Skip to content
Corporate & Investment Banking My progress Français

Home / Capital Markets & Securities Services / Securities Services / Module 16

16Trade Capture & Booking Models

How a trade becomes data — captured, enriched, booked and modelled into the golden record that feeds risk, P&L, settlement and reporting front to back

The moment a trade stops being a conversation and becomes data.

Reading time
2 h 00
Sections
79
Recall questions
33
Quiz questions
15
Simulators
1

Two threads run through this module

The business thread asks what a trade is, and where it belongs. A trade is an economic agreement — a price, a size, a direction, a counterparty. The moment it is struck, someone at the bank is exposed to the market moving. The business questions are: what did we actually agree, who bears the risk, which legal entity signed it, which desk manages it, and how does the profit and loss reach the people who are measured on it. None of that is technical. It is the commercial and legal reality of running a trading business, and it decides where money and accountability land.

The technology thread asks how that agreement becomes trustworthy data. One agreement enters through many doors — a voice call, an electronic venue, a bulk upload, a system-to-system feed — and must come to rest as a single canonical record that dozens of systems can read without arguing about the truth. That record has to be validated, enriched from reference data, versioned through its life, and made reconstructable for a regulator years later. This is a data-engineering problem of the first order, and it is where most of a markets technology budget is actually spent.

The two threads are inseparable. A booking decision that looks purely commercial — which entity faces the client — is enforced by system configuration and shows up in a regulatory report. A data model that looks purely technical decides whether risk and finance can ever agree on a number. Throughout the module we hold both threads at once.

How to read this module

Read it front to back the first time. Each part builds on the last: Part 1 defines the record and why it matters, Part 2 walks the record from capture to a fully booked object, Part 3 names the people who touch it, Part 4 opens up the systems that carry it, and Part 5 covers what goes wrong and how the business is changing. After that, the cheat sheet and glossary are built to be used on their own at a desk.

Coloured boxes carry specific kinds of content, and it is worth learning the legend once. A Key insight box crystallises an idea worth remembering. A Going deeper box adds mechanism for readers who want the next layer of detail. An IT & data lens box translates a business point into its systems and data consequences. A Watch out box flags a trap. A Worked example box follows our running case. An In a nutshell box recaps a part. A Check your understanding box lets you test yourself before moving on.

One running example threads through the whole module. It is a single, illustrative interest-rate swap — five years, fifty million euros, a European asset-management client — and we follow it from the moment it is struck to the moment a regulator receives a report generated from its record. Every figure attached to it is illustrative only; the point is the mechanism, not the numbers.

You will understand the trade record end to end

By the end you will be able to describe, without hand-waving, what a captured trade contains, how it is completed by enrichment, where it is booked, and how it feeds every downstream function. You will see capture not as data entry but as the hinge on which the entire markets business turns.

You will be able to map every actor to the record

You will know who originates a trade, who verifies it, who settles it, who values it independently, and who governs the reference data it depends on — and crucially, where a business analyst or technologist adds value in that chain.

You will understand where the truth actually comes from

You will understand why a bank fights to maintain one golden record rather than five conflicting copies, what a canonical trade model is, and why the books of record — trade, position and accounting — must forever reconcile to each other.

You will be able to trace data front to back

You will be able to follow a single field from the source system that created it, through the canonical model, to every downstream consumer that reads it — the source-to-target mapping that is the core artefact of markets technology change.

You will recognise the risk and transformation landscape

You will know the failure modes that recur across every bank, the controls that defend against them, the real incidents where those controls failed, and the transformation and AI agenda that is reshaping the discipline — with an honest view of what genuinely changes economics or risk.

1Context: the moment risk becomes data

A trade is an economic event that happens in a fraction of a second. A trade record is a structured object that outlives the event by years. This part explains the gap between the two, and why closing that gap correctly is the single most consequential thing a markets operation does.

Risk is real at execution, but the bank only knows it once the trade is captured

The economic risk exists the instant two parties agree a price. A trader says "done" on a fifty-million-euro swap and the bank is immediately exposed to interest rates moving. Nothing has been typed, filed or confirmed, yet the exposure is live and the profit and loss is already ticking. This is the uncomfortable truth at the heart of the discipline: the risk arrives before the record does.

The bank, as an institution, cannot act on risk it has not recorded. The trader's memory is not a system. Risk managers cannot hedge a position they cannot see; finance cannot book a profit it cannot measure; operations cannot settle a trade whose terms live only in a chat window. Until the deal is captured — recorded as a structured, machine-readable record — the bank is running exposure blind. Trade capture is the act that turns a private agreement into institutional knowledge.

Capture is therefore a control, not a formality. The faster and more accurately a trade is captured, the smaller the window in which the bank is exposed to risk it cannot manage. A trade struck at 09:41 and captured at 09:41 is a controlled position. The same trade captured at 16:00, or captured wrongly, is a latent incident. Speed and correctness at this one step ripple through everything that follows.

Trade capture is the hinge where execution ends and post-trade begins

Picture the markets business as a single flow from left to right. On the left, execution: a price is agreed by voice, on an electronic platform, or by an algorithm. In the middle, capture: the deal becomes a structured record. Then enrichment adds the reference data that completes it, the golden record lands in the book of record, and on the right the whole post-trade machine — risk, valuation, confirmation, settlement, reporting — reads from it. Capture is the hinge on which the left and right halves turn.

Figure 1 The front-to-back hinge — everything to the right of capture depends on the record being complete, correct and timely

Everything to the right depends on the hinge holding. Risk sees the position from the record. Finance books the profit and loss from the record. Operations confirms and settles from the record. Regulators are told about the trade from the record. None of these functions keeps its own independent copy of the truth. They are all consumers of one upstream object, which means the quality of that object sets a ceiling on the quality of everything downstream.

The hinge is fragile precisely because it is a convergence point. Many entry channels must land in one model. A trade must be captured near real time so intraday risk is right. A plain vanilla bond and a path-dependent exotic must both fit the same capture layer. The hinge carries the entire weight of the flow through a single narrow point, and narrow points are where things break.

For a technologist, capture is an ETL boundary made physical. Messy inputs arrive from many sources in many dialects. They are extracted, transformed into one canonical validated shape, and enriched from master data. That canonical record then becomes the master object every downstream pipeline joins to. It is, quite literally, the most-read data object in the entire markets business — which is why its schema, its latency and its data quality are architectural decisions, not implementation details.

A trade is not a row — it is a layered, stateful object

Beginners picture a trade as a line in a spreadsheet. It is closer to a small living organism. To understand everything that follows, read any captured trade as three layers stacked on top of each other. Miss any layer and you will misunderstand what the record is for.

The first layer is economics — the actual terms of the deal. Notional, price or rate, direction, index, dates. This is what determines the profit, loss and risk. It is the part the trader agreed and the part the market moves against. If the economics are wrong, every number computed from the trade is wrong.

The second layer is static and reference data — the identifiers that point to master data. The record does not carry the counterparty's full legal details; it carries a Legal Entity Identifier that points to them. It does not carry the swap's day-count and holiday rules; it carries a product identifier that points to them. Most of this layer is filled not by the trader but by enrichment, which looks up the master data behind each pointer. The trade is a set of keys into a much larger reference-data estate.

The third layer is lifecycle state — where the trade is in its life. A status of new or verified, a version number, timestamps, the channel it arrived through. A trade is not captured once and frozen. It is amended, it pays coupons, its floating rate is fixed against an index, it may be novated to a clearing house, and eventually it matures or is terminated. The record is a stateful object that moves through a defined lifecycle, and every transition must itself be captured.

Capture, enrichment and booking are three distinct acts

These three words are used loosely in practice and it causes real confusion. They name three different steps, and keeping them separate is the first sign you understand the domain.

Capture is the act of recording the executed terms into a system quickly and correctly. It is the data-ingestion event. Its job is to get the economics into the canonical model with minimal delay and minimal error. Capture answers: what was agreed?

Enrichment is the act of completing the skeleton. The captured record carries economics but is not yet processable. Enrichment attaches the product conventions, the settlement instructions, the counterparty's legal data, the book and the accounting defaults — mostly by automated lookup against reference data. Enrichment answers: what else does the record need to be usable?

Booking is the act of deciding how and where the trade is represented. Which trade template models it, which book holds it, which legal entity owns it, which accounting treatment applies. Booking is the subject of Part 2, and it is where commercial and legal judgement enters. Booking answers: how is this trade represented, and where does it live?

The range of what must be captured is enormous

A single capture layer has to model wildly different instruments. This is what makes the problem hard. A spot foreign-exchange trade is almost trivial — a currency pair, a rate, a value date. A vanilla interest-rate swap is moderate — two legs, an index, a schedule. A bespoke multi-asset structured note is genuinely difficult — a path-dependent payoff that no standard template quite fits. The capture layer must serve all of them, and the volume is distributed very unevenly across them.

Instrument class What it is Capture complexity Typical channel Where the difficulty lives
Spot FX Exchange one currency for another at today's rate Very low Electronic, STP Volume and speed, not modelling
Government / cash bond A standardised debt security bought or sold Low Electronic or voice Settlement and identifiers
Cash equity A share bought or sold, often in bulk allocations Low Upload / STP Allocation and give-up mechanics
Vanilla interest-rate swap Exchange fixed for floating cash flows Medium Electronic / voice Schedules, fixings, lifecycle events
FX forward / option A rate agreed for a future date, optionally optional Medium Electronic / voice Dates, optionality, conventions
Credit default swap Protection on a reference entity defaulting Medium-high Voice / platform Reference data, clearing, novation
Exotic / structured note A bespoke, often path-dependent payoff High Manual No standard template fits it

The distribution of volume is the key strategic fact. A small number of vanilla products — spot FX, government bonds, vanilla swaps — make up the vast majority of trades by count, and these flow through fully automated. A long tail of complex and exotic products is tiny by volume but large by value and by risk, and these often fall to manual booking or workarounds. The economics of a capture platform are dominated by pushing more of the volume into automated flow while safely containing the dangerous tail. We return to this coverage curve in Part 2, because the tail is where most incidents are born.

Why not simply build a bespoke system per asset class? Many banks did exactly that, and it is the root of the fragmentation problem this whole discipline wrestles with. A per-asset-class capture system is easy to build and impossible to reconcile: five systems each hold a different notion of what a trade is, and no one can aggregate risk across them without a nightly reconciliation battle. The modern answer is a single canonical model with product-specific templates hanging off a shared base — object-oriented modelling applied to trades. That design lets a new product become a new template without breaking the shared spine, and lets risk aggregate across the whole book on the common fields. Part 2 builds this idea properly.

Trades are the atoms; positions are what the desk actually manages

No trader thinks in individual trades for long. A trader thinks in positions — "I am long two hundred million of this bond", "my swap book has a net DV01 of minus forty thousand euros". A position is the aggregate: all trades in an instrument or book, netted along dimensions that matter — book, instrument, currency, maturity, risk factor. Risk is measured on positions, not trade by trade. Limits are set on positions. The desk manages positions.

But the position is entirely derived from the trades. It is a sum, not a source. If a trade is mis-captured, the position is wrong, and the desk is managing risk it does not really have or is blind to risk it does. This is why the identity position equals the sum of its trades is sacred and must always hold. A break between the trade-level view and the position-level view is not a rounding curiosity; it is a signal that capture or a feed has failed, and it must be investigated and cleared.

The trade-versus-position distinction is a data-architecture decision as much as a business one. Trades are the immutable, individually reportable atoms — captured once, versioned through their life, and the grain at which you report to a regulator. Positions are continuously recomputed projections, aggregated by book, instrument, currency and risk factor as trades flow in. Keeping the projection consistent with the atoms in real time is a genuine engineering challenge, and the reconciliation that proves position equals sum of trades is one of the most important controls a markets platform runs.

This is one of the largest and most critical IT estates a bank runs

None of this is small. Capture engines, canonical trade stores, enrichment services, reference-data platforms, and the feeds and mappings that connect them are collectively among the largest, most business-critical systems any bank operates. They run near real time, they cannot lose a message, and they must reconstruct any trade's full history years later for a regulator. The scale is not incidental — it is a direct consequence of the fact that every downstream function depends on this layer.

