Aller au contenu
Corporate & Investment Banking Mon parcours English

Accueil / Capital Markets & Securities Services / Fonctions transverses / Module 26

26Risk & P&L

Market Risk, the Greeks, and the Systems That Keep a Trading Book Honest

Les grecques, le risque de marché, et l’histoire quotidienne derrière les chiffres.

Lecture
2 h 50
Sections
71
Questions de rappel
28
Questions de QCM
13
Simulateurs
1

Le texte de ce module est encore en anglais : la version française arrive. Les outils interactifs, eux, sont déjà en français.

Two threads run through everything that follows

The business thread is about measurement and control. A desk takes risk deliberately; the firm's job is to size it, price it, and keep it inside agreed limits so a single bad day cannot sink the bank. That thread runs from the Greeks a trader watches in real time, through the aggregate measures a risk manager reports overnight, to the capital a regulator forces the firm to hold. It is, at heart, a story about how an institution stops itself from taking more risk than it can afford.

The technology thread is about data. Every risk and P&L number rests on positions, curves, and prices, and it is only as good as they are. A wrong mark, a missing trade, or a stale curve corrupts the risk, the P&L, and the capital at the same time. That is why risk & P&L is not primarily a mathematics problem — the quant methods are largely settled — but a data and systems problem, and why IT, business analysts, and data teams sit at its centre. This module is written to be equally useful to someone who wants to manage a book and someone who wants to build the platform that measures it.

How to read this module

Read it front to back the first time. Part 1 builds the fundamentals: what market risk is, the Greeks that name it, what P&L is and why it must reconcile, and the daily cycle that produces both. Part 2 moves from a single position to the whole book — VaR, expected shortfall, stress testing, the limit framework, and the FRTB capital regime that now governs it all. Part 3 introduces the people, the systems, and the regulation that make the numbers appear every morning. Part 4 shows what happens when it fails, where the real pain lives, and how cloud, real-time compute, and AI are reshaping the work.

A single fictional example runs through the whole module: a EUR interest-rate swaps and swaptions desk we will call the Kestrel book. Every figure attached to it is illustrative and chosen to teach, not drawn from any real institution. Watch it move through one trading day in Part 1, then reappear as the subject of a P&L investigation in Part 4.

Coloured boxes carry specific kinds of content, and it is worth learning the legend before you start.

Key insight boxes crystallise the single most important idea in a section. Going deeper boxes add nuance for readers who want the next layer of detail. IT & data lens boxes translate a concept into what it means for systems, pipelines, and data. Watch out boxes flag a common trap or a place where things quietly break. Worked example boxes carry the Kestrel book and other illustrative numbers. In a nutshell boxes recap a part in a few sentences. Check your understanding boxes let you test yourself before moving on, with the answers in italics.

You will understand what market risk actually is

You will be able to say precisely what market risk is, how it differs from credit, liquidity, and operational risk, and why a trading book is built to hold it rather than avoid it. The core intuition — that risk is a sensitivity multiplied by a plausible move — will let you see every later technique as a variation on one idea.

You will read the Greeks as the shape of a book

Delta, gamma, vega, theta, and DV01 will stop being jargon and become what they are: the slopes of value against each market variable. You will read a sensitivities ladder and know what curve and volatility risk a single net number is hiding.

You will know why P&L is a control, not just a score

You will understand the kinds of P&L, why an independent function computes it, and how P&L-explain decomposes a day into new trades, market moves, carry, and a tell-tale residual. You will see why chasing that residual to zero is the daily job, and why regulators now tie it to capital.

You will understand how risk becomes capital

You will be able to explain VaR and why the industry moved past it, what expected shortfall adds, why stress testing complements both, and how the FRTB regime turns the Greeks into the capital number that drives the bank's return on equity.

You will trace a risk number from a single mark to the regulator

You will be able to follow one figure from a raw market mark, through valuation, the Greeks, VaR, and aggregation, to a line in a regulatory report — and see why an error at the top cascades all the way down. That trace is the reason this whole discipline is, for a technologist, a data-quality and lineage problem first.

1The Fundamentals: What Risk and P&L Really Are

A trading book is a machine for holding risk on purpose and being paid for it. This part builds the two numbers that describe that machine every day — the risk it is running and the profit or loss it earned — from the ground up. We start with what market risk actually is, name it with the Greeks, read it off a sensitivities ladder, then turn to P&L: what it is, why it must reconcile to the last euro, and how the daily explain decomposes a day into its causes. Underneath all of it sits valuation and data, checked by an independent control. By the end you will be able to read a desk's day the way a controller, a risk manager and a regulator all read it — through the same lens.

Market risk is the risk you take on purpose

Market risk is the risk that the value of a position falls because a market variable moves — a rate, a price, an exchange rate, a credit spread, a volatility. That variable is outside your control; your position's exposure to it is not. The number that connects the two is the whole subject of this part. Everything that follows is a way of asking the same question with more precision: if the world moves, how much do I make or lose, and why.

What sets market risk apart is that a trading book exists in order to hold it. A lending business would rather a borrower never defaulted; credit risk is a cost of doing business. A trading desk is different. It is paid to warehouse market risk — to take the other side of a client's hedge, to make a price, to hold a position through a move — and the risk is the source of the return, not a byproduct of it. The job is therefore not to eliminate market risk. It is to size it, price it and control it: to hold exactly the risk you intend, in the amount you intend, and to know at all times what that risk is.

Three other risk types sit next to market risk and are often confused with it. Credit risk is the risk that a counterparty or issuer fails to pay what it owes — a borrower defaulting, a bond issuer missing a coupon. Liquidity risk is the risk that you cannot fund a position or exit it without moving the price against yourself. Operational risk is the risk that a process, a system or a person fails — a trade booked wrong, a settlement missed, a system outage. Each is a real discipline with its own models, its own teams and its own capital regime, and where they meet market risk is exactly where the hardest problems live. But they are distinct, and this module is about market risk.

Market risk shows up in four recognisable shapes

A book's market risk is not one undifferentiated lump; it arrives in shapes that repeat across every asset class. Learn the four shapes once and you can read any book, because a government-bond desk and an equity-options desk are running the same four exposures in different clothing.

Directional (delta) risk

Directional risk is the exposure you have simply because a price or rate moves up or down. The book is long or short; when the underlying rises, a long position gains and a short loses, in proportion. This is the first-order exposure — the biggest, most obvious risk in most books — and it is the one that gets hedged first. If you are long ten-year bonds and yields rise, you lose; that loss is directional risk expressing itself.

Curvature (gamma) risk

Curvature risk is the risk that your exposure itself changes as the market moves. A position whose directional risk is stable is easy to hedge once and leave. A position whose directional risk grows as the market moves away from you is not: the hedge you put on this morning is the wrong size by this afternoon. Small moves are fine; large moves hurt — or help — faster than the first-order number suggests. Options are the classic source, but any non-linear payoff carries it.

Volatility (vega) risk

Volatility risk is the exposure of an option's value to how much the market is expected to move, not to where it is now. An option is a bet on a range, so its value depends on the width of the range the market prices in — the implied volatility. That number moves on its own, driven by supply, demand and fear, even when the underlying is still. Any book holding options carries volatility risk whether or not the underlying has budged.

Basis and spread risk

Basis and spread risk is the risk that two things which should move together drift apart. A hedge is only a hedge if the hedging instrument tracks the position it protects. When a bond is hedged with a swap, or one index with a near-cousin, a residual exposure remains to the gap between them — the basis. Most of the time it is quiet. In a crisis, relationships that held for years break at once, and a book that looked hedged discovers it was not.

The core intuition: risk equals sensitivity times move

Every market-risk number in this module is built from one idea: market risk equals sensitivity multiplied by move. If you know how much your book's value changes for a one-unit move in a market variable — the sensitivity, which the market calls a Greek — and you have a view on how large a move is plausible, you can estimate the loss. Multiply the two and you have a euro figure. That is the entire mechanism. Greeks, value-at-risk and stress testing, which fill the rest of this module, are three different ways of choosing the "move" and combining it with the sensitivity.

In Europe, market risk is not only measured but capitalised under a specific rulebook, and the boundary of what counts is itself policed. European banks measure and hold capital against market risk under the Capital Requirements Regulation — the CRR, which is the EU's implementation of the global Basel standards — with the new Fundamental Review of the Trading Book rules, known as FRTB, phasing in for the trading book. Supervisors review both the models and the data behind them: the European Central Bank for the largest banks under the Single Supervisory Mechanism, alongside national authorities such as the Autorité de contrôle prudentiel et de résolution, the ACPR, in France. The line between the trading book and the banking book — which positions attract market-risk treatment at all — is a regulated, closely watched boundary, because which book a position sits in decides how its risk is measured and capitalised. Part 2 returns to FRTB and that boundary in detail; here it is enough to know the rulebook exists and that the perimeter is drawn deliberately.

Figure 1 Where risk and P&L sit: computed every day, on the whole book — the lens the desk, controllers and regulator all see the business through.

The Greeks name the risk

The Greeks are the sensitivities that turn a vague worry into a measured exposure. Each Greek answers one question: if this variable moves by one unit, how much does my book change? Give the four shapes of risk a number apiece and a book stops being a story and becomes a set of figures you can limit, hedge and reconcile. The names are Greek letters by convention; the meaning is plain.

Delta

Delta is the sensitivity of the book to a move in the underlying price or rate. If the underlying moves by one, the book makes or loses delta. It is the primary directional risk — the largest and first-hedged exposure in most books — and it is what a trader means when they say they are "long" or "short." A book with positive delta profits when the underlying rises and loses when it falls.

Gamma

Gamma is the rate at which delta itself changes as the underlying moves. It is the curvature Greek: the second derivative, the sensitivity of the sensitivity. High gamma means a hedge goes stale quickly, because a move that shifts the underlying also shifts your delta, so the position you re-hedged this morning is mis-hedged by the afternoon. Large moves are where gamma shows up, for good or ill — it is why big days are never just scaled-up small days.

Vega

Vega is the sensitivity of the book to a move in implied volatility. If implied vol rises by one point, the book makes or loses vega. It is the core risk of any options book, and it is invisible to anyone tracking only the underlying, because vol moves on its own. A short-vega book — one that has sold options — loses when volatility spikes, which is precisely when everything else is going wrong too.

Theta

Theta is the sensitivity of the book to the passage of time. Options lose value as expiry approaches, all else equal, so theta is the daily time decay a book earns or pays. A book that has sold options earns theta as the clock ticks down; a book that has bought them pays it. Theta is also where carry and roll-down show up — the quiet, mechanical part of a day's P&L that happens even if no market variable moves at all.

Rho and DV01

Rho is the sensitivity of the book to a move in interest rates, and in rates books its practical form is DV01. DV01 — the dollar, or euro, value of a one-basis-point move — is the P&L for a one-basis-point change in yield. In an interest-rate book this is the dominant Greek, the equivalent of delta for a bond or swap: it tells you, per basis point, how much the book makes or loses when the curve moves. The Kestrel book in the pages ahead is read almost entirely through DV01.

Cross and higher-order Greeks

Beyond the first- and second-order Greeks sit the cross and higher-order sensitivities that matter for large or exotic books. Vanna measures how delta changes when volatility moves; volga measures how vega changes when volatility moves; cross-gamma measures how the exposure to one factor changes when a different factor moves. They are real, and for a big options book they are not optional. For most purposes they stay in the background — worth naming so they are not a surprise, worth keeping intuitive rather than mysterious: each is still just a slope of value against a variable, only a subtler one.

The unifying idea is that a Greek is a partial derivative — the slope of value against one variable, holding the others still. That is all a Greek ever is. A trader manages a book by keeping each Greek within its limits and hedging the ones they do not want to hold, trimming delta here, buying back gamma there. The intuition and the calculus are the same picture drawn twice.

For a technologist, the Greeks are not abstractions — they are outputs of the pricing engine and inputs to almost everything downstream. The same engine that values a trade produces its Greeks as a by-product of the valuation. Those Greeks then feed the risk report, the P&L-explain and the regulatory capital calculation, three consumers that must all see the same numbers.

That is why clean, consistent Greeks are the whole game. If the risk system and the P&L system compute delta slightly differently — a different curve, a different bump size, a different model version — the two will not reconcile, and the mismatch surfaces every morning as an unexplained break. Getting one set of Greeks, computed once and consumed consistently, is more than half the engineering.

The risk ladder shows the book's shape

A sensitivities report — the risk ladder — lays out a book's Greeks so a trader and a risk manager can read its shape at a glance. A single net number tells you the size of the bet; the ladder tells you its structure. The distinction matters enormously, because two books with the same net exposure can carry completely different risks depending on how that exposure is distributed across the curve. The ladder makes the distribution visible.

The reference book for this module is the Kestrel book: a EUR interest-rate swaps and swaptions desk, used illustratively throughout. Its daily sensitivities report, at the close on 15 July, reads as follows. All figures are illustrative.

Line Value (EUR) Position Note
DV01 2Y -4,200 payer / short short the front end
DV01 5Y +11,800 receiver / long long the belly
DV01 10Y +18,500 receiver / long long 10s
DV01 30Y -6,100 payer / short short the long end
NET DV01 +20,000 — P&L per +1bp parallel shift
Gamma +2,300 per bp — DV01 change per 1bp move
Vega -85,000 per vol pt — swaptions, per 1 vol point
Theta +1,900 per day — carry plus roll-down

The DV01 ladder breaks the interest-rate risk into tenor buckets, and the shape it reveals is the point. Read down the buckets: the book is short the two-year, long the five- and ten-year belly, short the thirty-year. It is, in the trader's phrase, long the belly and short the wings. The net DV01 is +20,000 — meaning the book makes twenty thousand euros for a one-basis-point parallel rise in the whole curve — but that single number hides the structure entirely. A book could reach the same +20,000 net by being long a single tenor, and it would behave nothing like this one.

Net DV01 measures parallel-shift risk; the buckets measure curve risk, which a net number cannot show. If the entire curve moves up by one basis point, the Kestrel book makes +20,000 — that is the parallel risk. But curves rarely move in parallel. If the belly rallies while the wings sell off — a curve twist — the bucketed exposures move in opposite directions and the net number is silent about it. The ladder is what lets a risk manager see, and limit, the shape of the bet rather than only its size.

Gamma warns that the DV01 will not stay put as rates move. The +2,300 per basis point of gamma means that for each basis point the market moves, the book's DV01 changes by roughly that amount — so the +20,000 net is a snapshot, accurate now, drifting as the market travels. Vega, at −85,000 per vol point, is the swaptions exposure: the book is short volatility and loses eighty-five thousand euros for each one-point rise in implied vol. Theta, at +1,900 per day, is what the passage of time earns the book — carry and roll-down, the reward for holding the position quietly.

Risk limits sit on each line of the ladder, not only on the total. There is a limit on the two-year bucket, on the ten-year bucket, on net DV01, on vega — because a book can sit comfortably within a net-DV01 limit while running a dangerous curve or vol position. Limiting only the total would leave exactly the structural risks the ladder was built to reveal uncontrolled. The framework that governs these limits is a Part 2 subject; the ladder is where they land.

Figure 2 The Kestrel book's DV01 ladder — the curve risk a single net number hides (illustrative).

P&L is a control before it is a score

Profit and loss is the money a book made or lost — and, produced independently, it is a daily control on whether the book is understood at all. Everyone sees P&L as the scoreboard, and it is. But its deeper use is as a check: an independently computed P&L that ties to the desk's own tells you the book is understood end to end, and a P&L that does not tie tells you something is wrong before it becomes expensive. To use it that way you first have to know what it is made of.

P&L comes in four kinds

Realised P&L

Realised P&L is the cash actually made or lost on trades that have closed and on coupons that have paid. It is money that has moved — a position sold at a profit, a coupon received. There is no ambiguity about a realised number; it is settled fact. But it is only part of the day, because most of a live book's value is in positions still open.

Unrealised (mark-to-market) P&L

Unrealised P&L, also called mark-to-market or MTM P&L, is the change in value of open positions as their marks move. It is not yet cash — nothing has settled — but it is entirely real: it is the profit or loss you would crystallise if you closed everything at today's prices. For most trading books this is the largest and most volatile piece of the daily number, and because it depends on marks, it is only as trustworthy as the marks behind it.

New-deal (sales) P&L

New-deal P&L is the day-one margin captured when a trade is done away from fair value. When a desk sells a client a swap slightly above its model fair value, the difference is booked as profit on day one — the margin the desk earned for making the price and warehousing the risk. It is genuine value created the moment the trade is struck, distinct from any subsequent market move, and it is why a busy sales day can show a profit even when markets went against the book.

Carry and funding

Carry and funding is the cost or benefit of holding and financing the position over time. A position earns or pays carry as time passes — the coupon it accrues against the cost of funding it — independent of whether any market variable moved. It is the quiet, mechanical component of a day, closely related to theta, and it is the part of P&L that is most predictable in advance.

Total daily P&L is the sum of these four, and — this is the control — it is produced independently by product control, not by the desk. Product control is the finance-side function that recomputes the official P&L from marks and cash flows, separately from the trader's own running total. The independence is the whole point: two engines arriving at the same number is evidence the book is understood, and that evidence only means something if the second engine is not the trader's.

Why daily reconciliation is a control, not a formality

Reconciling the desk's P&L to product control's independent P&L every single day catches problems while they are small and cheap. The reconciliation is not bureaucracy; it is the earliest and most reliable warning system a trading floor has. Four things ride on it.

First, it confirms the book is understood. When the independently computed P&L ties to the desk's, it is strong evidence that positions, marks and models all agree across two separate systems. A break — a difference the two cannot explain — means an error somewhere, and the reconciliation is what surfaces it.

Second, it catches booking errors early. A trade booked with the wrong sign, a missing leg, a fat-fingered notional or a stale mark shows up the next morning as an unexplained P&L break. Without the daily tie-out, such an error can sit undetected for days, growing as the position moves, until it is a large and embarrassing loss rather than a same-day correction.

Third, it validates the risk. The P&L predicted from the Greeks should match the actual P&L. If the book made far more or less than its Greeks say it should have, the Greeks are suspect — a sensitivity is wrong, a model is misbehaving — and that matters, because those same Greeks drive the limits and the capital. The next section makes this validation explicit.

Fourth, it underpins trust and, now, capital. Regulators require P&L attribution to tie to risk — under FRTB this is formalised as the P&L-attribution test, which Part 2 covers — so daily reconciliation is no longer only good practice. It is a capital issue: a desk whose P&L and risk do not reconcile can be forced onto a more punitive capital regime. Reconciliation has become directly, expensively consequential.

