a diversified institution struggles to match.

The competitive consequence is that a bank's equities franchise cannot win on speed alone and rarely tries. It competes where it has structural advantages a pure-play electronic firm does not: a balance sheet that can commit capital to a block, a full-service relationship spanning research, financing and prime brokerage, global reach across products and regions, and the client trust that comes with being a regulated custodian of institutional relationships. Understanding this is essential to understanding why banks emphasise high-touch capability and capital commitment even as low-touch commoditises — those are precisely the areas the non-bank competitor cannot easily enter.

How the flow is actually won

Execution is bought through a formal, data-driven review, not a relationship

If execution quality is now measurable, the commercial process for allocating it has become formal to match. Most institutional clients run a periodic broker review — often called a broker vote — in which they score each counterparty and decide how to split next period's flow. Historically the vote weighted research, corporate access and relationship heavily. Post-unbundling, execution is voted on separately and largely on evidence: transaction-cost analysis, fill rates, venue quality, reliability of the algo suite, responsiveness when something breaks.

The output is a tiering: a small group of tier-one brokers receiving the bulk of the flow, a longer tail receiving occasional or specialist business. Getting into tier one is worth a great deal and losing that place is expensive, which is why desks monitor their own performance against client benchmarks obsessively — they are being graded continuously, whether or not anyone tells them.

Three mechanisms carry the money and the flow. The algo wheel we met in Part 2 automates part of the allocation entirely: eligible orders are distributed among brokers by rule, performance is measured, and weights adjust toward whoever executes best — a genuine meritocracy with no phone call involved. Commission sharing agreements let a client concentrate execution with the brokers who execute best while still directing research payments elsewhere, decoupling the two decisions that unbundling separated. And formal requests for proposal govern larger mandates such as portfolio transitions, where the client runs a competitive process on cost, capability and risk.

Mechanism What it allocates What it rewards
Broker vote / review The client's overall flow, by tier Measured execution quality and reliability
Algo wheel Individual orders, automatically Statistical performance against benchmark
Commission sharing agreement Payment, separately from execution Execution merit, decoupled from research
Request for proposal Large one-off mandates (transitions) Cost, capability, risk management

Information barriers: one institution that must not trade on what it knows

The problem: a bank knows things it must keep from itself

A large bank is many businesses under one roof, and some of them know things the others must never use. The research analyst forming a view, the corporate-broking team advising a listed company, the desk executing client orders, the prop or market-making book trading the firm's own capital — these functions cannot freely share what they know without creating abuse. If the execution desk knew what an advisory team knew about a company, or traded ahead of a client's large order, that is market abuse. The core problem of a trading floor is that a single institution possesses material information in one corner that it must be structurally prevented from using in another.

The classic answer is the information barrier, historically called a Chinese wall: a set of controls that separate functions so information does not leak between them. Four functions in particular are kept apart — research, corporate advisory, client execution, and the firm's own risk-taking — each walled from the others. When information genuinely must cross a wall, it happens through a formal, logged, controlled exception called wall-crossing: a named person is "brought over the wall" for a specific purpose, recorded, and restricted from trading the relevant name while they hold the information.

Why this is an engineering problem, not just a policy

A wall made of good intentions leaks. In a firm where every order, message and quote is electronic, the barrier has to be built into the systems themselves. Restricted and watch lists encode which names a desk may not trade or research at a given moment, enforced automatically in the order path. Access controls segregate systems and data so that people on one side of a wall cannot even see the other side's information. Wall-crossing logs record every exception with a timestamp and a reason. And trade surveillance watches the firm's own activity for the patterns that signal abuse — trading ahead of a client, coordinating across a wall, or manipulating a price.

Six roles inside the bank: the org chart of a software business

The desk has shifted from mostly traders to a blend of traders, quants, engineers and controllers

A modern equities desk looks less like a trading pit and more like a technology company that happens to trade. Six roles carry it.

Role Core job Primary technology
Sales-trader Client-facing: takes orders, advises strategy, works high-touch blocks, commits capital, protects information Sales-trader tools, IOI network, block venues
Algo / e-trader Runs the low-touch desk: monitors algos and flow, tunes parameters, handles exceptions and outages EMS, algo consoles, monitoring
Quant / algo developer Designs and calibrates the execution algos, SOR logic and TCA models — the R&D that creates the edge Modelling, backtesting, data platforms
Electronic trading COO Runs the e-trading business: platform strategy, venue connectivity, budgets, controls, change delivery Programme and vendor management
Market-maker / risk trader Prices and warehouses principal risk — capital commitment, program risk, ETF and single-stock market-making Risk book, pricing engines
Compliance / surveillance & IT Surveils trading for abuse, enforces best execution and controls; runs the low-latency stack Surveillance, reporting, infrastructure

The sales-trader is the human premium: they take the order and its context — urgency, price limit, discretion — advise on which algo and how aggressive, quietly source block liquidity through IOIs and relationships, commit capital when asked, and above all protect the client's intention from leaking. Their edge is judgement, discretion and a network. The algo or e-trader runs the machine floor: part trader, part operator, monitoring algos and flow, tuning parameters, and handling the exceptions and outages that the automation cannot. The quant or algo developer builds the edge itself — the execution algorithms, the routing logic, the TCA models — and is the reason one broker's execution is measurably better than another's. The electronic trading COO runs the e-trading business as a business: platform strategy, venue connectivity, controls and change. The market-maker or risk trader prices and warehouses the firm's own risk — this is the principal side, the balance sheet at work. And the compliance, surveillance and IT function enforces the rules and runs the infrastructure, carrying hard regulatory obligations that we return to in Part 4.

