Skip to content
Corporate & Investment Banking My progress Français

Home / Capital Markets & Securities Services / Cross-functional / Module 25

25Valuation & Product Control

How a Corporate & Investment Bank Knows What It Owns, What It Is Worth, and What It Made Today

What the bank owns, what it is worth, and what it made today.

Reading time
3 h 10
Sections
92
Recall questions
51
Quiz questions
15
Simulators
1

Why this module exists

Every other module in this series follows a deal. An equity capital markets module follows an initial public offering from mandate to listing. A debt capital markets module follows a bond from origination to settlement. This module follows something less visible and more consequential: the number that appears against a trading position at the end of each day, and the machinery that decides whether that number is allowed to stand.

The machinery matters because the number is not read off a screen. For a listed equity it very nearly is. For a ten-year interest rate swap it is calculated from a curve that someone had to build. For a bespoke credit tranche it comes out of a model that someone had to write, calibrated to inputs that may not be observable anywhere. In a modern Corporate & Investment Bank the majority of the trading balance sheet sits in that second and third category. Somebody has to decide whether the trader's view of that value is right.

That somebody is Product Control. In the language of the industry the function goes by several names — Product Control, Valuation Control Group, Independent Price Verification, Financial Control — and the boundaries differ from bank to bank. The work does not. It is the independent production, verification, explanation and reporting of trading profit and loss and of the fair value of trading positions.

Two threads run through this module

The business thread asks who decides what a position is worth, and why anyone should believe them. It covers the daily profit and loss cycle, the mechanics of price verification, the reserves and adjustments that sit between a model output and an accounting number, the governance committees that adjudicate disputes, and the incentives that make independence necessary in the first place. A reader who follows only this thread will finish able to sit in a valuation committee and follow the argument.

The technology thread asks what infrastructure makes that possible in the eight hours between a market close and a trading open. It covers trade capture, curve construction, pricing libraries, overnight batch architecture, reconciliation engines, the sub-ledger, and the messaging standards that carry a trade from execution to a regulatory report. A reader who follows only this thread will finish able to scope, assess or build a valuation control platform.

The two threads are inseparable. A control that cannot be executed within the batch window is not a control. A system that produces a number nobody can defend in front of an auditor is not a system.

How to read this module

Figure 1 The shape of the module, and the desk carried through all of it

The module opens with a Foundations part that builds the market vocabulary from nothing — unrealised profit, discounting, notional and basis points, swaps, options and implied volatility, and the Greeks. A reader who already trades derivatives can skip it. A reader arriving from consulting, technology or a non-markets part of a bank should not, because everything afterwards assumes it. The module then runs in five parts. Part 1 sets the context: what valuation means, why it is hard, where the function sits, and how the discipline was rebuilt after 2008. Part 2 is the core and the longest: the daily profit and loss cycle and the periodic verification cycle, stage by stage, with a single fictional desk carried through every stage. Part 3 maps the people. Part 4 maps the systems and the data. Part 5 covers what goes wrong, what has gone wrong in reality, and where the function is heading.

After Part 5 there is a one-page cheat sheet, a set of applied diagnostic cases that ask you to identify a fault from the evidence rather than restate a principle, a checklist for assessing a valuation platform, and a glossary.

Seven kinds of box interrupt the narrative. Key insight crystallises an idea worth carrying away. Going deeper adds technical detail for readers who want it and can be skipped without loss. IT & data lens names the systems and data behind the process just described. Watch out flags a common misunderstanding or a live risk. Worked example carries the fictional running case. In a nutshell closes each part. Check your understanding offers questions with answers, to be attempted before reading the answer.

Numbers in worked examples are illustrative unless a source is named. They are chosen to be realistic in shape and order of magnitude, not to describe any actual institution.

What you will be able to do

You will understand the daily profit and loss cycle end to end

From the moment a trade is executed to the moment a signed profit and loss figure reaches the desk head's screen, you will know every step, who performs it, what can fail, and how long each step has before the next one is blocked.

You will be able to map every actor around the valuation process

The trader, the product controller, the market risk manager, the model validator, the financial controller, the external auditor and the regulator all consume the same number for different purposes and apply different pressure to it. You will know what each one wants and why they disagree.

You will understand where the valuation actually comes from

Not in the abstract. You will be able to explain how a swap is discounted, how a volatility surface is calibrated, how an independent price verification level is set from competing sources, and how the gap between a model output and a realistic exit price becomes a reserve with a number attached.

You will be able to trace data front to back

From a trade record created at execution, through a curve node, a present value, a Greek, a profit and loss attribution line, a journal entry, and finally a regulatory capital figure. You will understand the cascade: which errors are contained and which are not.

You will recognise the risk and transformation landscape

Six risk types, five performance indicators that distinguish a controlled book from a lucky one, three real cases where valuation control failed with consequences measured in billions, and the technology changes — including generative artificial intelligence — that genuinely alter the economics rather than the appearance of the function.

Foundations — Seven ideas you need before anything else makes sense

This module talks about basis points, notionals, discount curves, implied volatility and unrealised profit as though everyone knows what they are. Most people arriving from consulting, technology or a non-markets part of a bank do not, and pretending otherwise wastes the rest of the module. This part builds each idea from nothing. A reader who already trades derivatives can skip to Part 1 without loss.

1. Owning a position is not the same as having made money

Start with the distinction the whole module rests on. A bank buys a bond for €100 million on Monday. On Friday, similar bonds are trading at a price that implies the bank's holding is worth €101 million. Has the bank made €1 million?

It has, and it has not, and both answers are correct in their own domain. No cash has moved. If the market falls back on Monday the €1 million disappears without ever having existed as money. But the bank genuinely holds something more valuable than it did on Monday, and refusing to say so would leave the accounts describing a world that no longer exists.

Accounting resolves this with two words. A gain is realised when the position is closed and the cash is in the bank: you bought at 100, you sold at 101, the €1 million is yours. A gain is unrealised — also called mark-to-market — when you still hold the position and the €1 million exists only as the difference between what you paid and what you believe you could sell for today.

Almost everything in this module concerns unrealised profit and loss. That is the entire point. Realised profit needs no control function: the cash either arrived or it did not, and the bank statement settles the argument. Unrealised profit is an assertion about a hypothetical sale that has not happened, made by the person whose bonus depends on the answer, about an asset that may have no observable price.

2. A euro tomorrow is worth less than a euro today, and the discount says how much

This is the mechanism behind almost every valuation in the module, and it has nothing to do with inflation. Suppose someone offers you €100 in one year, and suppose a completely safe investment currently returns 3% per year. How much would you pay today for that promise?

You would pay whatever amount grows to €100 in a year at 3%. That amount is €100 divided by 1.03, or €97.09. Put €97.09 in the safe investment today and you have €100 in a year, which makes you indifferent between the two. So the present value of €100 received in one year, at a 3% rate, is €97.09.

The factor 1 ÷ 1.03 = 0.9709 is called the discount factor for one year. For two years it is 1 ÷ 1.03², or 0.9426. For ten years, 1 ÷ 1.03¹⁰, or 0.7441. The further out the cash flow, the smaller the factor, and the effect compounds: a euro in ten years is worth about seventy-four cents today at 3%, and about sixty-one cents at 5%.

Figure 2 Ten identical promised cash flows; the filled portion of each bar is what it is worth today

Valuing anything that pays cash in the future is now mechanical. List every future cash flow with its date. Multiply each by the discount factor for that date. Add them up. That sum is the present value. Every bond valuation, every swap valuation and every derivative valuation in this module is a version of that three-step procedure, complicated only by the difficulty of knowing what the future cash flows will be and what the correct discount factors are.

Two things follow immediately, and both matter later. First, the discount factors have to come from somewhere, and the set of discount factors across all future dates is what the module calls the discount curve — hence the enormous attention paid to curve construction. Second, the longer the cash flow, the more the value depends on the discount factor rather than on the cash flow itself, which is why long-dated positions are disproportionately sensitive to curve choices that look like technical details.

3. Notional, basis points, and how a small number becomes a large one

Three pieces of vocabulary, all of which the rest of the module uses constantly.

A basis point is one hundredth of one percent. A rate that moves from 2.85% to 2.94% has moved nine basis points. The unit exists because rate markets move in increments too small to discuss comfortably in percentages, and because the difference between 2.85% and 2.86% is worth a great deal of money on a large position.

The notional is the reference amount a contract is calculated on, and it is usually not an amount anyone pays. In a €250 million interest rate swap, no one hands over €250 million. The figure exists only to scale the interest calculations: one party's payment is a fixed rate applied to €250 million, the other's is a floating rate applied to the same €250 million. What actually changes hands is the difference between the two, which might be a few hundred thousand euros a year. Confusing notional with exposure is the most common error made by people new to derivatives, and it runs in the dangerous direction — it makes books look larger than they are while hiding where the risk actually concentrates.

DV01 connects the two. It is the change in a position's value for a one basis point move in rates, and it is the standard unit in which rate risk is quoted. A ten-year swap on a €250 million notional has a DV01 of roughly €213,000 per basis point: rates move nine basis points and the position's value moves by about €1.9 million. That is how a number as small as a basis point becomes a number worth arguing about.

4. A derivative is a contract about something else, and a swap is the simplest useful one

A derivative is a contract whose value is determined by the behaviour of something outside the contract — an interest rate, an exchange rate, a share price, a company's creditworthiness. It derives its value from that other thing, which is the whole of the name. Nobody delivers anything at the start; the parties simply agree what will be owed under what conditions.

Why they exist is worth stating plainly, because it is easy to lose sight of. A company with a €250 million floating-rate loan does not know what its interest cost will be next year, and its board would like to. A bank can sell it a contract that converts that uncertainty into a fixed cost. The company is not speculating; it is buying certainty. The bank takes on the uncertainty, prices it, and hedges most of it elsewhere. That is the core economic function of the business this module controls.

Building a swap from nothing

An interest rate swap is an agreement to exchange two streams of interest payments on the same notional for a set period. One stream is fixed: the same rate every period, agreed at the start. The other is floating: recalculated each period from a published market rate.

Take the trade that recurs throughout this module. The bank pays fixed at 2.85% on a notional of €250 million for ten years, and receives floating, reset every three months. Each quarter, the bank owes 2.85% ÷ 4 × €250 million, which is €1.78 million. Each quarter it is owed the prevailing three-month rate ÷ 4 × €250 million, which changes every time. Only the net difference is actually paid.

Now the valuation question. At the moment the swap is struck, the fixed rate is set precisely so that the two streams are expected to be worth the same, which makes the contract worth zero to both parties. That is why no money changes hands at inception. But rates move. If market rates rise after the trade is struck, the floating stream the bank receives is now expected to be larger, while the fixed stream it pays is unchanged. The contract is now worth something positive to the bank, and exactly that amount negative to the counterparty.

To put a number on it, three things are needed, and none of them is quoted anywhere.

A projection of the floating rate at all forty future reset dates. These come from a forecasting curve built from market instruments that reference the same floating index.

A discount factor for each of the forty payment dates. These come from a discount curve, which for a collateralised swap is built from overnight interest rates rather than from the floating index — a distinction that did not exist before the financial crisis and that Part 2 explains.

An adjustment for the possibility that the counterparty does not pay. Which is the credit valuation adjustment.

That is the complete anatomy of the valuation problem. Two curves that have to be constructed, and one adjustment that has to be modelled. Everything in Part 2 elaborates on it.

What a curve actually is

A curve is the market's price of money at every future date, and it is built from a handful of observations. In euros, roughly eleven instruments are quoted with confidence: overnight and short-dated rates, then swap rates at one, two, three, five, seven, ten, fifteen, twenty and thirty years. From those eleven points, a value is needed at every date in between — and at forty specific quarterly dates for the swap above, almost none of which coincide with a quoted point.

Figure 3 Eleven observations, a continuous curve, and the two regions where judgement replaces evidence

Between the quoted points, the curve is interpolated. Different interpolation methods pass through the same observed points and disagree everywhere else. Beyond the last quoted point, the curve is extrapolated, which is a polite word for assumed. A position sitting on a quoted node is valued from evidence. A position sitting between nodes is valued from evidence plus a methodology choice. A position beyond the last node is valued from an assumption. Those three sentences explain most of what the fair value hierarchy in Part 1 is trying to capture.

5. An option is a right, not an obligation, and its price contains a forecast

An option gives its holder the right to do something, and the seller the obligation to honour it if the holder chooses. The buyer pays a premium up front for that asymmetry. A swaption — the instrument in Desk SRC-1's Book B — is an option on a swap: the right, at a chosen future date, to enter into an interest rate swap at a rate agreed today.

Why anyone wants one. An insurance company with long-dated liabilities wants protection against rates falling below a level that would make those liabilities unaffordable, without committing to a swap it might not need. It buys a swaption. The bank sells it, collects the premium, and now carries the obligation.

What determines the price is where it becomes interesting. An option is only worth exercising if the market moves in the holder's favour. So the price depends not on where the market is expected to go — that is already in the forward rate — but on how far it is likely to move in either direction. A market that is expected to sit still makes options nearly worthless. A market that could move violently makes the same options expensive.

That "how far it is likely to move" is called volatility, and it is the one input that cannot be observed. Rates can be observed. Dates can be observed. Strike prices are written in the contract. Volatility is a statement about the future, and it is the reason option books are the hardest part of a trading balance sheet to control.

Implied volatility: reading the forecast backwards

Here is the trick the market plays, and it is essential to understand it. A pricing formula takes a volatility assumption and returns an option price. But the market quotes prices, not volatilities. So the industry runs the formula backwards: take the observed price, and solve for the volatility that would produce it. The answer is the implied volatility — the market's collective expectation of future movement, extracted from what people are actually paying.

Implied volatility then becomes the currency in which options are quoted and compared. When the module says the consensus level for twenty-five-year euro swaption volatility is 21.4% against a desk mark of 20.6%, it means the market's implied expectation of annual movement is 21.4% and the desk is asserting a slightly calmer market. On a book with negative vega — a book that profits when volatility falls — asserting a calmer market is asserting a higher value for the book. Which is precisely why an independent check on that number exists.

The surface: implied volatility is not one number

Every combination of option expiry, underlying tenor and strike has its own implied volatility. A one-year option on a two-year swap and a twenty-five-year option on a thirty-year swap are different markets with different expectations. Laid out as a grid, the set of all these implied volatilities is called the volatility surface.

Figure 4 The volatility surface is densely quoted at the centre and empty exactly where the long-dated risk sits

The surface is quoted densely at the centre and sparsely at the edges, because that is where trading happens. Banks hold positions at the edges precisely because that is where clients want protection and where margins are widest. So the parts of the surface with the least market evidence carry a disproportionate share of the risk — a pattern that repeats throughout this module and that is, in a sentence, why valuation control is difficult.

5b. Credit is a third market, and its unobservable input is correlation

Rates and volatility have been built from nothing. Credit needs the same treatment, because the third book in the running example — the bespoke tranches — is the most Level 3 part of the balance sheet and the least intuitive.

A credit default swap is insurance on a company

One party pays a regular premium; the other pays out if a named company fails to meet its obligations. That is the whole contract. The premium is quoted in basis points per year on the protected notional and is called the credit spread. Protection on a company at 150 basis points means the buyer pays 1.5% of the notional each year, and receives compensation if the company defaults.

The spread is a price, and prices contain information. A market charging 150 basis points a year is making an implicit statement about how likely default is and how much would be lost if it happened. Two quantities separate that statement.

The recovery rate is what a creditor gets back after a default. Historically, for senior unsecured corporate debt, the market convention sits around 40%, meaning 60% is lost. That 60% is the loss given default.

The hazard rate is the annual probability of defaulting, given survival so far. The relationship between the three is close to arithmetic: over short horizons, spread ≈ hazard rate × loss given default. A 150 basis point spread with a 60% loss given default therefore implies a hazard rate of about 2.5% per year.

Compound that and you have the default probability over any horizon. Surviving five years at a 2.5% annual hazard has probability e^(−0.025 × 5) ≈ 88.3%, so the five-year default probability is about 11.7%. Notice what just happened: an observable market price was converted into a probability through two assumptions. The recovery assumption is one of them, and it is not observable. Changing recovery from 40% to 30% changes the implied hazard rate by a quarter.

A tranche slices the losses rather than the portfolio

Now take sixty companies instead of one. A portfolio of protection on sixty names, €10 million each, gives €600 million of exposure. Losses arrive one default at a time and accumulate.

A tranche is a horizontal slice of that accumulating loss, not a share of the portfolio. It is defined by two numbers. The attachment point is the level of cumulative loss at which the tranche starts to be hit. The detachment point is the level at which it is entirely wiped out. A 3–7% tranche absorbs nothing until portfolio losses reach 3% of €600 million, then absorbs everything between 3% and 7% — €24 million of exposure — and is exhausted beyond that.

Figure 5 Losses fill the stack from the bottom; a tranche is a slice of that stack, not a slice of the portfolio

This structure creates an input that exists nowhere else in finance. The value of a tranche does not depend only on how many defaults are expected. It depends on how they arrive together.

Default correlation, and why it cannot be looked up

Consider two extreme worlds, both with the same expected number of defaults.

In the first, defaults are entirely independent. Companies fail for their own idiosyncratic reasons. Over five years you get roughly the expected seven defaults, almost every time, clustered tightly around that number. Losses of around 7% of the portfolio are close to certain — which destroys a 3–7% tranche completely, and never touches a 7–15% tranche at all.

In the second, defaults are highly correlated. Companies fail together or not at all, because they fail for shared macroeconomic reasons. Most of the time there are very few defaults; occasionally there are very many. The average is the same seven, but the distribution has moved to the extremes. A 3–7% tranche now has a real chance of surviving untouched, and the senior tranche above it has a real chance of being hit.

Correlation therefore redistributes risk between tranches without changing the total. It is the single most important input to a tranche valuation, and — this is the point — it is not quoted anywhere. There is no market in default correlation. It is inferred from the prices of standardised index tranches, then mapped onto a bespoke portfolio through a procedure that is itself a modelling choice.

6. The Greeks are slopes, and they are the vocabulary of every explanation

A sensitivity — universally called a Greek — answers one question: if this one market factor moves by one unit, how much does the position's value change? Mathematically it is a derivative in the calculus sense, a slope. Practically it is the unit in which risk is discussed, hedged and, most importantly for this module, in which a day's profit and loss is explained.

Greek The factor that moves Reading it
Delta / DV01 The underlying rate or price The slope: value change per basis point
Gamma The delta itself The curvature: how much the slope changes as the market moves
Vega Implied volatility Value change per one point of implied volatility
Theta Time Value change from one day passing, everything else held still
CS01 Credit spread Value change per basis point of credit spread

Why they matter here rather than on the trading desk. If a book's delta is −€1.9 million per basis point and rates rose nine basis points, the book should have gained about €17.1 million from that cause alone. If it did not, either the delta is wrong, the rate move is wrong, or something happened that nobody has accounted for. The Greeks are what make a profit and loss figure testable rather than merely reported, and the test is the subject of Step 5 in Part 2.

Two cautions. A Greek is a local measurement: delta describes the slope where the market is now, and becomes progressively less accurate as the market moves away, which is what gamma corrects for. And a Greek is only as good as the model that produced it — a sensitivity computed from a mis-specified model will explain a profit and loss that never happened.

7. Nothing here is exact, and that is the working assumption

A final orientation, because it reframes everything that follows. Newcomers to this domain often expect that a sufficiently good system produces the right number. It does not, and it is not trying to.

Every valuation described in this module rests on a curve someone interpolated, a volatility someone extrapolated, a correlation someone assumed, and a model someone approved with documented limitations. The output is not a fact. It is a best estimate surrounded by uncertainty, and the professional skill lies in knowing how wide that uncertainty is and saying so in a number.