The business is shaped less by a market cycle than by a regulatory and transformation cycle. Where a deal-origination business rises and falls with issuance windows, the capture estate is driven by waves of regulation and consolidation. The post-crisis reforms mandated reporting of every derivative. The Legal Entity Identifier became mandatory — in practice, no LEI, no valid report, no trade. More recently the European reporting rewrite tightened data quality, mandated a standardised XML message format and a product identifier, and expanded the fields banks must report. Each wave forces the capture layer to get cleaner at source rather than patch reports downstream. The through-line is relentless: the record must become more accurate, more complete and more reconstructable with every cycle.

Why one record, and not five

The deepest principle of the whole discipline is that there must be exactly one authoritative version of each trade. It sounds obvious and it is the thing banks find hardest to achieve. The alternative — five systems each holding their own copy of the same trade — is not a hypothetical; it is the natural state of a large bank that grew its systems separately, and it is the source of most of the pain this module describes.

Multiple copies do not just cost storage — they manufacture disagreement. If risk holds one copy of a trade and finance holds another, the two will drift: an amendment lands in one and not the other, a reference-data value updates on one side, a feed lags. Now risk and finance report different numbers for the same book, and no one can say which is right without a reconciliation. Every extra copy of the truth is an extra thing to reconcile and an extra place for the copies to diverge. The golden record exists to collapse the copies into one, so that disagreement becomes impossible by construction rather than merely unlikely.

One record, many consumers: the front-to-back dependency

The reason capture is non-negotiable is that no downstream function keeps its own copy of the truth. They all read the one captured record, which makes capture completeness, correctness and timeliness a front-to-back concern rather than a front-office convenience. It is worth tracing the four dependencies explicitly, because seeing them together is what makes the phrase "single source of truth" concrete rather than a slogan.

Risk management computes everything from the captured trades. Positions, sensitivities, value-at-risk and limit checks are all derived from the records — there is no separate risk copy of the book. If a trade is missing or wrong, the risk system is confidently wrong, and the desk hedges a position that does not match reality.

Finance sources the profit and loss and the ledger from the same record. Daily profit and loss, valuation and the general ledger all read the trade. Finance's numbers are the official ones the bank reports externally, so a capture error becomes a misstatement in the accounts, usually surfacing days later at reconciliation when it is expensive to unwind.

Operations reads the record to move cash and securities. Confirmation, clearing, settlement and collateral all read the trade's terms and settlement instructions. A wrong term here is not an abstract data-quality issue — it is a failed settlement, a penalty and a counterparty dispute.

Regulatory reporting is generated from the record, not entered separately. EMIR, MiFIR and transaction reporting read the captured trade. Reporting is a by-product of good capture, which is exactly why the regulators push banks to fix data at source: a wrong record is a wrong report, automatically.

The golden record as a birth certificate

A useful way to hold the whole idea is that the trade record is the trade's birth certificate. A birth certificate does not create the child — the child exists first — but until the birth is registered, the state has no official knowledge of the person, and every later document, from a passport to a pension, is issued off the back of that one registration. If the certificate records the wrong date or name, the error is not local: it propagates into every document derived from it, and correcting it years later is painful precisely because so much was built on top.

The trade record works the same way. The trade exists the instant it is struck, but until it is captured the bank has no official knowledge of it, and every later artefact — the position, the P&L, the settlement instruction, the regulatory report — is issued off the back of that one record. A wrong field at capture is a wrong name on the certificate: it is inherited by everything downstream and expensive to unwind once the derived documents exist. The analogy also explains why there must be exactly one record. A person with two birth certificates that disagree has a problem no third document can resolve without going back to establish which is true.

One wrong field, four wrong answers: the blast radius

Because every consumer reads the one record, a single wrong field is wrong in four places at once. This is worth drawing, because it is the visual form of the module's central claim. There is no independent copy against which any consumer could catch the error, so the mistake does not surface as a disagreement between systems — it surfaces, if at all, only when someone reconciles against the outside world.

Figure 2 The blast radius of a single capture error — with no independent copy of the truth, one wrong field is wrong across risk, finance, operations and reporting simultaneously

Trace a single mis-keyed notional through the four consumers. Risk reads it and computes a position that is too large, so the desk hedges an exposure it does not have — a real, costly trade against a phantom. Finance reads the same notional and books a day-one P&L that is wrong, entering the official accounts. Operations reads it and instructs a settlement for the wrong amount, heading for a failed delivery and a penalty. Reporting reads it and files a regulatory report with the wrong notional — a reportable breach in itself. One keystroke, four failures, none of them visible to the others, because they are all looking at the same wrong number.

  1. Why does the bank own risk before it owns a trade record?

  2. What are the three layers of a trade record?

  3. How do capture, enrichment and booking differ?

  4. Why is the long tail of exotic products dangerous despite its small volume?

  5. Why must position equal the sum of its trades?

  6. Why is trade capture better described as an ETL boundary than as data entry?

1

  1. Why does the bank own risk before it owns a trade record?

  2. Which is NOT one of the three layers of a trade record?

  3. What does enrichment do?

2The flow: from a struck deal to a fully booked record

This is the longest part of the module, and the most important. We follow a single trade from the instant of execution to the moment it is a complete, validated, enriched, correctly booked record that risk, finance, operations and reporting can all read. Every stage has its own purpose, its own technology and its own way of going wrong.

The whole flow on one page

The journey from execution to a usable record is a short pipeline with a lot inside it. A trade is captured, enriched, validated, verified and confirmed, booked into the right model and location, distributed to consumers, and then maintained through its lifecycle until it settles and matures. Most of the pipeline is automated for vanilla flow and only exceptions need a human. But every stage is a place a trade can stall or corrupt, so it is worth seeing them laid out before we walk them.

Stage What happens Typical duration Output Who leads
1 Capture Executed economics recorded in the canonical model Seconds A skeleton trade record Front office / STP
2 Enrichment Reference and static data attached by lookup Seconds A complete record Automated rules engine
3 Validation Rules, tolerances and referential checks run Seconds A record cleared to book Capture system
4 Verify & confirm Middle office verifies; counterparty confirms Minutes to hours A trusted, matched record Middle office
5 Booking Representation and location fixed; internal legs created Seconds to minutes A booked, owned record Front / middle office
6 Distribution Risk, finance, ops and reporting read the record Continuous Positions, P&L, instructions, reports Downstream systems
7 Lifecycle Fixings, coupons, exercises, amendments, novations Days to years Versioned events Middle office / systems
8 Settlement Cash and securities exchanged against instructions T+1 / T+2 A settled, closed obligation Operations

Effort and risk are not spread evenly across the pipeline. They concentrate violently at the front. Capture, enrichment and validation happen in the first seconds, and it is precisely there that an error is cheapest to make and most expensive to discover later. A wrong notional keyed at capture is a decimal point; the same error found at reconciliation days later is a costly, multi-team investigation. The intensity curve peaks early, at capture and its immediate controls, and then settles into the long, lower-intensity tail of lifecycle management and settlement.

Figure 3 The capture-to-booking pipeline — effort and risk peak in the first seconds, at capture and its controls, then settle into the long lifecycle tail

Stage 1 — capture: many doors, one canonical model

Trades arrive through radically different channels, and all must land in the same shape. This is the first design principle of a capture layer: the downstream systems must not care how a trade arrived. Whether it came by voice or by API, the record that results is identical in structure, so risk and finance never have to know the trade's provenance to use it.

Voice and manual entry is the most flexible and the most dangerous door. A trader or middle-office analyst types a phone or chat trade into a capture screen. It handles anything, including the bespoke deals no automated channel can, but it is where fat-finger errors are born — a misplaced decimal, a wrong direction, a mistyped book. Every keystroke is a risk.

Electronic straight-through processing from venues is the cleanest door. Trades done on platforms — the electronic marketplaces where dealers and clients meet — flow in automatically through a standard messaging protocol, with no re-keying. Low touch means low error. The more flow that arrives this way, the cheaper and safer the operation.

Bulk upload, system-to-system APIs, and internal rebooking make up the rest. Files of trades load in bulk, common for equity allocations. Upstream engines — an algorithm, an electronic pricer, a structuring tool — write trades directly by API at high volume. And the middle office rebooks, splits and corrects trades, breaking a block into client accounts or transferring risk between desks. Each door is different, but each normalises into the same canonical record.

Stage 2 — enrichment: completing the skeleton from reference data

The captured skeleton is only half a trade. The trader supplied the economics; enrichment automatically attaches everything else the record needs to be processed, and it does so by looking up master data behind the identifiers the trade carries.

Enrichment is, mechanically, a series of data joins. It attaches the instrument and product terms — conventions, day-count, calendars, holiday rules. It attaches counterparty data — legal name, LEI, credit terms, netting and collateral agreements. It attaches settlement instructions, the standing details of where cash and securities actually go. It attaches static defaults — the book, cost centre, fees and accounting treatment. And it attaches the market-data links — the curves and fixings the trade will be valued against. Only after all of this is the record complete enough to value, confirm, settle and report.

Much of enrichment is rule-driven, and the rules are themselves an asset that must be governed. A rule might say "for this desk and this product, default the book to X and the fee schedule to Y". These rules are powerful and dangerous: a wrong rule enriches thousands of trades wrongly and silently. So the rules engine, and the reference data it reads, sit at the heart of the control framework.

Enrichment is where data quality either saves you or sinks you. Because it joins the trade to reference tables on keys — LEI, product identifier, book — a missing key is a hard stop: no settlement instruction, no enrichment, a stalled trade and an exception to work. Worse, a wrong key enriches successfully but incorrectly, so the trade looks complete while being subtly wrong downstream. This is why reference-data stewardship, covered as a role in Part 3 and as a platform in Part 4, exists as a first-class function. The backbone must be clean, governed and shared, because every trade in the bank leans on it.

The reference-data backbone the whole edifice stands on

Trades point to master data, and the whole structure is only as good as that data. It is worth naming the main reference domains, because a business analyst or technologist working in capture spends most of their time in exactly these tables.

The instrument and product master defines every tradable instrument — bond terms, swap conventions, option specifications, and identifiers like the ISIN and the product identifier. The counterparty and client master holds legal entities, LEIs, hierarchies, credit limits, and netting and collateral agreements — the who of every trade. The books and organisation master holds the trading books, desks, cost centres and legal entities and their hierarchy — the where it lives. Calendars and conventions hold holiday calendars, business-day rules, day-count conventions and fixing sources — the mechanics of dates and cash flows. Settlement instructions and accounts hold the standing details of where money and securities move. And market-data references hold the curves, indices and fixing sources the trade is valued and settled against.

Stage 3 — validation and the internal control gauntlet

Before a trade is trusted, it runs a gauntlet of layered controls. No single control catches everything, so a capture operation stacks several, each aimed at a different class of error. This is the internal governance of trade capture, and it is the direct equivalent of the committees and gates that govern a deal in an origination business.

Validation rules catch format and referential errors at the point of capture. Mandatory fields present, values in valid ranges, the book exists, the counterparty is active. These are the cheapest checks and they run instantly.

Tolerances and limits catch outliers. A notional ten times the desk's norm, a rate an order of magnitude off the market, a decimal-point slip — these trip a hard warning before the trade books. Tolerance checks are the single most important defence against fat-finger errors, and we will see in Part 5 exactly what happens when they are mis-configured.

Four-eyes verification catches judgement errors. A second person, usually in the middle office, verifies manual and large trades before they are trusted. Where a validation rule checks syntax, four-eyes checks meaning — is this trade plausible, does it match the term sheet, did the trader mean to do this. It is the human control on judgement, and it is why the maker and the checker must never be the same person.

Confirmation matching catches disagreement with the counterparty. The trade's terms are electronically matched against the other side's version through a matching platform. If the two sides disagree on notional, rate or dates, the match fails and the discrepancy surfaces early, before settlement, when it is still cheap to fix.

Reconciliation catches everything else. Position equals sum of trades, and the books reconcile to downstream systems daily. Breaks are investigated and cleared. Reconciliation is the backstop that assumes the other controls missed something, and it usually finds the residue they left behind.