The information barrier runs straight through this table

Notice where the wall falls. The sales-trader and the risk trader must be separated: the person working a client's block must not tip off the person running the firm's own book. The market-making function and the client-execution function are walled. And surveillance watches all of them. The org chart is not just a division of labour; it is a division of information, and the seams between roles are exactly where the controls of the previous section bite.

  1. Why does the buy-side taxonomy matter operationally, not just descriptively?

  2. When a sales-trader hunts a natural counterparty for a large buy order, which investor types are they most likely looking for, and why?

  3. What is the core problem that information barriers exist to solve?

  4. Why is the information barrier an engineering problem rather than only a policy?

  5. How does the information barrier show up in the desk's own org chart?

  6. Which role creates the desk's measurable execution edge, and how?

  7. How is a client's flow actually allocated between brokers, and why does the mechanism shape what the desk builds?

3

  1. A sales-trader hunts a natural seller for a large buy. Whom do they seek?

  2. Why does the buy-side taxonomy matter operationally?

  3. Why is the information barrier an engineering problem?

  4. What core problem does an information barrier solve?

4Systems, Data and the IT Stack

Equities is a software business, so its architecture is not a diagram on a wall — it is the product. This part builds the stack from first principles: what functions any trade needs, who supplies them, how one order travels front to back, the anatomy of the order object itself, the identifiers and messages that carry it, the integration patterns that connect the pieces, the resilience that must hold on the one irreversible moment, and the regulations that are really system specifications in legal language.

Functional architecture: what any trade needs, before naming a single vendor

Ask what functions a trade requires, then map systems to them

The cleanest way to understand a trading platform is to ignore vendor names at first and ask a simpler question: what functions does any equity order need on its way from a click to custody? The answer is a short chain of functional blocks, each doing one job.

Functional block What it does
Order management (OMS) The book of record: captures orders, manages positions, allocations and compliance checks
Execution management (EMS) The trading cockpit: real-time market data, algo selection, order routing
Algos & smart order router The execution engine: slices orders to a strategy, routes each slice to the best venue
Pre-trade risk & controls Real-time limit, fat-finger and kill-switch checks that clear every order before it leaves
Market data & connectivity (FIX) Real-time feeds in; orders and fills out over FIX to every venue and counterparty
TCA & analytics Captures every fill and timestamp; measures execution quality; feeds best-execution reporting
Post-trade operations Confirmation, allocation, clearing and settlement into custody and the books
Figure 15 The functional architecture: a chain of execution blocks, with cross-cutting capabilities — surveillance, best-execution monitoring, reference data and resilience — running underneath them all.

Two distinctions are worth fixing. The OMS versus EMS split is the one people confuse most: the OMS is the record-keeper — it holds the order, the positions, the allocations and the compliance state — while the EMS is the real-time trading tool that actually works the order in the market. The two are increasingly merged into a single "OEMS" that removes the reconciliation between them. And underneath every block runs a band of cross-cutting capabilities — surveillance and market-abuse controls, best-execution monitoring, reference data, and connectivity and operational resilience — that no single block owns but every block depends on.

The vendor landscape: where banks buy, and where they build

The buy-versus-build line is a map of competitive advantage

Most of the stack is bought. Banks license order and execution management systems, market data, and post-trade utilities from specialist vendors, because these are common plumbing where scale and standardisation beat in-house effort. The line where a bank builds instead of buys is exactly the line of competitive advantage. A bank builds its own algorithms, its own smart-order-routing logic and its own TCA models, because that is where measurable execution edge lives and where it will not hand its secret to a vendor who sells the same thing to rivals.

Block Typical vendors (illustrative) Buy or build
OMS / EMS Charles River, Bloomberg AIM/EMSX, FlexTrade, TradingScreen, Fidessa Mostly buy
Exchanges & MTFs Euronext, LSE, Deutsche Börse, Nasdaq; Cboe Europe, Turquoise, Aquis External venues
Algos & SOR The bank's own Build (the edge)
TCA & analytics Virtu (ITG), Abel Noser, Bloomberg BTCA — plus in-house Buy and build
Market data Bloomberg, LSEG (Refinitiv), FactSet, direct exchange feeds Buy
Post-trade LCH, Eurex Clearing (CCPs); Euroclear, Clearstream (CSDs); DTCC CTM (matching) External utilities

Vendor names consolidate constantly, and the European venue and post-trade layer is anchored by Euronext, LSEG and Deutsche Börse. Post-Brexit, the split between EU and UK venues and clearing arrangements adds a real layer of connectivity and reporting complexity that a platform must handle. Knowing the real names makes the architecture concrete — but the principle is what lasts: buy the plumbing, build the edge.

Market data and latency: the raw material and the arms race

Market data is both the desk's fuel and one of its largest costs

Every decision the stack makes runs on market data — the live stream of quotes and trades from every venue. It is the raw material of pricing, routing, algorithms and TCA, and it is also, quietly, one of the desk's biggest expenses. Exchanges sell their data, and the best, fastest feeds are costly; a large desk spends heavily just to see the market clearly. Data comes in tiers, and the tier you buy shapes what you can do. Direct exchange feeds are the fastest and most granular, delivering every order-book change with minimal delay. Consolidated feeds from vendors like Bloomberg and LSEG aggregate many venues into one stream, convenient but slower. And the depth of book you pay for — top-of-book only, or the full ladder — decides how well an algorithm can see the liquidity it is trading into.

There is a live European policy strand here worth knowing: the long-debated consolidated tape, a single aggregated post-trade (and eventually pre-trade) feed across all venues, intended to fix the opacity that fragmentation created. Its slow arrival is a reminder that fragmentation solved one problem — competition — by creating another: no one place shows you the whole market.