This is why the discipline distinguishes between an adjustment and a reserve, a distinction that recurs from Part 2 onward. When independent evidence shows the estimate is wrong, you correct it — that is an adjustment. When independent evidence shows only that the estimate could reasonably be anywhere within a range, you cannot correct it, so you quantify the range and hold money against it — that is a reserve. A control framework that produces only point estimates is not being precise. It is concealing what it does not know.

  1. Why does unrealised profit require a control function while realised profit does not?

  2. Why is the present value of a distant cash flow more sensitive to the curve than a near one?

  3. Why is a €1 billion notional book not necessarily riskier than a €100 million one?

  4. Why is an interest rate swap worth zero at inception but not thereafter?

  5. What does implied volatility actually measure, and why is it derived backwards?

  6. Why do banks hold most of their option risk in the least-quoted part of the volatility surface?

  7. Two portfolios have the same expected number of defaults but different default correlation. Why are their tranches worth different amounts?

  8. Why is a credit spread not the same thing as a default probability?

  9. What is the difference between an adjustment and a reserve, and why does it matter?

1Valuation is a manufactured number, and someone has to own the manufacturing

Before any process can be explained, one question has to be answered honestly: what does it mean to say a financial position is worth a particular amount? The answer is less obvious than it looks, and everything else in this module follows from it.

A price is an observation; a value is a construction

Start with the simplest possible case. A bank owns one thousand shares of a company listed on a large exchange. At the close, the last traded price was €42.30. The position is worth €42,300. Nobody needs a department for this. The price was observed, in size, in a market where anyone could have transacted at approximately that level. The number is not an opinion.

Now change one thing. The bank owns an interest rate swap: it pays a fixed rate of 2.85% annually on a notional of €250 million for ten years, and receives a floating rate that resets every three months. No exchange quotes this contract. It was negotiated bilaterally with a corporate client. There is no last traded price because it has never traded since the day it was struck.

The swap still has a value, but that value has to be built. Its worth today is the difference between what the bank expects to pay and what it expects to receive, over ten years, discounted back to today. To compute that, three separate constructions are needed. First, a projection of where the floating rate will reset at forty future dates — which requires a forward curve. Second, a set of discount factors to bring each future cash flow to a present value — which requires a discount curve, and since 2020 that curve is built from overnight index swap rates, not from the old interbank offered rates. Third, an assumption about whether the counterparty will actually pay, which introduces a credit adjustment.

None of those three things is quoted directly. Each is derived from other instruments that are quoted, through a calibration procedure that involves choices: which instruments to include, how to interpolate between them, how to extrapolate beyond the longest quoted maturity. Two competent banks running the same trade will produce present values that differ, and the difference will not be trivial on a ten-year notional of €250 million.

Now go one step further. The bank owns a bespoke single-tranche credit portfolio product referencing sixty corporate names, with attachment and detachment points negotiated for one client. Its value depends on the default probability of each name, on the correlation between those defaults, and on the recovery expected in default. Default correlation is not observable in any market. It is inferred from the prices of standardised index tranches, then mapped onto a bespoke portfolio through a procedure that is itself a modelling choice. There is no number to look up. There is only a model, an input the modeller selected, and a defence of both.

The accounting framework names the spectrum: the fair value hierarchy

International accounting standards give this spectrum a formal structure. Under IFRS 13, and its near-identical American counterpart ASC 820, fair value is defined as the price that would be received to sell an asset in an orderly transaction between market participants at the measurement date. That definition contains three words that do a lot of work: would be received, meaning an exit price, not an entry price or a hold-to-maturity value; orderly, meaning not a forced liquidation; and market participants, meaning what a knowledgeable third party would pay, not what this particular bank thinks the position is worth to itself.

The standard then classifies every position into one of three levels according to the observability of the inputs used.

Figure 6 The fair value hierarchy translates directly into how much control effort a position demands

Level 1 is a quoted price in an active market for the identical instrument. A Bund future. A share in a large listed company. A government bond that trades in size every minute. No adjustment is permitted; the quoted price is the fair value.

Level 2 uses observable inputs but requires a model. A vanilla interest rate swap sits here: no quote exists for the specific contract, but every input to its valuation — the swap rates, the basis spreads, the discount curve — comes from observable market instruments. A liquid credit default swap index sits here. So does a listed option on an underlying with a liquid, quoted volatility surface.

Level 3 means at least one input that materially affects the value is unobservable. The bespoke tranche is Level 3 because of the correlation assumption. A thirty-year currency option on an emerging market pair is Level 3 because nobody quotes volatility at that tenor. An illiquid loan held at fair value is Level 3 because there is no market.

The classification is not merely a disclosure requirement. It determines how much verification work the position attracts, how large the reserve against it must be, how much the external auditor will test, and how much regulatory capital the bank must hold against valuation uncertainty. The hierarchy is a work-allocation map disguised as an accounting note.

Not everything is fair-valued, and the accounting basis decides who does the work

A question that precedes every other question in this module: why is this position measured at a value at all? A bank holds a corporate loan and a corporate bond. Economically they are close cousins — both are money lent to a company against a promise to repay. One is carried at the amount lent, adjusted for expected credit losses, and never revalued to market. The other is revalued every night. The difference is not economic. It is a classification decision made under an accounting standard, and it determines whether Product Control ever sees the position.

Three measurement bases, one decision tree

Under IFRS 9, the standard governing financial instruments, a financial asset is measured on one of three bases, and the classification follows from two tests applied at initial recognition.

The business model test asks what the bank intends to do with the asset: hold it to collect contractual cash flows, hold it to collect and sell, or hold it for trading.

The cash flow characteristics test asks whether the contractual cash flows are solely payments of principal and interest on the principal outstanding — the criterion usually abbreviated to SPPI. A vanilla loan passes. A note whose coupon depends on an equity index does not.

Basis Applies when Revalued? Where gains and losses go
Amortised cost Held to collect, and cash flows are solely principal and interest No — carried at cost less impairment Income statement, via interest and credit losses
Fair value through other comprehensive income (FVOCI) Held to collect and sell, and cash flows are solely principal and interest Yes Equity reserve, recycled to income on sale
Fair value through profit or loss (FVTPL) Everything else: trading intent, or cash flows failing the SPPI test Yes Income statement, immediately

All derivatives are measured at fair value through profit or loss, without exception and regardless of intent, because a derivative's contractual cash flows never satisfy the principal-and-interest test. That single rule is why a derivatives desk generates a control burden out of all proportion to its balance sheet footprint: every contract it writes lands in the category that must be revalued nightly and explained daily.

The proportions across the European banking system are striking. Financial assets measured at fair value totalled approximately €7,279 billion at the end of 2018 — around a quarter of EU banks' aggregate balance sheet — against €17,613 billion carried at amortised cost, or 61%. Roughly three fifths of a European bank's assets are therefore invisible to the daily valuation process entirely. The control function's scope is defined not by what the bank owns but by how the accounting classified it.

The trading book boundary is a capital decision disguised as a bookkeeping one

Alongside the accounting classification sits a second, prudential one: is a position in the trading book or the banking book? The trading book contains positions held with intent to trade or to hedge trading positions; everything else is banking book. The two attract entirely different capital treatments — market risk capital on one side, credit risk capital on the other — and the difference can be substantial for the same instrument.

Which creates an obvious incentive and therefore an obvious control. Supervisors have treated boundary arbitrage as a recurring concern for two decades, and the Fundamental Review of the Trading Book tightened the definition specifically to address it: the boundary became rules-based rather than intent-based, reclassification between books requires supervisory approval and is expected to be exceptional, and any capital benefit arising from a reclassification is disallowed. For a control function, the practical consequence is that the trading-versus-banking-book flag on a trade record is not reference data. It is a controlled attribute whose changes must be evidenced and approved.

Hedge accounting: keeping the hedge and the hedged item in step

A structural mismatch arises whenever a derivative hedges something that is not fair-valued. A bank issues fixed-rate debt carried at amortised cost and swaps it to floating. The swap is at fair value through profit or loss and moves every day. The debt does not move at all. The result is reported volatility that reflects no economic exposure whatsoever — the two instruments offset in economic terms and diverge violently in accounting terms.

Hedge accounting is the remedy, and it is granted conditionally. Where a bank formally designates a hedging relationship, documents it at inception, and demonstrates that the hedge and the hedged item are economically offsetting, the accounting is adjusted so that the two move together: in a fair value hedge, the hedged item is revalued for the risk being hedged; in a cash flow hedge, the effective portion of the derivative's movement is deferred in equity.

The conditionality is where the operational burden sits. Designation must be documented before the fact, not asserted afterwards. Effectiveness must be demonstrated and monitored. Any ineffective portion goes straight to the income statement and must be measured, reported and explained — which makes it, in practice, a line in the profit and loss attribution that Product Control owns.

Its scale is modest but its failure mode is not: only about 6% of derivatives held by EU banks are designated for hedge accounting, yet a hedge relationship that fails its effectiveness test is de-designated, the accounting protection is lost retrospectively, and volatility appears in reported earnings that management has already told the market would not be there.

Where the function sits inside the bank

Product Control is the accounting and control layer that sits directly behind the trading desk. In organisational terms it belongs to Finance, reporting through a divisional chief financial officer to the group chief financial officer. It does not report to the head of trading. That reporting line is not an administrative detail — it is the entire basis of the function's usefulness, and Part 3 returns to it at length.

Around Product Control sit four adjacent functions that are frequently confused with it.

Market Risk measures the potential loss on the same positions under adverse scenarios, producing value-at-risk, stressed value-at-risk, sensitivities and limit utilisations. It uses many of the same inputs and often the same pricing library, but its question is forward-looking: what could happen. Product Control's question is backward-looking: what did happen, and what is it worth now.

Model Validation, sometimes called Model Risk Management, independently assesses whether a pricing or risk model is fit for purpose before it is approved for production use, and re-assesses it periodically thereafter. Product Control consumes approved models; it does not approve them. When a model is found deficient, the two functions negotiate a valuation adjustment to cover the deficiency until it is fixed.

Financial Control and Group Finance own the general ledger, the statutory accounts and the external reporting. Product Control feeds them a signed, substantiated valuation and profit and loss; they consolidate it, apply group accounting policy and publish it.

Operations, sometimes called Middle Office or Trade Support, owns trade lifecycle events: confirmation, settlement, collateral, corporate actions, and the reconciliation of the bank's records against the counterparty's. When Operations and Product Control disagree about what trades exist, the profit and loss cannot be signed.

These five functions typically share three pieces of infrastructure and diverge on everything else. They share the trade capture system (the books and records of what the bank owns), the market data platform (the curves and surfaces), and increasingly the pricing library (the analytics that turn one into a valuation). They diverge on the calculation engines, the reporting layers and the workflow tools. Where a bank has succeeded in making all five consume a single golden source of trades and a single golden source of market data, the reconciliation burden collapses. Where it has not — and most have not entirely — a substantial part of the control effort is spent explaining why two systems disagree about the same position.

The function has a product catalogue, and it is ranked by difficulty

It helps to think of Product Control as producing deliverables on a schedule, the way a trading desk produces trades. The deliverables differ in frequency, difficulty and consequence. Ranking them exposes where the intellectual and technological effort concentrates.

Deliverable Frequency What it answers Difficulty Principal technology
Flash P&L Daily, pre-open Roughly what did the desk make? Low Overnight batch, risk engine
P&L attribution (explain) Daily, T+1 morning Why did the desk make it? Medium Risk engine, sensitivity store, attribution tool
Balance sheet substantiation Daily to monthly Does the ledger agree with the trades? Medium Reconciliation engine, sub-ledger
Independent price verification Monthly (weekly for risky books) Is the desk's mark defensible? High IPV platform, consensus data, quote capture
Valuation adjustments and reserves Monthly What is the honest exit value? High Reserve calculators, XVA engine
Fair value hierarchy and disclosure Quarterly How observable is the book? Medium Classification engine, disclosure pack
Prudent valuation (AVA) Quarterly What number will the regulator accept for capital? High AVA calculator, regulatory reporting
New product and model sign-off Event-driven Can we book, value and control this at all? Very high Model inventory, NPA workflow

The first three are production; the last five are judgement. Production can be automated almost entirely and, in a well-run bank, is. Judgement can be supported by technology but not replaced by it, which is why the headcount and the seniority in modern Product Control functions have migrated towards independent price verification, reserves and new product approval, while daily profit and loss production has been industrialised, standardised and in many banks moved offshore.

The reference case for this module is a Level 3 derivatives book

The whole of Part 2 follows one fictional desk. Its name is Desk SRC-1, a structured rates and credit desk inside the markets division of a mid-sized European Corporate & Investment Bank. It runs three kinds of position: a large book of vanilla interest rate swaps used to hedge, a smaller book of Bermudan swaptions sold to corporate and insurance clients, and a residual portfolio of bespoke credit tranches inherited from a business the bank has since stopped writing.

That mix is chosen deliberately. The vanilla swaps are Level 2, high-volume and controllable by automation. The Bermudan swaptions are Level 2 in normal conditions but depend on a volatility surface whose long-dated corner is thinly quoted, and on a model choice — one-factor or two-factor — that materially changes the number. The bespoke tranches are unambiguously Level 3 and consume a disproportionate share of the control effort relative to their size. Every real trading book has this shape: a large, easy majority and a small, hard minority that generates most of the risk of being wrong.

What the function costs, and what being wrong costs

Product Control is a large, expensive and rarely discussed part of a Corporate & Investment Bank. A universal bank with a full markets franchise typically runs a Product Control and valuation function of several hundred to well over a thousand people globally, distributed across trading centres and offshore processing hubs. The function is a pure cost centre. It generates no revenue and its budget is under permanent pressure, which shapes almost every technology decision described in Part 4: automation is not pursued for elegance, it is pursued because the alternative is more headcount that the business will not fund.

The counterweight to that pressure is the cost of failure, and the cost of failure is not proportionate to the cost of the function. Three categories of consequence recur.

Restatement. If reported profit and loss was materially wrong, prior financial statements must be corrected in public. The direct accounting effect is often smaller than the loss of credibility with investors and the supervisory scrutiny that follows.

Regulatory penalty and capital add-on. Supervisors that lose confidence in a bank's valuation processes impose capital requirements against the weakness, which is expensive and persistent, and they can restrict the business until the weakness is remediated.

Direct loss. Positions carried at an optimistic mark eventually have to be exited at a real one. The gap between the two is realised as a loss on the day the position is closed, and because mismarking tends to grow while it is undetected, the realisation is usually large and sudden.

Where the balance sheet actually sits, in real numbers

The distribution of fair-valued assets across the hierarchy is public information, and it shapes everything about how the function is resourced. Across a sample of EU banks representing roughly 90% of the sector's total assets, the position at the end of 2018 was as follows.

Category Amount Share of total assets Share of fair-valued assets As % of CET1 capital
Level 1 €2,379bn 8.1% 33% 142%
Level 2 €4,600bn 15.7% 63% 275%
Level 3 €300bn 1.0% 4% 18%
All fair-valued €7,279bn 25% 100% —
Amortised cost €17,613bn 61% — —

Source: European Systemic Risk Board, Macroprudential implications of financial instruments in Levels 2 and 3 for accounting purposes, February 2020.

Four readings of that table matter.

Level 2 is where the volume is, at nearly two thirds of fair-valued assets and 275% of the sector's core capital. Level 2 means a model applied to observable inputs — which is precisely the category that automation can control at scale, provided inputs are captured cleanly and models are versioned and approved. The industrialisation of Level 2 valuation control is consequently the single largest technology programme most banks have run in this space.

Level 3 is small and moves fast. At €300 billion it was 4% of fair-valued assets and 1% of the balance sheet. But it had risen from €191 billion a year earlier — an increase of 57% in twelve months. A category that can grow by more than half in a year is not a stable residual to be managed by exception; it is the part of the book where the control framework has to be able to scale on short notice.

Derivatives and Level 2 are almost the same population. Some 95% of derivatives held by EU banks are classified Level 2, while 72% of debt securities are Level 1. This is why a derivatives franchise generates a control burden out of proportion to its size: it concentrates almost entirely in the category that requires a model, an input and a defence.

Concentration is extreme. In the corresponding supervisory transparency exercise, five banks accounted for half of all Level 2 assets in the sample, and ten banks for half of all Level 3. Most institutions in the sector carry very little of this problem; a small number of large dealers carry almost all of it, which is why valuation control capability is a defining characteristic of a Corporate & Investment Bank rather than a general banking competence.

What valuation uncertainty costs in capital

Prudent valuation puts a price on the whole problem, and the price is public. European banks must compute additional valuation adjustments quantifying valuation uncertainty and deduct them directly from common equity tier 1 capital. In the impact analysis underpinning the technical standards, the expected adjustment averaged 1.5% of an institution's CET1 capital, or approximately €227 million per institution.

Two thresholds define the regime in practice. Institutions with total absolute fair-valued assets and liabilities below €15 billion may use a simplified approach, computing the adjustment as 0.1% of the fair-valued amount. Everyone above that threshold uses the core approach, with its nine prescribed categories and a target certainty level of 90%.

An earlier data point puts the scale in perspective. The European Central Bank's 2014 comprehensive assessment of bank balance sheets identified €4.6 billion of additional valuation adjustments arising specifically from Level 3 assets, within a total of €37 billion of asset value adjustments across the exercise — alongside a finding that only 67% of banks applying international accounting standards had a clear written policy defining what constitutes an active market, and only 72% had one defining a Level 3 input.

The organisational question: where should the function live?

Banks have arrived at meaningfully different answers, and the differences have consequences.

Product Control inside Finance is the majority model and the one assumed throughout this module. The function reports to the chief financial officer, which gives it the independence that comes from a separate reporting line and ties it naturally to the general ledger and the statutory accounts. Its weakness is distance from the risk conversation: a controller who does not sit near the risk manager may verify a price without understanding the exposure behind it.

Valuation Control inside Risk places independent price verification and reserve setting in the risk division rather than Finance, on the argument that valuation uncertainty is a risk to be measured rather than an accounting entry to be booked. It integrates well with model validation and with prudent valuation, which is a capital calculation. Its weakness is a weaker link to the ledger and to audit.

The hybrid, now common at large institutions, splits the two: daily profit and loss production and substantiation sit in Finance, while independent price verification, reserves and prudent valuation sit in a Valuation Control Group that may report into Risk or into a joint governance structure. This resolves the trade-off at the cost of an interface that has to be managed.

The offshore or nearshore delivery model is orthogonal to the three above and now near-universal. Repetitive production — reconciliation breaks, standard reporting, first-level review — is performed in a lower-cost centre, while judgement, challenge and sign-off remain in the trading centre. It works when the process has been genuinely standardised and documented; it fails, expensively, when what was actually being exported was undocumented judgement.

Model Reporting line Strongest at Weakest at
PC in Finance Divisional CFO Ledger integrity, audit defence, independence Distance from the risk conversation
VCG in Risk Chief Risk Officer Model linkage, prudent valuation, capital view Weaker ledger and disclosure linkage
Hybrid (split) Both, with joint committee Covers both ends properly Interface management, duplicated data
Offshore delivery Any of the above Unit cost of production Fails where judgement was undocumented

The discipline has a cycle, and technology investment follows it

Valuation control as a distinct, heavily resourced discipline is largely a creation of the last two decades. Its trajectory explains why bank valuation infrastructure looks the way it does today, including the parts that look irrational.

Before 2007, the function existed but was thin. Traders marked their own books with light independent challenge. Complex structured credit positions were valued using models and inputs the desks themselves had specified. The industry assumption was that observable markets would remain observable.

The 2007–2009 crisis destroyed that assumption in the most direct possible way. Markets for structured credit stopped functioning; positions that had been marked from a model calibrated to trades that no longer occurred proved to be worth a fraction of their carrying value. Write-downs across the industry ran into hundreds of billions of dollars, and a material portion of them were valuation corrections rather than credit losses in the ordinary sense. Supervisors concluded that independent price verification, reserve policy and model governance had been structurally inadequate.