P&L-explain decomposes the day into its causes

P&L-explain — P&L attribution — breaks the day's total P&L into the specific causes that produced it, following one identity: new + market + carry + residual = total. The point is to show not just what the book made but why, euro by euro. New-trade margin, moves in each market factor, the carry of holding the position, and whatever is left over — the residual — must add up to the actual P&L. When they do, the day is understood. When the residual is large, the day is not, and the residual tells you where to look.

Here is the Kestrel book's P&L-explain for 15 July, in euros. All figures are illustrative.

Component P&L (EUR) Driver
New trades +145,000 day-one margin
Delta / rates -380,000 10Y sold off 2bp
Gamma +22,000 convexity gain
Vega / vol +64,000 vol down 0.8pt
Carry / theta +19,000 time decay plus funding
Explained -130,000 sum of the above
Actual (MTM) -134,500 product control
Residual -4,500 unexplained (3.4%)

Read the market lines as sensitivity times move, exactly the core equation. The ten-year sold off two basis points; against the book's long ten-year DV01 that cost the delta line −380,000, the dominant move of the day. Gamma added +22,000 as convexity worked in the book's favour on the size of the move. Vega contributed +64,000 because implied vol fell 0.8 of a point and the book was short vol — a fall in vol is a gain for a short-vega book. Carry and theta earned +19,000 for the passage of time. New trades brought in +145,000 of day-one margin. Sum them and the explained P&L is −130,000: that is what the Greeks say the day was worth.

The explained number is what the risk predicts; the actual is what product control computes from marks. Product control, revaluing the book independently from the day's marks, arrives at an actual mark-to-market P&L of −134,500. The two are computed by different engines from different starting points — one from Greeks and assumed moves, one from a full revaluation on marks — and they land 4,500 euros apart.

The residual is the tell, and its size is the verdict on the day. A small residual — here −4,500, some 3.4% of the actual — means the risk and the books agree: the Greeks explain almost the entire move, and the small gap is a benign effect rather than an error. In the Kestrel case it is a curve-interpolation effect, a minor artefact of how the curve is built between quoted points, not a mistake. A large residual is the opposite signal: it means a mismatch, and the usual suspects are a bad Greek that mis-predicts the move, a stale curve that mis-values the book, or an un-booked trade whose P&L appears in the actual but not the explain. The residual does not tell you the answer, but it tells you loudly that there is a question.

Figure 3 The Kestrel book's daily P&L explain: new + market + carry + residual = total (illustrative).

Chasing the residual to zero is the daily job. Every morning the explain is produced, the residual is measured against a tolerance, and if it is outside tolerance someone investigates until it is understood — a wrong mark corrected, a missing trade booked, a Greek re-examined. A book whose residual is reliably small is a book that is understood and trusted. A book whose residual is large or erratic is a book with a problem that has not yet been found, and finding it is the work.

The risk factors are what the book is exposed to

Market risk is organised by risk-factor class, and each class comes with its own curves, marks and Greeks. A risk factor is a market variable that a book's value depends on. Grouping them into classes — rates, equity, FX, credit spreads, volatility, commodity — is how banks, and regulators, keep the exposure tractable, because each class needs its own market data, its own curve construction and its own sensitivities.

Interest rates

Interest-rate risk — yield curves, swap rates and the bases between them — is the largest risk factor for most banks. It is measured by DV01 and bucketed by tenor, exactly as the Kestrel ladder showed. Any instrument with a fixed cash-flow schedule carries it, which is nearly everything on a bank's balance sheet.

Equity

Equity risk is exposure to single stocks and indices, carrying delta, gamma and vega on prices and dividends. It spans cash holdings and derivatives alike, and an equity-derivatives book runs the full set of Greeks on the price of each name and on the dividends it is expected to pay.

Foreign exchange

FX risk is exposure to currency pairs — spot delta and FX volatility. It runs through almost every cross-currency position in the bank, often as a by-product: a foreign-currency bond carries FX risk whether or not the desk wanted it, which is why FX exposure has to be swept up and managed even where it was never the point of the trade.

Credit spreads

Credit-spread risk is exposure to issuer and index spreads, measured by CS01 — the P&L for a one-basis-point move in spread. It lives in bonds, credit default swaps and credit derivatives. CS01 is the credit analogue of DV01: it isolates the risk that a name's or index's spread widens or tightens, separately from the underlying interest rate.

Volatility

Volatility is a risk factor in its own right, expressed through implied-vol surfaces across strikes and tenors. Every options book carries it via vega, and because vol moves independently of the underlying it cannot be folded into directional risk — it has to be marked, curve-built and limited on its own terms.

Commodity

Commodity risk is exposure to oil, gas, metals and power, usually with a term structure and its own quirks. Convenience yield, seasonality and physical-delivery constraints make commodity curves behave unlike financial ones, so the class carries idiosyncrasies — a summer power price, a storage cost — that the other factor classes do not.

A compact map of the classes shows how each ties to instruments, Greeks and data.

Risk-factor class Typical instruments Key Greek(s) Key market data
Interest rates Swaps, bonds, futures, swaptions DV01, gamma Yield curves, basis, swaption vols
Equity Cash equity, index futures, options Delta, gamma, vega Prices, dividends, equity vol surface
FX Spot, forwards, FX options Spot delta, FX vega Spot rates, forward points, FX vols
Credit spreads Bonds, CDS, index credit CS01 Issuer and index spread curves
Volatility Options across all classes Vega, volga Implied-vol surfaces (strike, tenor)
Commodity Futures, swaps, options Delta, vega Term-structure curves, seasonality

The taxonomy matters because risk, P&L and capital are all organised along these lines — and because a single trade can touch several at once. FRTB literally defines risk-factor buckets along this structure, so the classes are not just a convenience but the shape the capital calculation takes. And a structured trade — a cross-currency swaption, say — can carry rates, FX and volatility risk simultaneously, which is exactly why aggregating risk across factors is hard: the same trade lands in several classes, and the systems have to add it up consistently without double-counting or dropping a piece.

Mark-to-market and valuation are where the numbers come from

Every risk and P&L number rests on a valuation, and a valuation rests on marks, curves and the vol surface. Before you can compute a Greek or a day's P&L you have to value the position, and valuation is a chain: market data feeds curves and surfaces, which feed a pricing model, which produces a value and its sensitivities. Change any link and everything downstream changes with it.

The valuation chain has three links

Valuing a position takes three things, in order. First, market data — the marks: observed prices, rates and quotes for the inputs, snapped from the market at the close. Second, curves and surfaces: discount and forward curves built from those marks, and the volatility surface for options. Third, a pricing model: the calculation that turns the inputs into a present value and the Greeks. Feed a mark into a curve, feed the curve into the model, read off the value and the sensitivities. Change any input — a single mark, one curve point, a vol — and the value, the P&L and the risk all move together. Valuation is the one shared foundation the whole edifice stands on.

Valuation gets hard in four familiar places

Liquid versus illiquid marks

Liquid instruments have observable prices; illiquid ones must be interpolated or modelled, and they carry the most valuation risk. A liquid benchmark bond has a screen price you can trust. An off-the-run position, a bespoke structure or a long-dated option may trade rarely or never, so its mark is inferred — and an inferred mark is a judgement, which is where valuation error and disputes concentrate.

Curve construction

Building a consistent, arbitrage-free curve from scattered quotes is a genuine quant and data task. The market quotes a handful of points; the curve has to fill in everything between them smoothly and without allowing free money. Small errors in construction ripple everywhere, because every instrument priced off the curve inherits them — the Kestrel residual's benign curve-interpolation effect is exactly this, seen small.

The vol surface

Implied volatility varies by strike and by tenor — the smile and skew — so the vol surface is a two-dimensional object that has to be built and marked. It is central to any options book, because every option is priced off a point on the surface, and the surface itself moves and has to be kept arbitrage-free. Marking it well is a specialist discipline.

Valuation adjustments (XVA)

Fair value includes adjustments for credit, funding and capital — the XVA family — layered on top of the base valuation. Credit valuation adjustment, funding valuation adjustment and their relatives account for the fact that a derivative's true value depends on counterparty credit, funding cost and capital consumed. It is a whole discipline in its own right, sitting on top of the pricing model rather than inside it.

In Europe, how positions are valued — especially illiquid ones — is governed by accounting and prudential rules, and supervisors expect the process to be independent and documented. European accounting uses IFRS 13 fair value, which sets how positions are marked. On top of that, the CRR's prudent-valuation regime requires additional valuation adjustments — AVAs — that hold extra prudence against valuation uncertainty, reducing capital resources where marks are hard to trust. Supervisors expect an independent, well-documented valuation and marking process, which is exactly why the control described next is a formal, regulated function rather than an optional check.

Independent Price Verification controls the marks

Because the desk's own marks drive its P&L and its bonus, someone independent has to check them — that is Independent Price Verification, IPV. IPV is the periodic, independent check that the marks used to value the book are correct, performed by product control or a valuations function, never by the desk. It exists to remove a conflict of interest that would otherwise sit at the heart of the P&L.

The conflict is direct: a trader's P&L, and therefore their bonus, depends on the marks they use. Left unchecked, there is a standing incentive to mark optimistically — to nudge an illiquid position a little richer and book the difference as profit. No accusation of bad faith is needed; the incentive alone is enough to make unverified marks untrustworthy. IPV removes the incentive by taking the verification out of the hands of the person who benefits from it.

IPV works by comparing each mark to independent sources and booking a reserve where the desk's mark cannot be verified. The mechanism is comparison: line up the desk's mark against a price sourced from outside the desk, and where they agree within tolerance the mark passes; where they do not, the mark is challenged, adjusted, or a valuation reserve is held against the uncertainty. The reserve reduces the reported P&L, which is the point — it prices the doubt.

Independent sources

The benchmark comes from consensus services, broker quotes and exchange prices. Consensus pricing services — Totem is the best-known — pool marks from many banks and hand back an independent average against which each contributor is measured. Broker quotes and exchange closing prices supply further independent references. These are the yardsticks the desk's marks are held to.

Thresholds and tolerances

Each mark is tested against a threshold, and only those outside tolerance trigger action. A mark within an agreed tolerance of the independent source passes without fuss; a mark outside it triggers a challenge, an adjustment or a reserve. Setting the thresholds sensibly is its own craft — too tight and every mark is a dispute, too loose and the control means nothing.

Valuation reserves and AVAs

Where marks are genuinely uncertain, a prudent reserve is held, and under the CRR that flows through to capital. A valuation reserve reduces reported P&L to reflect the doubt; the CRR's additional valuation adjustments push the same prudence into the capital calculation, reducing capital resources. Uncertainty is thus not waved away — it is quantified and paid for.

A data-intensive control

IPV is, in practice, a data-management problem: broad, clean external price data mapped onto internal positions. To check thousands of marks you need thousands of independent prices, correctly matched to the right internal instruments, refreshed on schedule. Sourcing that data, cleaning it and mapping it is a serious and continuous task — the control is only as good as the data feeding it.

The daily risk and P&L cycle is a nightly pipeline

The numbers the desk trades on each morning are produced by a tightly choreographed process that runs from the close to the open — seven steps, every night. Nothing on the risk report or the P&L appears by magic; it is manufactured overnight by a sequence of steps that must complete, in order and on time, before the desk comes in. Walk the seven.

Step one — capture and close. At end of day, every trade is captured and booked, positions are frozen at the official close, and market data is snapped from the market. This is the starting line: an accurate, complete, frozen picture of what the book holds and what the market did. A trade missed here is a break tomorrow.

Step two — revalue and compute risk. An overnight batch revalues the whole book on today's curves and computes the Greeks and the risk numbers. This is the pricing engine run at scale across every position — the heaviest compute of the night.

Step three — produce P&L. Product control computes the official daily P&L independently, from marks and cash flows, using its own engine rather than the desk's. This is the second, independent view whose whole value is that it was produced separately.

Step four — P&L-explain and reconcile. The day's P&L is attributed to new, market and carry, compared to the risk-predicted P&L, and the residual is chased. This is where the two engines meet, where the explain identity is checked and where any break starts its investigation.

Step five — risk versus limits. Market risk checks the Greeks and the aggregate measures against the limits and escalates any breach for action. The risk is not just measured but tested against the pre-agreed boundaries, and a breach starts a defined process.

Step six — sign-off and distribute. P&L and risk are signed off and published to the desk, to management, and into the feed for regulatory reporting. Sign-off is the formal statement that the numbers are trusted enough to act on.

Step seven — the desk acts. The trader starts the day knowing the risk and the P&L, and hedges or adjusts accordingly. The whole pipeline exists to put a correct, reconciled, signed-off picture in front of the desk at the open.

The binding constraint is the deadline: all of this must finish before the desk opens, often within a few hours of the close. When the pipeline breaks or runs late, the desk trades blind — without a current view of its risk or a reconciled P&L — which is the situation the whole apparatus exists to prevent. That is why robustness, timeliness and data quality are the entire engineering challenge, not incidental to it.

Figure 4 The daily risk & P&L cycle — a nightly data pipeline from close to desk-open.

Strip away the finance vocabulary and the daily cycle is a nightly data pipeline with a hard SLA: snap, revalue, compute, reconcile, sign off, before the desk opens. Every property that matters to a data engineer matters here, and for the same reasons.

Robustness, because a batch that fails at 3am leaves the desk trading blind at 8am. Timeliness, because the deadline is immovable — the market opens whether or not the pipeline finished. Data quality, because a stale curve or a missing trade at snap time corrupts the risk, the P&L and the capital all at once, silently, and surfaces only as a break the next morning. The quant methodology is largely settled; making this pipeline run correctly and on time, every single night, is where the work and the value are.

A first worked example: the Kestrel book's day

Bring the whole part together on one day of one book — the Kestrel book, from the close to the reconciled, signed-off P&L. Everything above has been machinery; here is the machinery running.

In Europe, this exact reconciliation is now formalised under FRTB as the P&L-attribution test, with capital consequences. The PLA test compares product control's hypothetical P&L — a full independent revaluation — against risk's risk-theoretical P&L, the P&L predicted by the risk model's own factors. It is the Kestrel residual, made into a formal, statistical, capital-linked test. A desk whose residuals are too large or too volatile can be forced off its internal models and onto the more punitive standardised capital approach. The mechanics belong to Part 2; the point to carry forward is that the daily discipline of chasing the residual to zero is no longer only good housekeeping — in Europe it is directly linked to how much capital the bank must hold.

A trading desk makes money in three ways, and each one is a risk position

A book's P&L is not something separate from its risk report — it is the receipt for it. Every euro the Kestrel book expects to earn this quarter is compensation for holding some exposure that could just as easily lose money. Strip away the risk and the expected return goes with it; there is no such thing as riskless revenue on a trading desk. Understanding a book means asking, for every line of expected P&L, which risk is being paid for and whether the price is fair.

Bid-offer capture is paid for warehousing inventory

Market-making means quoting a two-way price and earning the gap between what you buy at and what you sell at. A client wants to pay fixed on a ten-year EUR swap; the desk quotes a bid and an offer around its view of fair value and captures the spread when the trade prints. That spread is not free money — the moment the trade is done, the desk owns the other side of it, and that position sits on the book until an offsetting trade, a hedge, or another client comes along to lay it off. Between execution and offload, the position is exposed to every basis-point move in the curve like any other position in the risk ladder. The wider the spread, the better the desk is paid for that waiting period; the longer the waiting period, the more market risk it is carrying to earn it. Bid-offer income, in other words, is a fee for warehousing inventory, not a reward that exists apart from the risk of holding it.

Carry and roll-down are the reward for holding a position through time

Carry is what a position earns simply from the passage of time, over and above the cost of funding it. A receiver swap collects fixed coupons that, in normal market conditions, exceed the cost of financing the position; roll-down adds to that, because as a bond or a swap ages, it slides down an upward-sloping curve toward a lower yield, and price moves inversely with yield. Together, carry and roll-down are the return earned by doing nothing except staying in the trade — no client flow, no repricing view, just time. This is precisely what Kestrel's theta line is telling the desk: a positive theta of roughly +1,900 per day (illustrative) says the book is structured to earn money as each day passes, all else equal, in the same way a landlord collects rent for owning a building rather than trading it. That daily carry is real income, but it is only earned because the desk is exposed to the risk that the curve does not behave as the roll-down assumes — a bear-steepening move can erase weeks of carry in an afternoon.

Warehousing a directional view is the most visible risk-for-return trade

The most transparent way a desk makes money is simply holding a view and being right about it. If the desk believes belly yields — the five-to-seven-year part of the curve — are going to fall, it can position DV01 (the euro change in value per one-basis-point move) to benefit from that fall and wait for the market to agree. Kestrel's net long-the-belly position, roughly +20,000 DV01 (illustrative), is exactly this kind of bet: a deliberate, sized exposure that pays off if belly yields drop and costs money if they rise. There is nothing hidden about this trade — it shows up directly on the risk ladder — but it is worth naming separately from market-making and carry because it is a pure opinion on where rates go, dressed up in no other justification. The desk is paid, when it is paid, precisely because someone had to be willing to hold that view and the risk that comes with it.

None of this means bigger bets are better bets. The job of risk measurement is not to stop the desk from taking these positions but to size them so that the expected reward justifies the capital they tie up — a question Part 2 picks up directly when it turns to capital consumption.

Economic P&L and accounting P&L answer different questions

Economic P&L asks what the position is worth today, in full; accounting P&L asks what the firm is currently allowed to recognise. The economic, or fair-value, number captures the entire change in the position's value on a mark-to-market basis — including any day-one margin earned at trade inception — net of the prudent valuation reserves and adjustments the desk holds against uncertainty. It is the desk's honest answer to "did we create value today," and it is the number risk managers and traders should be steering by hour to hour.

Economic (fair-value) P&L is the desk's real scorecard

Economic P&L moves whenever fair value moves, full stop. It does not wait for a rule to say the gain is allowed to be booked. This is the number that matches the risk ladder: if DV01 times the yield move explains the P&L, it is this economic number that the explain is reconciling against.

Accounting P&L follows recognition rules that can lag economic reality

Financial reporting under IFRS does not always let a desk book value the day it is created. The clearest case is day-one profit on instruments priced using unobservable, level-three inputs — where the model, not the market, sets the price. Accounting standards typically require that profit to be deferred and released into reported earnings over the life of the trade, or as inputs become observable, rather than recognised immediately, because the market has not yet validated the price. A desk can therefore be economically richer on day one than its accounts show.

The two views can diverge by a material amount, especially on structured or long-dated business, and neither one is "wrong" — they are answering different questions for different audiences. Reconciling the desk's real-time economic P&L with the firm's reported accounting P&L, and explaining the gap, is a core part of what product control does every close.