In electronic trading, speed is a measurable edge, so firms pay for microseconds

Once trading is electronic, being faster than the next participant is a real, monetisable advantage — you see a price change first, you reach the venue first, you win the queue. This is the latency arms race, and it has a specific vocabulary. Latency is the delay between an event and your reaction to it, measured in microseconds or nanoseconds. Co-location is renting rack space inside the exchange's own data centre so your servers sit physically next to the matching engine, shaving the speed-of-light delay of distance. And tick-to-trade is the time from receiving a market-data tick to firing an order in response — the number a low-latency desk obsesses over.

For a bank's agency desk, the goal is usually not to win the nanosecond race outright — that is the pure-play HFT firms' game — but to be fast enough that its routing and algos are not systematically picked off by faster participants. For its market-making, speed matters far more directly. Either way, latency is a budget line and an engineering discipline, not an abstraction.

Figure 16 A latency budget from market-data tick to order arriving at the venue. The pre-trade risk check is the segment that must be fast enough not to be resented. (Illustrative.)

Data lineage: following one order from click to reporting

An order is a chain of objects, each inheriting the last

The single most important idea in the whole stack is lineage: an order is not one static record but a chain of objects, each created from the one before it. Follow our Meridian order through the chain. The OMS creates the parent order — name, side, size, limits. The EMS and algo turn it into child orders — a sliced schedule. The router and FIX layer turn those into routed slices — venue, price. The venues return executions — price, time, venue. Allocation turns the fills into an average-price split across sub-accounts. Clearing and settlement turn that into a cleared and settled position. And reporting turns the whole thing into a regulatory record with an audit trail.

Figure 17 Data lineage: each object inherits from the one before it. A wrong field near the top cascades to allocation, settlement and reporting at the bottom.

The cascade principle is the reason data quality is not a back-office concern. An error introduced upstream — a wrong instrument identifier, a mis-keyed size, a bad venue code — is not contained. It flows into every downstream object, so a small mistake at the OMS reappears as a settlement fail and a reporting breach hours later. The chain has no natural firewall; the only defence is getting each field right at the point it is created.

The order data model: the atom of the whole system

The fields of the order object are what everything downstream depends on

If lineage is the chain, the order object is its first link, and its fields drive everything after. A parent order carries, at minimum: the instrument (which security, by identifier), the side (buy or sell), the quantity, the price constraints (limit, or none), the client and account it belongs to, the urgency or strategy, and a set of timestamps — arrival, first fill, last fill — that later become the benchmarks for measuring the trade. Each field has downstream consequences: the instrument identifier decides which venues and which settlement path apply; the account decides allocation; the timestamps decide the TCA verdict. Get the identifier wrong and the trade routes and settles against the wrong security.

Deduplication is a genuinely hard, mission-critical problem here. The same logical order can appear in multiple systems, and the same client can exist under several identifiers across the OMS, the EMS and the settlement layer; matching them correctly — recognising that two records are the same order or the same entity — is non-trivial and, done wrong, produces double-counting, mis-allocation and reporting errors. Clean, deduplicated reference data is the quiet foundation the whole stack stands on.

An order is not a record, it is a state machine

The data model tells you what an order contains; it does not tell you what an order does. And an order is not a static row — it is a living object that moves through a defined set of states, where every transition is a message, an event and an audit record. Understanding the order as a state machine is what separates someone who has read about trading systems from someone who can debug one.

Figure 18 The order lifecycle as a state machine. Every arrow is a FIX message and an immutable audit record; the awkward states are the transient ones in the middle.

The path is straightforward when everything works. An order is sent and sits pending new until the venue or broker acknowledges it, at which point it is new and live in the market. As executions arrive it becomes partially filled, accumulating quantity, and when the last share prints it is filled and done. Around this happy path sit the states that cause all the trouble. An order can be rejected outright — by a risk check, by the venue, by a limit breach. It can be amended, which means a cancel/replace that must be handled atomically. It can be cancelled, passing through a transient pending cancel while the request is in flight. And it can simply run out of time, ending expired or done for day with quantity unfilled.

The transient states are where the bugs live. Consider the classic race: a cancel request is sent while a fill is simultaneously coming back from the venue. Did the order cancel, or did it fill? For a few milliseconds the two systems disagree, and the resolution must be deterministic — the venue's version of events is authoritative, and the client system must reconcile to it. Get this wrong and you have phantom positions: shares the desk believes it does not own, or an order the desk believes is dead that is still working in the market. Every serious execution platform is, at its core, a careful implementation of this state machine with unambiguous rules about who wins each race.

Everything is timestamped, and the clocks must actually agree

A detail that sounds trivial until you think about it: every timestamp in this business is only meaningful if all the clocks agree. Transaction-cost analysis compares the moment an order arrived to the moment it filled. Surveillance reconstructs the sequence of events across venues to determine whether someone traded ahead of someone else. Best-execution evidence rests on knowing what the market looked like at the instant a decision was made. If two systems disagree about what time it is by even a few milliseconds, all of that analysis is quietly wrong — and events can appear to happen in the wrong order.

European rules take this seriously enough to specify it. MiFID II's technical standard on clock synchronisation requires firms to synchronise their business clocks to Coordinated Universal Time, with a maximum divergence and a timestamp granularity that tightens as the trading gets faster — down to the microsecond for high-frequency activity. In practice this means dedicated time sources, usually GPS-disciplined clocks or precision time protocol distributed across the trading estate, plus continuous monitoring to prove the divergence stayed within tolerance. It is an unglamorous piece of infrastructure that quietly underwrites TCA, surveillance and the entire audit trail.

Identifiers: naming the security, the entity and the instrument

Three codes answer three different questions, and each bites somewhere

An order references things that must be named unambiguously, and three identifier systems do the naming. Each answers a different question and each "bites" — causes a failure when wrong — in a different place.