Control Catches When it runs Human or system
Validation rules Format and referential errors At capture System
Tolerances and limits Outliers and fat-finger errors At capture System
Four-eyes / maker-checker Judgement errors Before trust Human (middle office)
Confirmation matching Disagreement with counterparty Post-capture System + human
Reconciliation Residual breaks and feed failures Daily System + human

Booking is two decisions: how the trade is represented, and where it lives

A booking model answers two questions, and they are genuinely different. The first is representation: how is the product modelled in the system — which template, which fields, which cash-flow structure, which valuation approach. The second is location: where is the trade booked — which book, which desk, which legal entity, facing which counterparty, in which system. The same economic trade can be booked several ways, and the booking model is the set of governed rules that decides which.

Representation and location each carry weight. Representation decides how the trade is valued and how its risk aggregates with everything else. Location decides which desk owns the risk and profit, which jurisdiction and regulator apply, and which balance sheet bears it. Get representation wrong and the numbers are computed wrongly; get location wrong and correct numbers land in the wrong place. Both must be consistent and governed — everyone must book the same product the same way, or aggregation and reporting break.

The product taxonomy: object-oriented modelling applied to trades

Different products are represented by different templates that share a common spine. Think of it as a class hierarchy. Every product — a bond, a swap, an option — inherits a common base: a trade identifier, a unique trade identifier, a notional, a counterparty LEI, a book, dates and a status. Each product type then adds its own specific fields: a bond adds coupon, maturity and face value; a swap adds two legs, an index and a tenor; an option adds strike, expiry and style.

This design is what lets the estate scale without breaking. A new product becomes a new template hanging off the shared base, which means you can onboard it without disturbing the existing model. And because every product shares the common fields, you can aggregate risk across the entire book on those fields regardless of asset class. The taxonomy is the quiet architectural decision that makes cross-asset risk possible at all.

Asset class Product Template adds
Rates Bond ISIN, coupon, maturity, face value
Rates Interest-rate swap pay leg, receive leg, index, tenor
Rates Swaption underlying swap, strike, expiry
Credit Credit default swap reference entity, spread, seniority
FX FX forward currency pair, forward rate, fixing date
FX FX option currency pair, strike, expiry, style
Equity Cash equity ISIN, quantity, price
Equity Total return swap underlying, financing leg

The canonical model and the golden source

A canonical trade model is a single, normalised representation independent of any one source system. Different capture systems — a swap platform, an equity order-management system, a structuring tool — each speak their own dialect. The canonical model is the common language they all map into. It is the holy grail of markets data architecture, and it is genuinely hard to retrofit onto a bank that grew its systems separately.

The golden source is the one authoritative store of that canonical trade. It is the book of record every downstream system trusts. Its value is precisely that it is singular: every consumer joins to one model instead of integrating with every source system separately, and a swap and a bond share the same core structure so risk and reporting aggregate cleanly across asset classes. Industry standards help — FpML gives a shared vocabulary for representing trades, and the ISDA Common Domain Model standardises how trades and their lifecycle events are represented across firms.

Why is the canonical model so hard to achieve in practice? Because legacy banks do not have one trade store — they have many, grown separately by product, region and acquisition, each with its own dialect and its own idea of what a trade is. Converging them onto a single canonical model is a multi-year transformation programme, not a project, and it is the single most valuable and most difficult thing a markets technology function can attempt. Every reconciliation break, every "why don't risk and finance agree" argument, and every slow regulatory-reporting fix traces back to the absence of one canonical truth. This is why consolidation onto a canonical model is the headline item of the transformation agenda in Part 5.

The books of record: one trade source, many derived views

Banks deliberately keep several books of record, and it confuses newcomers. Why not one? Because different consumers need different granularity, timing and controls, and forcing them into one store would serve none of them well.

Figure 4 One golden trade source, three derived books of record — all must reconcile to the same trades and events

The trade book of record — sometimes called the accurate-trade-record — is the authoritative store of every trade and its lifecycle, the operational source of truth for capture, confirmation and operations. The position book of record is the netted position view that risk and the desk manage, derived from the trades and aggregated by book, instrument and risk factor. The accounting book of record is the finance and general-ledger view, where trades and events are translated into accounting entries and official profit and loss. Each has different granularity and timing — real-time operations versus end-of-day accounting — but all must reconcile to the same trades.

The architecture principle is one golden trade source, many derived views. The trade book of record is authoritative; the position and accounting books are downstream projections of it. The permanent challenge is keeping all three reconciled — same trades, same events, consistent across real-time operational, risk and end-of-day accounting timelines. European supervisors expect a reconstructable audit trail from the trade record through to reported positions and accounts, so the books of record and their reconciliation are themselves part of the control framework.

Books, portfolios and hierarchies: the skeleton risk hangs on

Every trade hangs on an organisational hierarchy, and that hierarchy is itself reference data. The lowest-level container is the trading book — a desk's set of trades in a strategy or product, where risk and profit are first aggregated. Books roll up into a desk or portfolio, a group of books under a trading desk such as euro rates, where traders and limits are managed. Desks roll up into a cost centre or profit-and-loss unit, the finance view for management reporting. And everything sits within a legal entity, the regulated company that legally owns the trade and drives its capital, reporting and jurisdiction.

The hierarchy is the aggregation path every number is summed along. Book to desk to cost centre to entity is the tree up which every risk and profit number rolls. Mis-map a book to the wrong desk and the totals break silently. There is also a regulatory split cutting across the tree — the trading book versus the banking book, held-to-trade versus held-to-maturity, which changes capital treatment and which booking must respect.

Legal entity and booking location: a decision with capital and regulatory weight

A global bank has many legal entities across jurisdictions, and every trade must be booked in exactly one of them. That choice is not administrative. It determines the regulator and reporting regime that apply, which entity's balance sheet bears the capital and liquidity, which master agreement and collateral terms govern the trade, and the tax and booking-model policy — an area under increasing regulatory scrutiny.

Often the entity that faces the client is not where the risk should be managed. The client-facing entity is chosen for licensing, relationship and jurisdiction reasons — it is the entity legally allowed to sign with that client. The risk-holding entity is where the desk that actually manages the risk sits, which may be a different entity or country. When the two differ, the bank books an internal back-to-back trade to transfer the risk from the client-facing entity to the risk-holding one, keeping both books aligned.

Booking location became a supervisory battleground after the United Kingdom left the European Union. European banks re-booked European client risk into European entities in Paris, Frankfurt and Dublin, because the client-facing entity now had to be inside the bloc. Supervisors scrutinise "empty shell" arrangements and back-to-back models closely: the principle they enforce is that risk must be genuinely managed where it is booked, not parked in a letterbox entity while the real trading happens elsewhere. For a technologist, this means the booking model is not free configuration — it encodes a regulatory position that the bank must be able to defend, and the systems must make the risk-management reality match the booking reality.

The booking model, drawn

The client-facing and risk-holding split is clearer as a picture than as prose. The entity that signs the client sits between the client and the desk that actually manages the risk, and an internal back-to-back trade moves the risk from the former to the latter, so each entity's books hold what it is supposed to hold.

Figure 5 The booking model — the client-facing entity and the risk-holding entity, with an internal back-to-back trade moving the risk to where it is genuinely managed

Internal trades and transfer pricing: real bookings with the bank on both sides

When one desk originates a trade but another should manage the risk, the bank books an internal trade between them. It is a real record with the bank on both sides. The common patterns are a mirror trade, an equal-and-opposite booking so each desk sees the risk it owns; a back-to-back, passing risk from the client-facing entity to the risk desk; a hedge transfer, where a desk buys a hedge internally from the market-making desk; and treasury or funding trades that allocate the cost of balance sheet. Each internal trade is captured and booked exactly like an external one — same model, same lifecycle.

Internal trades introduce four things to watch, and all four are data-control problems. They must net to zero at group level — if a mirror is missing or mismatched, group profit and loss is misstated. The transfer price is the terms — it allocates profit between desks and must be fair and consistent, never a way to hide profit. They must be tagged internal versus external — reporting must exclude internal trades, which have no external counterparty and no reporting obligation, and a mis-tag causes false reports. And they multiply records — one client trade can spawn several internal legs, so the trade store holds far more rows than there are real trades.

Internal trades are the reason the trade count in a bank's store is much larger than the number of trades it has actually done with the outside world. This matters for any technologist sizing a system or reconciling a book: you are not counting deals, you are counting legs. The tagging of internal versus external is a mission-critical field — get it wrong and you either report a trade that has no counterparty, which is a false report, or fail to report one that does, which is a missing report. Both are breaches. The internal-external flag is a small field with an outsized blast radius.

Product coverage and the long tail: where risk concentrates

A small number of vanilla products make up the vast majority of trade volume, and they book straight through. Spot FX, government bonds, vanilla swaps — fully modelled, fully automated. The problem is the long tail: complex and exotic products, bespoke structures, multi-asset and path-dependent payoffs that are low in volume but high in value and risk.

The tail is dangerous because off-model booking is where control breaks down. If a product is not in the standard model, it is booked in a workaround — a spreadsheet, a proxy trade, a manual template — where risk and profit are less reliable. Manual booking means more errors and no STP. A proxy that books an exotic as a "close" vanilla mis-states its true risk because the model does not capture the real payoff. And valuation and reporting may be incomplete or wrong for those trades. The build-versus-cover decision — whether to properly model a rare product or handle it manually — is a constant cost-and-risk trade-off, and Part 3's new-product onboarding case shows how a bank drags a product out of the tail and into the controlled model.

Valuation is computed from the very same record

The record that captures the deal is also the input to pricing and risk. There is no separate valuation data entry. The pricing engine reads the trade's terms from the record and combines them with market data: the trade terms — notional, rate, dates, payoff — from the record; the market data — curves, volatilities, fixings — linked by reference data; and a pricing model, which the template specifies. Out comes present value, profit and loss, and sensitivities like delta, DV01 and vega — all derived from the same record that risk and finance read.

Figure 6 Valuation reads the captured trade terms, joins market data by reference, applies the template's model, and produces value and sensitivities

Lifecycle events: the trade lives, and every event must be booked

A trade is not done when it is booked. It lives, and its life generates events that must each be captured and re-booked. A floating rate is fixed against an index on schedule, updating the trade's cash flows. Coupons, interest and premium payments fall due, generating settlement instructions and realising profit and loss. An option is exercised or expires, creating or removing an underlying position. Terms are amended by agreement or to correct an error, producing a new version of the record with full audit trail. The trade is novated or assigned to a new counterparty, often a clearing house, and the counterparty leg is re-booked. And finally the trade is unwound early or matures, realising final profit and loss and closing the record and its cash flows.

Each event is a state transition that must be captured, timestamped and often re-reported. This is not optional housekeeping. Under the European reporting rewrite, lifecycle events — modification, termination, novation, error correction — each require an updated report using standardised action and event types, so the booking system must emit the right event type for every change. The industry Common Domain Model exists precisely to standardise how these events are represented across firms, so that both sides of a trade describe the same event the same way.

Settlement: the record becomes an exchange of cash and securities

Settlement is where the record finally moves money. Operations reads the trade's terms and settlement instructions and exchanges cash and securities with the counterparty. The governing principle is delivery versus payment — neither side hands over value without simultaneously receiving the other side's, which removes the risk that one party pays and the other defaults. Cash securities markets have compressed their settlement timelines toward one business day after trade, while many instruments still settle at two; the direction of travel is relentlessly shorter, which puts pressure on capture to be complete and correct sooner.

Settlement runs on standardised messaging. Post-trade instructions and confirmations move as standardised financial messages between banks, custodians and central securities depositories — historically the SWIFT MT message family, now migrating to the richer ISO 20022 XML standard. A wrong term or a wrong settlement instruction at capture causes a failed settlement — penalties, counterparty disputes and manual repair. This is the concrete, cash-moving consequence of a capture error, and it is why the settlement-instruction reference data is guarded as carefully as the economics.

The captured record, field by field

It helps to see a real record laid out, because every downstream number is a function of these specific fields. A canonical trade record is not abstract — it is a defined set of fields, each of which drives something concrete. Reading the record field by field is the fastest way to understand why capture quality is not negotiable: there is no field here whose corruption is harmless.