The decade that followed was a build-out. Independent price verification became mandatory in substance and in scope. Consensus pricing services, which pool anonymised marks from contributing banks and return statistical levels, moved from a curiosity to the primary evidence source for illiquid instruments. Prudent valuation entered European regulation as a hard capital deduction, forcing banks to quantify valuation uncertainty and subtract it from capital. Model risk management became a supervised discipline in its own right, driven in the United States by the SR 11-7 guidance on model risk management and in Europe by the targeted review of internal models. The XVA family — credit, debit, funding and capital valuation adjustments — moved from a niche derivatives-desk topic into mainstream valuation policy.

Since roughly 2015 the pressure has reversed. The controls are in place, the regulatory expectations are largely stable, and the question has become cost. The Fundamental Review of the Trading Book, developed through the 2010s and still being phased in across major jurisdictions, tightened the linkage between valuation quality and capital: its profit and loss attribution test makes the accuracy of a desk's risk-based explanation of its own profit and loss a direct determinant of whether that desk may use internal models for capital. A desk that cannot explain its profit and loss loses the capital benefit of its models. That is a rare and instructive case of a regulator putting a price on a control metric.

The implementation timetable has itself become a live issue. In the European Union the market risk requirements were postponed by two years to 1 January 2027, with temporary and targeted amendments applying for three years from that date until January 2030 — a response to the expectation that other major jurisdictions would delay further. The United Kingdom is implementing the wider framework from 1 January 2027 but delaying the internal model approach for market risk to 1 January 2028, while retaining the tightened trading book boundary and the standardised approach on the earlier date. In the United States the timetable remained unsettled through 2026. For a control function this uncertainty is not academic: the profit and loss attribution infrastructure has to be built and running in parallel production well before the date it starts to determine capital, and the date keeps moving.

The three questions Product Control answers every day

Reduced to essentials, the function answers three questions in sequence, and the order matters.

First: do we agree on what we own? Every valuation is worthless if the population of trades is wrong. A trade booked twice, booked with the wrong maturity, or missing entirely produces a valuation error that no amount of price verification will detect. Substantiation and reconciliation come first for a reason.

Second: what is each position worth, and how confident are we? This is the valuation question proper: the model, the inputs, the verification against independent evidence, and the reserve that quantifies what verification could not resolve.

Third: why did the value change since yesterday? This is the attribution question, and it is the one that catches errors the first two miss. If the desk's profit and loss can be fully explained by the movements in the market factors the desk is exposed to, the valuation process is probably working. If a large residual cannot be explained, something is wrong — a bad input, a missing trade, a model behaving unexpectedly, or a position nobody knew about.

The third question is the most powerful control in the entire framework, and it is the one this module treats in most detail. It is powerful because it is a closed test: the profit and loss has to reconcile to the sum of its explained parts, and the residual is a number that cannot be argued away.

  1. Why does a ten-year interest rate swap require a valuation construction while a listed share does not?

  2. What exactly does "exit price" mean in the IFRS 13 definition of fair value, and why is it the right basis?

  3. Why is Level 3 classification a property of the market rather than of the instrument?

  4. Where does most of the fair-valued trading balance sheet sit, and why does that determine the technology strategy?

  5. How does the Fundamental Review of the Trading Book put a price on a control metric?

  6. Why does an interest rate swap generate control work while an economically similar corporate loan may not?

  7. Why must substantiation precede price verification?

  8. What is the structural asymmetry in the economics of the control function?

1

  1. Why does a ten-year swap need a valuation construction when a listed share doesn't?

  2. What does “exit price” mean in IFRS 13?

  3. Why is Level 3 a property of the market, not of the instrument?

2The number is manufactured overnight, on a production line with no slack

This part follows one trading day and one month-end, stage by stage. The daily cycle has a hard deadline: the desk opens whether or not the control process finished. The monthly cycle has a softer deadline but a harder question. Together they constitute the whole of the process, and everything in Parts 3, 4 and 5 refers back to them.

The whole cycle in one table

Two cycles run simultaneously on the same positions. The daily cycle produces a profit and loss figure and an explanation of it. The periodic cycle — monthly for most books, weekly for the riskiest — verifies that the values feeding the daily cycle are defensible and quantifies what verification could not settle. The daily cycle is fast, automated and unforgiving of delay. The periodic cycle is slow, manual and unforgiving of imprecision.

Stage Timing Core activity Primary output Who leads
1. Trade capture Intraday, continuous Booking, amendment, confirmation Complete trade population Front Office / Operations
2. Market data close 17:00–19:00 local Snapshot, curve and surface construction Official closing market data Market Data / Quant
3. Overnight revaluation 19:00–04:00 Full portfolio pricing and sensitivity generation Position values and Greeks Batch / Risk engine
4. Flash P&L 04:00–08:30 Aggregation, first-pass validation Preliminary P&L by desk Product Control
5. P&L attribution 08:30–11:00 Decomposition into risk factors Explained P&L and residual Product Control
6. Sign-off and escalation 11:00–17:00 Desk review, investigation, adjustment Signed daily P&L PC and desk head
7. Ledger close and recon T+1 to T+2 Sub-ledger posting, reconciliation Substantiated balance sheet Financial Control
8. Independent price verification Month-end +1 to +5 Challenge of desk marks against evidence IPV differences and adjustments Valuation Control
9. Valuation adjustments Month-end +3 to +7 Reserve calculation and posting Adjusted fair value Valuation Control
10. Hierarchy, prudent value, disclosure Month-end +5 to +8 Classification, AVA, reporting Disclosures and capital numbers PC and Group Finance
Figure 7 Effort and risk concentrate twice: on the overnight critical path, and again at month-end verification

Where the pressure actually sits

The daily peak is not where most people expect it. Trade capture is continuous and rarely dramatic. The overnight batch is long but unattended. The pressure concentrates between roughly 04:00 and 11:00, when a small number of people have to determine whether an automatically produced number is right, find out why if it is not, and either fix it or explain it before the desk needs to trade on the basis of yesterday's position.

The reason this window is unforgiving is structural. The batch is a chain, and every link is upstream of a fixed deadline. If the market data close is late, curve construction is late. If curve construction is late, revaluation is late. If revaluation is late, there is no flash profit and loss, and the desk opens without knowing its position or its risk. The trading day does not move for the control process. This is the single most important operational fact about the function, and it drives most of the architecture decisions in Part 4.

The monthly peak is different in character. Nothing breaks if independent price verification runs a day late. What happens instead is that a wrong number goes into the accounts, and the wrongness is discovered later, in front of an auditor, when the cost of correction is far higher than the cost of doing it properly the first time.

The running case: Desk SRC-1 on 14 October

One desk, one day, carried through every stage. Desk SRC-1 is the structured rates and credit desk introduced in Part 1. On the morning of 14 October it holds 8,400 live trades with a mark-to-market value of €1.9 billion. Overnight, European rates sold off: the ten-year swap rate rose by nine basis points. Implied swaption volatility fell slightly. Credit spreads on the tranche portfolio were broadly unchanged, but one reference name in a bespoke tranche was downgraded.

By the time the day is over, the desk will report a profit of €1.64 million, of which €0.09 million cannot be attributed to any market factor. That residual is where the interesting part of the story lives. All figures in this running case are illustrative.

2

  1. Where does daily pressure concentrate in the valuation cycle?

  2. Who leads independent price verification?

  3. What does the periodic cycle verify?

Step 1 — Nothing can be valued until the bank agrees with itself about what it owns

The most expensive valuation errors are not wrong prices. They are right prices applied to the wrong population of trades.

The books and records are the foundation, and they are not automatic

A trade becomes real to the bank at the moment it is booked into the trade capture system. For an exchange-traded future, that happens automatically: the execution flows electronically from the exchange through the order management system into the position. For an over-the-counter derivative negotiated by voice with a corporate treasurer, someone types it in. The gap between the two is where the population goes wrong.

Three failure modes recur and each has a specific control.

A trade is booked with incorrect terms. A maturity date a year out, a notional with a misplaced decimal, a floating leg referencing the wrong index. The control is confirmation: the counterparty independently states its version of the trade and the two are matched. For standardised derivatives this happens electronically within hours, often through an industry matching platform. For bespoke trades it may involve a negotiated legal confirmation that takes days.

A trade is booked in the wrong book. This sounds administrative and is not. The book determines which desk owns the profit and loss, which risk limits apply, whether the position is in the trading book or the banking book for capital purposes, and how it is valued. A hedge booked into the wrong book leaves one desk apparently unhedged and another apparently speculating.

A trade exists in the front-office system and not in the ledger, or the reverse. This is the reconciliation problem, and it is the reason substantiation exists as a separate discipline. The front-office system is the trader's view of the world; the sub-ledger is the accountant's. They are fed from the same source in principle and diverge in practice, because of timing differences, failed feeds, manual adjustments and amendments applied to one and not the other.

The controls here are almost entirely systemic. Confirmation matching platforms compare the two counterparties' trade records field by field and raise exceptions. Position reconciliation runs daily between the trade capture system and the sub-ledger, and between the bank's records and the clearing house's or the prime broker's. Static data validation checks that every trade references a valid counterparty, a valid instrument, a valid book and a valid legal entity before it can be saved. A dedicated reconciliation platform typically sits across all of these, with configurable matching rules, tolerance thresholds and a break-management workflow.

The measure of maturity here is the age profile of unreconciled breaks, not their count. A thousand breaks all under one day old is a timing artefact. Forty breaks over thirty days old is an unquantified valuation exposure.

Amendments and cancellations are where profit and loss can be manufactured

A trade booked yesterday can be amended today, and the amendment changes yesterday's value. Most amendments are legitimate corrections. Some are not. The pattern that control functions look for is a trade cancelled and re-booked at a different price shortly after a loss appears, or a series of amendments to off-market rates that shift value between books or between periods.

The control is that amendments to prior-day trades generate a specific profit and loss line — usually called amendment or prior-day adjustment — which is reported separately and never buried in market movement. A desk with a persistent stream of amendment profit and loss is being asked why.

The internal gates before a new trade type can exist at all

Before Desk SRC-1 can book a product it has never traded, the product has to pass through a formal gauntlet. This is the valuation equivalent of the credit and commitment committees that govern a bank's lending decisions, and it exists for the same reason: to stop the front office from unilaterally creating an obligation the rest of the institution cannot manage.

New Product Approval

The New Product Approval process asks whether the institution can transact, book, value, risk-manage, settle, account for and report a proposed product. It is a cross-functional sign-off in which each function attests to its own readiness. Product Control's specific attestation is that the product can be valued with an approved model, that its inputs can be independently verified, and that a valuation adjustment framework exists for the parts that cannot be. A product that Product Control cannot verify does not get approved, or is approved with restrictions — a notional cap, a maturity limit, a requirement for a punitive reserve until an independent price source is established.

Model approval and the model inventory

No pricing model may be used in production until it has been independently validated and formally approved. Validation examines the mathematical derivation, tests the implementation against an independent benchmark, probes behaviour at boundaries and under stress, and documents known limitations. Approval is granted with a defined scope — this model, for these products, within these parameter ranges — and with an expiry requiring periodic revalidation.

The output that matters operationally is the model inventory: the authoritative register of every model in production, its version, its owner, its approval status, its limitations and the valuation adjustment held against those limitations. When a model is used outside its approved scope, that is a finding. When a model's limitations are known and no adjustment is held against them, that is a larger finding.

The Valuation Committee

Disputes that the daily and monthly processes cannot settle escalate to a Valuation Committee. It is chaired from Finance, includes senior representation from Product Control, Market Risk, Model Validation and the trading business, and meets at least monthly. Its agenda is the small number of positions where the desk and the control function genuinely disagree about value, where a reserve methodology is being changed, where a position is being reclassified between fair value levels, or where a model limitation requires a judgement call.

The committee is a decision-making body, not a discussion forum. Its minutes are evidence: when an external auditor or a supervisor asks why a €96 million tranche portfolio is carried at that number, the answer is a documented committee decision with a stated rationale, not a controller's professional instinct.

Gate Question it answers Who must sign Typical failure
New Product Approval Can we transact and control this at all? All functions, chaired by Risk or COO Product traded under a "similar existing product" exemption
Model approval Is the mathematics right and implemented correctly? Model Validation Model used outside approved scope
Model inventory review Do we know what is running in production? Model Risk Management Undocumented spreadsheet models in use
Valuation Committee Whose number stands where we disagree? CFO or delegate Decisions taken but not minuted

Step 2 — The closing market data is a decision, not an observation

Every position in the bank is valued against a single official set of curves and surfaces. Constructing that set is the most consequential and least visible step of the entire cycle.

Snapshot, not stream

At a defined moment each evening, the bank freezes a set of market observations and declares them official. This is the close snapshot. Everything valued that night is valued against it. The choice of moment is a policy decision with real consequences: a European bank typically closes its rates data at or shortly after the European market close, but its dollar positions continue to trade for hours afterwards in New York. A move in the American afternoon appears in tomorrow's profit and loss, not today's.

Consistency matters more than precision here. The valuation must use the same snapshot as the risk calculation, which must use the same snapshot as the profit and loss attribution. If the risk engine took a 17:00 snapshot and the valuation engine took 17:30, the attribution will show an unexplained residual that has nothing to do with any real error — a phenomenon so common it has a name in most banks, and so wasteful of investigative effort that eliminating it is often the single highest-return fix in the whole chain.

From quotes to a curve: the construction problem

A discount curve is not quoted. It is fitted. The raw material is a set of quoted instruments across maturities: overnight index swap rates at short tenors, then par swap rates at one, two, three, five, seven, ten, fifteen, twenty and thirty years. Between and beyond those points, the curve must be interpolated and extrapolated. The choices are real.

Which instruments enter the fit. Including a thinly traded twenty-five-year point may add noise rather than information. Excluding it may lose genuine signal.

How to interpolate. Linear interpolation on zero rates, linear on the logarithm of discount factors, or cubic spline on forward rates all produce the same values at the fitted points and different values everywhere else. A ten-year swap sitting exactly on a quoted node is insensitive to the choice. A structured trade with a cash flow at eleven years and three months is not.

How to extrapolate. Beyond the longest liquid quote — often thirty years in major currencies, much shorter elsewhere — the curve is an assumption. Positions in that region are candidates for Level 3 classification and for an extrapolation reserve.

How to handle the multi-curve reality. Since the financial crisis, and definitively since the transition away from interbank offered rates, a single curve no longer serves both purposes. Cash flows are projected off a forecasting curve specific to the floating index; they are discounted off a curve derived from the rate paid on posted collateral, which for a standard collateralised derivative is an overnight rate such as €STR or SOFR. A bank that discounts a collateralised swap off the wrong curve produces a value that is wrong by an amount that grows with maturity.

The volatility surface presents the same problem in two dimensions. For swaptions, implied volatility is quoted against option expiry, underlying swap tenor and strike relative to the forward. The quoted grid is dense in the middle — one to ten year expiries on two to ten year underlyings, near the money — and sparse at the edges. A bank holds positions at the edges precisely because that is where clients want protection and where margins are wider.

Filling the sparse regions requires a parametric model. The industry standard family for interest rate and foreign exchange volatility is the SABR model, which fits a smile through four parameters with intuitive meaning: the overall level of volatility, the exponent controlling how volatility responds to the level of the underlying, the volatility of volatility controlling the curvature of the smile, and the correlation between the underlying and its volatility controlling the skew. Fitting SABR to the quoted points and reading off the interpolated ones is a defensible construction. It is also a construction, and the difference between a SABR fit and an alternative parameterisation at a twenty-five-year expiry can be a full volatility point, which on Desk SRC-1's Book B is roughly €0.42 million per point of vega.

Market data infrastructure has four distinct components and banks routinely under-invest in the last two.

Acquisition: vendor feeds, broker pages, exchange feeds, internal contributions. Multiple sources per instrument is the norm, with a documented precedence.

Cleansing and validation: outlier detection against the prior close, arbitrage checks (a forward curve that implies negative forward rates where none should exist, a volatility surface that permits arbitrage between strikes), staleness detection.

Construction: the curve and surface builders, which must be versioned like code because a change to an interpolation method changes every valuation in the book.

Distribution and lineage: every downstream consumer must be able to answer, for any valuation, which snapshot and which construction version produced it. Without that, an investigation into an unexplained profit and loss cannot even start.

Two valuations, worked all the way through

Everything above is method. This section does it with numbers. The point is not the arithmetic, which is elementary. The point is to show exactly where a judgement enters, because every judgement is a place where the control function has something to verify.

A simplification is applied throughout for readability: annual rather than quarterly payments, and a flat curve at a single rate rather than a full term structure. Neither changes the structure of the calculation or any conclusion drawn from it. Real production valuations run forty quarterly cash flows against a curve with a different rate at every date.

The instrument

Desk SRC-1 pays fixed at 2.85% and receives floating on €250 million for ten years. The trade was struck when 2.85% was the market rate, so it was worth zero at inception. Today the market rate for the same ten-year swap is 2.94%. The question is what the contract is now worth.

Method one — discount every cash flow

The bank pays 2.85% and receives an expected 2.94%, so it nets 0.09% per year on €250 million: €225,000 a year for ten years. Discount each of those ten receipts at 2.94% and add them up.

Year Fixed leg paid (€m) Floating leg expected (€m) Net (€m) Discount factor Present value (€m)
1 −7.125 +7.350 +0.225 0.9714 0.219
2 −7.125 +7.350 +0.225 0.9437 0.212
3 −7.125 +7.350 +0.225 0.9167 0.206
4 −7.125 +7.350 +0.225 0.8906 0.200
5 −7.125 +7.350 +0.225 0.8651 0.195
6 −7.125 +7.350 +0.225 0.8404 0.189
7 −7.125 +7.350 +0.225 0.8164 0.184
8 −7.125 +7.350 +0.225 0.7931 0.178
9 −7.125 +7.350 +0.225 0.7704 0.173
10 −7.125 +7.350 +0.225 0.7484 0.168
Total 8.5564 1.925

The swap is worth €1.925 million to the bank. Nothing more mysterious than that is happening in a rates book.

Notice where the judgement sat. Not in the arithmetic. It sat in two places: the expected floating rate at each of the ten dates, which came from a forecasting curve someone constructed, and the ten discount factors, which came from a discount curve someone else constructed. Change the interpolation method between quoted nodes and the ten discount factors change slightly, and so does the answer.

Method two — the annuity shortcut

The same value can be obtained in one line, and the fact that it can is itself a control. Every net cash flow is identical, so the calculation collapses to the difference between the two rates, multiplied by the sum of the discount factors, multiplied by the notional.

The sum of the discount factors — 8.5564 in the table — is called the annuity factor. It is the present value of receiving one unit per year for ten years, and it is one of the most useful quantities in a rates book.

(2.94% − 2.85%) × 8.5564 × €250,000,000 = €1.925 million

The two methods agree exactly, and they must. In practice they are computed by different systems: the cash-flow method by the pricing library, the analytic method often by a risk system or a spreadsheet used for sanity-checking. When two independent implementations of the same mathematics disagree, one of them has a bug, a different curve, or a different day-count convention — and finding out which is a genuine daily activity in valuation control, not a theoretical exercise.

Where DV01 comes from

Rerun the same calculation with the market rate one basis point higher, at 2.95%, and take the difference. The annuity factor falls very slightly and the rate gap widens by a basis point. The net effect is €213,000.

That is the DV01. It was not assumed or looked up; it fell out of running the valuation twice. Nine basis points of rate movement therefore produce about €1.9 million of value change, which is precisely the number that appears in the attribution bridge for Desk SRC-1's Book A once scaled to the whole book.

Valuing the option, and why the volatility argument is worth money

Now the harder instrument. Desk SRC-1 has sold a payer swaption: the right, in five years, to enter a ten-year swap paying 2.85%. The forward rate for that swap is 2.94%, so the option is currently in the money for its holder.