Identifier Names Where it bites
ISIN The specific security (a given share or bond) Routing and settlement: a wrong ISIN trades or settles the wrong instrument
LEI The legal entity (the counterparty or client) Reporting and credit: transaction reporting requires the counterparty's LEI
FIGI The instrument, vendor-neutral Mapping across systems: reconciling one instrument across data sources

The ISIN — International Securities Identification Number — names the security itself, and a wrong one means the trade routes or settles against the wrong instrument. The LEI — Legal Entity Identifier — names the legal entity on each side, and European transaction reporting cannot be filed without it. The FIGI — Financial Instrument Global Identifier — provides a vendor-neutral mapping so that one instrument can be reconciled across data sources that each have their own internal codes. Together they let the system say precisely which security, which entity and which instrument every field refers to.

Integration patterns: batch, API and event streaming

How the pieces talk decides how much reconciliation you need

The functional blocks have to exchange data, and there are three patterns for doing so, each a generation apart. Batch processing — the legacy default — moves data in scheduled bundles, often overnight; it is simple but high-latency, and it forces reconciliation because two systems are only in sync at the end of each batch. APIs let one system query another on demand, in real time, which is the first step of modernisation. Event streaming — Kafka-style publish-and-subscribe — pushes every change to every interested system the instant it happens, in real time.

Immutable audit and the proprietary decision engine

Compliance records must be append-only, and the decisions stay in-house

Two architectural principles sit above the rest. First, compliance-critical records must be immutable — append-only, never edited in place — because surveillance, best-execution evidence and regulatory reporting all depend on a record that cannot be quietly changed after the fact. Governance and audit platforms exist precisely to store these records in a tamper-evident way. Second, the decision engines are always proprietary. A bank will license its order management from a vendor, but it builds its own smart order router and its own allocation logic, because these encode the choices that make its execution good and fair, and because handing them to a vendor would hand its edge to competitors. The record is open and auditable; the decision is closed and owned.

Messaging: two protocols for two worlds

FIX carries the front office, SWIFT and ISO 20022 carry the back

An order is carried by messages, and there are two messaging worlds that map exactly onto the front-office and back-office split. In the front office, the FIX protocol — Financial Information eXchange — carries new orders, execution reports and cancel/replace instructions between the buy-side, the broker and the venues. It is session-based and fast, working in milliseconds, and it is how an order actually enters and moves through the primary execution process. In post-trade, SWIFT messages carry confirmation, allocation and settlement instructions between brokers, custodians and depositories; the older SWIFT MT message types are migrating to the richer, structured ISO 20022 / MX standard, which carries more data in a machine-readable form.

Layer Standard Carries Timescale
Front office FIX New order, execution report, cancel/replace Milliseconds
Post-trade SWIFT MT → ISO 20022 Confirmation, allocation, settlement instruction Seconds to minutes

Building and releasing trading software is its own discipline

You cannot test in production when production moves real money

Most software teams can ship, watch, and roll back. A trading platform cannot: by the time you notice, the orders are in the market and the trades are done. This is why release engineering in trading is a distinct discipline with its own controls, and why the Knight Capital failure in Part 5 is best read not as a cautionary tale but as a specification for what follows.

Four testing layers stand between a change and the live market. Unit and integration testing covers the code in the ordinary way. Market replay and simulation is the layer specific to this business: recorded historical market data is replayed into the system so an algorithm can be run against a real day — including a violent one — and its behaviour observed without touching a venue. Conformance testing is mandated by the venues themselves: before a firm may connect a new or changed order-entry system, it must pass the exchange's certification suite proving it handles every message type and error condition correctly. And backtesting with honest assumptions validates an algorithm's logic, though it flatters itself unless it models the impact the algorithm's own orders would have caused.

Deployment carries its own controls. Canary releases send a small fraction of flow through the new version and compare its behaviour to the old before proceeding. Feature flags allow a behaviour to be switched off instantly without a redeployment — though, as Knight discovered, a flag whose meaning has been repurposed while old code still reads it is a loaded weapon. Version control and named ownership of every algorithm is a regulatory requirement in Europe, not merely good practice. And rollback must be tested, because a rollback path that has never been exercised is a hypothesis, not a plan.

Observability: you cannot control what you cannot see in real time

Because the market will not wait, monitoring in trading is not a dashboard reviewed each morning — it is a real-time control surface. The platform must continuously surface order flow health (rejection rates, unusual message volumes, orders stuck in transient states), latency distributions measured not by averages but by the tail, since the ninety-ninth percentile is where money is lost, fill quality drifting away from expectation on a given venue or algorithm, capacity and throughput against the message rates a volatile day produces, and position and risk against limits in real time. Alerts must be actionable and few; an operator who receives two hundred alerts a day will miss the one that mattered.

The Knight case again supplies the sharpest illustration. Alerts did fire during those forty-five minutes, but nobody could tell quickly which system was misbehaving or how to stop it, and the response made things worse before it made them better. Observability is not the ability to produce data about an incident afterwards; it is the ability to identify and stop the right thing during one.

Tick data: the largest dataset the desk owns

Market data is not only consumed in real time — it is stored, and the volume is formidable. Every quote update and every trade on every venue, retained for years, produces a dataset far larger than anything else the desk holds, and it must remain queryable because TCA, backtesting, surveillance reconstruction and regulatory enquiries all depend on it. Storing tick data well is a specialised problem, typically solved with time-series databases and columnar formats optimised for reading long, narrow slices of history rather than the row-by-row access a transactional database expects.

Two requirements shape the architecture. Normalisation: every venue publishes in its own format with its own conventions, so a canonical internal representation must be built and maintained, and every change a venue makes must be absorbed without corrupting history. Point-in-time accuracy: to reconstruct why a decision was made, you need the market as it appeared then — including the reference data as it stood then, not as it has since been corrected. A backtest run against retroactively cleaned data will produce results the live system could never have achieved.