Field Layer What it drives downstream
tradeId Lifecycle The primary key every consumer joins on
uti Static Pairs both sides' regulatory reports; deduplication
product / upi Static The template, the valuation model, product reporting
direction Economics Whether the desk is long or short the risk
notional Economics Position size, P&L magnitude, tolerance checks
fixedRate / floatIndex Economics Cash flows, valuation, sensitivities
tradeDate / effectiveDate / maturityDate Economics Schedules, tenor bucket, settlement timing
counterparty / cptyLEI Static Reporting, netting, credit, settlement routing
book Static The risk and P&L aggregation tree
legalEntity Static Jurisdiction, capital, regulator
status / version Lifecycle Gates the whole lifecycle; the audit trail
captureChannel / timestamp Lifecycle STP measurement, forensic reconstruction

Read the record as economics wrapped in pointers wrapped in state. The economics — notional, rate, index, dates, direction — are the deal. The static fields — the product and counterparty identifiers, the book and entity — are pointers into the reference-data estate, filled mostly by enrichment. The lifecycle fields — status, version, channel, timestamp — track the record as a living object. A single record of a swap might carry twenty or more fields, and the discipline of capture is that every one of them is either correct or a latent incident. There is no decorative field.

Confirmation and settlement mechanics, in detail

Confirmation is the moment the bank and its counterparty agree they captured the same trade, and it is a control worth understanding in its own right. After capture, each side holds its own record of the deal. Confirmation matches the two records against each other, field by field — notional, rate, dates, direction — so that any disagreement is caught before it becomes a settlement failure. For standardised products this is electronic and near-instant through a matching platform; for bespoke trades it may involve a legal confirmation document. Either way, an unmatched confirmation is a red flag that the two sides do not agree on what they traded, and it must be resolved before the trade is trusted.

A failed match is cheaper than a failed settlement, which is the whole point of confirming early. If the two sides disagree on a date and it is caught at confirmation, it is a phone call and an amendment. If the same disagreement is caught at settlement, it is a failed delivery, a penalty and a dispute, days later, with cash and securities in limbo. Confirmation exists to move the discovery of disagreement as early in the flow as possible, where it is a correction rather than an incident.

Settlement discipline is now enforced by regulation, which raises the cost of a bad record further. European settlement rules impose cash penalties on trades that fail to settle on time and require buy-in procedures for persistent fails, precisely to push the market toward cleaner, faster settlement. For the capture estate this means a wrong settlement instruction or a late-captured trade is not merely operationally annoying — it carries a direct financial penalty. The compression of settlement cycles toward one business day after trade tightens the screw again: there is less time to catch and fix a bad record before it must settle, so the record has to be right sooner.

Allocations and give-ups: one execution, many records

A single execution is often not a single trade — it is a block that must be split. An asset manager trading for twenty funds does not want twenty small orders in the market; it executes one large block at one price, then allocates it across its funds afterward. On the bank's side, that one execution becomes many booked records, one per client account, and the sum of the allocations must equal the block. This is a core mechanic in equities and listed derivatives, and it is a capture problem: many records, born from one execution, that must reconcile back to it.

Figure 7 Block to allocation to give-up — one execution becomes many booked records that must sum to the block, some given up to a clearer

A give-up adds a second hand-off on top of the allocation. In listed and cleared markets, the executing broker may not be the clearing broker. After execution, the trade is "given up" — transferred to the client's chosen clearing broker, who books it and takes it to the clearing house. So one execution can fan out into several allocations, and each allocation can be given up to a different clearer, multiplying the records again. Each hand-off is a place a trade can fall out, which is why allocation and give-up flows are a classic source of breaks.

Netting: why the exposure is far smaller than the notional

The gross notional a bank has booked with a counterparty overstates what it actually stands to lose, because of netting. Under a master agreement, the many trades between two parties are legally combined so that only the net amount is owed if one side defaults. A bank might have booked a hundred million of trades in the bank's favour and seventy million against, but its real exposure is the net thirty million, not the gross. Netting is the reason the eye-watering gross-notional figures quoted for derivatives markets are misleading: the exposure that matters is net, and it is a fraction of the gross.

Figure 8 Netting collapses many gross exposures under one master agreement into a single net figure — the real exposure the bank carries

Netting has two forms, and both depend on clean capture. Payment netting combines the cash flows due on the same day so only the net moves, cutting settlement traffic. Close-out netting is the important one: if a counterparty defaults, all trades under the master agreement are terminated and collapsed into a single net claim, rather than the defaulter cherry-picking the trades that favour it. But netting only holds if the trades are correctly captured under the right master agreement and the right netting set — a trade booked under the wrong agreement, or with the wrong counterparty entity, may not net, inflating the bank's exposure and its capital. Netting is a legal construct that only works if the data is right.

Collateral and margin: the trade keeps calling for cash

A booked derivative does not sit quietly — it calls for collateral as the market moves. Under a collateral agreement, the two parties post margin to cover their exposure to each other, and that margin is recalculated and exchanged continuously through the trade's life. This turns a single booked trade into an ongoing stream of collateral movements, each of which the capture and operations estate must process against the trade. Collateral management is a large operational function in its own right, and it reads directly from the captured records.

There are two kinds of margin, and they answer two different questions. Variation margin covers the move that has already happened: each day the position is marked to market, and the party that has lost value posts cash to the party that has gained it, so the current exposure is always covered. Initial margin covers the move that might happen: an amount held against the potential future exposure if the counterparty defaulted and the position had to be replaced in a stressed market. Variation margin flows daily with the market; initial margin is held for the life of the trade and sized by model.

Collateral is where a capture error becomes a cash dispute with a counterparty. Because margin is calculated from the trade record, a wrong notional or a wrong maturity does not just misstate risk internally — it produces a margin call the counterparty disputes, because their record says something different. Collateral disputes are, at root, reconciliation breaks between two banks' captured versions of the same trades, and resolving them means going back to the records and the confirmations. This is why clean capture and matched confirmations are the foundation of a quiet collateral operation: the margin numbers can only agree if the underlying trades agree first.

Booking and capital: where a trade lands decides what it costs

A trade consumes capital, and how much depends on how and where it is booked. Regulators require a bank to hold capital against the risk of its positions, and the booking model feeds directly into that calculation. The counterparty, the netting set, the collateral, the legal entity and the trading-versus-banking-book classification all flow from the captured record into the capital engines. Two economically identical trades booked differently can carry different capital charges — which makes booking not merely an operational choice but a financial one.

Two capital regimes bear directly on trade capture. The standardised approach for counterparty credit risk turns each derivative's terms — asset class, notional, maturity, collateral, netting — into a counterparty exposure and hence a capital number, so the fields captured at booking are the inputs to the charge. And the Fundamental Review of the Trading Book sharpens the capital treatment of market risk and hardens the boundary between the trading book and the banking book, so the book a trade is captured into changes how much capital it attracts. In both cases the capital calculation is only as good as the captured data: a wrong netting flag or a mis-classified book produces a wrong capital number, and capital is expensive.

The daily rhythm: marks, flash P&L and the T+1 sign-off

A markets business runs on a daily cycle, and the captured record is at the centre of it. Understanding this rhythm is understanding when errors surface and why timeliness at capture matters so much. The cycle repeats every business day and turns the continuous flow of trades into a discrete, signed-off daily truth.

The day ends with marking, a flash number and then an official one. At the close, positions are marked to market — valued against end-of-day market data pulled through the trade records. The front office produces a fast "flash" P&L, an early estimate for the desk, before the books are fully closed. Overnight and into the next morning, finance produces the official, reconciled P&L and signs it off, typically on a one-day lag. The gap between the flash and the official number is itself a control: a large, unexplained difference points to a capture, valuation or reference-data problem that surfaced overnight.

The trade as a state machine

The cleanest way to picture a trade's life is as a state machine — a defined set of states with defined transitions between them. A trade is born "new", becomes "verified" once the middle office and the counterparty agree, then lives as an active position through which events flow, and finally exits by novation, termination or maturity. Every transition is a captured, timestamped event, and the sequence of transitions is the audit trail a regulator can reconstruct years later.

Figure 9 The trade lifecycle as a state machine — new, verified, live with events flowing through it, then novated, terminated or matured; every transition is captured

Reading the diagram left to right is reading the trade's whole life. The record enters as new and is promoted to verified once trusted. In the live state it absorbs a stream of events — fixings, coupons, exercises, amendments — each of which updates it without changing its identity, which is why the self-loop matters: the trade stays the same trade while its terms and cash flows evolve. It exits either by being novated to a central counterparty, or by terminating early or maturing. A capture system that cannot represent every state and transition cannot support the products that use them.

Amend versus cancel-and-rebook: two ways to fix a trade, with different consequences

When a booked trade is wrong, there are two ways to correct it, and choosing between them is a real decision. An amendment changes the existing record in place, creating a new version while preserving the trade's identity and its full history — the trade was always this trade, now corrected. A cancel-and-rebook tears up the original record entirely and books a fresh one, breaking the identity chain. The two look similar on a screen and are not the same at all.

The choice matters because of what downstream reads. An amendment preserves the audit trail and the unique trade identifier, so the regulatory report is updated as a modification and both sides stay paired. A cancel-and-rebook generates a termination of one trade and the birth of another, which for a reported derivative means two report actions and, if mishandled, a broken pairing and a new identifier where the regulator expected continuity. As a rule, genuine economic changes and error corrections that preserve identity are amendments; only a trade booked so wrongly that it was never really the right trade should be cancelled and rebooked. Getting this wrong is a quiet but common source of reporting breaks.

Compression: shrinking the book without changing the risk

A live derivatives book accumulates offsetting trades, and periodically the bank compresses them away. Over months, a desk books hundreds of trades that partly cancel each other — a pay-fixed swap here, a near-identical receive-fixed swap there. The net risk is small, but the gross number of live trades is large, and every live trade carries operational cost, capital charges and reconciliation burden. Portfolio compression is the exercise, often run through a third-party service across many banks at once, of tearing up redundant offsetting trades and replacing them with a smaller set that carries the same net risk.

Compression is a mass lifecycle event, and the capture estate has to process it cleanly. It terminates many trades and books a few replacements in one coordinated run, each termination and each new trade a captured, reported event. Done well, it can remove a large share of a book's gross notional without moving its risk an inch — which is why it is a standing part of derivatives operations, not a one-off. For the capture layer it is a stress test: a bank that can compress cleanly has a lifecycle-event and reporting pipeline that works at scale.

  1. Why does effort and risk peak at the front of the flow?

  2. What is the difference between the representation and location dimensions of a booking model?

  3. Why do the controls have to be layered rather than single?

  4. What is a canonical trade model and why is it hard to retrofit?

  5. Why must internal trades net to zero, and why is the internal-external tag critical?

  6. Why is valuation consistent by construction when it reads the trade record?

  7. Why must every lifecycle event be booked, not just the original trade?

  8. What does delivery versus payment protect against at settlement?

2

  1. Why does effort and risk peak at the front of the flow?

  2. What is the location dimension of a booking model?

  3. Why must controls be layered?

3Roles: who touches the record, and where value is added

Trade capture has no single owner. It is a cross-functional discipline in which many teams read or write the same record, each with different needs and different fears. This part names the actors, maps their interactions, and shows where a business analyst or technologist earns their place in the chain.

Four camps surround the record, and no desk owns it alone

The whole front-to-back organisation touches trade capture, and it is useful to group the touchers into four camps. The first camp originates: front-office traders and sales strike the deal and capture its economics. The second camp guards and processes: the middle office and operations verify, enrich, confirm, settle and investigate breaks. The third camp consumes independently: risk and finance read positions, sensitivities and profit and loss, and check them independently of the front office. The fourth camp builds and stewards: reference-data teams and technology run the platforms, govern the master data, and deliver the change.

The record sits in the middle of every interaction between these camps. No team keeps its own copy of the truth. The front office writes the record; the middle office trusts and verifies it; operations acts on it; risk and finance measure from it; technology and data make it possible. That shared dependency is the defining social fact of the discipline: capture quality is a shared responsibility, and its systems are shared infrastructure that no single desk can claim.

Figure 10 The golden trade record at the hub — every team reads or writes it, and none owns it alone

The interactions, mapped

It helps to see the flows as a table of who sends what to whom, and at which stage. The pattern that emerges is that almost every arrow either writes the record or reads it — the record is one party to nearly every interaction.