The standard market approach applies the Black formula, which takes the forward rate, the strike, the time to expiry, the annuity factor of the underlying swap, and one further input — implied volatility. Everything except that last input is observable or contractual.

Input Value Source
Forward swap rate 2.94% Constructed from the curve
Strike 2.85% The contract
Time to expiry 5 years The contract
Forward annuity factor 7.4023 Constructed from the curve
Notional €250 million The contract
Implied volatility the argument Consensus, brokers, or the desk

Run the formula at the desk's mark and at the independently verified level.

Volatility used Swaption value (€m)
Desk mark: 20.6% 10.613
Verified consensus: 21.4% 10.984
Difference 0.371

On one trade, an argument about eight tenths of a volatility point is worth €371,000. The desk has sold this option, so a higher volatility means the liability is larger and the desk's book is worth less. The difference is not a rounding item and it is not a matter of opinion once independent evidence exists — which is exactly why the independent price verification process in Step 8 exists and why its output is posted as an adjustment rather than debated.

Dividing the difference by the volatility gap gives the vega: about €464,000 per volatility point on this single trade. Scale that across a book of 380 swaptions with concentrated long-dated positions and the reason Book B's monthly independent price verification produced a €1.9 million adjustment becomes arithmetic rather than assertion.

Valuing the tranche, and what the correlation argument is worth

The hardest instrument in the running example, worked with numbers. Desk SRC-1 holds a 3–7% tranche on a portfolio of sixty names, €10 million each, five years remaining. The spread on the underlying names implies a hazard rate of 2.5% a year; recovery is assumed at 40%.

Start with what is not in dispute. Over five years, each name has a default probability of about 11.75%. With 60% lost on default, the expected loss on the whole portfolio is 7.05% — roughly €42 million of the €600 million. That number falls out of the spreads and the recovery assumption and nobody argues about it.

Now the part that is in dispute. The tranche only absorbs losses between 3% and 7%. Whether it is nearly destroyed or partly intact depends on the shape of the loss distribution, which depends on correlation. Running the standard portfolio loss model across the defensible correlation range gives this.

Base correlation Expected loss on the tranche Value of the tranche exposure (€m)
0.38 — bottom of the defensible range 44.3% 10.63
0.42 — the desk's input 42.4% 10.18
0.47 — top of the defensible range 40.2% 9.64
Width of the range 4.1 points 0.99

Read the last row carefully, because it is the whole argument. On a single tranche with €24 million of exposure, moving correlation across a range that no reasonable person could rule out changes the value by €992,000 — roughly 4% of the tranche. The desk's 0.42 sits inside that range and cannot be called wrong.

So the control function does not call it wrong. It observes that the range is 4% wide on a position with no external price, and that the desk sits slightly towards the favourable end of it. A point estimate cannot be challenged; a range can be quantified. That quantification, aggregated across 120 tranches, is the €2.1 million model uncertainty reserve in the month-end stack.

Valuing the counterparty risk on the swap

Return to the ten-year swap and remove one assumption: that the counterparty pays. It is an uncollateralised corporate, so if it defaults while the swap is in the bank's favour, the bank loses the value it was owed.

The calculation multiplies three things together, year by year, and discounts the result. How much the bank expects to be owed at that time — the expected exposure. The probability the counterparty defaults in that year, given it survived until then. And the fraction lost if it does.

Figure 8 The exposure profile and the survival curve are the two inputs; CVA multiplies them together

The exposure profile has a characteristic shape and the shape carries the intuition. At inception the swap is worth zero, so there is nothing to lose. Exposure rises as rates have time to move away from the contract rate. It then falls, because fewer payments remain — a swap with one year left simply cannot be worth as much as one with nine. The peak sits early, at around a third of the life.

Year Average expected exposure (€m) Marginal default probability Discount factor CVA contribution (€m)
1 4.90 2.47% 0.9714 0.071
2 11.45 2.41% 0.9437 0.156
3 13.85 2.35% 0.9167 0.179
4 14.80 2.29% 0.8906 0.181
5 14.70 2.23% 0.8651 0.171
6 13.70 2.18% 0.8404 0.150
7 12.00 2.13% 0.8164 0.125
8 9.70 2.07% 0.7931 0.096
9 6.90 2.02% 0.7704 0.064
10 2.70 1.97% 0.7484 0.024
Total 1.217

The swap is worth €1.925 million on the assumption that the counterparty pays, and €1.217 million less once that assumption is priced. The credit adjustment is nearly two thirds of the trade's mid-market value, which is not an artefact of the illustration — it is what uncollateralised long-dated exposure to a moderately rated corporate actually costs.

Widen the counterparty's spread from 150 to 200 basis points and the adjustment rises to €1.565 million. That sensitivity is the reason CVA is hedged, why a dedicated desk owns it, and why credit spread moves appear in the daily attribution of a book that contains no credit instruments at all.

What the worked examples establish

Four things, each of which the rest of Part 2 assumes.

The arithmetic is never the problem. Both valuations above are a few lines of calculation. The difficulty lies entirely in the inputs: the curve that produced the discount factors, the forward rate, and the volatility.

Every input has a provenance, and provenance is what gets verified. The strike came from a contract and cannot be argued with. The volatility came from somebody's judgement about a market and can be argued with at a rate of €464,000 per point.

The nature of the uncertainty determines the control. The swap's inputs were fixed, the swaption's were disputable and observable, the tranche's were disputable and unobservable. Reconciliation, verification and reserving respectively.

Value and risk come from the same calculation. DV01 and vega were both produced by running the valuation twice. A bank that computes its risk in one system and its value in another is running the same mathematics twice with two chances of disagreeing — which is why the unexplained residual in Step 5 is so often a symptom of infrastructure rather than of trading.

Step 3 — Overnight, every position in the bank is priced from scratch

Between the evening close and the early morning, the bank recomputes the value and the risk of every position it holds. The batch is the industrial heart of the process.

What the batch actually does

The overnight batch performs three passes over the entire portfolio. The first prices every trade at the official close. The second prices every trade again under a set of perturbed market scenarios, producing sensitivities. The third prices the prior day's portfolio at today's market data, which is the input to profit and loss attribution.

The third pass is the one that people new to the domain do not anticipate, and it is essential. To decompose today's profit and loss into the part caused by market movement and the part caused by new trades, the engine must be able to hold the portfolio constant and move the market, then hold the market constant and move the portfolio. That requires re-pricing historical portfolio states, which doubles or triples the computational load.

Sensitivities: the vocabulary of the explanation

A sensitivity, or Greek, is the change in a position's value for a defined change in one market factor. The batch computes them by bumping the input and re-pricing — a full revaluation approach — or analytically where the model permits. Full revaluation is more accurate and vastly more expensive; the trade-off is one of the central architectural questions in valuation infrastructure.

Sensitivity Change measured What it explains on Desk SRC-1
DV01 / delta Value change per 1 basis point rate move The dominant driver on Book A
Gamma Change in delta per unit rate move Convexity in swaptions, matters on large moves
Vega Value change per 1 point of implied volatility The dominant driver on Book B
Theta Value change from one day of time passing Time decay and carry, always present
Credit spread sensitivity (CS01) Value change per 1 basis point spread move Book C and counterparty credit adjustment
Correlation sensitivity Value change per unit correlation move Book C only, unobservable, reserve-relevant
Cross-gamma Change in one sensitivity per move in another factor Structured positions; a common source of residual

Sensitivities are generated per risk factor, not per position, and the granularity is a design decision. A book's rate risk can be expressed as a single DV01, or as a vector of sensitivities to each curve node — a bucketed delta. The single number is easier to read. The bucketed vector is what allows the attribution to explain a day on which the curve steepened rather than shifted in parallel. Any bank running a serious attribution process buckets its sensitivities, and any bank that does not will find its unexplained profit and loss dominated by curve shape movements it cannot see.

The batch is the largest computational workload in most banks outside of risk simulation, and its architecture has changed more in the last decade than any other part of this chain.

The traditional pattern is a fixed on-premises compute grid running overnight, sized for the peak, idle for two-thirds of the day. Jobs are scheduled by dependency: market data first, then base valuation, then sensitivities, then attribution inputs.

The modern pattern is elastic cloud compute provisioned for the duration of the batch and released afterwards. This suits the workload almost perfectly, because it is embarrassingly parallel — trades price independently — and radically peaked in time. XVA calculations, which require Monte Carlo simulation of future exposure across thousands of paths, benefit most and were often the first workload migrated.

The hard constraints are not computational. They are data gravity — moving a full portfolio and market data set to and from a cloud environment nightly is non-trivial — and confidentiality, since the portfolio is among the most sensitive data a bank holds.

The dependency chain, and why it has no slack

The batch is a directed chain, and each stage has a completion time beyond which the next cannot start. A representative schedule for a European desk looks like this.

Figure 9 The overnight chain as a dependency graph: the bold path has no slack
Time Job Blocked if late
17:30 Trade population cut-off Everything
18:00 Market data snapshot complete Curve construction
19:00 Curves and surfaces published Valuation
19:00–01:00 Base valuation across all books Sensitivities
01:00–03:30 Sensitivity generation Attribution
03:30–04:30 Prior-day portfolio at today's market Attribution
04:30–06:00 XVA and portfolio-level adjustments Reporting
06:00–07:00 Aggregation and reporting layer refresh Flash P&L
08:30 Desk opens —

The consequence of this structure is that recovery time is the binding design constraint, not throughput. A batch that completes comfortably by 06:00 in the normal case but takes four hours to restart after a failure at 02:00 will miss the deadline. A batch that runs slower but restarts from checkpoints in twenty minutes will not. Part 4 returns to this under operational resilience, because it is the clearest example in the entire domain of resilience being an architectural property rather than an infrastructure purchase.

Step 4 — The flash profit and loss is a hypothesis to be tested, not a result to be reported

By early morning, a number exists. Nobody yet knows whether it is right.

What "flash" means and why it is provisional

Flash profit and loss is the first automated aggregation of the overnight batch output, available before the desk opens. It is the difference between the portfolio's value at last night's close and its value at the previous close, adjusted for cash flows and new trades. It exists because a trading desk cannot start the day blind: a trader needs to know their position and their approximate result before making the first decision of the morning.

It is explicitly provisional. Between the flash and the signed number, several things routinely change: late trades are booked, market data errors are corrected, valuation failures on individual trades are resolved, and manual adjustments are applied. In a well-controlled book the difference between flash and final is small and systematically small. Persistent large differences are themselves a control finding — they mean the automated process is not capturing the economics of the book.

The first-pass validation controls

Before a controller looks at anything, the system applies a battery of automated checks. The purpose is triage: to direct human attention to the small number of positions where something is genuinely wrong.

Completeness. Did every trade price? A trade that failed to value silently drops out of the portfolio and out of the profit and loss, which is a far worse outcome than a trade that failed loudly. The count of pricing failures is a headline control metric.

Threshold breach. Did any desk's daily profit and loss exceed a defined threshold, in absolute terms or relative to its typical distribution? Thresholds are set per desk and calibrated to the book's volatility, not set uniformly.

Sign and magnitude plausibility. Did the book move in the direction its risk profile predicts? Desk SRC-1's Book A has negative DV01, so a nine basis point rise in rates should produce a gain of roughly €1.9 million times nine, adjusted for the shape of the move. A loss on that day is a red flag before anyone examines a single trade.

Stale valuation. Did any position's value fail to change on a day when the factors it depends on moved materially?

Reconciliation status. Do the trade counts and notionals in the valuation engine match the books and records?

Step 5 — Attribution is the control that catches what price checking cannot

If the profit and loss can be fully explained by the risks the desk was known to be running, the process is probably sound. If it cannot, something is wrong somewhere, and the residual tells you how much.

The logic of the explain

Profit and loss attribution decomposes the day's result into causes. The method rests on a simple idea: if a position's value depends on a set of market factors, and those factors are known to have moved by known amounts, then the sensitivity of the position to each factor multiplied by that factor's move gives the predicted contribution of that factor. Sum the contributions, add the effects that are not market-driven — new trades, fees, funding, time decay — and compare the total to the actual change in value.

What remains is unexplained. It is the single most informative number in the daily pack.

Figure 10 The daily attribution bridge for Desk SRC-1: every component named, the residual isolated
Component What it captures Typical source of error
Delta / rate Sensitivity × curve move, bucketed by tenor Wrong bucketing, curve shape not captured
Gamma Second-order effect of large moves Omitted on days of large moves
Vega Sensitivity × implied volatility move Surface moves not decomposed by strike
Theta and carry Passage of one day, accrual, roll-down Day-count and calendar errors
New trades Trades executed today, valued at close Off-market trades, day-one profit
Fees, commissions, funding Non-market cash items Funding allocation methodology
Amendments and prior-day Corrections to previously booked trades Concentration here is a conduct flag
Unexplained residual Everything else The control metric

Where residuals come from

A residual is not automatically an error, and understanding the difference is most of the skill in this stage. Several sources of residual are structurally unavoidable.

Higher-order effects not captured by the sensitivities computed. If the batch produces delta and gamma but not cross-gamma, a day on which rates and volatility both moved significantly will leave a residual that is real, explainable in principle and simply not being explained.

Discretisation. Sensitivities are computed for a specific bump size. If delta is computed for a one basis point move and the market moved nine basis points, the linear approximation is imperfect even before gamma is considered.

Timing mismatches. The valuation snapshot and the risk snapshot taken minutes apart, as discussed in Step 2.

The sources that are errors are more interesting.

A missing or mis-booked trade produces a residual with no market explanation, often persistent day after day until found.

A market data error produces a residual concentrated in the positions dependent on the bad input.

A model behaving outside its calibrated region produces a residual that appears when the market moves into unusual territory and disappears when it moves back — the most dangerous pattern, because it is intermittent.

A deliberately mismarked position produces a residual that reverses when the mark is corrected, which is why unexplained profit and loss is a conduct control as well as an accuracy control.

The Fundamental Review of the Trading Book turned this internal metric into a regulatory test with capital consequences. Its profit and loss attribution test compares two series over a rolling twelve-month window: the hypothetical profit and loss, which is the actual change in the desk's value holding the portfolio constant, and the risk-theoretical profit and loss, which is what the desk's own risk model predicts given the same market moves.

Two statistics are computed on the difference between the series. The Spearman rank correlation measures whether the two series move together in ordering. The Kolmogorov–Smirnov statistic measures whether they share the same distribution. Desks are placed in a traffic-light zone: green desks retain internal model approval, amber desks retain it with a capital surcharge, and red desks are removed from the internal models approach and capitalised under the standardised approach, which is materially more expensive.

The regulatory design is deliberate and worth appreciating. It does not attempt to specify what a good risk model looks like. It specifies a test of whether the model explains the desk's own realised results, and prices failure. A bank whose attribution process is weak now pays for that weakness in capital, every quarter, in a number its chief financial officer can see.

Step 6 — Sign-off is where independence becomes an actual event

Everything before this point produces information. This is where someone takes responsibility for it.

The mechanics of the sign-off

The daily profit and loss is signed by the product controller and acknowledged by the desk head. Both matter and they are different acts. The controller attests that the number was produced in accordance with policy, that exceptions were investigated and resolved or disclosed, and that the residual is within tolerance. The desk head attests that the number is consistent with the trading activity they conducted — that they recognise their own day.

Disagreement between the two is a defined process, not an argument. If the desk asserts that a position is worth more than the independent valuation shows, the position is carried at the independent valuation and the difference is escalated. The default is not the desk's number. This inversion — that the burden of proof falls on the person closest to the position — is the operational meaning of independence, and it is worth stating plainly because it is frequently softened in practice.

The front-office estimate against the official number

Before the controller's number is signed, it is compared with the trader's own. Every desk keeps a running estimate of its result — on a blotter, in a spreadsheet, in the trading system's own position view — and that estimate is generally available before the official figure. Comparing the two is a distinct control with its own name, usually the front-office to Product Control reconciliation, and it is more powerful than it first appears.

Its power comes from the independence of the two errors. The desk's estimate is built from what the trader believes they did: the trades executed, the moves observed, the hedges applied. The official number is built from the books and records, the official market data and the pricing library. The two are constructed from different sources by different people with different information, so a difference between them is evidence that at least one of those sources is wrong — and which one it is usually becomes obvious quickly.

The recurring causes are worth memorising, because they map to the failure modes described earlier.

The desk knows about a trade the books do not. A late booking, a trade executed and not yet captured, an amendment agreed verbally. The desk's estimate includes it and the official number does not. This is a completeness failure and it is exactly what the reconciliation exists to catch.

The books know about a trade the desk does not. A duplicate booking, a trade allocated to the wrong desk, a lifecycle event applied automatically. Here the official number is the one that is wrong, and the desk is right.

Both agree on the trades and disagree on the marks. The desk used its own view of where the market closed; the official number used the official snapshot. This is the mark dispute proper, and it feeds directly into the escalation ladder.

Both agree on trades and marks and still differ. Then the difference sits in the calculation: a model, a convention, a day count, a discount curve. This is the most valuable finding of the four, because it points at infrastructure rather than at an event, and infrastructure errors are persistent.

The reconciliation is only useful if it is systematic, which requires the desk's estimate to be captured in a structured form rather than communicated verbally. Mature implementations take a feed of the desk's own position and estimate directly from the trading system, compare it automatically against the official figure at desk and sub-book level, and present the differences ranked by size with an ageing history.

The metric that matters is not the size of the difference on any one day but the persistence of a difference in the same direction on the same book. A one-day gap is an event. A gap that appears every day for a fortnight is a structural break — a systematically missing trade population, a convention mismatch, or a mark that has drifted — and it will not resolve itself.

Materiality and tolerance thresholds

Not every difference can be investigated, and a framework that pretends otherwise investigates nothing properly. Thresholds are therefore explicit, documented and calibrated.

Threshold type Set at Consequence of breach
Individual trade movement Absolute value, per book Trade-level review before sign-off
Desk daily P&L Multiple of historical daily volatility Desk head must comment in writing
Unexplained residual, daily Percentage of gross P&L or absolute Investigation logged, root cause required
Unexplained residual, cumulative Rolling monthly absolute Escalation to head of Product Control
IPV difference, per position Absolute and percentage Adjustment posted, Valuation Committee if disputed
Reconciliation break age Days outstanding Escalation at 5, 15 and 30 days

Calibrating thresholds to the book's own volatility rather than to a bank-wide number is the difference between a working framework and a theatrical one. A €500,000 daily movement on Desk SRC-1's Book A is unremarkable. The same movement on the run-off tranche book would be extraordinary. A single bank-wide threshold either floods the first book with false positives or lets the second book move freely.

The escalation ladder

Unresolved items climb a defined ladder with defined timescales. The controller investigates and resolves at desk level within the day. Items unresolved by the end of the day pass to the desk's Product Control manager with a written summary. Items unresolved after a defined number of days, or exceeding a defined size, reach the head of Product Control for the business and the desk's supervisor. Genuine valuation disputes go to the Valuation Committee. Anything with a conduct dimension — a pattern of amendments, a position marked away from independent evidence after being challenged, a trade booked to move profit between periods — leaves the ladder entirely and goes to Compliance and to the trading floor supervisor immediately.

Step 7 — The ledger close turns a management number into an accounting fact

Daily profit and loss is a management report. The general ledger is the legal record. Getting from one to the other is a distinct process with its own failure modes.

The sub-ledger and the general ledger

Trading positions are recorded in a sub-ledger that carries instrument-level detail, and summarised into the general ledger that carries the entity's accounts. The sub-ledger holds every trade with its fair value, its accrued interest, its valuation adjustments and its accounting classification. The general ledger holds the aggregate balances that appear in the financial statements.

The daily posting cycle moves value from one to the other. The essential control is that the two must agree, and that the sub-ledger must agree with the front-office system. Three-way reconciliation — front office to sub-ledger to general ledger — is the substantiation of the trading balance sheet, and it is run daily on the largest books and at minimum monthly everywhere.