Reconciliation and exception management: the work that never appears in a demo

Wherever two systems hold a view of the same reality, they will disagree, and something must find and fix the difference. Reconciliation compares positions, cash and trade records across the OMS, the clearing broker, the custodian and the books, and produces breaks — differences requiring investigation. Exception management is the workflow that carries each break to resolution: settlement fails chased before CSDR penalties accrue, unmatched confirmations resolved before the settlement deadline, allocation discrepancies corrected before they reach the client.

This is the least glamorous part of the stack and one of the most consequential, because it is where errors are caught before they become losses. It is also where the move to T+1 bites hardest: compressing the timetable does not reduce the number of breaks, it reduces the hours available to clear them, which is why exception management is being automated aggressively rather than simply staffed more heavily.

Model risk governance: an algorithm is a model, and models need owners

Long before generative AI arrived, the desk was already full of models: the impact model inside an Implementation Shortfall algorithm, the venue-selection logic in the router, the volume forecast a VWAP algorithm trades against, the anomaly thresholds in surveillance. Each of these is a model making consequential decisions, and each needs the governance that models require. That means a maintained inventory of every model in production with a named owner; independent validation by someone who did not build it, testing the assumptions rather than re-running the author's own tests; ongoing performance monitoring to detect drift as market conditions change; documented limitations, stating plainly where the model is not to be trusted; and change control so that a recalibration is a governed event.

The reason this matters here is that algorithm governance is often treated as a compliance formality — testing and version control to satisfy the regulator — when it is really the discipline that keeps a quietly decaying model from costing money for months before anyone notices. A model that was well calibrated to 2023's volatility regime may be systematically wrong in a different one, and nothing will announce it. Only monitoring will.

Operational resilience: the one irreversible moment has no second chance

Deal-critical moments cannot be rerun, so resilience is measurable, not aspirational

Most software can retry. A trading system, at certain moments, cannot. The closing auction happens once. A settlement instruction, once sent and matched, moves real cash and real shares. There is no second attempt at the moment that matters. That is what makes operational resilience in this business concrete rather than a slogan: the system must be provably available on the one irreversible moment, and "we were down for ten minutes" is not an incident report, it is a loss and a regulatory breach.

Resilience has precise components. Redundancy means having more than one of everything, so no single failure stops the system. Automatic failover means that when a component dies, a standby takes over without human intervention and without dropping orders. Disaster recovery means the whole environment can be reconstituted at a second site. And business continuity means the people and processes, not just the machines, can keep operating through a disruption.

Concept What it guarantees Why it matters on the critical day
Redundancy No single point of failure The order path survives a component dying mid-auction
Automatic failover Seamless takeover by a standby No dropped orders when a server fails at the close
Disaster recovery The environment can be rebuilt elsewhere Trading resumes after a site-level outage
Business continuity People and processes keep running The desk operates through a disruption, not just the systems

Five regulatory regimes, read as system specifications

The rules are really requirements on the platform

Regulation in equities is not background; it dictates architecture. Read as an engineer, each regime is a specification the platform must meet. Five matter most.

MiFID II / MiFIR is the master specification: best-execution monitoring, pre- and post-trade transparency, the systematic-internaliser regime, tick-size rules, and — heaviest of all in data terms — transaction reporting (RTS 22), a detailed report on every trade. The Market Abuse Regulation (MAR) requires the surveillance stack: the system must detect and evidence insider dealing and manipulation — spoofing, layering, wash trades. The Short Selling Regulation requires position monitoring and disclosure, and the ability to enforce bans. CSDR — the Central Securities Depositories Regulation — requires the post-trade discipline that makes settlement penalties calculable and avoidable. And DORA, with data-protection rules like GDPR alongside it, requires the resilience and data-governance controls above. Each is a legal document; each is also a set of features the platform cannot ship without.

  1. What is the difference between an OMS and an EMS?

  2. What is the cascade principle, and why does it make data quality a front-office concern?

  3. Why is moving from batch to event streaming more than a speed upgrade?

  4. What is the difference between FIX and SWIFT/ISO 20022, and what split do they embody?

  5. Why is it more useful to think of an order as a state machine than as a record, and why must the clocks be synchronised?

  6. What testing layers stand between a code change and the live market, and why can you not simply roll back?

  7. Why does an execution algorithm need model risk governance?

4

  1. What is the difference between an OMS and an EMS?

  2. Why is moving from batch to streaming more than a speed upgrade?

  3. Which standard carries execution reports in milliseconds?

  4. Why does the cascade principle make data quality a front-office concern?

5Risk, Real Failures and the Road Ahead

Everything so far describes how equity trading works when it works. This part is about the edges: the ways it fails, how quality is measured, three real events that broke firms and briefly broke the market, and where the business is genuinely changing — including what generative AI does and does not change. We end with the six ideas worth carrying out of the whole module.

A map of the risks, not just a list

Six risk types cover the field, and each one is somebody's job to control

Risk in trading is easy to gesture at and hard to manage unless it is organised. The useful move is to make the taxonomy mutually exclusive and collectively exhaustive — a map on which every real incident lands in exactly one place, and every place has an owner.

Figure 19 The six risk types of an equity-trading business. A map is more useful than a list because each type has a distinct owner and a distinct control.

Market and inventory risk is the principal book moving against the desk — the risk trader's warehoused position losing money before it can be unwound. Execution and impact risk is poor fills, information leakage and missing the benchmark — the everyday risk of the agency business. Operational and settlement risk is system failure, settlement fails and rogue algorithms — the risk that the machine or the plumbing breaks. Conduct and market-abuse risk is spoofing, layering, misuse of material non-public information and best-execution breaches — the risk that someone games the market. Technology and resilience risk is an outage on the one irreversible moment — the close, a settlement window. And regulatory and reporting risk is failed transparency, inaccurate transaction reports and surveillance gaps — the risk that the evidence trail is wrong. Each has a distinct owner, from the risk trader to compliance to IT, and a distinct control.