  1. What is market risk, in one sentence?

  2. What is the one equation underneath every market-risk measure?

  3. Why is market risk described as taken "on purpose," unlike credit or operational risk?

  4. What is the difference between net and bucketed DV01, and why does it matter?

  5. What does the residual in a P&L-explain tell you?

  6. What is the difference between realised and unrealised (MTM) P&L?

  7. What does IPV control, and why is it needed?

  8. What is the binding constraint on the daily risk and P&L cycle?

1

  1. What equation underlies every market-risk measure?

  2. Why is market risk taken “on purpose”?

  3. What is market risk?

2Measuring and Controlling Risk

From single-position Greeks to book-wide measures, limits, and the FRTB capital regime.

Value at Risk (VaR) puts a single number on a whole book

Every Greek in Part 1 tells you the risk to one variable; VaR tries to answer a harder question: how much could the whole book lose, from everything at once, tomorrow? Take the Kestrel book's dozens of DV01 buckets, its gamma, its vega, its cross-currency basis, and compress them into one euro figure that a trader, a risk manager and a board member can all read the same way — that compression is VaR's entire value proposition. A 1-day 99% VaR of EUR 10 million (illustrative) means that, based on the model and the data behind it, losses should stay under EUR 10 million on 99 days out of 100 — and on the remaining day, the loss could be worse, potentially much worse. The number says nothing about that hundredth day beyond the fact that it exists.

What the VaR statement actually promises

A VaR number is a probability statement, not a ceiling. It does not say "the book cannot lose more than EUR 10 million" — it says the loss is expected to exceed EUR 10 million on roughly one day in a hundred, under the assumptions baked into the model. Confidence level and horizon are both choices: a bank might quote 99% or 95%, one day or ten days, and the resulting numbers are not comparable across those choices without careful scaling. For the Kestrel book, a 1-day 99% VaR aggregates the DV01 ladder's curve risk, the gamma convexity and the vega exposure into one figure — but it is silent on which of those three drove the number, which is why the Greeks from Part 1 remain the tool for managing the book, and VaR the tool for summarising it.

Three methods compute that same statement, and they disagree with each other more often than a newcomer expects.

Historical simulation replays the market's own history

Take the last one to two years of real daily market moves and apply each of them, one by one, to today's book. Every historical day becomes a hypothetical P&L for the Kestrel book as it stands right now — reprice the whole book under yesterday's rate move, then the day before's, and so on across the window. Sort the resulting distribution of hypothetical P&Ls and read off the loss at the 99th percentile. It is intuitive, assumes nothing about the shape of the distribution, and is the most widely used method in practice — but it is only as good as the window: if the last two years were unusually calm, historical VaR understates the risk of a genuinely new kind of shock.

Parametric VaR trades realism for speed

Assume market moves follow a known statistical distribution — typically normal — and use volatilities and correlations to compute VaR with a formula rather than a simulation. Given each risk factor's volatility and the correlation matrix between them, the parametric (variance-covariance) method produces a VaR number almost instantly, which made it attractive when compute was expensive. The cost is the normality assumption itself: real market returns have fatter tails than a normal distribution predicts, so parametric VaR understates the likelihood of a large move — precisely the move that matters most.

Monte Carlo simulation generates its own scenarios

Instead of replaying history or assuming a formula, simulate thousands of random scenarios from a model of how the risk factors could move, then revalue the book under each one and read the percentile. Monte Carlo is the most flexible of the three: it can incorporate fat tails, non-linear payoffs and complex correlation structures the other two methods struggle with, particularly useful for a book like Kestrel's that carries swaptions and genuine gamma. That flexibility is paid for in compute — full revaluation across thousands of paths is expensive, and the result is only as good as the model generating the scenarios.

Method How it works Weakness
Historical simulation Replays 1-2 years of real market moves on today's book; reads the loss at the 99th percentile Assumes the recent past resembles the future; blind to shocks the window never saw
Parametric (variance-covariance) Assumes moves are normally distributed; computes VaR analytically from volatilities and correlations Understates fat-tail risk; breaks down for non-linear, optioned books
Monte Carlo Simulates thousands of scenarios from a model, revalues the book under each, reads the percentile Compute-heavy; only as good as the model generating the scenarios

Where VaR runs out of road

VaR's flaws are not implementation bugs — they are structural, and they are why the regulatory answer changed. First, VaR says nothing about how bad the tail day actually is: EUR 10 million and EUR 100 million are both "worse than VaR," and the number cannot tell them apart. Second, historical VaR assumes the future resembles the recent past, an assumption that fails exactly when it matters most — at the onset of a genuinely new crisis. Third, VaR tends to understate risk through a calm period and then jump sharply once volatility re-enters the estimation window, a lagging-indicator problem. Fourth, VaR is not always sub-additive: combining two books can, in some circumstances, produce a higher measured VaR than the sum of the two books' individual VaRs, contradicting the intuition that diversification should reduce risk. These four flaws, exposed brutally in 2008, drove the move to expected shortfall.

VaR was the heart of the old Basel internal-models market-risk charge, and European banks ran approved VaR models under the CRR (the EU's implementation of Basel) for decades. FRTB replaces regulatory VaR with expected shortfall for internal models, a direct response to VaR's tail-blindness. VaR has not disappeared from the toolkit: it remains widely used internally for limits and day-to-day management, precisely because it is fast, intuitive and comparable across desks — it has simply been retired as the number that determines regulatory capital.

Expected shortfall looks past the threshold into the tail

If VaR marks where the bad tail begins, expected shortfall (ES) measures how bad the tail actually is once you are in it. VaR asks what is the loss I will not exceed 99% of the time — a threshold question. ES asks a different question: given that I am already in the worst outcomes, what do I lose on average — a severity question. Formally, ES is the average of all losses beyond the VaR point, so it uses the same tail of the same distribution but summarises it completely rather than truncating it at one point. Two books can carry an identical VaR and very different ES: one might have a tail that drops off quickly, the other a long, fat tail of catastrophic outcomes, and VaR alone cannot distinguish them.

Why regulators chose ES over VaR for capital

FRTB's internal-models capital charge is built on a 97.5% expected shortfall, deliberately calibrated to a period of significant market stress rather than to recent, possibly calm, history. The choice of 97.5% rather than 99% is not arbitrary generosity — ES at 97.5% and VaR at 99% tend to produce broadly comparable capital levels for many distributions, while ES captures the tail shape that VaR discards. Calibrating to a stressed historical window, rather than the trailing one-to-two years typically used for VaR, means capital does not collapse to nothing during a calm market — exactly when a bank can least afford to be under-capitalised for the next shock.

It sees the tail

ES captures the severity of extreme losses, which is exactly the information VaR discards and 2008 punished. Where VaR treats every day beyond its threshold identically — "worse than X" — ES asks how much worse, on average, across all of those days, giving risk managers and regulators a genuine sense of catastrophic-scenario magnitude rather than just its frequency.

It is coherent, meaning sub-additive

ES respects the basic logic of diversification in a way VaR cannot guarantee. Combining two books under ES will never produce a higher measured risk than the sum of their individual ES figures — a mathematical property called sub-additivity, and one of the technical reasons regulators consider ES a more coherent risk measure than VaR.

Stressed calibration keeps capital honest through the cycle

Anchoring ES to a period of significant historical stress means the measure does not fall asleep in calm markets. A VaR model calibrated purely on the last two quiet years can drift toward reassuring but misleading numbers; a stressed-period ES resists that drift, holding capital closer to what a genuine crisis would demand even while the current market feels benign.

Liquidity horizons scale the measure to how fast you can actually exit

ES under FRTB is not a single one-day number — it is scaled by how long it would realistically take to exit or hedge each risk factor. A liquid interest-rate swap might carry a horizon of ten days; an illiquid credit or exotic-vol exposure might carry sixty or more. The longer the horizon, the larger the effective loss window and the larger the capital charge — a structural penalty for holding risk that cannot be exited quickly, a lesson the market learned when supposedly liquid instruments froze in 2008.

Figure 5 VaR marks where the tail begins; expected shortfall measures how bad it is once you are in it.

Stress testing asks what statistics alone cannot answer

VaR and ES are both calibrated to data that has already happened, and both are structurally blind to the shock that has never yet occurred. Stress testing is the deliberate complement: instead of asking how bad is a bad day, statistically, it asks what happens to this book if a specific, severe, hand-picked scenario unfolds, with no requirement that the scenario have ever been observed. For a book like Kestrel, long the belly and short vega, a plausible stress scenario is not abstract — a sharp curve-flattening shock combined with a vol spike would hit exactly the exposures the DV01 ladder and vega line are carrying.

Historical scenarios replay a real crisis on today's book

Take a genuine historical episode — Lehman's collapse in 2008, the COVID crash of March 2020, the 2022 global rates shock — and apply its actual market moves to the book as it stands today. These scenarios carry an immediate credibility hypothetical shocks lack: nobody can argue the scenario is implausible, because it already happened. Their limit mirrors their strength — a historical replay only tests shocks the market has already produced, and the next crisis is rarely identical to the last one.

Hypothetical scenarios probe shocks that have not happened yet

Design a plausible but unobserved shock — a sovereign default, a 200bp parallel jump in rates, an equity crash paired with a volatility spike — and run it through the same revaluation engine. Hypothetical scenarios let risk managers test combinations and magnitudes history has not yet delivered, filling a gap historical scenarios structurally cannot close, at the cost of requiring genuine judgment about what plausible-but-severe means this year.

Reverse stress testing starts from the wound and works backwards

Instead of picking a scenario and measuring the loss, start from a loss large enough to break the desk and ask what combination of market moves would cause it. This inversion is deliberately uncomfortable — it forces the risk team to imagine failure first and hunt for its causes, rather than testing a comfortable set of moves and hoping nothing worse exists. For a book carrying concentrated gamma or vega, reverse stress testing often surfaces the genuinely hidden vulnerability nobody thought to test directly.

Sensitivity and factor shocks isolate one driver at a time

Shock a single risk factor — rates up 50bp, equities down 10% — while holding everything else constant, to see which single move hurts the book most. This is the simplest and most diagnostic form of stress test: because only one variable moves, the resulting loss maps directly onto a single Greek or bucket, making it easy to prioritise which exposure to hedge first.

Regulatory stress tests are standardised across the industry

Supervisors — the EBA and the ECB in Europe — design and run stress scenarios across the whole banking sector, using the same scenario for every participating bank. Because every bank runs the identical scenario, results are directly comparable across the industry, making regulatory stress tests both a supervisory tool and a public disclosure exercise with real consequences for a bank's standing.

Correlation and concentration stress tests the crisis pattern itself

Test what happens when correlations that normally hold suddenly break, or when positions meant to be diversified all move together at once. This is the textbook crisis pattern — hedges that decouple from what they were meant to hedge — and it is precisely the failure mode a book's normal correlation assumptions, embedded quietly in VaR and the Greeks, cannot see coming.

Why stress complements, and cannot replace, the statistical measures

VaR and ES are calibrated to observed data, however cleverly resampled or scaled — they are structurally unable to imagine an event data has never contained. Stress tests are the deliberate answer: they probe the tail directly, through judgment and scenario design, not statistical extrapolation. The two are complementary rather than competing — a framework that only ran VaR and ES would be blind to novel shocks, while one that only ran stress tests would lose the daily, statistically grounded discipline VaR and ES provide.

Running a stress scenario is not a reporting exercise — it requires the ability to re-price the entire book, instantly, under an arbitrary and possibly extreme set of market moves.

That means the same revaluation engine that produces the daily Greeks and P&L must also be callable with scenario data instead of today's market snap — identical pricing logic, driven by a different, hand-picked set of inputs.

For a technologist, this is the real engineering requirement behind stress testing: not a separate system, but a revaluation engine flexible and fast enough to be driven by whatever scenario the risk team designs, on demand.

Limits turn a risk number into a boundary the desk cannot cross

Measuring risk is only half the job — without a pre-agreed boundary, a risk number is just a fact, not a control. A limit converts a Greek, a VaR figure or a stress-scenario loss into a line the desk cannot cross without escalation, and the layering of limits matters as much as their individual levels, because different layers catch different failure modes.

Greek and sensitivity limits are the most granular control

Caps on DV01, vega and gamma, set both per bucket and net, are the finest-grained and most real-time layer of the limit framework. For the Kestrel book this means a limit on the 5Y and 10Y DV01 buckets individually, not just the net +20,000 DV01 figure — a book can sit within its net limit while carrying dangerously concentrated curve risk in a single tenor, exactly what the bucketed ladder from Part 1 was built to reveal.

VaR and ES limits cap the aggregate statistical risk

A ceiling on the book's overall VaR or ES catches the risk that Greek limits, taken one at a time, can miss — the combined effect of many small exposures moving together. A desk can sit within every individual Greek limit and still run an aggregate VaR the risk appetite does not tolerate, once correlations between factors are accounted for — why this layer sits above the granular Greek limits rather than replacing them.

Stress and scenario limits cap the loss under a defined shock

Rather than limiting a statistical measure, this layer directly caps the loss the book would suffer under a named stress scenario — a Lehman replay, a hypothetical rates shock. It is the limit-framework equivalent of stress testing itself: a boundary on how bad a specific, severe, imaginable day is allowed to be, independent of what VaR or ES currently say about a normal day.

Notional and concentration limits are the simplest backstop

Plain caps on the size of a position and on exposure to any single name or issuer, set without reference to any risk model at all. Their simplicity is their strength — they catch outsized bets even if a pricing or risk model is wrong, stale or simply hasn't been built for a new product yet.

Stop-loss and P&L limits trigger action after the fact

A cumulative-loss trigger that forces risk reduction once a desk's P&L has fallen by a defined amount, regardless of what the forward-looking measures say. Stop-loss limits are the framework's last line of defence — they do not prevent a loss starting, but force a hard stop once losses accumulate past a threshold the desk cannot negotiate around.

How the limit hierarchy actually functions day to day

Limits are not set once and forgotten — they cascade from the board's stated risk appetite down through business lines, desks and individual traders into a coherent hierarchy. Utilisation against every limit is monitored continuously, typically through dashboards showing each desk exactly how much headroom it has left on every Greek, VaR figure and stress scenario, refreshed as trading and market moves change the picture. A breach triggers a defined process, not an informal conversation — notification, explanation, remediation, or a formally approved temporary excess, all documented. Limits are set and policed by market risk, a function independent of the desk that takes the risk, because a limit framework only works if the people enforcing it have no stake in the desk's P&L.

FRTB rebuilt how market-risk capital gets calculated

FRTB — the Fundamental Review of the Trading Book — is the Basel overhaul of how banks calculate the capital they must hold against trading losses, replacing pre-2008 rules that had been patched repeatedly rather than redesigned. Its goals were explicit: a clearer line between trading and banking books, more risk-sensitive charges, better tail capture via expected shortfall, and a credible standardised fall-back regardless of model sophistication. FRTB matters beyond compliance because capital is expensive — every euro held against market risk cannot be deployed elsewhere, so how a bank computes it is a direct lever on return on equity.

Two roads to the same capital number

Every bank has, at minimum, one route to FRTB capital; some banks, on some desks, run two in parallel. The Standardised Approach (SA) is a prescribed formula combining a sensitivities-based method, a default risk charge and a residual add-on for exotic risks the formula cannot otherwise capture — mandatory for every bank, and the fall-back for any desk without an approved internal model. The Internal Models Approach (IMA) lets a bank use its own expected-shortfall model, approved desk by desk, potentially delivering meaningfully lower capital than SA — but far more demanding to build, validate and keep compliant.

The Standardised Approach is mandatory, formulaic, and universal

Every trading desk at every bank must compute SA capital, even desks that also run IMA. SA combines three components — the sensitivities-based method (delta, vega and curvature run through prescribed buckets, risk weights and correlations), a default risk charge for jump-to-default exposure, and a residual risk add-on catching exotic payoffs the other two miss. Because it uses regulator-defined risk weights and correlations rather than the bank's own model, SA is transparent and comparable across institutions — the credible fall-back FRTB's designers wanted, one that does not depend on trusting any individual bank's modelling choices.

The Internal Models Approach rewards sophistication with lower capital, and demands more for it

IMA lets a desk use the bank's own expected-shortfall model rather than the prescribed SA formula, and the reward for doing it well can be substantially lower capital. The price of that reward is continuous proof: a desk must pass the risk-factor eligibility test on its modelled factors, clear backtesting thresholds, and pass the P&L-attribution test — all covered later in this Part — on an ongoing basis, not just at initial approval. Approval is granted, and can be withdrawn, desk by desk rather than bank-wide — a strong-performing rates desk and a struggling exotics desk at the same bank can sit on entirely different regimes. Even an IMA desk must still compute SA in parallel, because SA functions as a permanent floor and a comparability benchmark against which supervisors can sanity-check the internal model's output.

Figure 6 The FRTB capital regime: the standardised and internal-model approaches, and the IMA gates a desk must pass.

In the EU, FRTB is implemented through the CRR, as amended by CRR3, supervised by the ECB and national authorities — the ACPR in France. The EU has repeatedly re-timed the binding own-funds application of the FRTB capital rules, pushing back the go-live for IMA and SA capital requirements to manage alignment with other jurisdictions, even while the reporting elements of FRTB-SA phased in earlier and have applied for some time. Treat any specific go-live date as provisional and check the current CRR3 implementation timeline.

The trading book and the banking book are two different risk universes

The same instrument can sit in either book, and which one it sits in decides how its risk is measured, how it is capitalised, and how its P&L is recognised. The trading book holds positions taken with intent to trade — to profit from short-term price moves, or to hedge other trading positions — marked to market daily and carrying the full market-risk capital charge this Part has been building up: VaR or ES, stress, and FRTB SA or IMA. The banking book holds positions taken for banking purposes and generally held to maturity — loans, deposits, the treasury portfolio — accrual-accounted, carrying credit risk and interest-rate-risk-in-the-banking-book (IRRBB) treatment instead. Two identical cash flows, booked in different books, can produce very different capital numbers and P&L profiles, which is precisely why the boundary is an actively policed control point, not an administrative filing choice.

Arbitrage risk is the reason the boundary gets policed so hard

Before FRTB, firms could and did move positions between the trading book and the banking book to capture whichever capital treatment was cheaper — regulators call this regulatory capital arbitrage. FRTB was designed explicitly to close that door, tightening the rules on what can sit where and removing much of the discretion that made boundary-shifting profitable.

Assignment rules now prescribe, rather than suggest, where a position belongs

FRTB sets out a presumptive list of instrument types that must sit in the trading book, and requires documented justification for any exception. A bank can no longer simply declare intent and book a position accordingly — deviating from the presumptive list requires evidence a supervisor can review.