From To What flows Stage
Front office Capture system Executed economics Capture
Rules engine Trade record Enrichment from reference data Enrichment
Middle office Trade record Verification, amendments, breaks Verify
Bank Counterparty Confirmation for matching Confirm
Trade record Risk Positions and sensitivities Distribution
Trade record Finance Valuations and P&L to the ledger Distribution
Trade record Operations Terms and settlement instructions Settlement
Trade record Reporting Fields for EMIR / MiFIR / SFTR Reporting
Reference-data team Enrichment Governed master data Continuous
Technology All Platforms, feeds, mappings Continuous

The people you will actually work with

Behind the camps are specific roles, and knowing them is the difference between a consultant who understands the domain and one who does not. These are the people whose requirements you elicit, whose data you depend on, and whose sign-off you need.

The trade-support or middle-office analyst manages the daily lifecycle. They verify bookings, chase confirmations and clear breaks. They are the operational guardians of the record and, for anyone delivering change, the key source of requirements — they know exactly where the flow hurts.

The product controller, in finance, owns profit-and-loss attribution and valuation control. They care deeply about consistent, reconcilable trade data because their independent numbers must tie back to the front office's. When capture quality is poor, the product controller is the one who feels it at month-end.

The reference-data steward governs instruments, counterparties, LEIs, settlement instructions and hierarchies. They own the master data that enrichment depends on. Any change to capture that touches reference data lives or dies on the steward's cooperation.

The business analyst is the translator, and this is often where a consultant sits. The analyst elicits requirements, writes the source-to-target mappings, defines the validation and booking rules, and tests the flow end to end. Trade-capture change lives or dies on precise mappings, validation rules and product coverage — and the analyst who genuinely understands the trade model, the lifecycle and the downstream consumers is the person who makes a migration or a new-product onboarding actually work.

The developer or platform engineer builds and configures the systems — the capture screens, the trade store, the canonical model, the APIs and the vendor-platform configuration. And the data or architecture lead owns the canonical model, the lineage and the governance — the target state that every change is working toward.

The consumer taxonomy: four readers, four different needs

Just as an origination business must understand its investors, a capture business must understand its consumers. The record has four principal readers, and they do not want the same thing from it. Understanding what each consumes, on what timing, and at what granularity, is what lets you design a record that serves all of them — and explains why a defect that is invisible to one consumer is catastrophic to another.

Risk management reads positions and sensitivities, and it needs them near real time. Risk works on the aggregated position, not trade by trade: delta, DV01, vega, value-at-risk and limit checks, all computed from captured trades. Risk's fear is incompleteness and latency — a trade captured late or wrongly means the desk hedges the wrong amount, and a wrong direction or notional means real market losses can follow.

Finance and product control read valuations and profit and loss, and they need them reconcilable at end of day. Finance books the official profit and loss and the general ledger, and independently re-values to verify the front office's marks. Finance's fear is inconsistency — if the trade the front office sees is not the trade finance values, the numbers will not tie, and the break will surface expensively at reconciliation.

Operations reads terms and settlement instructions, and it needs them complete and correct by the settlement deadline. Confirmation, clearing, settlement and collateral all read the trade's terms and standing instructions. Operations' fear is a wrong or missing instruction — it causes a failed settlement, penalties and disputes.

Regulatory reporting reads a defined set of fields, and it needs them accurate and reconstructable. EMIR, MiFIR and securities-financing reporting are generated from the captured record. Reporting's fear is a wrong field that becomes a wrong report to a regulator — a reportable breach in itself.

Consumer What it reads Timing Why it depends on capture
Risk Positions, sensitivities, limits Near real time Hedges the wrong amount if capture is late or wrong
Finance / product control Valuations, P&L, ledger End of day Numbers fail to tie if the record is inconsistent
Operations Terms, settlement instructions By settlement deadline Failed settlement if instructions are wrong or missing
Regulatory reporting Defined regulatory fields T+1 and on events A wrong field is a wrong report — a breach

Segregation of duties: the information barrier inside the record

A markets business must keep certain functions apart, and the deepest reason is that one institution knows things it must not act on freely. The person who strikes a trade must not be the same person who verifies it, values it independently, or settles it. This is segregation of duties, and it is the trade-capture equivalent of the information barriers that separate advisory from trading. It is not bureaucracy — it is the structural defence against a single person being able to both create risk and conceal it.

The functions are deliberately separated, and each separation defends against a specific failure. The front office creates trades but cannot verify its own bookings — that is the middle office's job, which is why four-eyes requires a different pair of eyes. The front office marks its own positions but cannot be the final word on their value — that is finance's independent price verification. And the person who trades cannot also settle, because whoever controls both the record and the cash can hide a loss between them. Each wall turns a single point of failure into a two-person problem.

Segregation of duties is a technology problem as much as a policy one. The wall is enforced in software: entitlements and access control decide who can capture, who can verify, who can amend and who can settle; the maker-checker workflow is a system feature that refuses to let one user both book and approve; the audit trail records who did what and when, immutably; and surveillance systems watch for the patterns that segregation is meant to prevent. A policy that says "traders must not settle their own trades" is worthless unless the system makes it impossible. As trades and communications move to new channels — chat, messaging apps, new venues — the surface the wall must cover keeps expanding, and keeping entitlements aligned with it is a moving target.

The six teams around a trade, and the wall running through them

It helps to fix the operating model in a table: which team, in which division, doing what, on which technology. The information barrier — segregation of duties — runs horizontally through this table, keeping the creators apart from the verifiers and the settlers.

Team Division Core role Primary technology
Trading / sales Front office Originate and capture the trade Capture screens, e-trading, pricers
Middle office / trade support Middle office Verify, enrich, amend, confirm Trade store, matching platform, workflow
Operations / settlements Operations Confirm, clear, settle, collateral Settlement systems, SWIFT / ISO messaging
Risk management Risk Positions, sensitivities, limits Risk engine, position book of record
Finance / product control Finance Independent P&L, valuation, ledger Accounting book of record, IPV tools
Reference-data & technology Change / IT Steward master data, run platforms EDM platform, canonical store, feeds

Reading the table top to bottom is reading the trade's journey through the organisation. The front office creates it, the middle office trusts it, operations acts on it, risk and finance measure it, and technology and data make all of it possible. Every row reads or writes the same record, and the wall between the top rows and the control rows is what keeps the record honest.

What the middle office actually does all day

The middle office is the least understood and most important team in the flow, and it is worth seeing its day concretely. Its work is not glamorous and it is not optional: it is the human layer that turns a captured trade into a trusted one and keeps the record honest as it lives. Understanding this day is understanding where a change project's requirements actually come from.

The day is built around exceptions, not around the trades that go right. A high-STP operation means most trades book, enrich and confirm untouched — and the middle office never sees them. What reaches a human is the residue: a trade that failed enrichment because a settlement instruction was missing, a large manual trade that needs a second pair of eyes, a confirmation that would not match because the two sides disagree on a date, a reconciliation break where the position does not equal the sum of trades. Each of these is an exception in a queue, and clearing the queue is the job.

Break investigation is the core craft, and it is detective work. A break is a discrepancy between two records that should agree. The analyst's task is to find where the two diverged — a missed amendment on one side, a mapping error in a feed, a trade booked twice, a reference-data value that changed underneath a position. Resolving it means tracing the trade's lineage back to the point of divergence, which is exactly why the lineage and mapping artefacts from Part 4 matter operationally, not just architecturally. The middle-office analyst lives inside the lineage.

For anyone delivering change, the middle office is the richest source of requirements in the building, because they know precisely where the flow hurts. Every recurring break is a defect waiting to be fixed at source; every manual workaround is an automation opportunity; every "we always have to check this by hand" is a control that could be systematised. A business analyst who spends a week sitting with the middle office learns more about what the transformation should actually do than a month of stakeholder workshops, because the breaks queue is an unfiltered, ranked list of everything wrong with the current estate.

  1. Why does trade capture have no single owner?

  2. What is the business analyst's highest-leverage artefact, and why?

  3. How do risk and finance differ as consumers of the record?

  4. Why can a defect be invisible to one consumer yet catastrophic to another?

  5. What failure does segregation of duties defend against, and how is it enforced?

  6. Why is the Barings collapse the canonical lesson for this part?

3

  1. Why does trade capture have no single owner?

  2. What is the business analyst's highest-leverage artefact?

  3. What does finance fear most as a consumer of the record?

4Systems, data and IT: the machine that carries the record

This part opens the machine. We start with the functions any capture estate must perform, then name the vendors that fill them, follow a field from source to regulator, dissect the trade record itself, and cover the identifiers, integration patterns, messaging protocols, resilience and regulation that make the whole thing work. Read it as the architecture a technologist actually builds and runs.

Start with functions, not vendors

Before naming a single product, ask what functions any trade must pass through. Vendors come and go and every bank's estate looks different, but the functional blocks are stable. Naming them first gives you a clean map onto which any real system can be placed.

Figure 11 The functional architecture — a chain of blocks from capture to reporting, over a band of cross-cutting capabilities every trade depends on

Six functional blocks carry a trade from entry to reporting. Capture and e-trading is where the trade enters, from any channel, into the canonical model. Enrichment and reference data completes the skeleton by lookup. The trade store, the canonical golden record, is the book of record everything trusts. Valuation and risk computes present value, sensitivities and limits. Settlement and operations confirms, clears, settles and manages collateral. And reporting produces profit and loss, the ledger and the regulatory reports. Underneath all six runs a band of cross-cutting capabilities — compliance and segregation of duties, reference-data mastering, audit and lineage, and communication and workflow — that no single block owns but every block depends on.

Functional block What it does Consumes Produces
Capture / e-trading Records executed economics in the canonical model Channel messages A skeleton trade record
Enrichment / reference data Completes the record by lookup Master data A complete record
Trade store (golden record) Holds the authoritative trade and its lifecycle Enriched records The single source of truth
Valuation / risk Computes value, sensitivities, limits Record + market data Positions and P&L
Settlement / operations Confirms, clears, settles, collateral Terms + instructions Settled obligations
Reporting Produces P&L, ledger, regulatory reports The record + events Reports and the ledger

The vendor landscape: buy versus build as a map of advantage

A real bank estate is a mix of bought platforms and built components, and where the line falls is itself a strategy. Banks buy where a capability is undifferentiated commodity, and build where it is a source of competitive advantage. The pattern of buy-versus-build is a map of where a bank thinks its edge lies.

The dominant vendors are worth knowing by name. Murex, through its MX.3 platform, is the dominant cross-asset trading, capture and risk system — configurable trade templates, booking, risk and back office in one suite. Calypso, now part of Adenza under Nasdaq, is a major competitor, strong in rates, FX and clearing. Summit and similar treasury and capital-markets systems still run in many banks for specific asset classes. On the data side, enterprise data-management platforms such as GoldenSource master and distribute reference data — instruments, counterparties and prices. And Bloomberg is the ubiquitous source of instrument, pricing and identifier data, from tickers and ISINs to the feeds that enrich and value trades.

The real-world pattern is plurality, not a single system. A typical bank runs multiple capture systems — split by asset class, by region, by legacy acquisition — plus a reference-data platform, all feeding one aspirational golden trade store. Many banks build their own canonical trade store and capture layer around vendor pricing and risk, because the canonical model is where they see advantage and the pricing library is where they do not. Understanding how these vendors model and connect trades is core knowledge for anyone working in markets technology.

Why do banks so often build the trade store but buy the pricing engine? Because the two sit on opposite sides of the advantage line. A pricing library for a vanilla swap is a solved problem — every vendor's is much like every other's, so building your own wins nothing and costs a fortune to maintain. But the canonical trade model, the golden source and the way trades flow and reconcile across the estate is where a bank's operating leverage lives: a cleaner model means higher STP, fewer breaks and faster change. That is worth owning outright. The buy-versus-build map, read across a whole estate, tells you exactly where a bank believes its technology edge is — and where it has decided it has none.

Data lineage: following one field from source to regulator

Lineage is the traceable path of a field from the system that created it, through the canonical model, to every consumer that reads it. It is the single most useful artefact in markets technology, because it answers three questions that otherwise take days: where did this number come from, what breaks if I change this field, and where did this wrong value enter. A mapping document that captures lineage is, quite often, the actual deliverable of a change project.