Measuring execution quality: the KPIs that run the business

If you cannot measure the fill, you cannot sell the execution

Because execution is a product sold on provable quality, the metrics are not back-office reporting — they are the scoreboard the whole business runs on. Transaction-cost analysis produces a small set of key indicators, and every serious desk lives by them.

Implementation shortfall is the slippage from the arrival (decision) price — the purest measure of total cost, capturing both impact and drift. VWAP performance is the fill price versus the day's volume-weighted average — the benchmark many clients are measured against. Fill rate and completion is the share of the order actually executed, and how much was left undone. Market-impact and reversion measures how far the price moved during the trade and how much of that move bounced back afterward — separating temporary from permanent impact. And venue and price-improvement analysis measures how much better than the quoted price each venue delivered, which drives routing decisions. Together these feed the loop — execute, measure, diagnose, adjust — that improves the next order.

Figure 20 Execution quality is a feedback loop: every trade is measured, diagnosed and used to re-tune the algos and the router. The algo wheel turns this loop into a broker meritocracy.

Three real failures, three different lessons

The best teachers in this business are its disasters

Three real events, each a different kind of failure, together map the whole risk landscape. None is invented; the facts below are drawn from public regulatory findings and reporting.

The 2010 Flash Crash — when market structure itself failed

On 6 May 2010, beginning around 2:32 p.m. Eastern time, the US equity market briefly came apart. The Dow Jones Industrial Average fell roughly 1,000 points — about 9% — in minutes, and then recovered most of the loss within the same hour, in an episode that briefly erased close to a trillion dollars of value. Some individual stocks printed absurd prices, trading as low as a single cent or as high as $100,000, before the trades were later broken.

The trigger, according to the joint SEC-CFTC report, was a single large sell program: a fundamental trader (widely identified as the mutual-fund firm Waddell & Reed) sold 75,000 E-mini S&P 500 futures contracts, worth about $4.1 billion, to hedge an equity position. The order itself was legitimate. What turned it into a crash was how it was executed — an algorithm set to participate at a fixed 9% of trading volume, with no regard to price or time. In already-stressed, thinning conditions, the algorithm fed sell orders into the market as fast as volume rose, and volume was itself being inflated by high-frequency firms passing contracts back and forth. As prices fell, HFT market-makers — who have no obligation to keep quoting — hit their inventory limits and withdrew, and the liquidity that everyone had assumed was there simply vanished. The result was a self-reinforcing spiral into a vacuum.

The lesson is about market structure and liquidity. A fragmented, high-speed market can appear deeply liquid and then evaporate in seconds when the machines providing that liquidity step back at once. A blunt execution algorithm — no price limit, no time sensitivity — was the match, but the fragility was in the structure. The regulatory response reshaped the market: single-stock circuit breakers, then the limit-up/limit-down mechanism, and market-wide breakers, all designed to force a pause before a spiral can complete.

Knight Capital, 2012 — when a deployment killed a firm in 45 minutes

On 1 August 2012, Knight Capital — then the largest US retail equity market-maker — deployed new software to one of its trading systems to handle a new NYSE retail programme. The deployment was botched: an engineer's update was installed on only seven of the firm's eight servers, and a flag that had once controlled an old, dormant function called "Power Peg" was repurposed. On the eighth server, that flag switched the obsolete code back on. When the market opened, that server began firing millions of unintended orders into the market — buying high and selling low, over and over, across 154 stocks.

The numbers are staggering because of how fast they accumulated. In roughly 45 minutes, before the system could be stopped, Knight sent millions of erroneous orders, executed some four million trades across about 397 million shares, and built up around $7 billion in unwanted positions. Unwinding them cost the firm about $440 million — roughly three times its annual earnings and close to its entire market value. Its stock fell about 75% in two days; it survived only via emergency financing that handed most of the company to new investors, and it was soon absorbed by a competitor (Getco, forming KCG). A firm that had taken seventeen years to build was effectively destroyed before lunch.

The lesson is about technology and operational control. There was no market crash and no rogue human — just a deployment error and a missing safety net. The critical gap was the absence of an effective in-path pre-trade control and a fast kill-switch: nothing stood between a runaway algorithm and the live market. This is exactly the control the SEC's Market Access Rule (15c3-5) had been written to require, and exactly the "controls in the order path, not beside it" principle from Part 2. Redundancy, testing, deployment discipline and an instant kill-switch are not bureaucracy; they are what stands between a bug and insolvency.

Spoofing — when the conduct itself is the abuse

The third failure is not an accident but a deliberate manipulation, and it is the one the surveillance stack of Part 3 exists to catch. Spoofing — and its cousin layering — means placing large orders you never intend to execute, to create a false impression of supply or demand, then trading on the price move your fake orders caused, and cancelling them before they fill.

The emblematic individual case is Navinder Sarao, a London-based trader who, from around 2009 to 2014, used automated dynamic layering to spoof the E-mini S&P 500 futures market. He placed large resting orders he intended to cancel, moving prices to benefit his real trades on the other side. He was charged in 2015, extradited to the United States, and in November 2016 pleaded guilty to spoofing and wire fraud, consenting to forfeit close to $13 million; regulators found his order-book imbalance had contributed to the conditions of the Flash Crash itself. Spoofing is not only a lone-trader crime: in September 2020, JPMorgan entered a deferred-prosecution agreement and paid $920 million — the largest spoofing-related sanction on record, including a substantial criminal fine — to resolve manipulation across its metals and Treasury futures trading over roughly eight years, with prosecutors using racketeering law against the traders involved.