Re-designation between books is heavily restricted

Moving a position from one book to the other, once booked, is no longer a routine operational choice — it requires approval, and any capital benefit that would result is explicitly disallowed. This closes the loophole where a position could be reclassified once its risk profile made one book's treatment more attractive.

The boundary is, underneath the regulation, a data and tagging problem

Every single position needs a correct, auditable book designation, captured at booking and preserved through its life. That designation flows into which capital calculation, which P&L treatment and which supervisory scrutiny applies — a classification and lineage task that sits with the systems that book and track the trade, not just the desk that originated it.

The Sensitivities-Based Method turns Part 1's Greeks into a capital number

SBM is the engine room of the Standardised Approach — it takes exactly the delta, vega and curvature sensitivities Part 1 introduced and runs them through a regulator-defined formula to produce capital. Delta captures first-order risk to rates, credit spreads, equity, FX and commodity — the same directional exposure the Kestrel book's DV01 ladder measures, reframed for regulatory buckets. Vega captures risk to implied volatility, the book's -85,000 vega figure from Part 1. Curvature captures the extra risk from large moves — the regulatory cousin of gamma, via a prescribed shock-and-reprice formula rather than a simple second derivative.

How the sensitivities become a capital figure

For each sensitivity type, the bank computes the Greeks under the regulator's exact definitions, sorts them into prescribed buckets, multiplies by regulatory risk weights, and aggregates using regulator-set correlations within and across buckets. None of these steps is optional or bank-specific — bucket structure, risk weights and correlations are all fixed by the regulation itself, which is exactly what makes SBM standardised rather than a bank's own risk view running under a different name.

Sensitivities must match the regulator's definitions exactly, not the desk's

A trader's own Greek, computed for hedging, is not automatically the number the regulator wants for SBM — bucket boundaries, shock sizes and factor definitions are all prescribed. A desk's risk system typically needs a distinct calculation path, or at least a reconciliation layer, to produce SBM-compliant sensitivities alongside the risk-management Greeks the desk actually trades against.

The bucket and correlation aggregation is a heavy, precise calculation

SBM aggregates sensitivities across dozens of prescribed buckets, under several correlation scenarios specified by the regulation, not just one. Getting this right means implementing the full regulatory formula precisely, across every position and bucket, every single day.

Consistency between SA and the desk's own Greeks is a governance signal

The SA sensitivities computed for capital should, in principle, reconcile back to the risk-management Greeks the desk actually monitors and hedges against. A persistent divergence is not just an operational nuisance — it is a red flag that something in the pricing or booking chain is inconsistent, since SA capital and desk risk management are supposed to describe the same underlying exposure.

Every capital figure must be reproducible and auditable

A regulator must be able to trace any SBM capital number back to the position, the market data and the calculation that produced it. That lineage requirement is a data and systems discipline as much as a quant one, and it recurs throughout FRTB.

RFET decides which risk factors a bank is allowed to model

IMA does not let a bank claim expected-shortfall benefits for every risk factor it can dream up a model for — a factor only counts inside the internal model if it passes the Risk-Factor Eligibility Test. RFET is a data test, not a modelling test, underneath its regulatory name: to use its ES model on a given risk factor, a bank must demonstrate that the factor has enough real, verifiable price observations behind it — a minimum count per year, and no unreasonably long gap between them. The logic is simple: a model calibrated on a factor nobody can actually observe trading is not really a model of anything, and FRTB refuses to grant capital relief for pretending otherwise.

What happens to the factors that fail

A factor that passes RFET goes into the bank's expected-shortfall model and benefits from IMA's more favourable capital treatment; a factor that fails becomes a Non-Modellable Risk Factor, or NMRF. NMRFs are pulled out of the ES model entirely and capitalised separately via a deliberately punitive stress-scenario charge (stressed expected shortfall, SES) rather than the favourable diversified ES treatment modellable factors receive.

NMRF capital is punitive by design, not by accident

The stress-scenario charge applied to non-modellable factors is calibrated conservatively, and for many desks it ends up a disproportionately large slice of total IMA capital relative to the factor's actual size in the book. That asymmetry is intentional — FRTB's designers wanted a real capital incentive to source the data needed to make factors modellable, not a soft nudge.

Passing RFET is, in practice, a data-sourcing hunt

Clearing the eligibility threshold means finding enough real, verifiable price observations for every risk factor the desk wants inside its ES model — from the bank's own trades, vendor data, and industry data pools. For an illiquid tenor bucket or an exotic vol point, this can mean sourcing observations that simply do not exist inside the bank's own trading history — a genuine data-procurement exercise.

Data pooling is the industry's collective answer

Because no single bank trades every instrument often enough to clear RFET alone, industry data-pooling services aggregate anonymised price observations across multiple banks, letting more factors clear the threshold than any one bank's own data could support. Participation has become a meaningful, ongoing input into how much of a desk's risk can stay inside IMA.

NMRF exposure shapes whether IMA is worth running on a desk at all

If a desk's risk factors are heavily non-modellable, the punitive SES charge can erode or even reverse IMA's capital advantage over the standardised approach. How much of the desk's risk clears RFET, and how expensive the residual NMRF charge is, is exactly what determines whether pursuing IMA approval is worth the ongoing cost of running it.

RFET is not a one-off certification — a factor's eligibility can be reassessed as data availability changes, and the observations supporting it need continuous sourcing, cleaning and mapping to the exact risk factor the model uses.

That makes RFET, underneath the regulatory language, a live data-management pipeline: identifying which observations count as real, deduplicating them across internal and pooled sources, and keeping a defensible, timestamped record for every factor a desk wants to keep inside its model.

Lose that discipline for a quarter, and factors that were modellable can quietly become NMRFs — with the capital consequence landing before anyone notices the data gap that caused it.

Backtesting is where the model meets reality, every single day

A risk model earns the right to be used for capital only by proving, continuously, that its predictions match what actually happens — backtesting is the mechanism that proves it. Backtesting compares the model's predicted risk against realised outcomes: for a 99% 1-day VaR, losses should exceed the VaR figure on roughly one day in a hundred, so counting actual exceptions over a rolling year tells you directly whether the model is calibrated correctly.

The traffic-light zones and what each one means

Regulators sort backtesting outcomes into three zones based on the number of exceptions observed over the testing window. In the green zone, exceptions occur at roughly the expected rate and the model requires no capital penalty. In the amber zone, exceptions occur more often than predicted, and a capital multiplier is applied, proportional to how far the model has drifted from reality. In the red zone, exceptions occur far more often than the 99% confidence level implies, and the model's approval itself is put at risk.

Green zone means the model's calibration is holding

Roughly one exception in a hundred trading days is exactly what a well-calibrated 99% VaR model should produce, and a green-zone result confirms the model is doing what it claims.

Amber zone triggers a capital multiplier, not an automatic failure

More exceptions than expected, but not so many as to invalidate the model outright, moves a desk into the amber zone, where a multiplier increases the capital charge to compensate for the model's demonstrated understatement of risk.

Red zone puts the model's approval itself on the line

Too many exceptions signals the model is materially miscalibrated, and a red-zone result can trigger a formal review that ends in the desk losing IMA approval for that risk factor or that model entirely, forcing a fallback to the standardised approach.

Too few exceptions is also a warning sign, not a reassurance — a model producing almost none may be systematically over-conservative, which sounds safer but actually means it is not a faithful description of the book's real risk.

Model validation and governance keep the challenge continuous

Backtesting is the statistical half of model oversight; validation and governance are the institutional half, and FRTB requires both. Independent validation means a team entirely separate from the model's builders challenges every risk and pricing model before it goes live, and periodically thereafter — the same independence principle that separates the desk from market risk. Model risk management treats models as a distinct, inventoried risk type: every model is catalogued, rated for materiality, assigned an owner, and scheduled for periodic review. Ongoing monitoring — backtesting, benchmarking, performance metrics — runs continuously, because a model accurate a year ago is not guaranteed to be accurate today, and every assumption and limitation must be documented for a regulator or internal audit to challenge.

In the EU, model validation and governance expectations are set by the ECB, and the Targeted Review of Internal Models (TRIM) exercise scrutinised banks' internal models and the data behind them across the euro area. Ongoing supervision under the Single Supervisory Mechanism keeps models under continuous review, not one-time approval. Backtesting exceptions and validation findings feed directly into whether a desk keeps IMA approval and what capital multiplier it carries.

The P&L Attribution test is FRTB's toughest gate

PLA compares two independently produced daily P&Ls for every desk on IMA, and asks a simple but unforgiving question: does the risk model's view of the day match what the book actually made or lost? Hypothetical P&L (HPL) is product control's actual revaluation of the book — the same rigorous, independent P&L process Part 1 introduced. Risk-theoretical P&L (RTPL) is the P&L the risk model predicts, computed purely from the risk model's own risk factors and sensitivities — what the model thinks the day was worth, using only the inputs it considers relevant.

What close tracking, and divergence, actually mean

If HPL and RTPL track each other closely over time, the risk model is capturing everything that actually moves the book's value, and can be trusted for capital purposes. If the two diverge — if RTPL consistently misses moves that show up in HPL — the risk model is missing something real, and the desk fails the test. This is precisely the P&L-explain residual from Part 1, generalised and formalised: where Part 1 asked whether new-plus-market-plus-carry explained the day's actual P&L within tolerance, PLA asks the identical question at the level of the entire risk model, with a hard regulatory consequence attached.

Aligning two engines that were often built separately is the real work

The risk system and the P&L system frequently grew up as separate builds, owned by separate teams, running on different data feeds and sometimes different position snapshots. PLA forces those two engines onto genuinely consistent data and risk factors, because any inconsistency shows up directly as a divergence between RTPL and HPL — even when the difference is a data mismatch between the two systems, not a model flaw.

The traffic-light outcome has real capital teeth

A green PLA result keeps the desk on IMA; amber triggers warnings and closer scrutiny; a red result forces the desk onto the standardised approach and the correspondingly higher capital. Because SA capital can be materially more expensive than IMA capital for a well-modelled desk, a red PLA outcome is a direct, immediate cost, not a distant compliance risk.

The pass/fail line is drawn with precise statistical metrics

PLA agreement is not judged by eye — it is measured with prescribed statistical tests, including Spearman correlation between HPL and RTPL and a Kolmogorov-Smirnov test comparing their distributions, each against defined thresholds. These metrics leave little room for a desk to argue past a borderline result; the test is mechanical and repeatable, precisely so it cannot be gamed.

Passing PLA is, underneath the statistics, a data-lineage problem

Consistent positions, curves and risk factors across the risk system and the finance system are what actually make PLA pass — the statistics only measure the consequence of that consistency, or its absence. A desk chasing a PLA failure is very often chasing a data-lineage break between two systems, not a genuine flaw in the pricing model itself.

Risk data aggregation is the standard that made data a board issue

BCBS 239 is the Basel standard on risk data aggregation and risk reporting, written directly in response to a specific 2008 failure: many banks discovered, mid-crisis, that they could not quickly and accurately add up their own risk exposures. That inability to answer how much risk the group actually had, in a live crisis, is exactly what BCBS 239 was designed to prevent from recurring, and it sets out principles across four dimensions.

Accuracy and integrity mean the numbers must be correct and reconciled

Risk data feeding into VaR, ES, stress and FRTB capital has to be correct at the source and reconciled across the systems it passes through, not merely plausible-looking by the time it reaches a dashboard.

Completeness means capturing all material risk across the group

Every material risk, in every legal entity and every business line, has to be captured in the aggregated view — a partial picture that happens to look complete is arguably more dangerous than an obviously incomplete one.

Timeliness means producing the aggregated view fast, including during a crisis

The standard explicitly requires that aggregation remain fast enough to be useful in a crisis, not just during calm, routine reporting cycles — the exact capability that was missing in 2008 and the reason the standard exists at all.

Adaptability means answering new, unscripted risk questions on demand

Beyond the routine reports, the underlying data architecture has to support ad-hoc questions a crisis or a new regulatory request might throw up — total exposure to one counterparty across every desk and entity cannot wait for a quarterly reporting cycle to be built.

Why the standard is hard, and effectively never fully met

BCBS 239 is, underneath its principles, a governance standard about data — ownership, lineage, quality and architecture — and every one of those things is genuinely difficult at the scale of a global bank. Risk data lives across many fragmented source systems, organised differently by asset class, region and legal entity, and consolidating them into a single, trustworthy group-wide view is a permanent engineering programme, not a project with a finish line. Different systems frequently use different identifiers, hierarchies and taxonomies for what is, economically, the same position, so a single trade can look different depending on which system reports it — reconciling those differences consumes most of a risk-technology team's effort. Regulators expect to trace any reported number back to its source, with a named owner and a documented quality control at each step. Every measure covered in this Part — VaR, ES, stress, limits, FRTB capital — sits on top of this aggregation layer, so poor aggregation undermines the reliability of the entire risk stack built above it.

BCBS 239 rarely gets discussed as exciting work, and that is exactly why it matters so much.

VaR, ES and FRTB capital are the numbers everyone talks about; BCBS 239 is the plumbing that determines whether those numbers can be trusted at all — accurate positions, correctly mapped identifiers, a documented owner for every data feed, and the ability to produce a consolidated view fast enough to matter in a live crisis.

A bank can have a beautifully engineered expected-shortfall model and still fail its own risk management if the aggregation layer beneath it is fragmented, undocumented or too slow — precisely the failure BCBS 239 exists to close off.

Intraday and real-time risk push beyond the overnight batch

Everything this Part has described so far — VaR, ES, stress, the SBM, PLA, backtesting — has, by default, run on the overnight batch: positions frozen at the close, revalued overnight, signed off before the desk opens. That cycle produces a trustworthy, fully reconciled number once a day, but a single overnight snapshot can be badly out of date by lunchtime in a volatile market, and a growing part of risk management now happens intraday.

Pre-trade risk checks are the sharpest expression of this pressure. Before a large or automated trade executes, a system needs to check its impact against the desk's limits in real time — a genuine latency problem, since the check has to complete inside the trade's execution window or the limit control arrives too late to catch the breach.

Real-time Greeks and P&L extend the same logic to ongoing risk management. Desks increasingly want live Greeks and live P&L as the market moves through the day, not just the official end-of-day number, because a book's risk profile genuinely changes intraday.

None of this is free — full revaluation of a large book is computationally expensive, and running it continuously requires either substantial grid or cloud compute capacity or genuinely clever incremental revaluation methods that avoid recomputing everything from scratch.

Consistency between the fast intraday numbers and the official overnight numbers is a real design tension, not a solved problem. Two speeds, one truth is the aspiration — an intraday estimate a trader can act on immediately, and an overnight, fully reconciled number product control can sign off on — but keeping those views from quietly diverging is a genuine, ongoing engineering challenge.

The direction of travel is unmistakable: from a single overnight batch toward risk available on demand, intraday and increasingly close to real time. This is one of the biggest technology drivers in the risk discipline — grid and cloud compute, incremental revaluation, streaming architectures — and a natural bridge into the systems, vendors and transformation themes of Part 3.

Work the numbers once — VaR and expected shortfall by hand

Every VaR and ES definition in Part 2 sounds abstract until you rank a hundred numbers and count. The formulas are simple; the confusion usually comes from skipping the arithmetic and jumping straight to the concept. Do it once, on a tiny book, and the 99% VaR, the 97.5% ES, and the gap between them stop being abstractions and start being two numbers you could reproduce on a napkin — a sort, a count into the tail, and either a lookup or an average.

Real trading books scale this up, not change it. A desk-level VaR model runs on 250 to 500 days of history and thousands of risk factors instead of ten numbers on a page, but the arithmetic underneath — rank the outcomes, count into the tail, read off or average the threshold — is exactly what you just did.

The gap between the two numbers is the whole reason FRTB moved the internal-models charge from VaR to expected shortfall. VaR only ever reports the boundary of the tail; two books can share an identical 99% VaR while one has a mild tail just past that boundary and the other has a catastrophic one, and VaR alone cannot tell them apart. In the Kestrel example above, if the worst day had been −40.0m instead of −12.0m, the 99% VaR would still read about EUR 9.5m — completely unchanged, because VaR only looks at the second-worst day, never the worst one. ES would not stay still: the average of the worst two-to-three days would jump from about 9.8m toward something far larger, because ES is built from the tail values themselves rather than from the point where the tail begins. That sensitivity to the depth of the tail, not just its starting edge, is precisely why regulators consider ES the more conservative and more informative of the two measures — and why it, not VaR, sits at the center of the internal-models capital charge under FRTB. The window length matters too: a shorter history reacts faster to a recent stress period but is noisier day to day, while a longer history is more stable but slower to reflect a genuine regime change — trading desks live with that trade-off every time the model is recalibrated.

The standardised charge, computed on one bucket then two

The SBM formula looks forbidding because it is written for forty buckets at once; written for two, it is one line of arithmetic. Strip away the summation signs and what is left is a square root of squares plus a correlation cross-term, applied over and over across buckets, risk classes and correlation scenarios.

The two-bucket example is the whole SBM machine; a real book just runs it three more times and keeps the worst answer. The regulator does not hand a bank a single correlation number and leave it there — it hands three versions of every prescribed correlation, one scaled up toward 1.0 (the high scenario), one at the base value used above (medium), and one scaled down (low), each with its own floor so correlations never fall below zero or flip sign. A bank runs the full aggregation — within-bucket first, using each risk factor's own correlation to its neighbours inside the bucket, then across buckets using the cross-bucket correlation shown above — once under each of the three scenarios, and the capital charge for that risk class is whichever of the three scenarios produces the largest number. That "keep the worst" rule is deliberate: it stops a bank from being rewarded with a low charge purely because it happened to land on a correlation assumption that flattered its particular mix of positions. Scale the two-bucket example up to the Kestrel book's actual footprint — EUR delta, EUR vega, curvature across several tenor buckets, all under the GIRR (general interest rate risk) class — and the number of terms in the square root grows into the hundreds, but every one of those terms is still either a squared sensitivity or a correlation cross-term exactly like the 1.80 computed above.

Backtesting, counted over a year

The traffic-light system is a counting exercise, not a judgment call. A 99% 1-day VaR run over 250 trading days should, by construction, be exceeded on roughly 2.5 days — the loss beats the VaR number about one day in a hundred. Basel turns that expectation into zones:

Exceptions in the year Basel zone Consequence (illustrative)
0-4 Green Model accepted; no add-on
5-9 Amber Capital multiplier rises with each exception
10 or more Red Model's approval is in question

Three exceptions in a year is unremarkable. It sits comfortably in green, close to the 2.5 the model predicts. Seven exceptions moves the desk into amber, where the multiplier used to scale VaR into capital climbs step by step with each additional breach — the bank pays more capital for the same model rather than losing it outright. Eleven exceptions is red: the regulator's working assumption becomes that the model itself is broken, and the desk risks losing permission to use internal models altogether, falling back to the standardised charge for that book.