Source field (capture system) Canonical field Downstream consumers
Deal number trade.tradeId Risk, Finance, Operations, Reporting
Notional and currency trade.notional Risk (position), Finance (P&L)
Counterparty short name party.lei Reporting, Credit, Operations
Portfolio trade.book Risk and P&L aggregation tree
Fixed rate swap.fixedLeg.rate Valuation engine, P&L
Maturity trade.maturityDate Risk (tenor bucket), Settlement
Status trade.status All (lifecycle gating)
Generated at canonical layer trade.uti EMIR / MiFIR trade reporting

The lineage of our running swap reads left to right across the estate. Execution venue, then capture system, then the canonical trade store, then the risk, ledger and regulatory feeds. Each hop is a mapping, and each mapping is a place a field can be dropped, renamed or transformed wrongly. This is why most incidents happen at boundaries, not inside systems — a missed message, a mapping gap, a late feed. The engineering effort in a capture estate goes disproportionately into robust interfaces, monitoring and reconciliation at exactly these hops.

Figure 12 Data lineage — one field's journey from source system through the canonical record to every downstream consumer, with a reconciliation at each boundary

The anatomy of the trade record

It is worth looking closely at the atom itself, because every downstream number is built on its fields. A captured trade record, in a canonical view, carries a trade identifier and a unique trade identifier; the product and its product identifier; direction, notional, rate and index; trade, effective and maturity dates; the counterparty and its LEI; the book and legal entity; the trader; the status and version; the capture channel; and a timestamp. Each field drives something specific downstream: the notional drives the position and the profit and loss, the LEI drives the regulatory report and the netting, the book drives the aggregation tree, the status gates the whole lifecycle.

Identifiers: the keys everything joins on

A handful of standard identifiers are the join keys of the entire markets data estate. They are worth memorising, because enrichment, reporting and reconciliation all turn on them.

Identifier Full name Identifies Where it bites
UTI Unique Trade Identifier A single trade, shared by both sides Pairs the two sides' reports; deduplication
UPI Unique Product Identifier The product type Makes products comparable across firms
LEI Legal Entity Identifier A legal counterparty Mandatory to report — no LEI, no valid report
ISIN International Securities Identification Number A specific security Settlement and instrument reference
FIGI Financial Instrument Global Identifier An instrument, vendor-neutral Mapping instruments across data vendors

The UTI and the LEI carry the most operational weight. The UTI is generated and stored at capture so that both counterparties report the same trade under one identifier and the regulator can pair the two reports; it is also the backbone of deduplication. The LEI is simply mandatory — a trade without a valid LEI for each counterparty cannot be reported, and therefore cannot be booked in a controlled way. The others matter, but these two are the ones that stop a trade dead when they are missing.

Integration patterns: batch, API and event streaming

How systems talk to each other is not a detail — it decides how fresh, how reliable and how reconcilable the estate is. Three patterns dominate, and they represent three generations of architecture.

Batch is the legacy default. Files move on a schedule, often overnight. It is simple and robust but latent: risk built on last night's batch is a day stale, and every batch boundary needs a reconciliation to prove nothing was lost. Most legacy reconciliation burden exists because systems talk in batches and must be checked against each other.

APIs are the modernisation. A system requests data on demand and gets an answer now, so consumers are no longer waiting for the nightly file. APIs cut latency and coupling, and they are how most new integrations are built.

Event streaming is the frontier. Systems publish events — a trade captured, a trade amended — onto a stream, and every interested consumer subscribes and reacts in real time. This is more than a speed upgrade.

The migration from batch to event streaming is the one that changes the shape of the problem, not just its speed. Under batch, every consumer holds its own copy built from last night's file, and reconciliation exists to prove those copies still agree. Under a well-designed event stream, every consumer reacts to the same published events in real time, so the copies do not drift in the first place — and the need for much of the reconciliation simply disappears. That is why streaming is strategic: it does not make the old reconciliation faster, it removes the reason the reconciliation existed. Realising that benefit is hard, because it means re-plumbing an estate built on files, but the prize is a structurally simpler and more real-time bank.

Immutable audit and the decision engine

Two architectural principles recur across the compliance-critical parts of the estate. The first is that records which matter for compliance must be append-only. You never overwrite a trade's history; you add a new version, so the full sequence is reconstructable years later. Governance and audit platforms enforce this immutability, because a regulator's question is not "what is the trade now" but "what was it at every point in its life". The second is that the decision engines — the allocation, booking-rule and pricing logic that make judgements — are almost always proprietary, because they encode the bank's own policy and advantage and cannot be bought off the shelf.

Messaging protocols: two worlds, front and back

A trade lives in two messaging worlds, and the split between them is the front-office and back-office divide made concrete. In the front-office world, the FIX protocol carries orders and executions between venues, dealers and clients at very low latency. FIX is fast, standardised and session-based, and it is how an electronically executed trade enters the primary capture process in the first place. In the post-trade world, standardised financial messages move confirmations and settlement instructions between banks, custodians and central securities depositories — the SWIFT MT message family, now migrating to the richer, more structured ISO 20022 XML standard.

Figure 13 Two messaging worlds — FIX carries the trade in at the front, SWIFT and ISO 20022 carry it out at the back, with the canonical record in between
Layer Standard Carries Timescale
Front office FIX Orders, executions, quotes Milliseconds
Post-trade SWIFT MT Confirmations, settlement instructions Minutes to hours
Post-trade (modern) ISO 20022 / MX Richer confirmations and reports Minutes to hours

Operational resilience: the moments with no second chance

Some moments in a trade's life are irreversible, and the systems must not fail during them. A settlement deadline, an intraday risk cut during a market shock, a large batch of lifecycle events on a coupon date — these have no do-over. Resilience is the engineering discipline of surviving them, and it is where the abstract phrase "mission-critical" becomes something you can measure.

Concept What it guarantees Why it matters on the critical day
Redundancy No single component is a single point of failure A dead server does not stop capture during a market shock
Automatic failover Traffic switches to a healthy system without human action Seconds of downtime, not hours, at the settlement deadline
Disaster recovery A tested ability to restore in another location A data-centre loss does not lose the book of record
Business continuity People and process keep running through disruption The desk can still book when the primary site is down

An illustrative critical day: resilience made concrete

The abstract case for resilience becomes vivid on a specific bad day. Picture a morning of extreme market volatility — a central-bank surprise, a sharp move across rates and equities. Trading volume spikes, so capture throughput spikes; risk needs intraday recalculation because positions are moving fast; and a large batch of lifecycle events happens to fall due the same day. This is precisely the moment the capture estate is under maximum load and least able to tolerate a failure, and it is the moment resilience is designed for.

On such a day, each resilience concept earns its keep in a way you can point at. Redundancy means the surge in capture volume does not overwhelm a single server, because load is spread across several. Automatic failover means that when one component does fall over under the load, traffic switches to a healthy one in seconds, not hours, so capture never stops and risk is never blind. Rehearsed disaster recovery means that if the primary data centre is lost entirely, the book of record is restored elsewhere without losing a trade. And business continuity means the people — the middle office clearing the surge of breaks — can keep working even if their primary site is disrupted.

Five regulatory regimes read as system specifications

Regulation shapes every system in the chain, and a technologist should read each regime as a specification, not a legal summary. Five are worth knowing for a European capture estate, and each imposes concrete demands on the record.

EMIR requires every derivative and every lifecycle event to be reported to a trade repository, by both sides, the next day — so the record must carry the UTI, the UPI and both LEIs, and must emit a report on every event. MiFIR transaction reporting requires transactions to be reported to the regulator for market-surveillance purposes under a detailed field specification — so the record must capture dozens of precise fields correctly at source. SFTR extends the same logic to securities-financing transactions. The EMIR reporting rewrite of 2024 tightened data quality, mandated the ISO 20022 XML format and the UPI, and expanded the field set — pushing banks to fix capture data at source rather than patch reports downstream. And data-protection law governs the personal data that inevitably attaches to trades and counterparties — trader identifiers, client details — imposing retention and access constraints on the very audit trails the other regimes demand.

Reference-data mastering: the platform behind enrichment

Enrichment is only as good as the reference data it reads, so banks run a dedicated platform to master that data. An enterprise data-management platform sits behind enrichment as its supplier. It ingests instrument, counterparty and price data from many external and internal sources, resolves conflicts between them, applies governance and stewardship, and distributes one clean, agreed version to every system that needs it. Without it, each capture system would master its own reference data, and the same counterparty would have five slightly different records — the fragmentation problem in miniature.

Mastering means choosing a single golden version from many conflicting inputs. Three vendors might disagree on an instrument's maturity or a counterparty's legal name; the platform applies rules and human stewardship to pick the authoritative value, records why, and publishes it. This is genuine data engineering — matching and merging records, surviving source outages, versioning every change, and proving lineage back to the source. The stewards who run it are the reference-data role named in Part 3, and their work is invisible until it fails, at which point every enrichment in the bank inherits the error.

The reason a bank centralises reference data rather than letting each system fetch its own is the cascade. If enrichment reads a wrong counterparty record, it does not fail loudly — it enriches successfully with the wrong value, and that value flows into settlement routing, netting and the regulatory report before anyone notices. A centralised, governed platform turns reference data from a per-system afterthought into a controlled shared asset with one steward, one golden version and one audit trail. This is why "master the reference data" is a headline item of the transformation agenda: clean shared data is the precondition for high straight-through processing, and dirty data is the silent cause of a large share of breaks.

The canonical model in practice: FpML and the Common Domain Model

A canonical model is easier to describe than to build, and the industry has produced shared standards to help. FpML — Financial products Markup Language — is a long-established XML vocabulary for representing financial products and trades, particularly derivatives. It gives two parties, or two internal systems, a common language for describing the same trade, so a swap means the same thing on both sides of an interface.

The ISDA Common Domain Model goes further by standardising events, not just trades. Where FpML describes what a trade is, the Common Domain Model describes what happens to it — how a novation, a partial termination, an amendment or a fixing is represented — so that both counterparties, and every internal system, describe the same lifecycle event the same way. This matters enormously for reporting: if both sides of a trade model an event identically, their reports reconcile by construction rather than by a painful bilateral investigation. Adopting these standards is how a bank makes its own canonical model interoperable with the rest of the market instead of a private dialect no one else speaks.

Why do industry standards matter for an internal transformation? Because a canonical model built in isolation solves the internal fragmentation problem but leaves the external one untouched. Two banks reporting the same trade to the same repository must describe it identically, or the repository cannot pair their reports — and a large share of reporting breaks are exactly these bilateral mismatches. Aligning the internal canonical model to FpML and the Common Domain Model means the bank's model is compatible with its counterparties' and its regulators' by design. The pragmatic path most banks take is to build an internal canonical model that maps cleanly to these standards, getting the internal benefit of one truth and the external benefit of interoperability at the same time.

Straight-through processing and where the flow breaks

Full straight-through processing is a trade that travels from execution to reporting with no human touch. Venue, capture, enrich, validate, book, then risk, confirmation, settlement and reporting — every step automated, with humans handling only exceptions. A high STP rate means lower cost, fewer errors and faster, more accurate risk, which is why it is a headline metric for the operation. But STP is fragile precisely because it is a chain: every hand-off is a place a trade can fall out into a manual exception.

Figure 14 Full straight-through processing and its break points — every hand-off is a place a trade can fall out into a manual exception

The break points are predictable, and each maps to a specific weakness. Enrichment fails when a settlement instruction, product term or counterparty mapping is missing, and the trade stalls waiting for reference data. A product-coverage gap sends an exotic the system cannot model to manual booking, with no STP and high error risk. A reference-data miss — a new counterparty with no LEI, an instrument not yet set up — stops enrichment completing. And an interface or mapping error drops a message or maps a field wrongly, so the trade arrives late, missing or subtly wrong downstream.