The lesson is about conduct and market abuse. Spoofing is exactly the pattern that MAR bans and that trade-surveillance systems are built to detect: orders placed with no intent to trade, layered to move a price, cancelled once they have done their work. The enforcement wave — from a lone trader to a global bank — is why surveillance is a first-class system embedded in the pipeline, not an afterthought.

Three cases, three boxes on the risk map

Case Failure mode Where it lands on the map The lesson
Flash Crash (2010) Market-structure fragility, liquidity evaporation Execution / technology & resilience Liquidity can vanish; blunt algos plus fragile structure spiral
Knight Capital (2012) Technology & deployment failure Operational / technology & resilience Controls must sit in the order path, with a real kill-switch
Spoofing (Sarao; JPMorgan) Deliberate conduct / market abuse Conduct / market abuse Surveillance must detect intent-to-cancel manipulation

Five forces reshaping the business

The direction of travel is automation, data and resilience

The equity-trading business is not static, and five trends are reshaping it at once. AI-assisted execution and analytics — machine learning tuning algos, predicting short-term liquidity and reading TCA — is moving from pilot to production. Real-time, cloud and API platforms are replacing overnight batch with streaming architectures that a T+1 world requires. RegTech and automated surveillance apply the same analytics to compliance, catching abuse and best-execution failures faster and more cheaply. Settlement compression to T+1, and the operational automation it forces, is remaking post-trade. And continued electronification and consolidation push ever more flow through low-touch channels and concentrate it among the largest, most automated players.

Trend What it changes Why it matters
AI-assisted execution & analytics Smarter algos, liquidity prediction, adaptive routing Potential edge in the core execution product
Real-time cloud & API platforms Batch gives way to streaming Enables T+1, removes reconciliation, cuts cost
RegTech & automated surveillance Compliance becomes data-driven Catches abuse and best-ex gaps faster and cheaper
T+1 settlement Halves the post-trade window Forces same-day affirmation and STP
Electronification & consolidation More low-touch, fewer players Scale becomes the decisive advantage

Generative AI: real uses, honest limits

Where GenAI genuinely helps, and where it must be fenced

Generative and machine-learning AI is the loudest current theme, and it does have real uses in this business — alongside real limits that matter more here than almost anywhere else. Four use cases are genuine. Execution analytics and summarisation: turning mountains of TCA and market data into plain-language diagnostics a trader can act on. Predictive and adaptive algorithms: models that forecast short-term liquidity and volatility and adjust the schedule accordingly. Smarter routing: learning from past fills which venues perform for which orders. And surveillance and operations: spotting abuse patterns and settlement exceptions faster than rule-based systems, and drafting the first version of routine reports.

But the caveats are not boilerplate; in trading they are the whole game. Data quality is the gate: a model trained on messy, poorly-labelled trade data will produce confident, wrong answers — the cascade principle again, now with a model on top. Edge decays and models overfit: a strategy that worked in backtest can fail live as the market adapts, and an overfit model mistakes noise for signal. Explainability is a governance requirement, not a nicety: a regulator asking why an algo routed an order, or why surveillance flagged (or missed) an abuse, will not accept "the model decided" — the decision must be explainable and auditable. And confidentiality is acute: client orders and positions are exactly the material non-public information the information barrier exists to protect, so feeding them to a general-purpose model is a serious risk. The honest posture is augment, not replace: AI makes traders, quants and compliance officers faster and sharper, but the accountable human — and the auditable decision — stays.

GenAI use case The gain The governing risk
Execution analytics & summarisation Faster, clearer diagnostics from TCA data Data quality; misleading confidence
Predictive / adaptive algos Better liquidity and volatility forecasts Edge decay, overfitting, backtest illusion
Smarter routing Learned venue selection Explainability for best-execution proof
Surveillance & operations Faster abuse and exception detection False positives/negatives; auditability

Further out: three changes that would alter the structure itself

The five trends above are extrapolations — these three are discontinuities

Everything in the previous section is the present accelerating. It is worth separating out a small number of developments that, if they happen, would change the shape of the market rather than its speed. Each is genuinely uncertain, and the honest framing is conditional.

Tokenisation and DLT settlement would compress the post-trade layer rather than automate it. The logic of T+1 — and eventually T+0 — runs toward atomic settlement, where the exchange of securities and cash happens simultaneously and irrevocably on a shared ledger, removing the settlement window and much of the machinery built to manage risk inside it. The European DLT Pilot Regime already provides a supervised sandbox for market infrastructures to try exactly this. But the obstacles are structural rather than technical: netting, which we saw collapses 1,800 settlements into one, is a feature of the current model that atomic settlement gives up, and pre-funding every trade in full would consume enormous liquidity. Expect the change to arrive first where netting matters least, and slowly.

Extended and continuous trading hours are being pushed by demand for round-the-clock access, particularly from retail investors and from Asian participants trading US names. Some venues have moved toward overnight sessions, and the question of a genuinely twenty-four-hour equity market is live. The barrier is not technology but the day itself: the closing auction is the anchor of index valuation, benchmark pricing and fund accounting, and a market with no close has no natural reference price. Any move here forces a redesign of things — benchmarks, net asset values, batch cycles — that quietly assume a day has an end.

A structurally larger European retail participation would change the flow mix rather than the infrastructure. Retail investing in Europe has historically lagged the United States, and both policy — the Savings and Investment Union agenda — and commercial pressure from low-cost brokers are pushing it up. More retail flow means more uninformed volume, which improves conditions for market-makers, intensifies debate about order routing and payment for order flow, and raises the political salience of market structure. It is the change most likely to happen and the least likely to be described as a technology story.