Zero exceptions over a whole year looks safe and is not. A model calibrated correctly should be breached roughly 2.5 times; a model with no breaches at all is very likely overstating risk, holding a VaR number well above what the book's actual behaviour justifies, and burning capital the desk did not need to hold. Backtesting punishes both directions — a model that is too loose and a model that is too tight are both mis-specified, just in opposite ways.

Exceptions are not all the same, and the pattern matters more than the count. A single breach on a day when a genuine shock moved every book is the tail arriving on schedule — a 99% model is meant to be beaten about once in a hundred days. A cluster of breaches tied to one risk factor is different: it points to a gap in the model itself, a curve segment or a correlation the model no longer captures. The multiplier bites at once — the internal-models charge is typically VaR times a factor that starts near 1.5 in the green zone and steps up through amber — so a desk drifting from three exceptions to seven pays more capital every day until backtesting improves, rather than losing its model outright.

Clean, hypothetical and actual P&L are three different numbers

The word "P&L" hides three distinct numbers, and the tests of a risk model turn on telling them apart. Backtesting and the PLA test deliberately compare like with like, which means stripping out everything that is not a pure market move.

Hypothetical (clean) P&L revalues yesterday's book on today's market

Take exactly the positions held at yesterday's close and revalue them on today's market data — no new trades, no fees, no intraday activity. That like-for-like number isolates the effect of the market moving, which is why it is the P&L backtested against VaR and the hypothetical P&L (HPL) compared in the PLA test.

Actual P&L is the full accounting number

Actual P&L includes everything the day really produced: new deals and their day-one margin, fees and commissions, reserve movements, intraday trading. It is the number the desk is paid on, but for model-testing it is contaminated by activity that has nothing to do with whether the risk model is right.

Risk-theoretical P&L is what the model predicts

Risk-theoretical P&L (RTPL) is the P&L rebuilt from the risk model's own factors and sensitivities — the RTPL in the PLA test. If the model captures the book faithfully, RTPL tracks the hypothetical P&L closely; where they drift apart, the model is missing something real.

P&L type What it includes What it is used for
Hypothetical / clean Yesterday's positions revalued on today's data only Backtesting; the HPL in the PLA test
Actual Everything: new trades, fees, reserves, intraday The desk's real result; not used to test the model
Risk-theoretical (RTPL) P&L rebuilt from the risk model's own factors The RTPL in the PLA test

Using the clean numbers is the whole point. A model test contaminated by fees and new trades would measure the desk's activity, not the model's accuracy, so the tests that decide capital insist on the like-for-like P&L that moves only when the market moves.

From expected shortfall to the capital number

The internal-models charge is a sum of parts, and the expected-shortfall number is only the largest of them. The modelled ES is aggregated across liquidity horizons and risk classes into the internally-modelled capital charge (the IMCC); to it is added a stressed add-on for the non-modellable risk factors that failed RFET (the SES), and a default risk charge (the DRC) for jump-to-default exposure, with a multiplier set by backtesting. The standardised charge assembles differently but in the same additive spirit: the sensitivities-based method (SBM), plus the DRC, plus a residual risk add-on (the RRAO) for exotic payoffs the other two miss.

Figure 7 The FRTB capital charge assembled: standardised (SBM + DRC + RRAO) and internal-model (stressed ES + NMRF + DRC) stacks, with the output floor (illustrative, simplified).

The output floor stops internal models from straying too far below the standardised number. Under CRR3, a bank's internal-model capital cannot fall below 72.5% of what the standardised approach would produce for the same book, phased in over several years. The floor means SA is never merely a fall-back a desk can forget once it has model approval — it is a permanent benchmark that sets a hard minimum on the whole firm's capital, which is exactly why banks run both engines side by side forever.

Which trade is using my risk? Incremental and component VaR

Every morning, a risk manager and a trader ask the same question from opposite ends of the desk: of everything sitting on the book, what is actually eating the risk budget? A single book-level VaR number cannot answer it. VaR is not additive — positions diversify against each other, so the risk of a position measured on its own is almost never the risk it actually adds to the book. Kestrel's book is the illustration: a 10Y receiver leg and a 30Y payer leg, each with real standalone risk, sit together and partly cancel. Knowing the book's total VaR tells you nothing about which of the two legs — or which desk, or which currency bucket — is responsible for it, and it tells a limit-setter nothing about who should be asked to trim size if the book's total VaR breaches its ceiling. Three related measures exist precisely to open that black box, and they answer three different questions: what would a specific trade change, how fast is risk changing right now for a given exposure, and how should a fixed total be fairly split among everyone who holds a piece of it.

Incremental VaR answers "what does this trade cost me in risk?"

Incremental VaR is the change in the book's total VaR from adding or removing one specific position. It is computed by running VaR twice — once on the book as it stands, once on the book with the candidate trade added (or stripped out) — and taking the difference. This is the number a trader actually wants before doing a hedge, unwinding a position, or pricing a new deal: not "how risky is this trade in isolation" but "how much will my limit usage move if I do this." A trade that looks large on a standalone basis can have a small, even negative, incremental VaR if it happens to offset existing exposure; a trade that looks small in isolation can have an outsized incremental VaR if it happens to load onto a risk factor the book is already leaning on heavily. Incremental VaR is the honest, decision-relevant answer to a decision-relevant question, which is why it sits in pre-trade checks and limit-utilization screens, not just in end-of-day reporting.

Marginal VaR is the rate, not the change

Marginal VaR asks a narrower, more technical question: if I increase a position by a small amount, how fast does VaR move? Where incremental VaR measures the effect of a full trade — potentially a large, discrete change — marginal VaR is a per-unit sensitivity, the slope of VaR with respect to position size at the current point on the book. It is useful for a different purpose: ranking exposures by how VaR-expensive they are at the margin, right now, given everything else on the book. Two positions can have similar current VaR contributions but very different marginal VaR — one might already be near a point of diminishing diversification benefit, where adding one more unit costs relatively little risk, while another is on the steep part of the curve, where the next unit is disproportionately expensive. Desks that actively manage risk budget, rather than just report it, watch marginal VaR to decide where the next unit of size is cheap and where it is not.

Component VaR is the one that has to add up

Component VaR — sometimes called contribution VaR — takes the total book VaR and splits it into pieces that sum exactly back to that total, one piece per desk, book, or risk factor. This additivity is the entire point: incremental and marginal VaR are excellent for asking "what if" about one trade, but neither is designed to allocate a fixed total across everyone who holds a piece of it. Component VaR is. It is the number that feeds limit monitoring by desk, capital allocation across business lines, and any internal charge-back of risk cost to the people who created it — because unlike standalone VaR, the components genuinely exhaust the total with nothing left over and nothing double-counted. Add every desk's component VaR together and the answer is the book's total VaR, exactly, which is the property that makes it defensible to hand each desk a number and say "this is your share."

The subtlety worth sitting with is this: because of diversification, a position's standalone VaR — its risk measured alone, ignoring everything else on the book — is almost always larger than its component VaR, its true contribution once its interaction with the rest of the book is accounted for. A well-hedged position can even carry a negative component VaR: adding it to the book reduces total VaR, because it offsets risk that was already there. Charging that position for its standalone risk, as if it sat on its own book, would overstate what it actually costs the firm.

From risk to return — a desk is judged on the capital it consumes

A desk is not judged on P&L alone but on P&L relative to the capital its risk forces the firm to hold. Two traders can each report the same profit for the year and still have delivered very different value to the bank, because the capital consumed to make that profit is not the same, and capital — as Part 2 established through FRTB's SA and IMA charges — is expensive. It sits on the balance sheet earning nothing for shareholders until it is called on to absorb a loss; every euro of it has an opportunity cost. A book that produces a given P&L on less capital is, quite literally, worth more to the firm than a book that produces the same P&L on more.

RAROC makes capital the denominator of performance

Risk-adjusted return on capital, RAROC, is the ratio that forces this comparison to be made explicitly: P&L divided by the capital the position or desk consumes. It reframes the question from "how much did we make" to "how much did we make for every unit of the scarce resource we used to make it." A trading floor is not actually short of trading ideas; it is short of capital to deploy against them, and RAROC is how the firm decides which desks deserve more of it and which do not. Raise a desk's RAROC and it can grow within its existing capital allocation, or free capital for someone else's better idea; let it drift down and the desk is, in effect, asking the firm to hold expensive capital against increasingly mediocre returns.

Desk Annual P&L (illustrative) Capital consumed (illustrative) Return on capital
Desk A EUR 50m EUR 200m 25%
Desk B EUR 50m EUR 400m 12.5%

Same P&L, same headline number on the trading floor screen, and yet Desk A is twice as valuable to the firm per euro of capital tied up. This is exactly why regulatory capital calculations are not a back-office afterthought bolted onto trading — they set the denominator that determines which desks look good and which look mediocre, regardless of how much money crossed the P&L line. A desk that runs a cleaner FRTB profile, with fewer NMRFs and a tighter IMA/SA gap, is quietly buying itself a better return on capital without changing a single trade.

Economic capital and regulatory capital are related, but they are not the same number

Regulatory capital follows the rulebook — FRTB's standardised approach or internal models approach, computed exactly as the regime prescribes, because that is what supervisors require the bank to hold. Economic capital is the firm's own internal estimate of the buffer its risk actually needs, built from the bank's own view of tail risk, correlations across business lines, and the confidence level management wants to defend against. The two usually move together but rarely match exactly: a regime can be conservative in places the firm's own models find benign, or lenient in places the firm's own risk view finds dangerous. A desk that is well-behaved under the regulatory number can still be flagged internally if its economic capital consumption looks high, and the reverse happens too. Desks are watched on both, because regulatory capital determines what the bank must hold and economic capital determines what management believes it should hold — and a firm that only manages to the regulatory number is trusting the rulebook to be a perfect proxy for real risk, which it is not designed to be.

Reserves and valuation adjustments move P&L today, not just protect it tomorrow

A prudent reserve is not simply a conservative habit — it is a lever on reported profit. An additional valuation adjustment, a model reserve held because a pricing model is known to be imperfect in some regime, a bid-offer reserve reflecting the cost of exiting at the quoted mark rather than the mid — each reduces today's reported P&L in exchange for protection against tomorrow's mark being wrong. The trade cuts both ways: building a reserve lowers this period's profit, and releasing one raises it, with no new trade done. A desk under pressure to show a strong number has a structural incentive to see reserves released early or set thin, which is precisely why reserve policy is governed independently of the desk that benefits from it, with its own sign-off, periodic review, and audit trail back to the same data quality BCBS 239 demands.

  1. What does a 1-day 99% VaR of EUR 10m actually tell you, and what does it not tell you?

  2. Why did FRTB replace VaR with expected shortfall for internal models?

  3. What is reverse stress testing, and why is it useful?

  4. How does the risk-limit framework cascade through an organisation?

  5. What is the difference between the Standardised Approach and the Internal Models Approach under FRTB, and how is IMA approved?

  6. What is a Non-Modellable Risk Factor, and why does it matter so much?

  7. What does the P&L Attribution test compare, and what happens if a desk fails it?

  8. Why is BCBS 239 described as a foundation the rest of the risk stack sits on?

2

  1. What does a 1-day 99 % VaR of EUR 10m NOT tell you?

  2. Why did FRTB move to expected shortfall?

  3. What is reverse stress testing?

3The People, the Systems, and the Rules

Who runs risk & P&L, on what systems and vendors, under which rules.

Independence, Not Hierarchy, Is the Organising Principle

A trading book is only as trustworthy as the separation between the people who take its risk and the people who measure it. Parts 1 and 2 built the mathematics — the Greeks, P&L-explain, VaR, expected shortfall, and FRTB's SA and IMA engines. None of that mathematics protects anyone unless it is produced by someone who has no incentive to make the number look better than it is. Every bank organises risk & P&L around this single idea: the desk that owns the position must never also be the desk that certifies the risk or the profit it reports. Six groups of actors sit around the Kestrel book, the illustrative EUR rates swaps and swaptions desk carried through this module, and each has a distinct mandate.

Front-office desk risk

Traders and the risk/strategy staff embedded in the desk take, hedge and manage market risk in real time. They own the position — a swaption struck this morning, a swap unwound this afternoon — and they run the desk's own view of its Greeks and P&L intraday, because they are the ones who must act on it before the market moves again.

Independent market risk

A separate department, reporting up to the Chief Risk Officer rather than to the trading floor, computes risk independently of the desk. It sets and monitors limits, investigates breaches, and challenges the desk's own numbers with its own VaR, ES and stress runs, built from the same market data but calculated on a different, controlled engine.

Product control

Sitting on the finance side, product control produces the official daily P&L independently of the desk. It runs P&L-explain, owns the reconciliation between risk-explained and actual P&L, and books valuation reserves when a mark looks aggressive. The number product control signs is the number the bank reports; the desk's own P&L view is informative but not final.

Risk quants (methodology)

Quants build and own the pricing and risk models that everyone else consumes — the VaR and ES engines, the sensitivity libraries, the FRTB SA and IMA calculators from Part 2. They decide how risk is calculated, document the models for validation, and are the people who translate a regulatory text into a formula.

Risk IT / BA / data

This is the module's core audience: the people who build and run the engines, the market-data pipelines, the aggregation layer and the reporting that turns methodology into numbers that are correct, on time and reconciled, every single day without exception. A brilliant model that runs three hours late or silently drops a book is worse than no model at all.

Regulators & validators

The ECB and national supervisors, together with an independent internal-model-validation function inside the bank, approve models, review data quality and police the boundary between the trading book and the banking book. They are the last line of the same separation principle, checking that the checkers are doing their job properly.

Figure 8 The risk & P&L actor map — those who take the risk are separated from those who measure and challenge it.

Four Roles Carry Risk & P&L Day to Day

A technologist working in this module will spend most of their time with four of the six actors above. The table below sets out what each actually does, day to day, and where a build, a feed or a reconciliation touches their work.

Role Mandate Key outputs Where IT/BA/data plugs in
Market Risk Manager Measure and control the desk's market risk independently of the front office Limits, VaR/ES/stress reports, breach investigations, risk sign-off Consumes the independent risk engine's output; needs it correct and on time before the desk opens
Product Controller Produce and reconcile the official daily P&L Daily P&L, P&L-explain, reconciliation to the desk, valuation reserves Consumes the P&L system's attribution output; drives the residual-investigation tooling
Risk Quant Define how risk and P&L are calculated Pricing models, sensitivity and VaR methodology, FRTB SA/IMA specifications, validation documentation Hands specifications to engineering; reviews implementation against the spec
Risk IT / BA Make the numbers appear correctly, on time, every day Risk engine, market-data feeds, aggregation pipelines, reporting, reconciliation controls Owns the build end to end — this is the seat a technologist actually sits in
Market Risk Manager

The market risk manager is the CRO's eyes on the book, and their currency is the limit. They own the limit framework set against VaR, ES, sensitivities and stress scenarios from Part 2, and when a desk breaches one, it is the market risk manager who investigates it before the desk can trade further. They sign the daily risk report that goes up the chain — a number that, if the systems behind it are late or wrong, leaves the CRO blind on a day the market moves.

Product Controller

The product controller's signature is what turns a P&L number from an opinion into a fact. They run the P&L-explain waterfall from Part 1 every morning, chase whatever residual is left after new trades, market moves and carry are accounted for, and book a reserve if a mark looks stretched relative to independent price verification. The desk trades on the risk it sees intraday, but the number the bank reports, and the number that feeds capital, is the one product control certifies.

Risk Quant

A risk quant's real output is a specification someone else has to implement exactly. They decide how a Greek is computed, how VaR aggregates across risk factors, how the FRTB sensitivities-based method buckets and correlates delta, vega and curvature — and then they write that decision down clearly enough for both engineering to build it and for an independent validator to challenge it. A model that only the quant who wrote it understands has already failed the validation test that Part 2 and this Part both return to.

Risk IT / BA

Everything above depends on this role actually working, every night, at scale. Risk IT and business analysts build the engines that revalue the book, the feeds that bring in curves and prices, the aggregation pipeline that consolidates desks into one firm-wide view, and the controls that catch a bad feed before it reaches a report. Where the quant defines the twenty-line formula, risk IT builds and operates the system that runs it correctly across a hundred thousand trades before the market opens — this is where the discipline actually lives.

Four Layers Turn Positions Into Capital, Every Night

Positions do not arrive at risk & P&L by accident — they flow in from the trade and booking systems covered in Module 19, and everything downstream is one continuous overnight, and increasingly intraday, processing cycle. Four cooperating system layers turn a booked trade into a risk number, a P&L line and a capital charge.

Market-data platform

Curves, prices and volatility surfaces are snapped at end of day, or throughout the day, cleansed of stale or obviously wrong quotes, and published to every downstream system. This is the fuel: a swaption book cannot be revalued without a vol surface, and the daily cycle from Part 1 begins here.

Risk / pricing engine

The engine that revalues the book, computing the Greeks, VaR, ES, stress scenarios and the FRTB SA and IMA charges from Part 2. It is the single largest compute consumer in the chain, because it must reprice every position under both current markets and every scenario the desk and the regulator require.

P&L systems

Product control's own engine, separate from the risk engine, that produces the daily P&L and runs the P&L-explain attribution — splitting the day's move into new trades, market movement, carry and residual, exactly as taught in Part 1.

Risk-data aggregation

The layer that consolidates output across desks, risk factors and legal entities into one consistent view of the firm's risk. A single swaption book is simple; a global rates business spanning dozens of books is not, and this layer is where that complexity is absorbed.

Reporting & capital

Dashboards for the desk and for management, limit monitoring, and the regulatory reports — including the FRTB capital figures — that go to the supervisor. This is the visible end product, but it is only as good as the four layers feeding it.

Figure 9 The risk & P&L system landscape and its cross-cutting capabilities.

Why two engines, not one

Every bank of any size runs at least two engines that compute the same risk, deliberately. The front-office risk system — often the trading platform's own risk module, such as Murex MX.3 or Calypso — is used live by the desk to price and hedge intraday. A separate, independent risk and P&L engine, owned by risk and finance, recomputes the same positions against the same, or independently sourced, market data to challenge the desk's own view. The two are not supposed to agree by construction; they are supposed to agree by outcome, and when they do not, someone has to explain why. Reconciling the front-office and independent numbers on the Kestrel book — a vega difference here, a curve-mapping difference there — is not a one-off project. It is a permanent programme, staffed continuously, because new trade types, new curves and new model versions constantly reopen small gaps between the two.