Cash, collateral and the settlement reality check

Reconciliation to cash is the hardest control to argue with. A derivative's fair value is a model output and can be debated. A margin call settled in cash is a bank transfer that either happened or did not.

For collateralised derivatives — now the overwhelming majority of the interbank book — daily variation margin is exchanged based on each party's calculation of the position's value. The margin call is therefore an independent daily valuation of the same portfolio, produced by a counterparty with an opposite economic interest. When the bank's value and the counterparty's differ beyond a tolerance, a formal dispute process begins. Persistent, one-directional margin disputes on a portfolio are among the strongest available evidence that a book is mismarked, and Product Control functions that do not receive collateral dispute data are missing a free control.

The post-trade layer runs on standardised messaging. Settlement instructions, confirmations and margin calls travel over SWIFT, historically as MT-series messages and increasingly as ISO 20022 MX messages as the industry migration proceeds. Cleared derivatives generate margin and position reports from the clearing house. Bilateral collateral is increasingly managed through electronic margin call platforms that exchange calls, agree amounts and record disputes in a structured form.

The value of the structured form is that it can be consumed as a control input. A margin dispute recorded as a free-text email is invisible to the valuation process. The same dispute recorded as a structured record with a position identifier, a claimed value and a difference is an automated exception feeding directly into the price verification population. Turning post-trade exhaust into control evidence is one of the cheapest genuine improvements available in this domain, and it is available precisely because the messaging is standardised.

Step 8 — Independent price verification is where the desk's mark is put on trial

Once a month, every material position is tested against evidence the desk did not produce.

The principle: evidence outranks assertion

Independent price verification, universally abbreviated to IPV, is the periodic comparison of the values used in the accounts against independently sourced evidence of what those positions are worth. The word doing the work is independently. Evidence obtained by the desk, from a source the desk selected, is not independent. The control function obtains its own evidence, from sources it has assessed, with a documented order of precedence.

Figure 11 IPV in one picture: independent sources establish the level, and what they cannot resolve becomes a reserve

The source hierarchy

Sources are ranked, and the ranking is policy. A controller does not choose the most convenient source or the one closest to the desk's mark.

Exchange and clearing house settlement prices

The strongest evidence available. A clearing house settlement price is a real, binding number: margin is exchanged on it. For any position with an exchange-traded equivalent, this is the first source consulted.

Consensus pricing services

For over-the-counter derivatives, the industry solved the evidence problem collectively. Contributing banks submit their own marks for a defined set of instruments to a third-party service. The service cleanses the submissions, discards outliers, and returns statistical output: a consensus mean or median, a standard deviation across contributors, and the number of contributors. Participation is reciprocal — a bank must contribute to receive.

The output is powerful for three reasons. It is genuinely independent of any single contributor. It comes with a dispersion measure, which is a direct quantification of market uncertainty and feeds straight into reserve calculation. And the contributor count tells the controller how much weight the level deserves: a consensus level built from fifteen contributors is evidence, and one built from three is an indication.

Contribution creates an obligation the control function must police. A desk that submits its own optimistic mark to the consensus service and then verifies against the resulting consensus has closed a circle around itself. Submissions are therefore controlled: made from the verified book where possible, subject to review, and monitored for systematic bias relative to the consensus.

Broker and dealer quotes

Executable quotes from market makers, obtained by the control function directly. Quality varies enormously. A firm two-way quote in size from an active market maker is strong evidence. An indicative level provided by a broker as a courtesy, at a price the broker would not trade, is weak. Policy typically requires multiple quotes, records the identity of the provider, distinguishes firm from indicative, and captures the bid-offer spread as an input to the close-out reserve.

Observed transactions

A trade in the same or a closely comparable instrument is direct evidence, and it is rare in illiquid markets. Size matters: a small odd-lot trade in a thin market may reflect a specific liquidity need rather than a level. The bank's own trades count, with caution — a trade executed by the desk being verified is not independent, though it is informative.

Model with independently verified inputs

The last resort, and for Level 3 positions, the only option. Where no external evidence exists for the instrument, the control function verifies the inputs independently instead: the correlation assumption is benchmarked against index tranche implied levels, the extrapolated volatility is compared against a range of alternative extrapolation methods, the recovery assumption is compared against historical experience and against what other market participants disclose. The output is not a verified price but a verified range, and the width of that range determines the reserve.

Source Independence Typical coverage Weakness
Exchange / CCP settlement Highest Listed and cleared products Limited to standardised instruments
Consensus service High Broad OTC derivatives coverage Requires contribution; can be thin at edges
Broker quotes Medium to high Instrument-specific Indicative quotes carry little weight
Observed transactions High when genuine Sparse in illiquid markets Size and motivation unknown
Model with verified inputs Lowest Universal fallback Verifies the input, not the price

Coverage, sampling and the honest metric

Not every one of 8,400 trades is verified every month, and pretending otherwise is a common form of control theatre. Sound policy defines a verification population by materiality and risk: all Level 3 positions, all positions above a size threshold, all positions in books with elevated residuals, plus a rotating sample of the remainder such that every position is verified at a defined minimum frequency.

The metric that matters is coverage by value, not by count. A function reporting that it verified 94% of trades has said almost nothing if the unverified 6% contains the entire Level 3 book. Coverage should be reported as the percentage of fair value verified, split by hierarchy level, with the unverified residual explicitly named.

Step 9 — Reserves convert what you do not know into a number in the accounts

A model produces a mid-market value under idealised assumptions. The accounts must show what the position would realistically fetch. The difference is a stack of adjustments, each answering a specific question.

Figure 12 Each adjustment answers one question the mid-market model value does not

Bid-offer and close-out: you cannot exit at mid

Every model prices at mid. Real exit happens at bid if you are long and at offer if you are short. A book carried at mid is carried above its realisable value by roughly half the bid-offer spread on its net risk.

The reserve is computed on net risk, not gross notional, and that distinction is the whole of the mechanics. A desk with a thousand trades that net to a small residual exposure faces a small exit cost; a desk with a hundred trades all pointing the same way faces a large one. The calculation therefore proceeds by risk factor and bucket: take the net sensitivity in each bucket, multiply by half the observed bid-offer spread for that factor and bucket, and sum.

Concentration deserves an explicit uplift. If the desk's net position in a bucket represents a meaningful fraction of the daily traded volume in that instrument, exiting will move the market against it, and the effective spread is wider than the quoted one. Policies capture this by scaling the spread once the position exceeds a defined multiple of normal market size.

Model uncertainty: the model might be wrong

Where a valuation depends on a model whose limitations are known, a reserve covers the potential error. The number comes from one of three approaches, in descending order of preference.

Alternative model comparison. Value the portfolio under the approved model and under a credible alternative — a two-factor rather than a one-factor short-rate model, a different correlation mapping — and reserve some portion of the difference.

Parameter range. Where the model is agreed but an input is uncertain, revalue across the defensible range of the input and reserve to a conservative point within it. This is what Desk SRC-1 does with base correlation.

Model validation finding. Where validation has identified a specific deficiency with a quantified impact, that quantification is the reserve until the deficiency is remediated. This creates a direct incentive: the reserve is a visible charge against the desk's profit and loss that disappears when the model is fixed.

The XVA family: mid-market ignores counterparty, funding and capital

A derivative's model value assumes both parties pay and that the trade is costlessly financed. Neither is true, and a family of portfolio-level adjustments corrects for it. These are computed centrally, usually by a dedicated XVA desk, because they are portfolio effects that cannot be decomposed cleanly to individual trades.

Credit valuation adjustment (CVA) reduces the value of a derivative asset to reflect the possibility that the counterparty defaults before the trade matures. It is computed by simulating the bank's future exposure to the counterparty across time, applying the counterparty's default probability derived from its credit spread, and applying an expected recovery. It is materially larger for uncollateralised counterparties — typically corporates that have negotiated an exemption from posting collateral — than for collateralised financial counterparties.

Debit valuation adjustment (DVA) is the symmetric effect on the bank's own liabilities: because the bank itself might default, its derivative liabilities are worth less than face. DVA is required under accounting standards and disliked by essentially everyone, because it produces the counter-intuitive result that a deterioration in the bank's own creditworthiness generates accounting profit. Regulators exclude it from regulatory capital for that reason.

Funding valuation adjustment (FVA) reflects the cost of funding an uncollateralised derivative position. If the bank has an in-the-money uncollateralised trade with a corporate, it has an asset generating no cash, and it must fund the hedge on which it is posting collateral. That funding has a cost above the risk-free rate, and FVA capitalises it.

Capital and margin valuation adjustments (KVA and MVA) extend the same logic to the cost of holding regulatory capital against the position over its life, and to the cost of funding initial margin. Their inclusion in accounting fair value is contested; their inclusion in trade pricing is not.

XVA is where valuation control meets serious computational load. Computing CVA requires simulating the exposure profile of an entire netting set — every trade with a given counterparty, under the applicable netting agreement — across thousands of Monte Carlo paths and dozens of future time points, then integrating against a default probability curve.

Two consequences follow. First, XVA is the workload that most often justified a move to elastic or accelerated compute, since it is parallel by path and peaked in time. Second, XVA sensitivities — how the adjustment moves when a credit spread or a rate moves — are needed for hedging and for attribution, and computing them by brute-force bumping multiplies an already large calculation by the number of risk factors. Adjoint algorithmic differentiation, which computes all sensitivities in a small multiple of the cost of a single valuation, moved from academic technique to production necessity for precisely this reason.

Day-one profit and the deferral rule

A trade can be profitable the instant it is struck, and accounting standards restrict when that profit may be recognised. If a bank sells a structured product for €10 million and its model values it at €9.2 million, the model says €800,000 of profit exists on day one.

The rule is that day-one profit may be recognised immediately only if the valuation is based on observable inputs — a Level 1 or Level 2 valuation. Where the valuation depends on unobservable inputs, the day-one profit is deferred and released over time, typically over the life of the trade or as the inputs become observable.

The economic logic is sound. A day-one profit derived from an unobservable input is a profit the bank has asserted on the basis of its own assumption. Recognising it immediately would let a bank book profit by choosing a model. Deferral forces the profit to be earned through the passage of time and the resolution of uncertainty.

The operational consequence is a schedule that must be tracked per trade for years, with release triggers, and with a control to ensure that a trade terminated early releases its remaining deferral correctly rather than silently.

Prudent valuation: the number the regulator will accept

Accounting fair value and regulatory prudent value are different numbers, deliberately. European capital regulation requires banks to compute additional valuation adjustments — AVAs — that reduce the fair value carried in the accounts to a level with a high degree of certainty of achievement, expressed in the regulation as a 90% confidence standard. The resulting difference is deducted directly from common equity tier 1 capital.

Nine categories of AVA are prescribed, covering market price uncertainty, close-out costs, model risk, concentrated positions, unearned credit spreads, investing and funding costs, future administrative costs, early termination and operational risk. Banks below a size threshold may use a simplified approach; larger banks must use the core approach with its prescribed methodology and aggregation.

Step 10 — Classification and disclosure make the uncertainty public

The final step converts internal knowledge about valuation quality into information the outside world can act on.

Assigning and moving between levels

Every fair-valued position is classified into the hierarchy at each reporting date, and the classification is reassessed rather than inherited. The test is whether the inputs that are significant to the valuation are observable. A position with one unobservable input that barely moves the value stays in Level 2; a position with one unobservable input that dominates the value is Level 3.

Transfers between levels are disclosed, with the reason. A transfer into Level 3 signals that observability was lost. A transfer out signals that a market re-established itself, or that the bank found a source it previously lacked.

What must be disclosed for Level 3

Level 3 positions attract the most demanding disclosure in the entire valuation note, and the demands are informative in themselves. Banks must disclose the valuation techniques and the unobservable inputs used, with a quantitative range for each significant input. They must provide a reconciliation of Level 3 balances from opening to closing, separating purchases, sales, issues, settlements, transfers in and out, and gains and losses split between realised and unrealised. And they must provide a sensitivity analysis showing the effect on fair value of changing unobservable inputs to reasonably possible alternatives.

That last requirement is the one that binds. It forces the bank to state publicly how much its own Level 3 valuations would move under a plausible alternative assumption. It is, in effect, a published version of the model uncertainty reserve analysis, and a reader comparing the disclosed sensitivity against the disclosed reserve learns a great deal about how conservatively a bank marks.

The lifecycle of an adjustment does not end when it is posted

Reserves and adjustments have a life, and managing that life is a control in its own right. Four events recur.

Release on exit. When a position is closed, the reserve held against it must be released. Getting this right is essential: a reserve released into profit when the position exits is legitimate, but a reserve released while the position remains is profit taken from the control framework.

Release on observability. When a market re-establishes itself and an input becomes observable, the uncertainty that justified the reserve has been resolved. Deferred day-one profit is released on the same logic.

Release on remediation. When a model deficiency is fixed and revalidated, the model reserve held against it is released. This is the incentive mechanism described earlier working as intended.

Model change profit and loss. When an approved model is replaced by a new version, the portfolio's value changes on the day of the change, and that change is profit and loss caused by nothing that happened in the market. It must be identified separately, explained, approved and disclosed — never absorbed into market movement. A book whose profit is materially attributable to model changes is a book whose valuation methodology is being tuned towards a result, and separate reporting of model change profit and loss is what makes that visible.

  1. Why must the overnight batch price the prior day's portfolio at today's market data?

  2. A book's daily attribution shows a residual that is negligible on most days and large on a few. Why is that more concerning than a consistently moderate residual?

  3. Why is the bid-offer reserve computed on net risk rather than gross notional?

  4. When does IPV produce an adjustment, and when does it produce a reserve?

  5. Why is day-one profit deferred when the valuation depends on unobservable inputs?

  6. Why do collateral margin disputes constitute valuation evidence?

  7. What distinguishes accounting fair value from regulatory prudent value?

  8. Why must model change profit and loss be reported separately?

  9. Two systems value the same swap and disagree by a small amount. Both used the same curve. What are the likely causes?

  10. Why is the front-office to Product Control reconciliation more powerful than an internal consistency check?

  11. Why is recovery time, rather than total runtime, the binding design constraint on the overnight batch?

3The same number serves six masters, and they want different things from it

Valuation looks like a technical exercise until you notice that everyone involved has a stake in the answer. Understanding who wants what, and why they are not aligned, explains the entire institutional design.

Four camps surround every valuation

The people involved in a bank's valuation process sort into four groups with genuinely different objectives. They are not villains and heroes. Each is doing its job, and the jobs conflict.

The front office — the trading desks — creates the positions and forms the first view of what they are worth. Traders have more information about their own book than anyone else in the institution: they know why a position was put on, what the client wanted, where the market was when they last transacted, and what it would take to get out. That knowledge is genuinely valuable and the control framework should use it. Traders are also compensated on the reported value of their book, which is the reason the framework cannot rely on it alone.

The control functions — Product Control, Valuation Control, Market Risk and Model Validation — produce and challenge the number independently. Their objective is defensibility. They are measured on whether the numbers survive scrutiny, whether errors are caught before they become losses, and whether the process meets policy. They are not measured on the size of the profit.

Finance and assurance — Financial Control, Group Finance, Internal Audit and the external auditor — consume the number and attest to it publicly. Their objective is that the published financial statements are right and defensible. The external auditor's opinion is a legal artefact with consequences for the audit firm, which makes it the most genuinely independent challenge in the chain.

Market infrastructure and data providers — exchanges, clearing houses, consensus services, brokers and vendors — supply the external evidence without which independence is impossible. They have no stake in the bank's result, which is exactly what makes them useful.

Figure 13 Product Control sits between the people who make the number and everyone who relies on it

Every interaction, mapped

From To What flows When
Front office Trade capture Trade bookings and amendments Continuous
Front office Product Control Desk marks, commentary on P&L, dispute arguments Daily
Operations Product Control Confirmation status, settlement breaks, collateral disputes Daily
Market data team Valuation engine Official close snapshot, curves, surfaces Daily, evening
Batch / risk engine Product Control Values, sensitivities, attribution inputs Daily, overnight
Product Control Front office Flash and signed P&L, exception queries, adjustments Daily
Product Control Market Risk Unexplained P&L, valuation issues affecting risk factors Daily
Model Validation Product Control Approved models, documented limitations, quantified findings Event-driven
Product Control Valuation Committee Disputed marks, reserve methodology changes, level transfers Monthly
Consensus service Valuation Control Consensus levels, dispersion, contributor counts Monthly
Valuation Control Consensus service The bank's own contributed marks Monthly
Brokers / dealers Valuation Control Quotes, firm and indicative On request
Clearing house Operations and PC Settlement prices, margin, position reports Daily
Product Control Financial Control Signed P&L, fair values, reserves for posting Daily to monthly
Financial Control Group Finance Sub-ledger balances, disclosure inputs Monthly to quarterly
Group Finance External auditor Financial statements, valuation evidence, committee minutes Quarterly, annually
Product Control Regulatory reporting Prudent value, AVA, PLA test results Quarterly
Internal Audit All Control testing, findings, remediation tracking Cyclical

The taxonomy of the number's consumers

A single fair value figure is consumed by at least six distinct audiences, each of which applies pressure in a different direction. Understanding this taxonomy explains most of the conflict in the process, and it is the valuation analogue of understanding investor types in a capital markets transaction.

The trader

Wants the number high, on their own positions, now. A trader's compensation is a function of reported profit, and reported profit is a function of the marks. The pressure is rarely dramatic and almost never explicit; it takes the form of a persistently optimistic view of where the market is, a preference for the most favourable source, and energetic argument at the margins. The horizon is short: the current performance period.

The desk head and the business

Wants the number stable and explicable. A desk head is less concerned with a single day's mark than with volatility they cannot explain to their own management. This creates a pressure that is subtly different from the trader's: not upward, but towards smoothness. Smoothing is its own hazard, because a book marked to reduce apparent volatility understates risk in exactly the periods when risk is rising.

The chief financial officer and Group Finance

Wants the number right and, above all, not restated. The asymmetry is severe: a slightly conservative valuation costs some reported profit, while a valuation that has to be publicly corrected costs credibility, supervisory standing and often people's jobs. This is the strongest structural force for conservatism in a bank, and it is why the reporting line of Product Control into Finance is not an accident.

Market Risk and the chief risk officer

Wants the number consistent with the risk representation. Risk limits, value-at-risk and stress results are computed from the same positions and often the same models. A valuation that disagrees with the risk model's view of the same position means one of the two is wrong, and neither function can rely on its own outputs until the disagreement is resolved.

The regulator and the capital calculation

Wants the number conservative and reproducible. Prudent valuation deducts valuation uncertainty from capital. The profit and loss attribution test conditions internal model approval on explanatory quality. A supervisor is not primarily interested in whether a single mark is accurate; they are interested in whether the process that produced it would catch an error, and whether the bank has quantified what it does not know.

The external auditor and, behind them, the investor

Wants the number evidenced. An auditor cannot verify a Level 3 correlation input by knowing the market; they verify that a documented methodology was applied consistently, that independent evidence was obtained where it existed, that judgements were escalated and minuted, and that the disclosed sensitivities are consistent with the internal analysis. The auditor is testing the process as a proxy for the number, which is why documentation quality is not administrative overhead but the substance of what is being examined.

Consumer Wants the number to be Horizon Pressure exerted
Trader Favourable Current performance period Upward on own positions
Desk head Stable and explicable Quarterly Towards smoothness
CFO / Group Finance Right and never restated Multi-year reputational Strongly conservative
Market Risk Consistent with risk view Daily to monthly Towards internal coherence
Regulator Conservative and reproducible Structural Towards quantified uncertainty
External auditor Evidenced and documented Annual Towards process discipline

Independence is architectural, not organisational

The reason a bank cannot let traders mark their own books is not that traders are dishonest. It is that the institution has placed a person with superior information about a position in a role where their compensation depends on the value assigned to it. That is a structural conflict, and structural conflicts are managed by structure.