  1. Why is a risk map more useful than a risk list?

  2. Why can beating VWAP hide a bad execution?

  3. What distinct failure mode does each of the three cases represent?

  4. What was the critical control missing at Knight Capital, and where should it have sat?

  5. Why is explainability a governance requirement for AI in trading, not just a nicety?

  6. What is the single filter to apply to any claimed transformation?

A final exercise: put the whole module to work on one order

Reading this module is not the same as being able to use it

Everything so far has been explained. This last exercise asks you to decide. Work through it before reading the commentary — the value is in committing to an answer and then finding out where your reasoning was thin.

5

  1. Why can beating VWAP hide a bad execution?

  2. Which failure mode is Knight Capital?

  3. What control was missing at Knight Capital?

  4. Why is a risk map more useful than a list?

Module Summary — The Six Ideas That Matter Most

If you remember nothing else from this module, remember these six. Each compresses one of its parts, and together they are the whole business in miniature.

One — trading is execution, not stock-picking

The desk does not decide what to own; it converts someone else's decision into filled, settled shares while giving up as little as possible to impact, timing and cost. Access, skill, capital and evidence are what it sells. Everything else is in service of doing that well and proving it.

Two — the market is fragmented by design, so software reassembles it

One stock trades across a dozen lit and dark venues at once, a deliberate product of regulation. Smart order routing exists because no human could find the best price and size across all of them in real time. Fragmentation is the environment, not a defect, and the machine is the answer to it.

Three — execution quality is made at the order book and measured against the right benchmark

The central problem is impact versus timing, with an efficient-frontier optimum. Algorithms decide how much and when, routers decide where, and the true cost is the slippage from the arrival price — not the headline commission, and not a flattering VWAP number.

Four — compliance and control are core architecture, not overhead

The information barrier, pre-trade risk in the order path, the kill-switch, surveillance and transaction reporting are built into the pipeline and must run as fast and reliably as the trading. Knight Capital and the spoofing cases show what happens when a control is missing or a manipulation goes undetected.

Five — the economics are thin margins and a scoreboard, so scale and evidence win

Fractions of a basis point per trade force automation, cost discipline and consolidation, while a data-driven broker scoreboard rewards provable quality and pushes execution toward commoditisation. The business is cyclical, so fixed cost must be carried through every phase of volatile revenue.

Six — real transformation moves cost, quality or risk — including, carefully, with AI

Streaming that removes reconciliation, T+1 that forces automation, and AI that measurably sharpens execution or surveillance change the business; dashboards that only change the appearance do not. AI raises the value of clean data and good governance rather than replacing them, and it augments the accountable human rather than standing in for them.

Process Cheat Sheet — The Whole Trade on One Page

One page to hold the entire module in view: the seven stages of a trade with their timing, activity, output and owners, then the numbers worth carrying in your head. Use it as a map to place any detail from the deep dives.

The seven stages at a glance

Stage Timing Core activity Key output Lead & systems
1. Order received Seconds Capture name, side, size, limits, urgency; onboard/KYC checked Validated parent order Sales-trader / OMS
2. Strategy & routing Minutes Choose agency vs principal; pick algo, block or capital commitment; slice Execution plan Sales-trader, e-desk / EMS
3. Execution Minutes–days Work the order: algo + SOR across lit and dark; seek blocks; watch impact Fills across venues Algo, SOR / EMS, FIX
4. Allocation Trade day Split average-priced fills across sub-accounts fairly Allocated positions Middle office / OMS
5. Confirm & affirm T / T+0 Match trade details broker-to-client, same-day Affirmed trade Middle office / matching
6. Clearing (CCP) T / T+1 Novation and netting; default-risk management Cleared position CCP / operations
7. Settlement (CSD) T+1 Delivery-versus-payment; record in custody Settled, owned shares CSD, custody / SWIFT-ISO

The numbers worth memorising

All figures below are illustrative teaching values except where marked real; they are orders of magnitude, not quotes.

Quantity Rule-of-thumb value Why it matters
Reference order size 2,000,000 shares ≈ 3× daily volume Too big to trade at once — the whole reason for algos and blocks
Low-touch commission A fraction of a basis point to a few bps Thin margin forces automation and scale
Implementation shortfall (example) ~60 bps vs arrival price (€240k on the order) The true, all-in cost of working the order
IS decomposition (example) ~25 bps impact + ~30 bps drift + ~5 bps fees Impact is controllable; drift is the price of time
VWAP vs arrival (example) Beat VWAP +3 bps but −60 bps vs arrival The benchmark defines what "good" means
Block cross (example) 800k of 2m shares crossed dark at midpoint A block does the heavy lifting; the algo cleans up
POV participation ~10% of live volume, typical Adapts to real liquidity but leaves a footprint
Settlement cycle T+1 — EU/UK/CH target October 2027 (real) Removes the overnight window; forces STP
Flash Crash 6 May 2010: DJIA ~−9% in minutes (real) Liquidity can evaporate in a fragmented, fast market
Knight Capital 1 Aug 2012: ~$440m lost in ~45 min (real) Controls must sit in the order path, with a kill-switch
Spoofing sanction JPMorgan 2020: $920m (real) Market abuse is caught by embedded surveillance
All-in cost split (example) ~3 bps commission vs ~85 bps everything else The commission is a rounding error
Transaction tax Tens of bps in FR/IT/UK regimes Explicit, large, and it changes routing
Spread (example) €0.04 on €20.00 = 20 bps round trip The quoted spread prices a small trade only
Days to trade at 25% POV 2.8 days of volume → ~11 trading days Speed and discretion cannot both be had
Closing auction share Often ~20% of daily volume Why a whole algo family targets the close
Tick-to-trade (example) ~90 µs end to end Controls must clear without adding latency

The one-sentence version

An equity trade is an intention turned into settled shares: captured as an order, worked across a fragmented market by algorithms and routers that trade off impact against timing, measured against the arrival price, allocated fairly, and cleared and settled through a CCP and CSD — with controls, surveillance and reporting built into every step.