The aggregation layer is the hard part

Any single desk's risk is a tractable engineering problem: one book, one set of curves, one overnight run. The genuinely hard problem is consolidation — taking hundreds of books, thousands of risk factors and multiple legal entities and turning them into one consistent, reconciled, firm-wide view of risk, fast enough that market risk can sign off before the desk opens. A currency mismatch, a double-counted hedge across two entities, or a risk factor mapped under two different names in two systems can each silently distort the firm-wide number without breaking any single desk's own report. This is precisely the problem BCBS 239, covered later in this Part, was written to regulate, and in most banks it consumes more engineering effort than the pricing models themselves — aggregation logic, reconciliation tooling and lineage tracking, not derivatives mathematics, is where the bulk of the risk-IT headcount sits.

One Trade Becomes Six Objects — and One Bad Input Cascades Through All of Them

Follow a single object end to end and the whole landscape from the previous section becomes concrete. Take one trade on the Kestrel book: a ten-year EUR payer swaption struck this morning.

Stage System Data object created
Booking Trade & booking system (Module 19) Trade record: book, instrument, notional, counterparty, curve mapping
Market-data snap Market-data platform Mark or curve point: a curve node, a vol-surface point
Revaluation Risk / pricing engine Present value and Greeks: delta, gamma, vega, theta
Risk aggregation Risk-data aggregation A DV01, a VaR or ES contribution, attributed to book and factor
Capital calculation Risk / pricing engine (FRTB module) An FRTB SA or IMA capital charge for that position
Regulatory reporting Reporting & capital A line in a regulatory return or a management report

That trade record turns into a mark, the mark turns into a present value and a set of Greeks, the Greeks turn into a DV01 and a VaR contribution, and the VaR contribution turns into an FRTB capital charge that ultimately becomes one line in a regulatory report the bank submits to its supervisor. Six objects, six systems, one underlying position.

The cascade principle is that an error never stays local. A stale mark on that one ten-year vol point does not just misprice one swaption — it corrupts the desk's P&L-explain residual, distorts the vega bucket in that day's VaR, and, if it survives long enough, feeds a wrong number into the FRTB capital calculation and the regulatory line that follows it. Risk, P&L and capital are not three independent checks on the same trade; they are three sequential consumers of the same six objects, so an error at the top of the chain reappears, disguised, at every stage below it.

Figure 10 Data lineage: one object from position to capital, and why an error at the top cascades all the way down.

Data model anatomy

The trade or position record sitting at the top of that chain carries a specific set of fields, and each one drives a different downstream calculation. The book and desk identifier routes the trade to the right risk-aggregation bucket and the right limit. The instrument type and ISIN, where one applies, determine which pricing model and which FRTB risk class apply. Tenor and notional set the size and shape of the sensitivity. Currency drives FX conversion and cross-currency basis risk. The counterparty field is what links this same trade record into the counterparty-credit-risk calculation covered later in this Part. The trading-book flag is the input to the trading/banking-book boundary test that FRTB explicitly requires. And the curve mapping — which discount curve, which forward curve, which vol-surface node this trade is priced off — is the field that most often breaks silently, because it is metadata rather than an economic term of the trade, and nobody notices it is wrong until the risk number looks strange.

Deduplication of that record is not a housekeeping detail; it is a mission-critical control. The same swaption can arrive at the risk-data aggregation layer twice — once from a live front-office feed and once from an end-of-day batch reconciliation extract — and if the two are not recognised as the same trade, the book's risk is silently doubled. Recognising that two records with slightly different timestamps, slightly different formatting, or routed through different feeds represent one economic trade is a genuinely hard matching problem, and getting it wrong in either direction, missing a duplicate or wrongly merging two distinct trades, corrupts the same cascade described above.

Without Consistent Identifiers, Aggregation Is Impossible

The aggregation layer's hardest problem, described above, is really an identification problem: two systems can only agree that they are looking at the same instrument, the same legal entity or the same risk factor if they name it the same way. A handful of identifier standards exist precisely to make that possible.

ISIN

The International Securities Identification Number names the security — a bond, an equity, a listed option — uniquely across the industry. It is the reference most systems use to agree on what instrument a trade actually is, but plenty of OTC derivatives, including much of the Kestrel book's swaptions, do not naturally carry one, which is exactly where the harder identification problems begin.

LEI

The Legal Entity Identifier names the counterparty or the booking entity, not the instrument. It is what lets counterparty-credit-risk and legal-entity-level aggregation, covered above in the aggregation layer, correctly attribute exposure to one legal entity rather than splitting it silently across two records that both refer to the same institution under slightly different names.

FIGI

The Financial Instrument Global Identifier is a vendor-neutral instrument identifier that maps across data providers, useful precisely where an ISIN does not exist or where a single instrument needs to be matched consistently across Bloomberg, Refinitiv and an in-house system without depending on any one vendor's proprietary code.

Beyond these three, a curve or risk-factor identifier is the least standardised and most consequential of all: the string a risk engine uses internally to mean "the ten-year EUR swap curve" or "the five-year EUR swaption vol surface, at-the-money." Two systems computing VaR for the same book must resolve to the same identifier for that risk factor, or their sensitivities will not reconcile even when every trade and every market input is otherwise identical — a mismatch that shows up as an unexplained VaR difference between the front-office and independent risk engines discussed earlier in this Part.

Getting Data In Is Mostly Ingestion and Reconciliation, Not Origination

Risk & P&L does not originate trades — it sits downstream of the trade and booking systems from Module 19, so its integration story is overwhelmingly about receiving, validating and reconciling data that was created somewhere else, not creating new data of its own.

Two feeds dominate. Position feeds carry booked trades in from the trade and booking systems: every new deal, amendment and cancellation on the Kestrel book has to arrive complete and on time, or the overnight revaluation is working from an incomplete book. Market-data feeds carry curves, prices and volatility surfaces in from the market-data platform described earlier, sourced ultimately from the vendors covered in the next section. Further upstream still, the front office's own execution flow often runs over the FIX protocol — a fast, machine-readable messaging standard that carries an order or an execution into the primary trading process the instant a trade is done, which is one of the reasons a position can, in principle, reach the front-office risk system intraday rather than only overnight. On the post-trade side, confirmation and settlement messaging has historically run over SWIFT's MT message formats and is progressively migrating to the richer, structured ISO 20022 standard — a shift that matters to risk & P&L because more structured, better-tagged post-trade messages make trade and counterparty data easier to validate and reconcile automatically further down the chain.

How that data actually moves follows one of three integration patterns.

Batch

The legacy default: positions and market data are extracted, transferred and loaded on a schedule, typically overnight. It is simple to reason about and easy to control, but its weakness is latency — a batch that runs late, or fails partway through, leaves the desk trading without a current risk view until it is fixed and rerun.

API

An on-demand call that returns data or triggers a calculation when something needs it, rather than waiting for the next scheduled batch. This is the modernisation step most banks are mid-way through: pricing a what-if scenario or pulling a single position's current Greeks no longer needs to wait for the next overnight cycle.

Event streaming

A Kafka-style architecture where a system publishes an event — a new trade booked, a curve updated — the moment it happens, and every downstream consumer picks it up in near real time. This is the pattern behind intraday and real-time risk, and its side effect is important: because every consumer reads the same event at the same time, the class of reconciliation break that comes from two systems processing the same batch file at slightly different times, or with slightly different logic, structurally shrinks.

The move from batch to streaming is not simply a speed upgrade — it changes what reconciliation even means. In a batch world, the front-office and independent risk engines each pull their own extract on their own schedule, and disagreements between them are partly just timing artefacts layered on top of genuine data or model differences. In an event-streamed world, both systems consume the identical event at the identical moment, so a remaining difference is far more likely to be a real methodology or model difference — the kind of difference risk teams actually want to see, rather than noise they have to filter out first.

The Vendor Landscape: Buy, Build, or Both

No bank builds every layer of the risk & P&L stack from scratch, and none buys every layer off the shelf either. The pattern across the industry is a mix, and which layer gets built in-house versus bought is itself a signal of where a bank believes its competitive advantage lies.

Murex (MX.3)

The dominant cross-asset trading, risk and P&L platform in the industry. Many banks run their front-office desk risk, their market-data management and even parts of their P&L natively inside MX.3, rather than stitching together separate systems for each function.

Calypso (Adenza/Nasdaq)

Murex's major rival, particularly strong in rates, FX and cross-asset trading, offering trading, risk and collateral management in a single stack. The platform has changed ownership more than once — Calypso became part of Adenza, which was subsequently acquired by Nasdaq — a reminder that the brand on a vendor platform can outlive several corporate parents.

In-house risk engines

Large banks frequently build their own independent risk and P&L engines, and their own FRTB calculators, rather than relying on a vendor platform for the independent challenger function described earlier in this Part. The reasons are control, flexibility to adapt quickly to a new regulatory requirement, and intellectual property the bank does not want embedded in a shared vendor product.

Risk-analytics vendors

Specialist libraries and platforms — Numerix, FINCAD, MSCI/RiskMetrics among them — supply pricing, VaR and analytics capability that a bank can embed inside its own engine rather than building every pricing model from first principles.

Market-data providers

Bloomberg, Refinitiv, now part of LSEG, and ICE supply the prices, curves and volatility surfaces that feed the market-data platform, while consensus services such as Totem aggregate contributed prices across multiple banks to produce the independent benchmarks that IPV, from Part 1, checks the desk's own marks against.

Grid / compute & cloud

Compute grids, and increasingly cloud infrastructure and big-data stores, provide the raw processing power that runs the overnight, and increasingly intraday, revaluation at the scale a firm-wide book requires.

The buy-versus-build pattern maps fairly closely to where a bank believes it has an edge. Trading and desk risk, where speed and product coverage matter most to the front office, are commonly bought from Murex or Calypso, because building a competitive cross-asset trading platform from scratch is an enormous, continuous undertaking few banks choose to fund. The independent risk and P&L engine, by contrast, is often built in-house precisely because that engine's whole purpose is to be a genuinely independent, bank-specific challenge to the vendor-run front-office numbers — buying the same vendor platform twice would undermine the independence the second engine exists to provide. Analytics libraries and market data, meanwhile, are almost always bought, because pricing formulas and market prices are not where any single bank differentiates itself.

Europe/France specifics: Murex is a French firm, headquartered in Paris, and it is heavily used across European banks — a technologist working in risk & P&L in the region is very likely to work in or alongside MX.3 at some point. Vendor names in this space change through acquisition on a regular cycle, as the Calypso-to-Adenza-to-Nasdaq history shows, so it is worth treating any specific brand name as a moving target and focusing on the underlying capability — trading, risk, P&L, collateral — that platform provides, rather than memorising a name that may belong to a different owner in a few years.

On the Critical Day, There Is No Second Attempt

Some moments in the risk & P&L cycle cannot simply be retried later. The sign-off deadline before the desk opens, month-end and quarter-end capital reporting, and a crisis morning when every desk wants an updated risk view at once are all moments where a system outage does not just delay a report — it leaves the bank trading, or reporting to its supervisor, blind. Four concepts describe how the systems in this Part are built to survive those moments.

Redundancy

Running duplicate infrastructure — a second server, a second data centre, a second copy of a critical feed — so that the failure of any one component does not, by itself, take the system down. Redundancy is the raw capacity that everything else in this section depends on.

Automatic failover

The mechanism that detects a failure and switches processing to the redundant component without a human having to notice and intervene first. A risk engine with redundant hardware but no automatic failover still fails the moment its primary instance goes down overnight, because nobody is watching at that hour to switch it manually in time.

Disaster recovery (DR)

The planned, tested process for restoring a system after a major failure — a data-centre outage, a large-scale infrastructure event — including how quickly the system must be back, the recovery time objective, and how much data loss is tolerable, the recovery point objective. DR is what turns "we have a backup somewhere" into a number a bank can actually commit to and test against.

Business continuity (BCP)

The broader plan for keeping the business itself running through a disruption — not just the technology, but where staff work from, which processes can run manually or on a reduced basis, and in what order functions are restored. BCP is the organisational layer sitting above DR's technical recovery.

Concept What it guarantees Why it matters on the critical day
Redundancy No single point of failure in the infrastructure Keeps one server or feed outage from becoming a firm-wide outage
Automatic failover Continuity without waiting for a human to intervene Keeps an overnight batch running through a hardware failure at 2am
Disaster recovery A tested recovery time and recovery point after a major failure Turns "we can recover" into a committed, measured number
Business continuity The wider business keeps functioning, not just the systems Keeps sign-off, reporting and trading decisions possible even if a site is unusable

Every System Is Built to a Regulatory Specification

Read the regulation shaping risk & P&L not as a legal summary but as a specification the systems above are built against — each rule below dictates something concrete about how a system must behave.

FRTB (via CRR/CRR3 in the EU)

The Fundamental Review of the Trading Book is the market-risk capital regime built around the standardised approach, SA, and the internal-models approach, IMA, plus a defined trading-book/banking-book boundary, covered in depth in Part 2. As a system specification, it requires an engine that can compute delta, vega and curvature sensitivities bucketed exactly as the rule prescribes, aggregate them with the correlations the rule sets out, and separately support an IMA path with expected shortfall, liquidity horizons, and the RFET, NMRF and PLA machinery from Part 2. In the EU, FRTB lands through the Capital Requirements Regulation, in its CRR3 form.

Basel market-risk framework

The global Basel Committee on Banking Supervision standard that FRTB implements at EU level: sensitivities-based capital, expected shortfall as the tail measure, liquidity horizons that scale by risk factor, and the desk-level model-approval structure that determines which desks may use IMA at all. As a specification, it is the reason the risk engine's FRTB module has to be organised desk by desk, not book by book — approval, and therefore capital treatment, is granted or withdrawn at the desk level.

BCBS 239 (risk-data aggregation & reporting)

A set of principles for effective risk-data aggregation and reporting, built around four qualities: accuracy, completeness, timeliness and adaptability. As a specification, it governs precisely the aggregation layer and the data-lineage chain walked earlier in this Part — it is the reason a bank must be able to show, for any number in a risk report, exactly which trades, marks and calculations produced it, and how quickly it could be reproduced if challenged.

Model validation / ECB TRIM

Internal models used for regulatory capital must be independently validated inside the bank and approved by the supervisor before they can be relied on. The ECB's Targeted Review of Internal Models, TRIM, exercise reviewed internal models and the data feeding them across euro-area banks. As a specification, this requires every model change described in this Part — a new VaR methodology, an updated FRTB IMA calibration — to be documented, tested and re-approved through a formal governance process before it reaches production, not adjusted quietly by whoever built it.

IFRS 13 & prudent valuation / CRR AVAs

IFRS 13 sets the accounting fair-value standard behind mark-to-market from Part 1, while the EU's prudent-valuation requirement under the CRR asks banks to hold Additional Valuation Adjustments, AVAs, against uncertainty in the marks themselves, particularly for illiquid positions where independent price verification cannot fully confirm the desk's own mark. As a specification, this is a direct requirement on the valuation and IPV chain: a system must be able to quantify, position by position, how uncertain a mark is, not just what the mark itself is, and translate that uncertainty into a capital deduction.

Europe/France specifics: EU banks implement FRTB through the CRR, in its CRR3 form, under the ECB's Single Supervisory Mechanism, with national regulators such as France's ACPR, the Autorité de Contrôle Prudentiel et de Résolution, supervising individual institutions alongside the ECB. The reporting version of FRTB-SA has applied for some time, giving supervisors and banks visibility into the standardised numbers, while the capital go-live — the point at which FRTB actually drives the capital requirement rather than just a reporting figure — is phased in over a longer timeline that has shifted before and should always be checked against the current EU implementation date rather than assumed.

Market Risk Is One Risk Among Several — Know the Boundary

Market risk, the subject of this module, is one member of a family of risk types a bank must measure, capitalise and manage, and knowing where its boundary sits is what prevents a bank from double-counting the same exposure twice or, worse, missing a risk entirely because each team assumed another team owned it.

Credit risk

Loss because a borrower or bond issuer fails to pay. It is the banking book's core risk, with its own capital regime, its own rating and provisioning models, and its own dedicated teams — distinct from anything the market-risk desks in this module compute.

Counterparty credit risk (CCR)

Loss if a derivatives counterparty defaults while a trade the bank holds against them has positive value — a swaption on the Kestrel book that is deep in the money is worth nothing if the counterparty that owes on it cannot pay. CCR is the bridge between market risk and credit risk, and it is what drives credit valuation adjustment, CVA, and the broader family of XVAs, which price exactly this default risk into a derivative's value.

Liquidity risk

The inability to fund a position or exit it without materially moving its price. It splits into funding liquidity, whether the bank can finance its balance sheet, and market liquidity, whether a position can be sold without a large price impact, and it is governed by its own framework, the Liquidity Coverage Ratio and Net Stable Funding Ratio, entirely separate from the VaR and FRTB machinery covered in Part 2.

Operational risk

Loss from failed processes, systems, people or external events. This is where a booking error on the Kestrel book, a market-data feed outage, or the kind of system failure described in the resilience section above would ultimately be captured and capitalised, if it caused a loss.

Model risk

Loss because a model — a pricing model, a VaR methodology, an FRTB calculator — is wrong or is used outside the conditions it was validated for. It cuts across every risk type above rather than sitting beside them, and it is governed by the model-validation and TRIM-style review described in the previous section.

A closely related concept worth naming here is interest-rate risk in the banking book, IRRBB — the market-risk-like exposure of banking-book positions, such as loans and deposits, to interest-rate moves. It uses techniques that echo the sensitivities and stress testing built for the trading book in Part 2, but it sits in a separate framework because the banking book is managed and capitalised differently from the trading book this module has focused on.

Figure 11 Market risk among its neighbours — each with its own capital regime, models and teams.

The boundaries matter because each of these risk types carries its own capital regime, its own models and its own dedicated team, and a position that is not clearly assigned to exactly one of them either gets capitalised twice or, more dangerously, not at all. The place these boundaries genuinely meet — a swaption counterparty defaulting while the trade is deep in the money — is not a gap between market and credit risk to be argued over; it is a defined, owned discipline in its own right, CVA/XVA, built specifically to live at that junction. Clean boundaries elsewhere in the risk taxonomy exist so that this kind of deliberately built bridge remains the exception, not the norm.