The four functions that must be separated

Four activities must not be performed by the same person or under the same reporting line, and the separation is the foundation of everything else.

Execution — deciding what to trade and at what price. Front office.

Recording — determining what was traded and capturing it in the books. Operations and trade capture, independent of execution.

Valuation — determining what the position is now worth. Product Control and Valuation Control, reporting to Finance.

Verification of the method — determining whether the model and the process are sound. Model Validation and Internal Audit, independent of all three.

Collapsing any two of these creates a specific failure mode. A desk that records its own trades can create positions that do not exist or suppress ones that do. A desk that values its own positions can report profit that has not been earned. A function that both builds and validates a model has no independent check on its own mathematics.

Where the conflict actually shows up

The pressure on a valuation is rarely a confrontation. It appears in four recognisable forms, and a controller learns to identify each.

Source shopping. The desk proposes a valuation source that happens to be the most favourable, presented as the most appropriate. The defence is a documented source hierarchy set in advance of any specific dispute.

Methodology drift. A methodology is revisited, with plausible technical justification, at a moment when the change happens to help. The defence is that methodology changes go to committee and their profit and loss impact is quantified and reported separately.

Threshold management. Positions marked just inside a materiality threshold, or a position split so that neither part breaches. The defence is monitoring the distribution of differences against thresholds — a book whose IPV differences cluster just below the threshold is telling you something.

Timing. A loss recognised slowly and a gain recognised quickly. The defence is symmetric policy applied without regard to direction, and monitoring of the asymmetry.

Independence has to be enforced in software, because policy alone does not survive contact with a busy trading floor.

Entitlements and segregation of duties. The trade capture system must prevent the same user from booking and approving a trade, and prevent front-office users from amending valuation adjustments. This is configured, tested and re-tested — entitlement drift as people change roles is one of the most common audit findings in the domain.

Marking rights. The system must record who is permitted to set a mark for each instrument, and enforce it. Where the desk marks and the control function verifies, both values must be stored, not just the final one, so that the size and direction of the difference is a queryable data set.

Immutable audit trail. Every mark, adjustment, override and approval is written append-only with user, timestamp and reason. Nothing may be edited in place. This is what makes a two-year-old decision reconstructable.

Override workflow. Where a controller accepts a desk value over an independent source, that override must be a first-class object with a reason, an approver and an expiry — not a silent edit to a cell.

Surveillance analytics. Patterns rather than instances: marks that persistently sit at the favourable end of the verified range, IPV differences clustering below threshold, amendment profit and loss concentrated in specific traders or specific days of the month, reserve releases coinciding with periods of weak performance.

Figure 14 The escalation ladder, and the one signal that does not use it

The teams that do the work

A markets valuation function of any size is built from six recognisable teams. Titles differ between institutions; the division of labour does not.

Team Sits in Core role Primary technology
Daily P&L / Product Control Finance Produce, validate, explain and sign daily P&L per desk Risk engine output, attribution tool, P&L reporting layer
Valuation Control Group Finance or Risk Monthly IPV, reserve calculation, fair value hierarchy IPV platform, consensus data, quote capture, reserve calculators
Financial Control Finance Sub-ledger, GL posting, substantiation, disclosure Sub-ledger, GL, reconciliation platform
Market Data and Quantitative Analytics Markets or Technology Curve and surface construction, pricing library ownership Market data platform, curve builders, analytics library
Model Validation / Model Risk Risk Independent validation, model inventory, limitation quantification Model inventory, benchmarking environment
XVA and Central Valuation Markets or Finance CVA, DVA, FVA computation and hedging Monte Carlo XVA engine, elastic compute

The independence boundary runs through the middle of this table. The first three report through Finance and are structurally independent of the trading business. Model Validation reports through Risk and is independent of both. Market Data and Analytics frequently sits inside the markets division or in technology, which is a recognised tension: the team that builds the curves is close to the desks that depend on them. Banks manage this with independent review of curve methodology changes and by placing the official close data under Finance's control even where its construction sits elsewhere. The XVA function sits in different places at different institutions precisely because it is genuinely half a trading desk and half a control function.

Mapping the teams to the process

Stage from Part 2 Owning team Escalates to
1. Trade capture and reconciliation Operations with Financial Control Head of Operations
2. Market data close and curves Market Data and Analytics Head of Markets Technology
3. Overnight revaluation Technology, run for Risk and PC Batch operations
4. Flash P&L and validation Daily P&L / Product Control PC manager
5. Attribution Daily P&L / Product Control Head of PC for the business
6. Sign-off and escalation Product Control with desk head Valuation Committee
7. Ledger close and substantiation Financial Control Divisional CFO
8. IPV Valuation Control Group Valuation Committee
9. Reserves and XVA VCG with XVA team Valuation Committee
10. Hierarchy, AVA, disclosure VCG with Group Finance Audit Committee

  1. Why does the desk head's pressure on the valuation differ from the trader's, and why is it also dangerous?

  2. How does the identity of the consumer change what counts as a good valuation?

  3. What are the four activities that must be separated, and what fails if valuation and execution are combined?

  4. What is source shopping and what defends against it?

  5. Why should both the desk mark and the independently verified value be stored, rather than only the final number?

  6. Why is the reporting line of the market data and analytics team the largest unmanaged independence exposure in many banks?

3

  1. Why is the desk head's push for smoothness dangerous?

  2. What does an auditor accept as justification?

  3. What fails if valuation and execution are combined?

4The control framework is a data pipeline, and its weakest link is upstream of everything

A valuation control function is a set of judgements resting on a set of numbers produced by a set of systems. This part works through the systems from the outside in: what functions are needed, who supplies them, how data flows between them, and what breaks.

Start with functions, not with vendors

Before naming a single product, ask what any bank must be able to do to control the valuation of a trading book. The answer decomposes into six sequential capabilities and a band of cross-cutting ones. This decomposition is stable across institutions and across decades; the products that fill the boxes are not.

Figure 15 Six functional blocks and the capabilities that cut across all of them
Functional block What it must do Failure if absent or weak
Trade capture and books Hold the authoritative, complete record of every position with full economic terms Valuation applied to the wrong population — undetectable by price checking
Market data and curve construction Acquire, validate, snapshot and build official curves and surfaces Every downstream number wrong simultaneously
Pricing and analytics library Turn instrument terms plus market data into a value and sensitivities Inconsistent valuation between risk and finance views
Risk and P&L engine Aggregate values, generate sensitivities, decompose P&L No attribution, therefore no closed test on the number
Valuation control platform Manage IPV workflow, evidence capture, reserve calculation, adjustments Verification exists as spreadsheets and cannot be evidenced
Sub-ledger, GL and reporting Post to the accounts, substantiate, classify, disclose Management number never becomes an accounting fact

The cross-cutting capabilities are the ones that get cut first and cost most later. Reference and static data — the instrument, counterparty, book and legal entity master — underpins everything and is nobody's favourite programme. The model inventory and approval workflow is a governance system whose absence only becomes visible under supervisory review. The reconciliation engine is the connective tissue between every pair of systems. The immutable audit trail is worthless until the day it is the only thing standing between the bank and an unanswerable question. Entitlements and segregation of duties enforce the independence described in Part 3. Workflow and sign-off turn a process into a record.

Who supplies what, and where banks build

The buy-versus-build pattern in this domain is unusually informative, because it maps precisely onto where banks believe their competitive advantage lies. Naming the products makes the map concrete. The names below are illustrative of each segment rather than exhaustive, and none is an endorsement.

Figure 16 The buy-build map is a map of where judgement lives

Trade capture and cross-asset trading platforms are bought. The established cross-asset platforms — Murex MX.3, Adenza Calypso, FIS Summit and Sophis, ION and Openlink in commodities and treasury — dominate rates, credit and foreign exchange derivatives because building a system that handles every product a markets division trades, with full lifecycle events, is enormous and confers no advantage. Banks differentiate on what they trade, not on where they record it.

Market data is bought; curve construction is usually built. Raw data comes from Bloomberg, LSEG, ICE Data Services and S&P Global Market Intelligence, from exchanges directly, and from inter-dealer brokers. The construction of official curves and surfaces from that raw material is typically proprietary, because the methodology is a genuine differentiator and because it must be tightly coupled to the pricing library.

The pricing analytics library is mixed, and this is the most contested boundary in the stack. Commercial libraries — Numerix, FINCAD, and the analytics embedded in the trading platforms themselves — provide broad, validated, documented coverage of standard products and are heavily used. Open-source libraries such as QuantLib appear widely for benchmarking and for less exotic products. Banks build proprietary models where their franchise is distinctive or where a commercial library's approach does not match their risk view. The long trend is towards commercial coverage for vanilla and proprietary code for the exotic tail, with an important consequence: the boundary between the two must be actively managed, because a book valued half from one library and half from another will not attribute cleanly.

Risk and profit and loss engines are bought as part of the trading platform or as a separate risk system, then extended heavily. Almost no bank runs a vendor risk engine unmodified.

Independent price verification splits into evidence and workflow, and banks treat them differently. The evidence is bought and there is little choice about it: IHS Markit Totem is the dominant consensus service for over-the-counter derivatives, alongside ICE Data Services and Bloomberg BVAL for evaluated pricing of securities. The workflow — the population selection, the challenge process, the sign-off, the audit trail — is the least standardised block in the whole stack. Some banks buy a dedicated valuation control product; many build; and a disturbing number still run material parts of it on spreadsheets, which is the most common significant finding in supervisory reviews of valuation control.

The reserve and adjustment engine is essentially always proprietary. Reserve methodology is bank-specific, changes with the book, and encodes judgement no vendor can generalise. This mirrors a pattern across capital markets technology: the calculation that embodies the institution's own judgement is the one it writes itself.

Reconciliation is bought. Duco, Gresham and SmartStream are the recognisable names; the market is mature precisely because the problem is generic — compare two files, apply tolerances, manage the breaks.

The general ledger is bought universally, overwhelmingly from SAP or Oracle, with the sub-ledger sitting between it and the trading systems as either a vendor product or a build.

Block Pattern Representative suppliers Why
Trade capture and books Buy Murex, Calypso, Summit, Sophis, Openlink Records facts; no advantage in building
Market data acquisition Buy Bloomberg, LSEG, ICE Data, S&P Global Commodity input, licensed not created
Curve and surface construction Build Proprietary, on vendor data Methodology is a differentiator
Pricing analytics Mixed Numerix, FINCAD, QuantLib + in-house Vendor for vanilla, proprietary for exotics
Risk and P&L engine Mixed Vendor core, extended in-house Aggregation logic is bank-specific
IPV evidence Buy IHS Markit Totem, ICE, Bloomberg BVAL Independence requires an external source
IPV workflow Mixed Vendor, in-house, or spreadsheets Least standardised block in the stack
Reserve and adjustment engine Build Always proprietary Encodes the institution's own judgement
Reconciliation Buy Duco, Gresham, SmartStream Generic problem, mature market
Sub-ledger and GL Buy SAP, Oracle Accounting is standardised

Follow one trade all the way through

The clearest way to understand the architecture is to trace a single object from creation to regulatory report. Take one trade on Desk SRC-1: a ten-year euro payer swaption, notional €250 million, executed on 14 October.

Figure 17 One trade, six systems, six data objects — and a dependency that only runs one way

At execution, the trade is captured with its full economic terms: counterparty, notional, currency, option expiry, underlying swap tenor, strike, exercise style, premium, collateral agreement, book, legal entity and a unique trade identifier. This record is the atom of everything that follows.

At the close, the market data platform snapshots the swaption volatility surface and the euro discount and forecasting curves. The relevant curve nodes and surface points become inputs bound to this valuation.

Overnight, the pricing library takes the trade terms and the market data and produces a present value, a vector of bucketed rate sensitivities, a vega by expiry and strike, and second-order sensitivities. These are written to the risk store with references to both the trade version and the market data version used.

In the early morning, the risk and profit and loss engine differences today's value against yesterday's, allocates the change across risk factors, and produces the attribution line items that appear in the desk's pack.

During the day, the value and the associated reserves are posted to the sub-ledger as a fair value balance and an adjustment balance, and summarised into the general ledger.

At quarter-end, the position contributes to the fair value hierarchy disclosure, to the additional valuation adjustment calculation, and to the regulatory reports covering financial and capital positions. The same trade also appears in transaction reporting under derivatives reporting obligations, keyed on its unique trade identifier.

The cascade principle

Dependency runs in one direction only, and that asymmetry is the central fact of the architecture. An error introduced at trade capture propagates to every subsequent stage. An error in the curve propagates to valuation, sensitivities, attribution, ledger and capital. An error in the ledger posting affects only the accounts.

The practical consequence is that control effort should be weighted towards the front of the chain, and it usually is not. Downstream errors are visible: a reconciliation breaks, a report fails to balance, someone notices. Upstream errors are silent, because everything downstream agrees with everything else — they were all fed from the same wrong input. A perfectly reconciled set of systems all agreeing on a wrong curve is the most dangerous state the architecture can be in, and it produces no exceptions at all.

The defence against silent upstream error is not reconciliation, which by construction cannot detect it. It is independent verification against an external source — which is exactly what independent price verification is, and which explains why IPV cannot be replaced by better internal controls no matter how good they get.

The anatomy of the two atoms

Two data objects carry the entire process: the trade record and the market data point. Understanding their fields is understanding what can go wrong.

The trade record

Field group Examples What it drives downstream
Identity Internal trade ID, UTI, version number Lineage, reporting, amendment tracking
Economic terms Notional, currency, dates, rate, strike, index The valuation itself
Counterparty Legal entity, LEI, netting set, CSA reference CVA, collateral, credit limits
Booking Book, desk, legal entity, trading vs banking book P&L ownership, capital treatment
Classification Product type, model assignment, FV hierarchy level Which model prices it, how much IPV it attracts
Lifecycle Trade date, effective date, amendment history, status Prior-day P&L, amendment reporting
Valuation metadata Marking source, last verified date, reserve allocation IPV population, reserve attribution

Three fields carry disproportionate risk. The netting set and collateral agreement reference determines whether a position is treated as collateralised, which changes the discount curve and the credit adjustment — a trade with a missing collateral agreement reference is valued as uncollateralised and can be wrong by a large margin. The trading versus banking book flag determines the entire capital treatment and is a supervisory focus, because the boundary between the two is a place where value can be moved to reduce capital. The model assignment determines which pricing model runs, and an instrument silently defaulting to a generic model because its specific one was not mapped is a failure mode that produces plausible numbers.

The market data point

Field Example Why it matters
Instrument / factor key EUR 10Y par swap rate Identity of what is being observed
Value 2.847% The observation
Source Named vendor, broker, exchange or internal Precedence and independence assessment
Snapshot timestamp 14 Oct, 17:00:00 CET Consistency across engines
Quality flag Verified / stale / interpolated / extrapolated Whether it is evidence or assumption
Construction version Curve builder v4.2, SABR fit Reproducibility of the valuation
Contributor count and dispersion 11 contributors, σ = 0.9 vol points Direct input to reserve sizing

The quality flag and the construction version are the two fields most often absent from legacy market data platforms, and both are essential. Without a quality flag, an interpolated point and an observed point are indistinguishable downstream, which makes it impossible to determine automatically whether a valuation rests on evidence or assumption — and therefore impossible to automate fair value hierarchy classification. Without a construction version, a valuation cannot be reproduced after a methodology change, which makes any investigation of a historical number a forensic exercise.

Deduplication is a real and unglamorous problem

The same instrument arrives from multiple sources under multiple identifiers, and resolving them to one entity is harder than it sounds. A single corporate bond may appear with an ISIN from one vendor, a CUSIP from another, a proprietary code from a broker feed and an internal instrument code in the bank's own master. A single counterparty may appear as three subsidiaries in the credit system and one consolidated group in the collateral system.

The consequences of getting it wrong run in both directions. Failing to recognise that two records are the same instrument double-counts the position, or values the same exposure twice with two different marks. Wrongly merging two distinct instruments — two bonds from the same issuer with different seniority, two contracts under different netting agreements — creates an exposure that nets when it should not, which is a real economic error and not merely a data one.

This is why the instrument and counterparty masters, unglamorous as they are, sit at the top of any credible architecture assessment. They are the place where identity is established, and every downstream aggregation assumes identity has been established correctly.

Identifiers: the vocabulary of identity

Identifier What it names Where it bites
ISIN A specific security, globally Bonds and equities; not defined for most OTC derivatives
CUSIP A security, North American convention Mapping to ISIN required for cross-region books
FIGI An instrument, vendor-neutral and openly licensed Resolving the same instrument across vendor feeds
LEI A legal entity Counterparty identity, netting, regulatory reporting
UTI A single derivative transaction, agreed between both parties Trade reporting, reconciliation with the counterparty
UPI A derivative product type, not an individual trade Aggregating OTC positions for reporting and analysis
Internal instrument ID The bank's own master key The join point for every internal system

The identifier problem in derivatives is structurally different from the one in securities. A bond has an ISIN because it is issued once and traded many times by many parties. A bespoke swap is a bilateral contract that exists in exactly two systems, and identity has to be agreed between two counterparties rather than assigned by an issuer. The unique transaction identifier exists precisely to solve that: both parties reference the same string, which makes reconciliation and regulatory reporting possible. Where UTI generation and exchange is weak, portfolio reconciliation with counterparties degrades, and portfolio reconciliation is one of the few genuinely independent checks on the completeness of the trade population.

How the systems talk to each other

Three integration patterns coexist in every bank, representing three technology generations, and each has a distinct effect on the control framework.

Batch file transfer is the legacy default and still carries the majority of the volume. A file is produced at a point in time and consumed downstream. It is simple, auditable and easy to reconcile. Its weakness is latency and its deeper weakness is that it creates states — the downstream system holds a copy of the upstream truth as of a point in time, and the two can diverge. Every reconciliation in a bank exists because of a state copy somewhere.

Application programming interfaces are the modernisation pattern. A downstream system requests the current value when it needs it rather than holding a copy. This removes a class of reconciliation entirely, because there is no second copy to diverge. It introduces availability coupling: if the pricing service is down, the consumer cannot proceed.

Event streaming pushes changes as they happen. A trade booking, an amendment, a market data update is published once and consumed by every interested system. Consumers maintain their own projections but derive them from a single ordered log rather than from independent extracts.

Figure 18 Removing the manual hand-offs removes the copies, and removing the copies removes the reconciliation

The migration from batch to streaming is usually justified as a latency improvement. That undersells it substantially.

Batch architectures require reconciliation by construction. If two systems each hold their own copy of the trade population, refreshed independently, they will differ, and the bank must run a process to find and explain the differences. That process consumes real headcount and produces breaks that must be aged, investigated and cleared.

A single ordered event log removes the reason the copies could differ. Every consumer sees the same events in the same order. Divergence becomes a bug rather than an expected operational condition. The reconciliation does not get faster — it becomes unnecessary.

The same log also solves the lineage problem almost for free. An append-only event stream is an audit trail: the sequence of events that produced any state is recoverable by replay. A bank that has invested in event streaming for latency reasons has usually acquired an audit capability it has not yet realised it owns.

The practical obstacle is rarely technical. It is that migrating requires the downstream systems to change how they consume data, and those systems are numerous, old and owned by different budgets.

Immutability and the adjustment engine

Two specific properties of the valuation control platform deserve separate treatment because they are frequently under-specified.

Records that support a financial statement must be append-only. A mark that was overridden, an adjustment that was reversed, a reserve that was released — each must remain visible with its original value, its change, the user who made it, the timestamp and the stated reason. Editing in place destroys the evidence that a control operated. When a supervisor asks why a position was carried at a particular level eighteen months ago, the answer must be reconstructable from the system rather than from the memory of someone who may have left.