Break point Cause Consequence
Enrichment failure Missing SSI, product terms or mapping Trade stalls as an exception
Product coverage gap Exotic the model cannot represent Manual booking, no STP, high error risk
Reference-data miss New counterparty, no LEI, unset instrument Enrichment cannot complete
Interface / mapping error Dropped message, wrong field mapping Late, missing or subtly wrong downstream

Raising the STP rate is therefore an attack on these four break points at once. Cleaner reference data cuts enrichment failures and reference-data misses; wider product coverage shrinks the manual tail; and robust, monitored interfaces cut mapping errors. This is why the transformation agenda in Part 5 treats STP not as a single lever but as the visible result of fixing fragmentation and data quality underneath.

Clearing and novation: when a counterparty becomes a CCP

A large share of derivatives no longer face their original counterparty — they face a central counterparty. After the financial crisis, regulators mandated central clearing for many standardised derivatives to reduce the web of bilateral exposures that made the system fragile. Central clearing changes the trade record in a specific, bookable way, and the capture estate has to handle the transition cleanly.

Clearing works by novation, and novation is a lifecycle event the record must capture. When a trade is cleared, the original bilateral trade between the bank and its counterparty is legally replaced by two trades, each facing the central counterparty — the clearing house steps into the middle and becomes the counterparty to both sides. The original counterparty leg is torn up and re-booked against the central counterparty. This is not a cosmetic change: the counterparty LEI changes, the collateral and margin terms change, and the regulatory report must reflect the novation with the correct event type. A capture system that cannot represent novation cannot support clearing.

Central clearing also introduces daily margining, which generates its own record traffic. The central counterparty marks every cleared position to market daily and calls variation margin to cover the move, plus initial margin to cover potential future exposure. Each margin call and movement is an event that operations must process against the trade, and the collateral must reconcile to the cleared positions. For the capture estate, clearing turns a single bilateral trade into an ongoing relationship with a clearing house, with novation at the start and a stream of margin events through its life.

Why does clearing matter so much for trade capture specifically? Because it is the clearest example of a trade that is booked several ways over its life against several counterparties, all of which must reconcile. The trade is captured bilaterally, novated to the central counterparty, margined daily, and eventually terminated or matured — and at every step the record must emit the correct lifecycle event, re-book the correct counterparty leg, and keep the internal, external and cleared views reconciled. A bank whose capture layer handles clearing cleanly can handle almost any lifecycle complexity; one that cannot is exposed precisely where post-crisis regulation concentrated the volume.

Who generates the UTI? The generation waterfall

A single trade needs one shared identifier, which raises a deceptively hard question: which side creates it? The unique trade identifier only works if both counterparties report the same value, so exactly one of them must generate it and communicate it to the other in time. When two banks, or a bank and a client, trade with each other, they cannot both invent an identifier independently — that would give the regulator two unpaired reports for one trade. So the industry defined a waterfall: a deterministic set of tie-break rules deciding who generates the identifier.

The waterfall walks down a priority order until responsibility lands on one party. If the trade is executed on a venue or cleared, the venue or clearing house generates the identifier. If not, and one side is the kind of entity obliged to report while the other is not, the reporting side generates it. If both are equivalent, a further tie-break — often based on sorting the two legal entity identifiers — decides. The point of the waterfall is determinism: from the same facts, both parties reach the same conclusion about who generates the identifier, so there is never a clash and never a gap.

The UTI waterfall is a small rule with a large operational tail. Because the identifier must reach the other side before the reporting deadline, the capture estate has to generate it, or receive it, and stitch it onto the trade in time — and if it arrives late, is malformed, or the two sides disagree on who should have generated it, the reports do not pair and a break lands at the trade repository. This is why identifier generation, which sounds like a triviality, is a genuine engineering concern: it is a distributed agreement problem between two firms who must independently compute the same answer, under a deadline, for every trade.

Bitemporality: two clocks on every record

The hardest and most important idea in trade data is that every record lives on two clocks at once. One clock is valid time — when something was true in the business: the date a trade was effective, the date a term applied. The other is system time — when the bank knew it: the moment the record was captured or corrected. Confusing the two makes it impossible to answer the questions a regulator and an auditor actually ask, and keeping them separate is what makes a trade store trustworthy.

Figure 15 Two clocks on every record — valid time (when it was true) and system time (when we knew it) — the key to reconstructing what the bank believed on any past date

The two clocks are needed because corrections travel backward in valid time but forward in system time. Suppose a trade is booked today, and next week the bank discovers a term was wrong and corrects it, effective from the original date. In valid time the correction reaches back to the trade's start; in system time it happened only when the correction was booked. A single-clock system cannot represent this: it either overwrites history, losing the fact that the bank once believed something different, or it cannot apply the correction to the past. A bitemporal system keeps both, so it can answer "what is true now" and "what did we believe on any given past date" — which is exactly what audit reconstruction and dispute resolution require.

Why does bitemporality matter so much for a regulator? Because the regulator's question is almost never "what does the trade look like today". It is "what did you report on the fifteenth, and was it correct given what you knew then". Answering that requires replaying the record as of a past system time — seeing the world through the bank's eyes on that date, before later corrections were known. A trade store that only holds the current truth cannot do this; it has thrown away the history of its own beliefs. Bitemporal modelling is the price of being able to reconstruct not just what was true, but what you knew and when — and in a regulated business, being able to prove what you knew and when is the whole game.

Event semantics: idempotency, ordering and replay

Moving from batch files to a real-time event stream solves one problem and creates three new ones, and they are the ones that keep engineers awake. When a trade is published as a stream of events that many consumers subscribe to, the estate becomes real-time and reconciliation-light — but only if it handles duplication, disorder and recovery correctly. These three concerns are the price of admission for event-driven architecture, and getting them wrong reintroduces exactly the breaks streaming was meant to remove.

Idempotency means processing the same event twice has the same effect as processing it once. Networks retry, systems replay, and the same "trade amended" event may arrive more than once. A consumer that is not idempotent will double-apply it — booking the amendment twice, doubling a position. The fix is that every event carries an identity, and consumers recognise one they have already seen and ignore the repeat. This is deduplication in its streaming form, and it is why the unique identifiers matter as much to the pipeline as to the regulator.

Ordering means events are applied in the sequence they truly occurred. A trade might be captured, then amended, then the amendment corrected — and if a consumer receives the amendment before the original, or the correction before the amendment, it builds a wrong picture. The stream must either guarantee order or carry enough information — sequence numbers, timestamps — for consumers to reassemble the true sequence. And replay means that after an outage, a consumer can rebuild its state by re-reading the stream from a known point, which is only safe if processing is idempotent and ordered.

The reason these three semantics are non-negotiable is that an event stream is a promise about truth, and a broken promise is a silent corruption. Under batch, a bad file is usually obvious — it is late, or it fails to load. Under streaming, a duplicated or out-of-order event does not announce itself; it quietly leaves one consumer holding a position that no longer matches the trade store, and the break only surfaces at the next reconciliation. This is the irony of the migration: streaming can remove the need for reconciliation, but only if idempotency, ordering and replay are engineered correctly — and if they are not, streaming manufactures exactly the drift it promised to abolish. The promise is real, but it has to be earned.

Inside the reconciliation engine

Reconciliation is named constantly in this module, so it is worth opening the engine that does it. A reconciliation engine compares two sets of records that should agree — the trades in the capture system against the trades in a downstream consumer, or the position against the sum of trades — and reports where they differ. It is the backstop control that assumes the other defences missed something, and in a fragmented estate it runs everywhere, at every boundary.

The engine works by matching, tolerating and classifying. First it matches records across the two sides on keys — a trade identifier, or a composite of counterparty, product and date when identifiers differ. Then it applies tolerances, so that trivial differences — a rounding, a timing lag — do not raise a false break. What remains is a break, and the engine classifies it: a trade present on one side and missing on the other, a trade present on both but with a mismatched field, a timing difference that will clear tomorrow. The classification routes the break to the right team and, increasingly, to automated resolution for the well-understood types.

Break type What it means Typical cause
Missing one side A trade in one system, absent in the other Dropped feed, unbooked leg, timing
Field mismatch Present both sides, a field differs Mapping error, unpropagated amendment
Timing difference Agree once both sides catch up Batch lag across a boundary
Duplicate The same trade booked twice Failed deduplication, replayed event

Migrating the estate: parallel run and shadow booking

The hardest thing a markets technology function ever does is move trades from an old system to a new one without breaking anything, and there is a disciplined way to do it. A trade store cannot simply be switched off and replaced — it holds live positions that risk, finance and reporting depend on every second. So migrations are done gradually, under reconciliation, with the old and new systems running side by side until the new one is proven. This is the change project a business analyst spends a career on, and its techniques are worth naming.

Parallel run and shadow booking are the core techniques. In a parallel run, both the old and new systems process the same trades for a period, and their outputs — positions, valuations, reports — are reconciled every day; the new system is only trusted once it matches the old one within tolerance for long enough. Shadow booking is the same idea applied to new capability: a new product or a new booking model is booked in the new system in parallel with the existing workaround, and the two are compared before the workaround is retired. In both, reconciliation is not a check at the end — it is the mechanism of the migration itself, the thing that proves the new system is safe before anyone relies on it.

Why is reconciliation-driven migration the only safe way to move a trade store? Because the alternative — a big-bang cutover — bets the bank's ability to see its own risk on the new system being perfect on day one, and no system is. A parallel run turns that bet into a measured comparison: for weeks, every discrepancy between old and new is a break to investigate, and each investigation either finds a bug in the new system or a quirk in the old one that needs deliberate handling. By the time the old system is retired, every difference has been explained, not assumed away. This is slow and expensive and it is why large migrations take years — but the failure mode of the fast alternative is a bank that cannot value its book, which is not a risk any markets business can take.

Measuring data quality: the five dimensions

"Data quality" is too vague to manage until it is broken into dimensions that can each be measured. A capture operation that wants to improve its data cannot act on a slogan; it needs to know which aspect of quality is failing and by how much. Five dimensions cover the field, and each maps to a concrete metric on the captured record.

Dimension The question it answers Failure looks like
Completeness Are all required fields present? A trade missing its SSI or LEI, stalling enrichment
Accuracy Do the values match reality? A notional that disagrees with the term sheet
Timeliness Is the record captured soon enough? A late booking, so intraday risk is wrong
Consistency Do copies and related fields agree? Position not equal to the sum of trades
Uniqueness Is each trade represented exactly once? A duplicate double-counting risk and P&L

Naming the dimensions turns data quality into an engineering target. Completeness is measured as the share of records with all mandatory fields; timeliness as the lag from execution to a complete record; uniqueness as the duplicate rate the deduplication catches. Each dimension has an owner and a defence from earlier in this module — completeness leans on validation, accuracy on four-eyes and confirmation, consistency on reconciliation, uniqueness on deduplication and the unique identifier. Measuring them separately tells an operation not just that its data is imperfect but exactly where to spend the next unit of effort.

The reason to decompose data quality into dimensions is that the fixes are different for each, and a single quality score hides that. A completeness problem is solved by better reference data and validation; an accuracy problem by tighter controls at the point of entry; a timeliness problem by more straight-through processing; a consistency problem by consolidating copies; a uniqueness problem by fixing deduplication. Report one blended "quality is 94 percent" number and you know nothing actionable. Report the five dimensions separately and you have a work programme, ranked by which dimension is costing the most breaks. This is the data-quality version of the module's recurring lesson: the useful metric is never the single blended number.

  1. Why name functions before vendors?

  2. Why do banks often build the trade store but buy the pricing engine?

  3. Why do most incidents happen at boundaries between systems?

  4. What is deduplication and why is it hard?

  5. How is the migration to event streaming more than a speed upgrade?

  6. What is the difference between FIX and SWIFT, and where does each sit?

  7. Why is redundancy alone insufficient for operational resilience?

  8. Why is regulatory reporting quality now a trade-capture problem?

4

  1. Why do banks build the trade store but buy the pricing engine?

  2. Where do most incidents happen?

  3. Why name functions before vendors?

5Risks, failures and the transformation agenda

This part is about what goes wrong and what is being done about it. We map the risks that recur across every bank, the metrics that tell a healthy operation from a fragile one, three real incidents where the controls failed, the transformation programme reshaping the discipline, and an honest view of where artificial intelligence genuinely earns its place.

The six risk types, mapped for controllability

The risks in trade capture form a small, mutually exclusive set, and naming them cleanly is the first step to controlling them. Six types cover the field. Read the table as a map of controllability — each risk has an owner and a main defence, which means each is a problem you can engineer against, not a hazard you simply endure.