Inside the risk engine — how the numbers are actually produced

Strip away the dashboards and the risk engine is a machine with one job: reprice the whole book, many times, under many market states, then read risk and P&L off the results. Everything discussed so far — VaR, ES, SBM, PLA, backtesting — is a way of slicing the output of that machine. The machine itself has a shape worth understanding, because its constraints explain why FRTB felt like an engineering project as much as a risk-methodology one.

Figure 12 Inside the risk engine: scenario generation feeds a revaluation grid whose output is the risk cube, all wrapped in the overnight orchestration.
Bump-and-revalue is simple and slow; adjoint differentiation is fast and essential

Bump-and-revalue is the intuitive method: move one risk factor a little, reprice the trade, read off the sensitivity as the change in value. It is easy to explain and easy to trust, which is why it dominated for decades. Its cost is the problem. One bump means one full repricing, so a book sensitive to thousands of risk factors — an FRTB desk under the standardised approach's bucket structure, say — needs thousands of pricings just to produce a single day's sensitivities, then the same again for every scenario in a VaR or ES run. On an illustrative desk with, say, five thousand live risk factors and a Monte Carlo book behind them, bump-and-revalue can mean millions of individual pricing calls for one overnight cycle — a volume that scales roughly linearly with the number of factors, which is exactly the dimension FRTB inflated.

Adjoint algorithmic differentiation, AAD, inverts the economics: instead of repricing once per factor, it computes the entire gradient — every sensitivity, to every factor, simultaneously — in roughly the cost of a single pricing run (illustratively, a small constant multiple of one pricing, rather than one pricing multiplied by the factor count), by propagating derivatives backward through the same computational graph used for the valuation itself. That is not an incremental speed-up; it is a different order of complexity — the cost stops scaling with the number of risk factors at all. It is why AAD moved, within a few years, from a niche technique used by a handful of quant desks pricing exotics to something close to mandatory across the risk engine once FRTB multiplied the number of buckets a bank has to report against. The trade-off is implementation cost: AAD requires the pricing library itself to be written, or rewritten, in a way that supports differentiating through it, which is a multi-year engineering undertaking for a legacy pricing stack, not a configuration change.

The revaluation grid and the risk cube are the shape of the compute

The revaluation grid is the distribution layer — it takes the list of (trade, scenario) pricing jobs and spreads them across as many compute nodes as the firm can afford, because the volumes involved are embarrassingly parallel: each pricing is independent of every other one. The output, once collected, is best pictured as a cube rather than a table — P&L indexed along three axes: scenario, position, and risk factor. That geometry matters because it explains why the same run underpins so many different numbers. VaR and ES are read by slicing the cube along the scenario axis and looking at the P&L distribution. Sensitivities and SBM charges are read by slicing along the risk-factor axis. Desk-level or business-line aggregation is read by slicing along the position axis. Nobody runs a separate calculation for each report; they query the same cube differently, which is also why an error in the underlying revaluation shows up simultaneously across VaR, stress, and Greeks — one bad number, many downstream symptoms. An illustrative sense of scale: a few thousand positions against a few thousand historical scenario days already produces tens of millions of cube cells before a single risk-factor breakdown is added — which is why the storage and query design of the cube is as much an engineering decision as the pricing logic that fills it.

One engine, many scenario sets

The revaluation machinery itself does not know or care whether it is computing VaR, a Monte Carlo exotic price, or a stress test — it only knows how to take a scenario and produce a P&L. What changes between use cases is purely the scenario data fed in: a rolling historical window of actual market moves for historical-simulation VaR, thousands of simulated forward paths for Monte Carlo pricing, or a handful of hand-specified shocks — a 2008-style credit seizure, a rates-up-200bp move — for stress testing. Building one well-engineered revaluation engine and treating scenario generation as a swappable input, rather than building separate calculation stacks for VaR, stress, and pricing, is the design choice that keeps the whole system maintainable. It is also the reason a bug in the revaluation core is so consequential: it does not just corrupt one report, it corrupts every report that engine feeds.

Incremental revaluation and caching are how the run fits the night

None of this matters if the run cannot finish before the desk opens, and the brute-force version of "reprice everything under every scenario every night" does not fit inside eight hours for a large book. The practical answer is to reprice only what changed: if a trade's terms, its market data inputs, and its model version are all identical to last night's run, its valuation and sensitivities are almost certainly still valid, and recomputing them is wasted machine time. Incremental revaluation identifies that overlap and skips it; caching keeps intermediate results — a discount curve build, a calibrated volatility surface — available for reuse rather than rebuilding them from scratch inside every pricing call. Neither technique is exotic, but together they are frequently the difference between a batch that lands with margin to spare and one that blows through the deadline on a day when nothing about the methodology itself changed — only the volume.

Orchestrating the overnight run is a systems problem in its own right

The hardest part of producing tomorrow's risk numbers is rarely any single calculation — it is sequencing hundreds of interdependent steps so the whole chain lands by a fixed clock time. Picture the run as a dependency graph: snap the market data close, build the curves and surfaces off that snap, revalue the book against those curves, compute risk and P&L off the revaluation, aggregate up to desk and firm level, reconcile against yesterday's numbers and the front-office P&L, and only then sign off for the risk report to go out. Every arrow in that chain is a dependency. A market-data feed that arrives fifteen minutes late, or a curve build that fails silently, does not cost fifteen minutes — it starves every downstream step of its input, and the delay compounds as it cascades through revaluation, aggregation, and reconciliation. This is the same architecture pattern the module met earlier in the two-engine landscape and the aggregation layer: independent engines each run their own version of this graph, and the aggregation step at the top of Part 3's system landscape is itself just another node waiting on both of them to finish cleanly.

Concern What it means in practice
The SLA The entire chain — data snap through sign-off — must complete before the desk opens; there is no partial credit for a risk report that arrives at 9am when trading starts at 7am
Restart and replay A single failed node in the graph must be re-runnable on its own, picking up from where it broke, because re-running the full batch from the start after a late-stage failure almost always blows the deadline
Observability The run needs to be legible at 3am — which nodes have completed, which are running, which are at risk of missing their window — rather than discovered as "will not land" only once the deadline has already passed

This is why risk platforms increasingly borrow the tooling of modern data engineering rather than building bespoke batch schedulers from scratch: orchestration frameworks that express the run as an explicit dependency graph, lineage tracking that shows which output came from which input and which scenario, and monitoring that surfaces a stalled node in real time rather than in the post-mortem. A risk run is, mechanically, a production data pipeline — the only difference from an ordinary overnight ETL job is that this one has a deadline the business cannot move.

The governance chain above the desk turns numbers into decisions

A risk number that stays on a desk's screen has not done its job — it has to climb a governance chain that turns "the book's ES is such-and-such" into "the desk is allowed to run this much risk, funded by this much capital."

The CRO and the board risk committee own the appetite

Independent risk reports up through the Chief Risk Officer, not through the front office, and that reporting line is what makes the number credible rather than merely convenient. Above the CRO sits the board's risk committee, which sets the firm's overall risk appetite — a ceiling, typically expressed across several dimensions at once (VaR, stress loss, concentration), from which every desk limit and every capital allocation ultimately cascades. Nothing below this level invents risk appetite; it only distributes a ceiling that was set here. The independence this module opened Part 3 with — the actor ecosystem where risk sits apart from the desks it measures — is what gives this chain its authority: a limit set by the people whose bonus depends on the desk's P&L would not be a limit at all.

ALCO and capital allocation make risk compete for a scarce resource

The appetite set at board level does not stay abstract for long — it turns into desk-level limits and capital budgets, typically through an asset-liability committee (ALCO) or an equivalent capital-allocation forum. Because capital is finite and, from the firm's point of view, expensive to hold, desks are effectively competing with each other for a share of it. The risk-and-return metrics built up in earlier parts of this module — VaR relative to capital consumed, ES relative to revenue generated — are precisely how that competition gets judged: a desk that produces strong returns per unit of capital gets more of it next cycle; one that does not, gets less, independent of how much absolute P&L it books.

New product approval keeps un-modellable risk off the book

A genuinely new product — a new payoff structure, a new underlying, a new market — cannot simply start trading because a salesperson found a client for it. It has to clear a new-product approval process that requires sign-off from risk, valuation, operations, and model validation before the first trade books. The purpose is narrow but critical: stop the firm from taking on a risk it cannot yet measure, price defensibly, or hold capital against. A product that has not been through this gate is, by construction, a product the risk engine described above does not know how to reprice correctly.

Stress testing feeds capital planning through the ICAAP

Firm-wide stress and reverse-stress results are not run for their own sake — they feed the Internal Capital Adequacy Assessment Process, the exercise that translates a hypothetical scenario into a decision about how much capital the firm actually holds. A stress scenario, in this light, is not an academic what-if confined to a risk report; it is a direct input into the capital plan the firm takes to its regulator and its board. That is the thread connecting the stress-testing methodology from Part 2 to a real balance-sheet decision: the scenario only matters because ICAAP gives it somewhere to land.

Surveillance and the human risk are the last control

The same numbers built to measure market risk are also the firm's earliest warning system against its own people. An unexplained P&L residual, or a mark that cannot be independently verified, is very often the first visible trace of mismarking or unauthorised trading — long before any investigation confirms it as such. That is why P&L-explain and independent price verification double as fraud detection, not just accuracy checks: a control built to catch a stale model or a data error catches a concealed position for exactly the same reason — both leave a residual that should not be there.

Three controls sit alongside those numeric checks. Mandate monitoring asks whether a desk is trading what it is actually authorised to trade — the products, the markets, the size — rather than drifting into activity nobody approved; a desk quietly trading outside its mandate is often the earliest and least dramatic sign that oversight has slipped. Trade surveillance watches trading behaviour itself for patterns that look like concealment, unauthorised risk-taking, or market abuse — cancelled trades booked and reversed just before a reporting cut-off, positions that grow just under a limit threshold, activity concentrated around known control gaps. Segregation of duties is the structural backstop underneath both: the front office that takes a position must never be the same function that confirms, settles, or independently values that position, because a trader who can mark their own book can hide a loss inside the mark, and once that single control is bypassed, everything downstream of it — the P&L, the VaR, the capital charge — inherits the same fiction. Every rogue-trading and large mismarking loss this module has referenced traces back, at the point of concealment, to exactly one of these controls being absent, overridden, or quietly bypassed — not to a market move nobody could have foreseen. The pattern repeats closely enough across the well-known cases that it is worth naming directly: it is rarely the model that failed, it is the independence around the model.

  1. Why does independence between front office and risk/product control matter more than efficiency?

  2. Who owns the official daily P&L number the bank reports?

  3. Why do banks run two risk engines instead of one?

  4. Why is the aggregation layer harder to build than any single desk's risk calculation?

  5. What does the LEI identify, and what does the ISIN identify?

  6. Where does market risk sit relative to credit risk and counterparty credit risk?

  7. What is the difference between disaster recovery and business continuity?

3

  1. Who owns the official daily P&L?

  2. Why run two risk engines?

  3. Why does independence matter more than efficiency?

4When It Fails, Where It Hurts, and Where It's Going

The failures that shaped the discipline, the recurring pain that never quite goes away, and how cloud, real-time compute and AI are reshaping the work.

Every Failure in Risk and P&L Fits One of Six Boxes

Risk and P&L break in a small number of recognisable ways, and naming them is the first step to defending against them. Parts 1 and 2 built the toolkit — Greeks, P&L-explain, VaR, expected shortfall, FRTB, BCBS 239. Every one of those tools exists because something, somewhere, once failed without it. Sorting the failures into a small set of modes turns a long list of scary anecdotes into something a team can actually organise its controls around. Six modes cover the space: a model that cannot see the risk, a mark that is wrong, an input that is dirty, a firm that cannot add its risk up fast enough, a limit that is breached without anyone noticing, and a person or a process that bends the rules. Each has a distinct owner and a distinct defence, which is exactly why splitting them out matters more than lumping them into a single word like "risk management failure."

Measurement failure
Failure mode What it is Who owns the defence Main defence
Measurement failure The risk model is blind to a risk that is actually there — a tail it doesn't capture, a factor it doesn't include, a correlation it assumes wrongly Risk quants, market risk Expected shortfall, stress testing, backtesting, model validation
Valuation/marking failure
Failure mode What it is Who owns the defence Main defence
Valuation/marking failure A position is marked wrong — carelessly or deliberately optimistically — so P&L and risk are both built on a false value Product control, valuations Independent Price Verification (IPV), valuation reserves/AVAs, IPV thresholds
Data-quality failure
Failure mode What it is Who owns the defence Main defence
Data-quality failure An input feeding the chain is stale, spiked or mis-scaled — a bad curve point, a broker quote in the wrong units, a missing feed Risk IT/BA, data teams, market risk Snap-time validation rules, market-data quality checks, P&L-explain
Aggregation & timeliness failure
Failure mode What it is Who owns the defence Main defence
Aggregation & timeliness failure Positions and risk factors are correct individually but the firm-wide view arrives too late, or doesn't reconcile, to act on Risk IT/BA, market risk BCBS 239 principles, aggregation layer, timeliness SLAs
Limit & control failure
Failure mode What it is Who owns the defence Main defence
Limit & control failure Risk runs past its agreed boundary and nobody catches it in time — a limit not monitored, a breach not escalated Market risk, desk management Limit framework, real-time monitoring, breach escalation process
Governance & conduct failure
Failure mode What it is Who owns the defence Main defence
Governance & conduct failure A model is misused outside its design, a mark is manipulated, or oversight is bypassed or overridden Senior management, compliance, model validation, internal audit Independent validation, segregation of duties, escalation culture, audit

Three Real Collapses Show What Each Failure Mode Looks Like

Case studies exist because the six failure modes are abstractions until they cost someone billions. Three episodes, each well documented in regulatory reports, litigation and public post-mortems, make the abstractions concrete. Each illustrates a distinct failure mode from the table above, though in practice a real collapse is rarely pure — most involve two or three modes reinforcing each other.

The model that couldn't see the risk — JPMorgan's "London Whale" (2012)

JPMorgan's Chief Investment Office ran a synthetic-credit derivatives book — large positions in credit-index instruments — that grew to a size out of proportion to any hedging rationale. Through 2011 and into 2012, the book's value-at-risk model was changed; the new model, later found to be under-calibrated and poorly reviewed, reported roughly half the risk the old one did, even as the underlying positions grew larger and more concentrated. As the book moved against the desk, traders are alleged to have marked positions away from the middle of the market to understate the scale of the losses building up, delaying the moment the true number became visible to management and to the regulator. By the time the position was unwound, the loss had reached roughly USD 6 billion — a figure confirmed across JPMorgan's own task-force review, the US Senate Permanent Subcommittee on Investigations report, and the settlements JPMorgan reached with US and UK regulators. Three failure modes stack here: measurement (a VaR model that understated the risk it was supposed to catch), valuation (marks moved away from fair value to hide the scale of the loss), and governance (a model change and a growing concentrated position that senior oversight failed to challenge in time). The episode is a large part of why the P&L-attribution test and independent model validation, from Part 2, exist in their current, strict form — a risk model that quietly diverges from the P&L it is meant to predict is precisely the failure PLA is built to catch before it reaches this scale.

The measure that missed the tail — the 2007-08 crisis and VaR's blindness

Through the summer of 2007 and into 2008, several major banks reported daily trading losses that their own value-at-risk models had described as near-impossible — moves their historical or parametric VaR, calibrated on the calm years that preceded the crisis, assigned a probability of essentially zero. A senior Goldman Sachs executive famously described seeing "25-standard-deviation moves, several days in a row" — a public illustration of how far realised losses had drifted from what a VaR model calibrated to normal markets could conceive of. The problem was not that any single bank's VaR was miscoded; it was structural. VaR marks a threshold — the loss you should not exceed most of the time — and says nothing about how bad the rare bad day actually is, and a model trained on quiet years has no memory of a once-a-decade shock. That blindness to the tail, replicated across the industry, is a direct cause of the Basel Committee's decision to replace VaR with stressed expected shortfall as the internal-models metric under FRTB, exactly as covered in Part 2: ES is calibrated to a period of significant stress precisely so that capital does not evaporate in the calm years before the next crisis. This is a pure measurement failure — no single mark was wrong, no single limit was breached in isolation, but the industry's core risk number was looking at the wrong thing.

The concentration that broke the desk — Amaranth Advisors (2006)

Amaranth Advisors was a multi-strategy hedge fund whose energy desk, run by a single trader, built an enormous position in natural-gas futures and calendar spreads — bets on the difference between prices for delivery in different months. The position was large relative to the market's own liquidity, and it was concentrated in a way that meant Amaranth was effectively unable to unwind it without moving the market against itself. When gas prices moved sharply against the position in September 2006, the fund could not exit fast enough; losses over the following days reached roughly USD 6.6 billion, and the fund collapsed within weeks. This is a limit-and-control failure with a liquidity dimension: the position sizes and concentration should have been capped and monitored well before the market turned, and once the market did turn, the position's illiquidity turned a large loss into a fatal one. It is worth a one-line comparison to an older, differently shaped failure: Barings Bank's collapse in 1995, when a single trader in Singapore ran unauthorised, unhedged derivatives positions that were concealed through a misused internal account — the classic rogue-trading control failure, where segregation of duties between the front office and settlement/back office simply did not exist. Amaranth is concentration and liquidity; Barings is governance and segregation of duties — different mechanisms, same lesson, that limits only work if someone independent is actually watching them.

Case Failure mode What broke Lesson
JPMorgan "London Whale" (2012) Measurement + valuation + governance An under-calibrated VaR model and optimistic marks hid a growing loss of roughly USD 6bn Independent validation and marking must keep pace with a book's growth, not lag it
2007-08 crisis Measurement (tail-blindness) VaR calibrated to calm markets said little about the scale of the crash Statistical models need a stressed, tail-aware measure alongside them — the direct root of FRTB's ES
Amaranth Advisors (2006) Limit & control (concentration/liquidity) A concentrated, illiquid natural-gas spread position of roughly USD 6.6bn in losses could not be exited as the market turned Position and concentration limits must account for how long a position takes to unwind, not just its size

A −1.9m Residual Shows How a Break Gets Hunted Down

Most P&L breaks look, for the first hour, like a trading disaster, and turn out to be a single dirty input a few hours later. That gap between how a break first feels and what it actually is is exactly the value of P&L-explain as a diagnostic: it does not just tell you the day went wrong, it tells you where in the chain — new trades, market moves, carry, or an unexplained residual — the story stops adding up, and a residual concentrated in one sub-book and one Greek line narrows the search from "the whole book" to "one vol point" in three steps. The discipline is not glamorous, but it is what turns a −1.9m panic into a fifteen-minute fix.