The adjustment and reserve engine is where the bank's own judgement is codified, and it must be treated as such. It takes sensitivities, market data, dispersion measures, model validation findings and policy parameters, and produces reserve numbers that reduce reported profit. Three properties follow. It must be versioned, because a change in methodology changes the number and the change must be attributable. It must be reproducible, meaning that re-running a prior period reproduces the prior number exactly. And its parameters must be governed — the bid-offer spreads assumed, the confidence levels applied, the concentration thresholds — because a parameter change is a methodology change wearing a disguise.

Messaging: four standards, four clock speeds

Data crosses institutional boundaries through standardised messages, and the standard in use tells you which layer of the bank you are in.

Figure 19 The standards form a chain from execution to the accounts, and each runs on a different clock
FIX — the front-office execution layer

The Financial Information eXchange protocol carries orders, executions and confirmations between trading counterparties and venues. It is terse, session-based and designed for low latency: messages are short tag-value pairs, and the protocol has been the electronic trading backbone since the 1990s. In the valuation context FIX matters because it is how executed trades reach the trade capture system for listed and electronically traded instruments, which is to say how the trade population is built in the first place.

FpML — the over-the-counter derivatives layer

Financial products Markup Language is the XML-based standard for describing over-the-counter derivative contracts in full economic detail. Where FIX describes an execution, FpML describes a contract: every leg, every schedule, every optionality, every collateral term. This distinction is fundamental. A swaption cannot be represented as a price and a quantity; it needs a structured description of its exercise schedule and its underlying.

FpML is the language of confirmation matching, of portfolio reconciliation with counterparties, and of trade reporting to repositories. For a valuation function it matters because it is the standard in which the trade population can be exchanged and compared with a counterparty — which is the only truly independent check on completeness available.

SWIFT MT and ISO 20022 — the post-trade layer

Settlement, custody, collateral and payment instructions travel over SWIFT. The legacy MT message series is being progressively replaced by ISO 20022 MX messages, which are XML-based, carry substantially richer structured data, and are being adopted through a coordinated multi-year industry migration.

The richness matters for control purposes, not just for operations. An MT message carries limited structured reference data, so information that a control function might want — the specific position a margin call relates to, the reason for a dispute — often arrives as unstructured text or not at all. ISO 20022 carries it as structured fields, which means post-trade exhaust can be consumed automatically as control evidence rather than read by a person.

The ledger feed — the accounting layer

The final hop, from sub-ledger to general ledger, uses no industry standard at all. It is a proprietary interface between two internal systems, typically a batch journal feed. This is the least standardised and most bank-specific link in the whole chain, and it is where a surprising amount of accounting complexity is encoded in transformation logic that nobody has documented.

Layer Standard Carries Timescale
Execution FIX Orders, executions, listed confirmations Milliseconds to seconds
OTC contract FpML Full contract economics, confirmations, reconciliation Minutes to hours
Market data Vendor APIs and file feeds Quotes, curves, consensus levels Real time to overnight
Post-trade SWIFT MT → ISO 20022 Settlement, collateral, margin, custody T+1 to T+2
Accounting Proprietary journal feed Fair value, reserves, accruals Daily to monthly

Operational resilience: the batch that cannot be late

Resilience in this domain has an unusually crisp definition, because there is exactly one irreversible moment: the market opens at a fixed time whether or not the control process finished.

The consequences of missing it are concrete. A desk that opens without a validated position and risk view is trading blind. Risk limits cannot be monitored against a position that has not been computed. Margin calls cannot be validated. And an intraday failure to produce a position for a regulated activity is a reportable operational incident in most jurisdictions.

Concept What it guarantees Why it matters on the critical night
Redundancy A spare component exists Necessary but insufficient: a spare that must be activated manually costs the time you do not have
Automatic failover The spare takes over without human action Removes the detect-decide-act delay, which is where most of the outage actually is
Checkpointing The batch restarts from the last completed step Converts a full re-run into a partial one, which is usually the difference between meeting and missing the open
Idempotency Re-running a step produces the same result Makes restart safe; without it, restart risks double-posting or double-counting
Disaster recovery Processing can move to another site Covers site loss, but recovery time objectives are typically measured in hours
Business continuity The process can run in degraded mode The documented fallback: which books can be valued at prior-day marks, who approves it, what is disclosed
Rehearsal The above have been tested under realistic conditions An untested failover is a hypothesis, and the cutover is exactly when hypotheses fail

Two of these deserve emphasis because they are architectural rather than infrastructural. Checkpointing and idempotency are properties of how the batch is written, not of what hardware it runs on. A batch built as a single long-running job that must complete or restart from zero cannot be made resilient by buying more machines. A batch built as a sequence of idempotent, checkpointed steps recovers in minutes on modest infrastructure.

The degraded-mode plan is the control the auditors will actually ask about. When the batch genuinely fails, someone must decide what happens: which books are valued at prior-day marks, what materiality threshold triggers a manual valuation instead, who approves the fallback, and how the resulting profit and loss is flagged and corrected. A bank without a documented, exercised degraded-mode procedure will improvise under time pressure, and improvisation at 06:00 is how a temporary outage becomes a misstatement.

Changing the machine without breaking the number

Everything above describes the system running. This section describes changing it, which is where a disproportionate share of valuation incidents originate. A curve methodology is improved, a pricing model is upgraded, a vendor release is applied, a book is migrated to a new platform. Each of those moves every valuation it touches, on a day when nothing happened in the market.

The parallel run is the control, and it is expensive

No material valuation change goes live without both versions producing numbers side by side for a defined period. The old system remains authoritative; the new one runs alongside; the differences are analysed daily.

What the parallel run is actually testing is not whether the new number is different. It will be different — that is usually the point of the change. It is testing whether the differences are explained, bounded and expected. A revised interpolation method should move long-dated positions and leave positions on quoted nodes untouched; if it moves everything, something else changed too.

Question the parallel run answers What a bad answer looks like
Which positions moved, and do they match the change's intended scope? Movements in books the change should not have touched
Is the size of the movement within the pre-agreed tolerance? A tail of positions moving far more than expected
Do the differences net to something explicable, or do they cancel? A small net difference hiding large offsetting errors
Does the new version reproduce the old on unchanged inputs? Any difference at all — this is a defect, not a change
Do the sensitivities move consistently with the values? Values move, Greeks do not, or vice versa

The fourth row deserves emphasis. A regression test that feeds identical inputs to old and new and demands identical output is the cheapest and most powerful test available, and it is skipped constantly because it seems redundant when the whole point was to change the answer. It is not redundant: it separates the intended change from everything else that came with the release.

Duration, and who signs

The length of the parallel run is a policy parameter, not a project decision, and it should scale with the materiality of the change and with the volatility of the affected books. A vendor patch to a vanilla pricer might parallel for a week; a replacement of the curve construction framework, for a full month-end cycle including an independent price verification and a reserve calculation, because those are the processes that consume the output.

The bascule itself is a controlled event. Someone signs that the differences were analysed and accepted. The resulting one-off profit and loss is quantified before the change, disclosed separately as model change profit and loss, and never absorbed into market movement. And the old version is retained in a runnable state, because the most common discovery after a migration is that a position class was mishandled and the prior number has to be reproduced.

Where the work is actually done

The functional description in Part 3 says nothing about geography, and geography shapes the control framework more than most organisation charts admit.

The daily cycle follows the sun, and the handover is a control point. A global book is controlled by teams in each major trading centre, and the position at the Asian close becomes the opening context for Europe, which hands to the Americas. Each handover transfers not only work but unresolved exceptions, and a break passed across three time zones with three partial investigations is a recognisable failure mode. Mature functions treat the handover as a formal artefact — a structured record of what is open, what was concluded and what the next region must decide — rather than an email.

Production has largely moved to lower-cost centres; judgement has not. The split that works in practice divides activities by whether the answer is determinable from documented rules.

Activity Typically located Why
Reconciliation break clearing Offshore delivery centre Rule-driven, high volume, documentable
Standard reporting and pack production Offshore delivery centre Repetitive, deadline-driven
First-level exception review Offshore, escalating onshore Triage is rule-based; resolution often is not
Daily sign-off and desk challenge Trading centre Requires relationship, context and authority
IPV judgement and source assessment Trading centre Requires market knowledge and independence
Reserve methodology and committee Trading centre, senior Institutional judgement, minuted

The model fails in a specific and predictable way. It works when the process being relocated has genuinely been reduced to documented rules. It fails when what was actually exported was undocumented judgement — a controller who "knew" which breaks mattered, applying criteria nobody had ever written down. The test before relocating any activity is whether the procedure can be executed correctly by someone who has never met the desk. If it cannot, the activity is not ready to move, and moving it anyway converts a working control into a queue.

Measuring the data, not just moving it

BCBS 239 requires accuracy and completeness; it does not define them. Operationally, four dimensions cover the ground, and each has a metric that can be computed automatically and trended.

Dimension The question A usable metric
Completeness Is everything that should be here, here? Trades priced ÷ trades expected; curve nodes populated ÷ nodes required
Accuracy Are the values right? IPV difference distribution; count of corrections post-sign-off
Timeliness Did it arrive in time to be used? Batch stage completion versus target; feed arrival time distribution
Consistency Do the copies agree? Reconciliation break count and ageing; risk-versus-finance position variance

One further dimension matters specifically in this domain and is usually missed: provenance. Can the source of every input be identified, and can the valuation be reproduced from it? That is not a data quality attribute in the classical sense — it is a property of the architecture — but it is the one the supervisor tests, and it is unmeasurable unless the quality flag and construction version described earlier are captured at the point of use.

The practical implementation is a data quality dashboard fed from the same pipeline that produces the valuations, not a separate reporting exercise. Every batch stage emits its own metrics: rows expected and received, nodes stale, trades failed, breaks opened and closed, stage completion time against target. Thresholds raise exceptions into the same queue as valuation exceptions, because a controller investigating a residual and a controller investigating a late feed are frequently investigating the same incident.

The measure that distinguishes a real framework from a reporting one is whether any metric has ever blocked a process. A dashboard that only informs is a dashboard. A completeness check that halts the ledger posting until it passes is a control.

Six regulatory regimes, read as system specifications

Regulation shapes valuation infrastructure more directly than it shapes almost any other part of a bank, because these rules specify outputs that systems must produce. Read as engineering requirements rather than as legal texts, six regimes matter.

IFRS 13 and ASC 820 — the fair value measurement standards

As a system specification: the platform must classify every fair-valued position into a three-level hierarchy at each reporting date, based on the observability of significant inputs; produce a full roll-forward of Level 3 balances distinguishing purchases, sales, settlements, transfers and realised versus unrealised gains; and generate a sensitivity analysis of Level 3 valuations to reasonably possible alternative inputs.

What it implies architecturally: the market data quality flag must be captured at the point of use and bound to the valuation, because classification cannot otherwise be produced systematically; and the alternative-input sensitivity must be computable, meaning the pricing library must support revaluation under specified parameter overrides.

Prudent valuation under European capital regulation

As a system specification: compute nine categories of additional valuation adjustment to a 90% certainty standard, aggregate them under the prescribed method, and deduct the result from common equity tier 1 capital, quarterly, with full supporting evidence.

What it implies architecturally: valuation uncertainty must be quantified per position and per risk factor rather than assessed qualitatively; dispersion data from consensus services must be retained as a first-class input rather than consulted and discarded; and the calculation must be reproducible for supervisory inspection long after the reporting date.

The Fundamental Review of the Trading Book

As a system specification: maintain a formal trading-desk structure; compute hypothetical and risk-theoretical profit and loss series per desk daily; run the attribution test statistics over a rolling window; assess each risk factor for modellability against real-price observation criteria; and capitalise non-modellable factors separately.

What it implies architecturally: this is the most demanding of the five. Daily production of two parallel profit and loss series per desk requires the risk and finance views to be reconciled at desk granularity every day, which many banks did not previously do. Real-price observability assessment requires the bank to track, per risk factor, whether committed quotes or executed transactions were observed — a data-collection obligation that did not exist before and drove the growth of industry data-pooling initiatives.

Model risk governance

As a system specification: maintain a complete inventory of models in production with version, owner, approval status, scope, limitations and expiry; enforce that no model runs outside its approved scope; support independent validation with access to benchmark environments; and track the quantified impact of known limitations.

What it implies architecturally: the model inventory must be integrated with the pricing library rather than maintained alongside it as a document. An inventory that is a register maintained by hand will diverge from what is actually executing, and that divergence is exactly the finding a supervisor is looking for.

BCBS 239 — risk data aggregation and reporting

As a system specification: be able to aggregate risk data accurately and reliably across the group, on demand and under stress, with complete and traceable lineage from source to report; produce reports that are accurate, comprehensive, clear and timely; and demonstrate that the aggregation is largely automated rather than assembled manually.

Why it belongs in this list even though it is a risk standard. The fourteen principles of the Basel Committee's standard on effective risk data aggregation and risk reporting are, read carefully, a specification for exactly the data architecture this part has described. They require a single authoritative source for each data element, an unambiguous taxonomy, reconciliation between risk and accounting data, and the ability to reproduce a reported number from its inputs. A valuation control function cannot satisfy the standard's lineage requirement without the market data quality flag and construction version described earlier, and cannot satisfy its aggregation requirement without a resolved instrument and counterparty master.

What it implies architecturally: manual workarounds are explicitly treated as a weakness, not a mitigant. A number that is correct but assembled by a person from four extracts fails the standard even though it is right. This is the clearest regulatory statement anywhere that the method of production is part of the control, not merely a means to it — and it is the reason the unglamorous data programmes in Part 5's transformation filter have a compliance justification as well as an economic one.

Supervisory assessments of compliance with the standard have consistently found the industry short of the expected level more than a decade after publication, with data architecture and lineage the most persistent gaps. That durability is informative: the requirement is not conceptually difficult, it is expensive and it competes with everything else for budget.

Internal control over financial reporting

As a system specification: identify the controls on which the financial statements depend, document them, test their operation and evidence the testing. For valuation this means the reconciliations, the sign-offs, the segregation of duties, the approval workflows and the completeness of the trade population.

What it implies architecturally: controls must produce machine-readable evidence of their own operation. A control that operated but left no trace is, for these purposes, a control that did not operate — which is why workflow, sign-off and immutable audit are not conveniences but requirements.

Regime Core system obligation Hardest technical implication
IFRS 13 / ASC 820 Hierarchy classification, Level 3 roll-forward and sensitivity Data quality flags bound to each valuation
Prudent valuation Nine AVA categories to a 90% certainty standard Position-level quantified uncertainty, reproducible
FRTB Daily HPL and RTPL per desk, PLA test, NMRF Risk and finance reconciled daily at desk level
Model risk governance Live model inventory with enforced scope Inventory integrated with execution, not documented alongside
BCBS 239 On-demand, traceable, largely automated aggregation Manual workarounds count as weaknesses, not fixes
ICFR / SOX Documented, tested, evidenced controls Controls emitting machine-readable evidence

  1. Why can reconciliation not detect an upstream market data error?

  2. What is the difference between FIX and FpML, and why does an OTC derivatives book need the second?

  3. What does FIX carry that SWIFT does not, and vice versa?

  4. Why is redundancy alone insufficient for the overnight batch?

  5. Why do the quality flag and construction version fields matter so much on a market data point?

  6. Why does event streaming remove reconciliation rather than merely accelerate it?

  7. Which of the five regulatory regimes is architecturally hardest, and why?

  8. In a parallel run, why is it a defect rather than a success if the new version differs from the old on unchanged inputs?

  9. What is the test for whether an activity can be relocated to an offshore delivery centre?

  10. Why is a reserve calculation in a spreadsheet a control problem rather than a tooling preference?

4

  1. Why can't reconciliation detect an upstream market-data error?

  2. Why does an OTC derivatives book need FpML?

  3. What does SWIFT carry that FIX does not?

5What goes wrong is knowable in advance, because it has already happened

The failures in this domain are not exotic. They repeat, with the same three shapes, across institutions and across decades. This part names the risks, provides the metrics that distinguish a controlled book from a lucky one, examines three real cases, and assesses where technology genuinely changes the economics.

Six risks, and they differ in how controllable they are

A complete taxonomy of what can go wrong in valuation and product control has six categories. They are mutually exclusive and collectively exhaustive: every incident in this domain falls into exactly one, and the categorisation determines who is accountable and what defence applies.

Risk type What it is Who owns it Main defence
Valuation and model risk The model is wrong, mis-specified or used outside its scope Model Risk Management with Valuation Control Independent validation, model inventory, model reserve
Data and input risk The inputs are stale, wrong, misapplied or unavailable Market Data with Product Control Source hierarchy, staleness detection, IPV, dispersion-based reserves
Process and operational risk The process fails: batch, reconciliation, completeness, timing Technology and Operations Checkpointing, automated exceptions, degraded-mode procedure
Conduct and mismarking risk Someone deliberately marks a position away from its value Front office supervision with Compliance Independence, override workflow, surveillance analytics
Reporting and disclosure risk The number is right internally but wrongly classified or disclosed Group Finance Classification controls, ICFR testing, external audit
Capital and regulatory risk Valuation weakness translates into capital cost or model loss CFO and CRO jointly Prudent valuation, PLA test monitoring, remediation tracking
Figure 20 Six risks placed by how far engineering can take you

How to tell a controlled book from a lucky one

A trading book that has not blown up is not evidence of control. Most books do not blow up in most years. The metrics below distinguish a process that would catch an error from one that has simply not yet been tested.

Indicator What it measures What good looks like
Unexplained P&L ratio Daily residual as a share of gross P&L, and its stability Small and systematically small; no fat tail on volatile days
IPV coverage by value Percentage of fair value independently verified, split by hierarchy level High and stable; Level 3 at or near 100%; unverified residual named
IPV difference profile Size, direction and distribution of desk-versus-verified differences Centred near zero, symmetric, no clustering just below threshold
Reconciliation break ageing Age profile of unreconciled items, not their count Very few items over 30 days; a stable ageing curve
Reserve adequacy and stability Reserve as a share of Level 3 value, and its volatility Moves with the book and with market conditions, not with results
Batch completion and recovery Time to complete, and time to recover from a mid-run failure Consistent completion with margin; rehearsed recovery inside the window
PLA test zone Regulatory attribution test outcome per desk Green, stable, with early warning on desks trending amber

A second warning applies to reserves. A reserve that is stable in absolute terms while the book grows is a reserve being managed towards a result rather than towards the risk. A reserve that falls in a period of weak trading performance deserves specific examination, because reserve release is the easiest source of profit available to a desk under pressure, and it requires no market to cooperate.

Three cases, three failure modes

These three are chosen because they are real, well documented, and cover the three distinct ways valuation control fails. One is deliberate mismarking. One is a control structure that existed but was ineffective. One is a market that stopped providing the evidence the whole framework depends on.

Credit Suisse, 2007–2008: deliberate mismarking of an illiquid book

What happened. Traders on Credit Suisse's structured credit desk held a portfolio of asset-backed securities and related derivatives — a book referred to internally as ABN1, containing roughly $3.5 billion of subprime mortgage bonds alongside credit default swaps, valued at approximately $4.7 billion at the end of 2007. As the subprime market deteriorated through late 2007, the traders marked the positions above their actual value in order to conceal losses, protect year-end bonuses and, in one case, secure a promotion.

How it unravelled. The mismarking was detected in early 2008, and in February 2008 Credit Suisse disclosed approximately $2.65 billion of additional subprime-related losses attributable to the misconduct. The United States Securities and Exchange Commission subsequently charged four former employees, including the global head of structured credit trading and the head of hedge trading; criminal proceedings followed, with guilty pleas from several of those involved.

The valuation-control reading. This is the pure conduct case. The instruments were illiquid, external evidence was thin and deteriorating, and the desk's own marks carried disproportionate weight because there was little to challenge them with. The failure was not that a model was wrong. It was that a book whose value depended almost entirely on assertion was not subject to independent challenge proportionate to that dependence — and the moment when the market for external evidence dried up was exactly the moment when the control needed to intensify rather than relax.