Under FRTB, this kind of break carries a consequence that goes beyond the morning's sign-off. Residuals like this one are exactly what feeds the P&L-attribution test from Part 2 — a desk whose hypothetical and risk-theoretical P&L repeatedly diverge by more than the prescribed statistical thresholds risks being pushed off internal models and onto the standardised approach, at a real cost in capital. That is why European desks, operating under CRR3 and ECB supervision, invest heavily in market-data validation at the exact moment data is snapped overnight — catching a mis-scaled broker quote before it enters the revaluation engine is cheaper, in capital terms, than explaining it away after the fact.

Building FRTB Capital Is Mostly a Data Programme

A bank's FRTB implementation programme has a deceptively simple mandate: go live with market-risk capital computed under CRR3, using the standardised approach (SA) across the whole trading book and the internal models approach (IMA) for a handful of strategic desks where it is worth the extra effort. The programme draws together risk quants, risk IT/BA, data teams, product control and model validation, and it runs for years rather than months. The tension that shapes every decision inside it is that IMA needs far less capital than SA for a well-modelled desk, but it demands a level of data completeness, lineage and P&L discipline that SA simply does not ask for — a desk only earns the cheaper capital treatment by proving, continuously, that its data and its models deserve the trust.

The programme breaks into five workstreams, and they run largely in parallel rather than in sequence. First, the SA sensitivities engine: computing delta, vega and curvature for every trade against the FRTB-prescribed buckets, then aggregating with regulatory risk weights and correlations — the mechanical core that every desk needs regardless of whether it ever qualifies for IMA. Second, the trading-book boundary: classifying every position correctly between trading book and banking book, and building the controls and documented evidence a supervisor will test directly, because a mis-classified position is a capital arbitrage the regulator is specifically looking for. Third, for the IMA desks, building expected shortfall with liquidity horizons and running the Risk-Factor Eligibility Test to separate modellable factors from non-modellable ones, capitalising the non-modellable factors through the punitive stress-scenario charge covered in Part 2. Fourth, standing up the P&L-attribution and backtesting engines per desk — the daily gate, described in the case study above, that decides whether a desk keeps its internal-models approval. Fifth, data lineage and sign-off: BCBS 239-grade traceability from every capital number back to its source positions and market data, model-validation packs, and ultimately ECB approval before a desk can rely on IMA at all.

The lesson that surfaces once a bank has actually run this programme is that the quant maths — the SBM formulas, the ES calculation, the correlation aggregation — is a comparatively small and well-understood part of the effort. Roughly 80% of the work is data and systems: sourcing enough real price observations to pass RFET, building lineage a supervisor can inspect, reconciling SA sensitivities against risk-management Greeks, and keeping PLA statistics inside their thresholds day after day. Whether IMA is worth pursuing for a given desk is decided by data quality and PLA performance, not by whether the desk's pricing models are sophisticated enough.

European banks deliver this under CRR3 and ECB supervision, working alongside national regulators such as the ACPR in France. A common strategic pattern has been to launch SA first — its reporting obligations applied earlier than the binding capital requirement — and pursue IMA selectively, only on the desks where the capital saving is large enough to justify the ongoing burden of NMRF capitalisation, PLA discipline and model validation. That is a cost-benefit decision made by business and risk management together, not a purely technical one, and it is a decision every large European trading desk has had to make explicitly.

The Pain Points Are Almost Never About the Maths

Risk-data aggregation

Consolidating hundreds of books, thousands of risk factors and multiple legal entities into one consistent, reconciled firm-wide view — with full lineage back to source — is the BCBS 239 problem from Part 2, and it is never fully solved. Systems are added, acquired banks bring their own architectures, new products bring new risk factors, and the aggregation layer has to absorb all of it while still finishing before the desk opens.

Overnight compute crunch

Full revaluation, VaR/ES, stress testing and the FRTB capital calculation all have to complete in the same few-hour overnight window. A run that breaks or overruns does not just delay a report — it leaves the desk trading the next morning without a current view of its own risk, which is the exact condition the whole daily cycle from Part 1 exists to prevent.

FO-vs-risk consistency

The front-office risk system and the independent risk engine rarely agree exactly, because they were often built at different times, on different platforms, with slightly different models, data snapshots and timing conventions. Reconciling the two — the "two engines, one truth" principle from Part 3 — is not a project with an end date; it is a permanent, ongoing job.

Model risk & change

Every pricing or risk model can be wrong, or can be used outside the conditions it was validated for. Every change to a model — a recalibration, a new curve-construction method, a parameter update — has to be validated, versioned and made reproducible, so that a number from six months ago can still be explained today. The governance overhead is heavy because the cost of getting it wrong, as London Whale showed, is not hypothetical.

Market-data quality

Stale, spiked or mis-scaled marks silently corrupt risk, P&L and capital at the same time, because all three are computed from the same inputs. The Kestrel vol-surface break above is the pattern in miniature; validating data at the moment it is snapped, before it ever reaches the revaluation engine, is a constant, unglamorous battle that never really ends.

Reconciliation everywhere

Position reconciliations, P&L reconciliations, risk reconciliations across systems consume a disproportionate share of the total effort in a risk & P&L function — and they are exactly where breaks and errors first become visible, well before anyone has diagnosed the root cause.

Line up all six and a pattern falls out immediately: almost none of them is a maths problem. The quantitative methodology behind VaR, ES, the Greeks and the FRTB formulas is, by this point, well established and broadly settled across the industry. The actual difficulty — and the actual value a technologist adds — is in making that settled methodology run correctly, on time, reconciled across systems, and auditable back to its source, every single day, without exception. That is a data and systems discipline, not a research problem, and it is where a career in this space is genuinely built.

Cloud and Real-Time Compute Turn the Batch Into a Service

Cloud & elastic grid

A bank can burst thousands of compute cores for the nightly revaluation and release them once the run completes, paying for the peak capacity it needs without owning and maintaining it year-round. This is the most direct technology answer to the overnight compute crunch above — elastic capacity turns a fixed, expensive on-premises grid into a cost that scales with the size of the job.

Intraday / real-time risk

Greeks, and increasingly VaR itself, are refreshed through the trading day rather than only at the overnight close, so the desk sees its risk move as it trades rather than working from a snapshot that is already hours old by lunchtime. This builds directly on the intraday-risk direction introduced in Part 2.

On-demand what-if

A trader or a risk manager can price a prospective hedge or stress a new scenario in seconds, instead of waiting for the next overnight batch to find out what a trade would do to the book's risk.

Adjoint / AAD Greeks

Adjoint algorithmic differentiation computes a full vector of sensitivities — every Greek, across every risk factor — in roughly the cost of one pricing run, rather than one run per bump. For a book that has to report sensitivities against the dozens of FRTB buckets from the sensitivities-based method in Part 2, this is a genuine step-change in what is computationally affordable.

None of this matters unless speed converts into more time for control, not less. The honest case for faster compute is that finishing the overnight run sooner leaves more of the window for reconciliation, investigation and sign-off before the desk opens — the batch stops being the binding constraint on the whole daily cycle. FRTB makes elastic scale close to a necessity rather than a luxury, because expected shortfall with liquidity horizons and thousands of stress scenarios multiplies the compute bill well beyond what plain VaR ever required. Moving positions and market data into a cloud environment also raises real data-residency, latency and control questions, particularly for European institutions operating under EU data-protection and outsourcing supervision — these are not solved by the technology itself, they are solved by the governance wrapped around it. And underneath all of it, correctness still rules: a risk number that arrives faster but is wrong is still wrong, and every migration has to prove that reconciliation, lineage and reproducibility survive the move intact.

The filter that separates a real upgrade from a shinier dashboard is simple to state and hard to apply honestly: does this change move the economics or the risk, or does it just move the appearance of speed? A grid that lets a desk run stress scenarios in minutes instead of hours changes what the desk can actually do — it can test more scenarios, catch more concentration risk, react faster to a market move. A dashboard that displays the same overnight numbers with a nicer refresh animation changes nothing about the underlying risk position or the capital it attracts. Every cloud, real-time or compute investment in this space should be tested against that question before it is funded, because elastic infrastructure is expensive, and "faster" is only valuable when the extra speed buys a genuinely different decision.

AI Earns Its Keep on Data, Not on Judgment

Market-data anomaly detection

Machine learning models can flag stale, spiked or mis-scaled marks at the moment they are snapped, before they ever reach the revaluation engine — this is exactly the class of failure that produced the Kestrel vol-surface break, and it is the most natural fit for AI anywhere in this discipline, because it is a pure pattern-recognition problem on structured, historical data.

P&L-explain & residual triage

Rather than a human scanning every desk's residual every morning, a model can cluster and rank breaks by size, persistence and pattern, so analysts spend their time on the handful of residuals that actually matter rather than working through the whole list in order.

ML pricing surrogates

A fast machine-learning approximation of a slow exotic pricer can accelerate VaR, XVA and what-if calculations that would otherwise require thousands of full revaluations — with the exact model kept running as a backstop and a periodic check, rather than being retired.

Data quality & reconciliation

Models can auto-match breaks across position, P&L and risk reconciliations, suggest mappings between inconsistent identifiers and taxonomies, and summarise exceptions for a human reviewer, cutting down the reconciliation effort identified as a pain point above.

The honest caveats matter as much as the use cases, because this is a control function, not a product feature. Explainability is not optional here: a risk or capital number has to be defensible to independent model validation and to the ECB, and a model whose reasoning cannot be reconstructed cannot be used to compute capital, however accurate it appears to be on average. A pricing surrogate needs guardrails — it has to be bounded and continuously monitored against the exact pricer it approximates, because it accelerates the calculation, it does not replace the model of record. Model risk management, from Part 2, applies to machine-learning models exactly as it applies to a pricing model or a VaR engine: they have to be inventoried, validated, versioned, backtested and governed, and the rules do not relax just because the model happens to be a neural network rather than a closed-form formula. And across every one of these use cases, the operating principle is the same — AI assists and triages, it does not automate the decision away. A human still decides and signs off, because risk and P&L remain accountable, auditable controls, and accountability cannot be delegated to a model.

Generative AI Helps With Words, Not With Numbers

Large language models are reaching into risk and P&L functions from a different angle than the machine-learning use cases above — not by computing a number, but by drafting, summarising and interpreting text around the numbers. That distinction matters, because it defines both where GenAI is genuinely useful and where it needs the tightest governance.

Model and control documentation

Drafting model-validation packs, control narratives and daily P&L commentary from structured inputs — Greeks, P&L-explain output, reconciliation exceptions — can save a meaningful amount of the writing effort that currently falls on risk quants and product controllers, provided a human reviews the draft against the underlying numbers before it goes anywhere near a regulator.

Code and test generation

Accelerating the development of risk-engine components, generating test cases against known scenarios, and drafting data-mapping logic between systems with different identifiers and taxonomies — work that is currently a large, unglamorous share of any risk-IT team's time, and where a generated first draft reviewed by an engineer is a genuine productivity gain.

Regulatory-text interpretation

Mapping FRTB and BCBS 239 requirements to specific controls and data fields, and answering ad-hoc questions like "does this new instrument type fall inside the FRTB sensitivities-based method, and if so, which bucket?" — a genuinely hard, text-heavy task that a model can accelerate as a first pass, with a compliance or regulatory-affairs specialist confirming the answer before it is relied on.

Three risks sit underneath all three use cases and they are not symmetric with the benefits. MNPI and confidentiality: position data, client data and non-public information cannot leak into a model's context or training data, which constrains where and how any GenAI tool can be deployed inside a bank at all. Hallucination and accuracy: a large language model producing a confident, fluent, wrong number or a wrong regulatory citation is a genuinely dangerous failure mode inside a control function, precisely because fluency is easy to mistake for correctness. And explainability and model-risk governance: the moment GenAI output feeds into a risk or capital number, or into a control narrative a regulator will read, it is a model in the Part 2 sense of the word, and it has to be governed as one — inventoried, validated, monitored, and owned, not treated as a convenience tool outside the model-risk framework because it happens to produce text rather than a number.

Use of GenAI The gain The governing risk
Model & control documentation Faster drafting of validation packs, narratives and commentary Confidentiality of position/client data; accuracy of drafted content
Code & test generation Faster risk-engine development and data-mapping logic Code correctness; still requires full engineering review and testing
Regulatory-text interpretation Faster first-pass mapping of FRTB/BCBS 239 requirements to controls Hallucinated or outdated regulatory interpretation; needs specialist sign-off

4

  1. Which failure modes stacked in the London Whale case?

  2. Building FRTB capital is mostly…

  3. Where does AI earn its keep in risk and P&L?

  4. What does cloud compute change?

Six Ideas Carry the Whole Module

1. Risk is sensitivity times move

Every measure this module covered — a single Greek, a bucketed DV01 ladder, VaR, expected shortfall, a stress scenario, the FRTB sensitivities-based method — is a variation on the same combination: how much the book's value changes for a given move, multiplied by how large a move you are prepared to assume. Different techniques choose different assumptions about the move; none of them escape the underlying structure.

2. P&L is a control, not a score

A good day is not "we made money" — it is "we understand every part of the P&L, and the risk numbers predicted it." The residual in P&L-explain is the tell that separates a book that is understood from one that is not, and under FRTB that residual now feeds directly into the P&L-attribution test, which decides whether a desk keeps the cheaper internal-models capital treatment at all.

3. Data cascades, so errors are never contained

A single wrong mark or a single mis-scaled input does not stay contained to one number — it corrupts the risk calculation, the P&L calculation and the capital calculation simultaneously, because all three are built from the same underlying positions and market data. The Kestrel vol-surface break shows the pattern at a small, recoverable scale; London Whale shows it at a scale that reached the front pages.

4. Independence is the core control

Two engines compute the same risk and the same P&L from the same underlying positions — one owned by the front office, one owned by product control and market risk — precisely so that neither can silently drift from reality without the other catching it. Reconciling the two, day after day, is not bureaucratic overhead; it is the mechanism that makes the whole system trustworthy.

5. Compliance is core architecture

FRTB, BCBS 239 and the resilience of the daily cycle are not compliance add-ons bolted onto a working system after the fact — they are system specifications that shape how the risk engine, the aggregation layer and the data pipeline are actually built, from the FRTB bucket structure down to the lineage every number must carry.

6. Real transformation moves cost or risk

Cloud compute, real-time risk and generative AI are only worth building when they change what a desk or a control function can actually do — process more scenarios, catch a break faster, free up analyst time for the breaks that matter — and every one of them, including GenAI, is governed as tightly as it is useful, because none of them are exempt from the model-risk and data-governance principles that run through the entire module.

  1. Why are most P&L breaks data breaks rather than trading losses?

  2. Why is an FRTB implementation programme described as roughly 80% data and systems work?

  3. What are the three failure modes the case studies illustrate?

  4. What do adjoint (AAD) Greeks actually buy a desk?

  5. Why must generative AI output that feeds a risk or capital number be governed as a model?

Cheat Sheet — Risk and P&L on One Page

A single reference for the whole discipline: the daily cycle that produces the numbers, the measures that describe the risk, and the rules of thumb worth carrying in your head. All illustrative figures are marked as such.

The daily cycle — from close to desk-open

Step Timing Core activity Key output Lead and systems
Capture & close End of day Freeze positions at the official close; snap market data A clean, frozen book + a market-data snapshot Desk + booking systems (Module 19), market-data platform
Revalue & compute risk Overnight Reprice the book on today's curves; compute the Greeks, VaR, ES, stress, FRTB Sensitivities, VaR/ES, stress, capital numbers Risk / pricing engine (Murex, Calypso, in-house)
Produce P&L Overnight Compute the official daily P&L independently from marks and cash flows The certified daily P&L Product control + P&L systems
P&L-explain & reconcile Early AM Attribute P&L to new + market + carry; compare to risk-predicted P&L; chase the residual An explained P&L, residual within tolerance Product control + risk
Risk vs limits Early AM Check Greeks, VaR and stress against limits; escalate breaches Limit-utilisation report, breach escalations Independent market risk
Sign-off & distribute Pre-open Sign off P&L and risk; publish to desk, management and regulatory feeds Signed-off numbers, feeding capital reporting Market risk + product control
Desk acts Market open Trade the new day knowing the risk and P&L; hedge or adjust Hedged, informed book The desk

The measurement toolkit at a glance

Measure What it answers Where it bites
The Greeks How much value changes for a one-unit move in each variable Real-time desk management; the atoms of everything else
VaR The loss not exceeded at a confidence level over a horizon Internal limits and day-to-day management
Expected shortfall (ES) The average loss once you are in the tail FRTB internal-model capital (97.5%, stressed)
Stress & scenarios The loss under severe, defined shocks beyond the statistics Probing the never-yet-seen; regulatory stress tests
SBM (FRTB-SA) Capital from prescribed sensitivities, buckets and correlations The mandatory floor and fall-back for all
PLA test Do the risk model and the official P&L agree, per desk Whether a desk keeps internal models or drops to SA

Key numbers and rules of thumb

Metric Rule of thumb (illustrative)
DV01 P&L for a 1bp yield move; the dominant Greek in a rates book
CS01 P&L for a 1bp credit-spread move
VaR confidence / horizon Typically 99%, 1-day for internal use
Expected shortfall (FRTB) 97.5% confidence, calibrated to a stressed period
Backtesting exceptions About 1 in 100 days for a 99% 1-day VaR; green / amber / red zones
P&L-explain residual Small (within tolerance) = risk and books agree; large = an error to hunt
FRTB effort split Roughly 80% data and systems, 20% quant maths
IMA approval Granted and withdrawn per desk; SA is the permanent floor
Sign-off deadline The whole cycle must finish before the desk opens — often within hours of the close
Most P&L breaks Are data breaks, not trading losses

The minimal math — one equation per concept

Concept The one relationship (illustrative)
DV01 P&L for a +1bp move = change in present value per 1bp of yield
Delta P&L P&L ≈ delta × the move in the underlying
P&L explain identity total P&L = new trades + market (Greek × move) + carry + residual
VaR the loss at the chosen percentile of the P&L distribution (e.g. the 99% loss)
Expected shortfall ES = the average of the losses beyond the VaR point
SBM aggregation K = the square root of (sum of weighted-sensitivity² + correlation cross-terms)
Backtesting expected exceptions ≈ (1 − confidence) × trading days (≈ 2.5 at 99% over 250)
RAROC return on capital = P&L ÷ capital consumed

The whole discipline in one sentence

Every day, independently of the desk, revalue the whole book on fresh data, decompose the P&L until the risk explains it, check the risk against its limits, turn it into the capital the regulator demands, and sign it off before the desk opens — and do it correctly, on time, and auditably, every single time.