JPMorgan, 2012: a valuation control function that existed and did not work

What happened. Traders in JPMorgan's Chief Investment Office in London built a very large synthetic credit portfolio. When positions moved against them in early 2012, the traders departed from their prior marking practice in order to reduce the reported losses. Because most positions had a bid-offer spread, marking towards the favourable end of the spread rather than at mid produced a materially better result without any single mark being obviously indefensible.

The control that should have caught it. The Chief Investment Office had its own Valuation Control Group, whose explicit function was to test trader marks against external mid-market benchmarks within pre-established thresholds calibrated to bid-offer spreads. It did not catch the problem. Subsequent examination found the group was ineffective and insufficiently independent from the traders it was meant to police, and that flaws in the valuation controls dated back several years.

The consequences. JPMorgan restated its first-quarter 2012 results, reducing consolidated net revenue by $660 million and after-tax net income by $459 million. Trading losses on the portfolio ultimately ran to approximately $6 billion. The firm paid roughly $920 million in penalties across regulators in the United States and the United Kingdom, including a $200 million settlement with the Securities and Exchange Commission in which it admitted wrongdoing, and a fine of over £137 million from the United Kingdom's Financial Conduct Authority.

The valuation-control reading. This is the most instructive of the three, because the control existed on paper. The specific mechanism of failure — marking within the bid-offer spread — is the hardest form of mismarking to detect, because no individual mark is outside a defensible range. Detecting it requires looking at the pattern: consistent positioning at the favourable end, across positions, over time, in one direction. That is a data analysis, not a price check. And the structural failure was the one described in Part 3: a control function embedded within the business it was controlling, dependent on it, and lacking the independence its function required.

March–April 2020: the market stopped supplying evidence

What happened. As the pandemic shock hit financial markets in March 2020, liquidity withdrew across asset classes with extraordinary speed. Corporate bonds that had traded daily stopped trading. Exchange-traded funds holding fixed income traded at large discounts to the calculated value of their underlying holdings, raising the question of which of the two was the real price. Basis relationships between instruments that normally track each other dislocated and stayed dislocated. In April, the front-month contract for West Texas Intermediate crude oil settled at a negative price for the first time, breaking pricing models that assumed prices could not go below zero.

The valuation-control reading. No one mismarked anything, and no model was mis-specified in the ordinary sense. The framework's foundation simply withdrew: consensus services thinned as contributors themselves lost confidence, broker quotes became indicative rather than executable, observed transactions ceased, and positions that had been comfortably Level 2 became Level 3 in a matter of days. Banks with the ability to reassess observability dynamically, widen reserves in response to dispersion, and reclassify positions promptly navigated it. Banks whose hierarchy classification was a static attribute set at onboarding, and whose reserve parameters were fixed constants reviewed annually, found their control framework silently ceased to function at the moment of maximum need.

Case Failure type Step it maps to Lesson
Credit Suisse 2007–08 Deliberate mismarking of an illiquid book Step 8 — IPV Where external evidence is weakest, independent challenge must be strongest, not most accommodating
JPMorgan 2012 Ineffective and non-independent control function Steps 6 and 8 — sign-off and IPV Marking within the spread is invisible per position and obvious in the pattern; independence must be structural
March–April 2020 Loss of observable inputs across markets Steps 2, 9 and 10 — data, reserves, classification Observability is a live variable; a framework with static classification and fixed parameters fails exactly when needed

Generative artificial intelligence: real uses, real constraints

Large language models have found genuine footholds in valuation control, and the pattern of where they work is consistent. They are useful where the output is language and a human verifies it. They are dangerous where the output is a number and nobody does.

Drafting the explanation, not producing the number

A substantial part of a product controller's day is writing commentary: explaining a day's profit and loss to a desk head, documenting why a residual arose, summarising a month's IPV findings, preparing the narrative for a valuation committee pack. A model with access to the attribution output, the market moves and the exception log can draft this in seconds, and the controller edits rather than composes. The number is produced by the risk engine as it always was; only the prose is generated.

Investigating exceptions faster

When a residual appears, the controller's task is a search: which positions contributed, what changed, has this happened before, what was concluded last time. A model with access to the historical exception log, prior investigation notes and the current day's data can propose candidate explanations ranked by plausibility. The controller tests them. This compresses the slowest and least structured part of the daily cycle.

Reading and comparing documentation

Model validation and new product approval involve reading long technical and legal documents and identifying what matters. A model can extract the stated limitations from a validation report, compare a new product's term sheet against the approved scope of an existing model and flag the differences, or check that a set of control procedures still describes what the systems actually do. The judgement remains human; the reading is accelerated.

Where it must not go

Producing or adjusting a valuation directly is out of scope, and the reason is not conservatism. A pricing model must be deterministic, reproducible, independently validatable and explainable in mechanical terms — the requirements set out in model risk governance. A large language model satisfies none of these for numerical output. It is not reproducible, its reasoning cannot be audited in the required sense, and it is capable of producing a plausible number with no derivation behind it. That combination is precisely the failure mode the entire model governance framework exists to prevent.

Use of generative AI The gain The governing risk
Drafting P&L and IPV commentary Hours of controller time daily Fluent narrative that misdescribes the numbers; requires review against source
Ranking candidate explanations for a residual Compresses the slowest part of the cycle Anchoring: a confident wrong hypothesis crowds out the right one
Extracting limitations from model documentation Faster, more consistent review Omission of a material limitation; needs verification against the source
Comparing new products against approved model scope Earlier detection of scope breach False assurance; the sign-off must remain human
Summarising committee material Faster preparation, better consistency Loss of the specific detail that made the judgement defensible
Producing or adjusting a valuation Not applicable Excluded: not reproducible, not validatable, not explainable

Three risks govern every use above. Confidentiality: a trading portfolio is among the most sensitive data a bank holds, and prompts containing position detail must not leave a controlled environment — which for most institutions means privately hosted models with no retention. Accuracy: these models produce fluent text regardless of whether it is correct, and fluent text about numbers is unusually persuasive, so every generated statement about a figure must be traceable to the figure. Explainability and model governance: any model influencing a financial reporting output falls within the model risk framework, which means inventory, validation, defined scope and periodic review — the same regime applied to a pricing model, applied to a language model.

  1. Why is conduct risk the only one of the six categories that cannot be addressed by better engineering alone?

  2. Why is the average unexplained P&L a misleading metric?

  3. What made the JPMorgan 2012 mismarking so hard to detect at the position level?

  4. What is the specific lesson of March–April 2020 for the design of a control framework?

  5. Why is straight-through processing described as the highest-return and least fashionable trend?

  6. Why can a large language model draft the commentary on a valuation but not produce the valuation?

  7. What does the "economics or appearance" filter reject, and what does it favour?

5

  1. Why can't engineering alone address conduct risk?

  2. Why is average unexplained P&L misleading?

  3. Why was the 2012 JPMorgan mismarking hard to detect?

The six ideas that matter most

If everything else fades, these six should remain.

Idea 1 — The product defines the control effort, not the size

A €10 million bespoke tranche can consume more control resource than a €1 billion government bond position, and correctly so. Effort follows unobservability, not notional. This is why the fair value hierarchy is best understood as a work-allocation map, why coverage metrics must be reported by value and by level rather than by trade count, and why a control function organised around business size rather than valuation difficulty will be systematically misallocated. On Desk SRC-1, 1.4% of the trades generated 62% of the reserves and consumed half the month-end effort.

Idea 2 — Data cascades, so errors upstream are never contained

Dependency in this architecture runs one way. A trade record error reaches every subsequent system. A curve node error reaches valuation, sensitivities, attribution, the ledger and the capital calculation simultaneously. And because every downstream system was fed from the same wrong input, they all agree — which means reconciliation, the control most banks rely on most heavily, cannot detect the problem by construction. The atoms are the trade record and the market data point, and the fields that determine whether the cascade is traceable are the ones most often missing: the quality flag that distinguishes an observation from an assumption, and the construction version that makes a valuation reproducible.

Idea 3 — Independent price verification and reserves are where the value is actually made

The desk marks the book; the control function determines what the accounts say. The distinction between an adjustment and a reserve is the intellectual core of the discipline: where independent evidence establishes a level, post an adjustment; where evidence establishes only a defensible range, quantify the width of the range and post a reserve. Everything downstream — the reported profit, the fair value in the balance sheet, the prudent value supporting capital — is determined at this step. On Desk SRC-1 the distance between the desk's mid-market model number and the prudent value was 4.4%, which is the control function's contribution expressed as a single figure.

Idea 4 — Independence is architecture, not an organisation chart

The reporting line is necessary and nowhere near sufficient. JPMorgan's Chief Investment Office had a Valuation Control Group whose stated function was precisely to test trader marks against external benchmarks within bid-offer-calibrated thresholds, and it did not work, because it was embedded in and dependent on the business it controlled. Independence is realised in entitlements that prevent one person from booking and approving, in marking rights that record who may set what, in storage of both the desk mark and the verified value so that the difference is a queryable data set, in override workflow with reason and approver and expiry, and in an immutable audit trail that makes a two-year-old decision reconstructable. Each of those is a system property. None of them is an organisational one.

Idea 5 — Compensation on marked profit quietly distorts every judgement in the chain

A trader is paid on a number they influence, before that number has been tested by an exit. This is not a scandal; it is the standard structure of the industry, and it is the reason the entire control apparatus exists. Its effects are subtle rather than dramatic: source shopping presented as source selection, methodology revisited at a convenient moment, positions marked just inside a materiality threshold, losses recognised slowly and gains quickly, reserves released in a weak quarter. Each of these is individually defensible and collectively diagnostic. The defences are correspondingly structural — a source hierarchy fixed in advance, methodology changes routed to committee with their profit impact separately reported, monitoring of the distribution of differences against thresholds, and release of a reserve governed as tightly as its establishment.

Idea 6 — Real transformation moves cost or risk, and generative AI is no exception

The filter applies uniformly: does the change alter the economics or the risk, or only the appearance? Straight-through processing removes reconciliation work permanently. A single golden market data source with lineage makes every downstream control cheap. Machine-learned anomaly detection finds patterns that thresholds cannot. Elastic compute matches a workload that is peaked and parallel by nature. Each changes something real. Generative artificial intelligence changes something real too, at a specific and bounded place: it accelerates the production of language and the search through evidence. It must not produce the number, because model risk governance requires reproducibility, validatability and mechanical explainability that generative output cannot supply. The human reviewing that output is not a limitation awaiting removal. It is the control.

Process Cheat Sheet — The Whole Cycle on One Page

A single reference for the entire valuation and product control process, daily and periodic.

Stage Timing Core activity Key output Lead and systems
1. Trade capture Continuous Booking, amendment, confirmation, reconciliation Complete trade population FO / Ops — trade capture, confirmation matching, recon engine
2. Market data close 17:00–19:00 Snapshot, validate, build curves and surfaces Official close data with quality flags Market Data — vendor feeds, curve builders
3. Overnight revaluation 19:00–04:00 Price all trades, generate Greeks, price prior-day book Values and sensitivities Technology — pricing library, compute grid
4. Flash P&L 04:00–08:30 Aggregate, run automated validation checks Preliminary P&L plus exception list Product Control — P&L engine, exception workflow
5. Attribution 08:30–11:00 Decompose P&L by risk factor Explained P&L and unexplained residual Product Control — attribution tool
6. Sign-off 11:00–17:00 Investigate exceptions, resolve or escalate, sign Signed daily P&L PC and desk head — workflow, audit trail
7. Ledger close T+1 to T+2 Post to sub-ledger and GL, three-way reconcile Substantiated balance sheet Financial Control — sub-ledger, GL, recon
8. IPV M+1 to M+5 Test marks against ranked independent evidence IPV differences and adjustments Valuation Control — IPV platform, consensus data
9. Reserves and XVA M+3 to M+7 Bid-offer, model, CVA/DVA/FVA, day-one deferral Adjusted fair value VCG and XVA — reserve engine, Monte Carlo
10. Classification and disclosure M+5 to M+8 Hierarchy levels, AVA, regulatory and public reporting Disclosures and prudent value VCG and Group Finance — classification, reporting

Key numbers and rules of thumb

Metric Rule of thumb
Batch window Roughly 12 hours from close to desk open; recovery time, not runtime, is the binding constraint
Balance sheet split Level 2 dominates fair-valued trading assets; Level 3 typically low single-digit percent
Level 3 effort A small percentage of trades routinely drives the majority of reserves and month-end effort
Unexplained P&L Judge the distribution, not the average; a fat tail on volatile days is the warning sign
IPV coverage Report by value and by hierarchy level; Level 3 at or near 100%; name the unverified residual
Consensus dispersion Standard deviation across contributors feeds reserve sizing directly; contributor count sets the weight
Bid-offer reserve Computed on net risk by bucket × half-spread, with an uplift where position exceeds normal market size
Day-one profit Recognised immediately only on observable-input valuations; otherwise deferred and released over time
Prudent valuation AVAs computed to a 90% certainty standard and deducted from common equity tier 1
PLA test Spearman and Kolmogorov–Smirnov on HPL versus RTPL; red zone means loss of internal models for that desk
Reconciliation breaks Judge the ageing profile, not the count; items over 30 days are unquantified valuation exposure
Swap value (market rate − contract rate) × annuity factor × notional
Vega scale On a €250m 5y-into-10y swaption, roughly €0.46m per volatility point
Balance sheet split ~25% of EU bank assets fair-valued, ~61% amortised cost; of fair value, ~63% Level 2 and ~4% Level 3
Prudent valuation cost AVAs averaged ~1.5% of CET1, ~€227m per institution; simplified approach below €15bn at 0.1%
Source precedence Exchange/CCP → consensus → broker quotes → observed trades → model with verified inputs
Adjustment vs reserve Evidence establishes a level → adjustment. Evidence establishes a range → reserve
Model change P&L Always reported separately, never absorbed into market movement

The one-sentence version

Every night the bank re-prices everything it owns against a curve someone constructed, explains the resulting profit and loss down to a residual it must justify, and once a month puts the desk's marks on trial against independent evidence — converting what the evidence settles into an adjustment and what it cannot into a reserve, so that the number reaching the accounts is not the truest one available but the most defensible one.

Applied diagnostics — six situations, and what you should conclude

Every question so far has asked you to restate a principle. These ask you to use one. Each case gives you what a controller would actually have on screen. Work out what is happening before reading the diagnosis.

Case 1 — The residual that only appears sometimes

What you have. A rates desk's daily unexplained profit and loss over ten business days, alongside the size of the day's move in the ten-year swap rate.

Day Move in 10y rate (bp) Gross P&L (€m) Unexplained (€m)
1 +2 0.41 0.01
2 −1 −0.22 0.01
3 +3 0.58 0.02
4 −22 −4.60 1.41
5 +1 0.18 0.00
6 +4 0.79 0.03
7 −2 −0.38 0.01
8 +19 3.95 1.02
9 −3 −0.61 0.02
10 +2 0.40 0.01

The desk's position is a large book of swaps and sold swaptions. The average residual over the ten days is €0.25 million, comfortably inside a €0.4 million tolerance.

Case 2 — The book that reconciles perfectly

What you have. A structured credit book of ninety positions. Every control is green.

Control Result
Front office to sub-ledger reconciliation 0 breaks
Sub-ledger to general ledger 0 breaks
Trade count, risk engine versus books Matched
Pricing failures overnight 0
Unexplained P&L, 30-day average 0.3% of gross
IPV coverage by trade count 100%
IPV differences posted last month €0

The head of Product Control asks you whether this book is well controlled.

Case 3 — The desk that is always slightly right

What you have. Twelve months of independent price verification differences on one desk, expressed as the desk's mark minus the verified level, as a percentage of position value. The desk's IPV threshold is 0.50%.

Month J F M A M J J A S O N D
Difference (%) +0.44 +0.47 +0.41 +0.48 +0.39 +0.46 +0.43 +0.49 +0.42 +0.45 +0.47 +0.44

No adjustment has been posted in twelve months. Every month passed.

Case 4 — The morning the batch did not finish

What you have. It is 07:10. The batch failed at 02:40 during sensitivity generation and restarted from the beginning. Current estimate for completion is 10:15. The desk opens at 08:30.

Book Fair value (€m) Level Prior-day marks acceptable?
A — vanilla swaps 1,640 2 Rates moved 11bp overnight
B — swaptions 190 2 and 3 Volatility broadly unchanged
C — bespoke tranches 96 3 No market activity

Case 5 — The reserve that fell

What you have. A desk's model uncertainty reserve and its trading performance over four quarters.

Quarter Level 3 fair value (€m) Model reserve (€m) Reserve as % of L3 Desk revenue (€m)
Q1 92 2.3 2.5% 41
Q2 95 2.4 2.5% 38
Q3 101 2.4 2.4% 12
Q4 104 1.6 1.5% 9

The desk explains that the Q4 reduction follows a methodology refinement approved during the quarter.

Case 6 — Two systems, one swap, small difference

What you have. The pricing library and the risk system value the same ten-year swap and disagree by €4,100 on a present value of €1.925 million — a difference of about 0.2%. Both used the same curve version and the same trade record.

Assessing a valuation platform — the questions that separate a real framework from a performed one

For readers evaluating, building or inheriting a valuation control platform. The questions are ordered by how much they reveal per minute spent asking them.

The five questions to ask first

These five identify most weaknesses faster than any document review.

1. Show me the unexplained profit and loss for this desk over the last ninety days, as a distribution rather than an average. A flat distribution is a working process. A fat tail on volatile days is an unmodelled effect. A suspiciously smooth series may be a plugged residual.

2. For this position, reproduce the valuation as at a date six months ago. This tests market data snapshotting, construction versioning and model version control in one request. Institutions that cannot do it usually discover the fact during the attempt.

3. What percentage of fair value — not trade count — was independently verified last month, split by hierarchy level, and what was the unverified residual? The answer distinguishes coverage from the appearance of coverage.

4. Where is the reserve calculation performed, and can you re-run last quarter's and reproduce the number exactly? A spreadsheet that has been overwritten cannot, which answers the question about whether the calculation is part of financial reporting infrastructure.

5. What happened the last time the batch failed after midnight, and what was the recovery time? Every institution has an answer to this. The quality of the answer is the quality of the resilience.

The full checklist by functional block

Block What to ask
Trade capture Is there a single authoritative trade record, or several? What is the ageing profile of unreconciled breaks, not the count? Is portfolio reconciliation with counterparties performed, and on what frequency? Can an amendment to a prior-day trade be made without generating a separate P&L line?
Market data Is every input tagged as observed, interpolated or extrapolated at the point of use? Is staleness detected automatically? Is the construction methodology versioned like code? Do the valuation and risk engines consume the identical snapshot?
Pricing analytics Is there one library or several, and do they agree on the same trade? Is the model inventory integrated with what actually executes, or maintained as a document alongside it? Can a model run outside its approved scope, and would anyone know?
Risk and P&L Are sensitivities bucketed or single-figure? Is prior-day portfolio revaluation produced, and does attribution use it? Are the hypothetical and risk-theoretical series produced daily at desk level?
Valuation control Is the source hierarchy documented in advance of disputes? Are both the desk mark and the verified level stored, or only the outcome? Is an override a first-class object with reason, approver and expiry, or an edit? Are consensus dispersion and contributor counts retained as reserve inputs?
Ledger and reporting Is three-way reconciliation performed daily on the largest books? Are collateral disputes consumed as structured control evidence? Can hierarchy classification be produced systematically from data quality flags, or is it assembled by hand?
Cross-cutting Are records append-only? Can segregation of duties be demonstrated by configuration rather than by policy? Has failover been rehearsed under realistic conditions? Has any data quality metric ever blocked a process?