Risk type What it is Who owns it Main defence
Input / capture error Wrong economics entered — notional, rate, direction Front office Tolerances, four-eyes, confirmation matching
Booking / representation error Right trade, wrong book, entity or template Front / middle office Booking rules, validation, governance
Valuation / model error The record is marked or modelled wrongly Finance / risk Independent price verification, model governance
Operational / settlement error Breaks, failed settlement, unreconciled positions Operations Reconciliation, delivery versus payment
Regulatory reporting error A wrong or missing report to a regulator Compliance / operations Source data quality, reconciliation to reports
Conduct / segregation failure One party creates and conceals risk Risk / compliance Segregation of duties, surveillance, audit

Telling a good operation from a lucky one

A capture operation that has had no incident this month is not necessarily well run — it may just be lucky. To tell the difference you measure. Five metrics, read together, distinguish a genuinely healthy operation from one that is one bad day away from a headline.

Metric What it measures Healthy signal
STP rate Share of trades booked with no manual touch High and rising
Break rate Reconciliation breaks per thousand trades Low and stable
Capture timeliness Lag from execution to a complete record Near real time
Breaks outstanding Ageing of unresolved reconciliation breaks Low, and clearing fast
Reporting accuracy Share of regulatory reports accepted first time Very high

Three real failures, three distinct modes

The failure cases that matter are not random — they are the three ways the capture process breaks. The input can be wrong, the representation can be wrong, or the control structure can be wrong. Each of the three real incidents below is a clean example of one mode, and together they cover the field.

Citigroup, 2022 — when the input is wrong and the control is mis-set

A single mis-keyed order triggered a European flash crash. In May 2022 a Citigroup trader in London intended to sell a basket of equities worth roughly fifty-eight million dollars. A manual input error instead created an order with a notional of around four hundred and forty-four billion dollars. A pop-up warning was generated but overridden, and a preventive hard-limit control that should have blocked the order was not configured to catch it. Around 1.4 billion dollars of the order executed across 284 orders before it was stopped, sending several European indices down more than four percent for a few minutes and costing the bank roughly forty-eight million dollars. Because it was a bank-holiday period, the team that flags trading errors was short-staffed and did not notice for about fifteen minutes. In May 2024, United Kingdom regulators fined Citigroup a combined sixty-two million pounds — about seventy-nine million dollars — the Financial Conduct Authority and the Prudential Regulation Authority splitting the penalty for failures in trading controls.

The lesson is that a control is only as good as its configuration. Citigroup had the right controls in principle — a warning and a hard block — but the warning was overridable and the block was mis-set, and the human backstop was thinly staffed on the day. This is the input-error mode from Part 2's control gauntlet, and it shows that tolerances and hard limits are not a checkbox but a living configuration that must actually catch the errors it is meant to.

JPMorgan, 2012 — when the representation lies

The London Whale losses came from a book that was marked to conceal them. In 2012, traders in JPMorgan's Chief Investment Office, led by the trader nicknamed the London Whale, built enormous credit-derivative positions. As the positions moved against them, the book was marked not at the midpoint of the bid-ask spread but at the favourable extreme, understating the losses. A United States Senate investigation later found that on selected positions the office had booked values at the boundary of the bid-ask spread, and that using midpoints would have swelled the reported losses by hundreds of millions of dollars on those marks alone. The total loss reached at least 6.2 billion dollars. The Senate report described mismarked valuations, hidden losses, weak controls and a risk-model change that understated the danger.

The lesson is that the golden record can be corrupted at valuation, and only independent verification catches it. The economics were captured correctly; the marks were wrong, and wrong on purpose. This is the valuation-error mode, and its defence is precisely the independent price verification described in Part 2 — finance re-valuing the book against independent market data, refusing to accept the front office's marks on trust. When that control is weak or captured, a book can lie about its own value for months.

Barings, 1995 — when the control structure is absent

The oldest merchant bank in Britain was destroyed by one trader who controlled his own checks. Nick Leeson, a trader in Barings' Singapore office, had the unsupervised ability to both trade and settle his own trades — the front office and the back office in one pair of hands. He hid mounting losses in an error account, numbered 88888, that was meant to be cleared daily, while reporting profits to London. An internal audit in 1994 flagged that combining trading and settlement exposed the bank to error and fraud and recommended separating the roles; because Leeson appeared highly profitable, the recommendation was not enforced. Warning letters and a funding discrepancy were not adequately followed up. When the market moved against his hidden positions, the losses reached about 827 million pounds — twice the bank's trading capital. In February 1995 Barings collapsed and was sold to ING for one pound.

The lesson is that segregation of duties is not optional, and profitability is no excuse to suspend it. No amount of validation or tolerance-checking helps when the same person creates the risk and controls its record. This is the conduct-and-segregation mode, and it is the reason Part 3 treats the wall between creators, verifiers and settlers as a structural, software-enforced control rather than a policy preference.

Case Failure mode Step it maps to Lesson
Citigroup 2022 Input / capture error Validation and tolerance controls A control is only as good as its configuration
JPMorgan 2012 Valuation / representation Independent price verification The record can lie; only independent checks catch it
Barings 1995 Conduct / segregation Segregation of duties Never let one person create and conceal risk

Reading the metrics together, not one at a time

No single metric tells you whether a capture operation is healthy, and reading them in isolation is how operations fool themselves. The metrics only mean something in combination, because each one can be gamed or can hide a problem that another reveals. An illustrative walk through a health picture shows why.

A high straight-through-processing rate paired with a low, fast-clearing break backlog is genuine health. Trades flow through untouched and correct, and the few that break are resolved quickly. This is the picture a well-run, consolidated estate produces, and it is cheap to run because humans handle only true exceptions.

A high straight-through-processing rate paired with a growing, ageing break backlog is a fragile operation borrowing against its future. The trades book fast, but a rising pile of unresolved breaks says they are booking fast and wrong, or that the reconciliation catching the errors is falling behind. This picture looks good on a single-metric dashboard and is one bad day from a headline, which is exactly the trap the Citigroup case illustrates — automation without a matching control is speed pointed at a wall.

A low straight-through-processing rate with a low break rate is expensive but honest. Lots of manual touching, but the touching is catching the errors. This is a costly operation — often a legacy estate with heavy product-coverage gaps — but it is not lying to itself, and its transformation path is clear: automate the manual flow without losing the control the manual flow provides.

Picture STP rate Break backlog Read
Genuine health High Low, clearing fast Consolidated, well-run, cheap
Fragile High Growing, ageing Automating errors; one bad day from a headline
Expensive but honest Low Low Legacy, manual, costly; clear transformation path

The economics of poor capture: what a break actually costs

It is worth putting rough numbers on why any of this matters, because the case for clean capture is ultimately economic. The costs of poor capture are usually described qualitatively — "expensive", "risky" — but they are real line items, and seeing them stacked up explains why banks spend so heavily on this unglamorous layer. The costs fall into a few buckets, and they compound.

The direct cost of a break is the labour to investigate and clear it. A reconciliation break takes a skilled analyst time to trace, resolve and document, and a large fragmented estate produces breaks in volume — the "armies of reconciliation" that are a standing line in a markets operation's budget. Multiply a per-break cost by a daily break count across every boundary in the estate, and the operational cost of fragmentation is a serious number before anything has actually gone wrong.

The tail costs are larger and rarer. A failed settlement carries a direct penalty under settlement-discipline rules and ties up cash and securities. A mis-hedge from a wrong position is a real market loss. A wrong regulatory report is a reportable breach that, if it persists, draws enforcement — and the three real cases in this part show the extreme tail, where control failures cost tens of millions in fines, billions in losses, or the whole institution. The distribution is skewed: most capture errors cost a little, a few cost catastrophically, and the rare catastrophe justifies the whole control estate.

Generative AI: where it genuinely helps, and where it must be governed

Artificial intelligence has real, specific uses in trade capture, and it is worth being precise about them rather than either dismissive or breathless. Four use cases are already delivering value. Machine learning can suggest missing enrichment — a likely settlement instruction, book or term inferred from patterns in past trades — lifting the STP rate on trades that would otherwise stall. Anomaly detection models flag bookings that look unusual — an odd notional, a wrong book, a rare product — before they cause damage, a statistical complement to fixed tolerances. Language models extract structured trade terms from unstructured term sheets, confirmations and emails, cutting manual re-keying at exactly the point where fat-finger errors are born. And break-prediction models forecast and prioritise likely reconciliation breaks so they are resolved faster.

Use of GenAI The gain The governing risk
Auto-enrichment suggestions Higher STP, fewer stalled trades Wrong suggestion booked without a human check
Anomaly detection on bookings Catches odd trades before damage False positives, and gaming; needs explainability
NLP on term sheets and confirms Less manual re-keying, fewer errors Extraction errors on bespoke wording
Break prediction and triage Faster reconciliation Model drift; over-reliance on the prediction

  1. Why is the risk taxonomy described as a map of controllability?

  2. Why can the STP rate alone flatter a fragile operation?

  3. How do the three real cases map to distinct failure modes?

  4. What filter separates real transformation from cosmetic change?

  5. What are the three non-negotiable governance caveats for GenAI in capture?

Dragging a product out of the long tail — an illustrative onboarding

The routine way a bank reduces its risk is to bring a dangerous product into the controlled model. Consider an illustrative case: a desk wants to trade a new structured note — an autocallable, whose payoff depends on an underlying crossing barriers on observation dates. Today it would be booked in a spreadsheet, off-system, with unreliable risk, profit and reporting. Onboarding it means bringing it into the booking model so it captures, enriches, values and reports straight through, and it is a classic business-analyst project that touches every theme in this module.

The onboarding runs through six steps, and it is as much a data project as a pricing one. First, define the product: the analyst works with the desk to specify the payoff, terms, cash flows and lifecycle events precisely. Second, model and template it: extend the taxonomy with a new template or configuration and define the valuation model. Third, set up enrichment and reference data: product terms, identifiers, calendars, settlement instructions and any new reference data. Fourth, map downstream: a source-to-target mapping to risk, profit, settlement and regulatory reporting, confirming every consumer can handle it. Fifth, test and define controls: exercise capture, valuation, lifecycle and reporting end to end, and set validation and tolerances. Sixth, go live and monitor: release with limited volume and watch STP, breaks and reconciliations before scaling.

5

  1. Why can the STP rate alone flatter a fragile operation?

  2. Which case is an input error with a mis-set control?

  3. How is the risk taxonomy ordered?

The six ideas that matter most

If you keep six things from this module, keep these. Each compresses a part of the argument into a sentence you can carry to a desk.

Idea 1 — Capture, not execution, is where risk becomes knowable

The economic risk is real the instant a deal is struck, but the bank can only see, hedge, settle and report it once it is captured as a structured record. That record is the single source of truth for everything downstream. Capture is therefore a control, not a clerical step — the hinge on which the whole front-to-back flow turns.

Idea 2 — Data cascades, so a capture error is never contained

The trade is a layered object — economics, static reference data, and lifecycle state — and every downstream function reads the one record without keeping its own copy of the truth. One wrong field corrupts risk, profit, settlement and reporting at the same time. Deduplication, the internal-external tag, and clean reference data are the unglamorous fields with the largest blast radius.

Idea 3 — Booking is where the trade gets its place, and its place has weight

Booking is two governed decisions: how the trade is represented, and where it lives. The template decides how it is valued and aggregated; the legal entity decides which regulator, capital and jurisdiction apply. Get representation wrong and the numbers are computed wrongly; get location wrong and correct numbers land in the wrong place — which is why booking is rules, not free choice.

Idea 4 — Compliance is core architecture, not an afterthought

Segregation of duties, immutable audit, the identifiers, the messaging standards and operational resilience are not bolted on — they are the shape of the estate. Reporting is a by-product of a correct record, not a separate data-entry task, and the 2024 rewrite pushes the fix to source. The Barings, London Whale and Citigroup failures are each a compliance-architecture failure as much as a trading one.

Idea 5 — Fragmentation is the quiet tax that distorts everything

Most pain traces to one root: too many systems holding overlapping, inconsistent trade and reference data. Fragmentation is what makes risk and finance disagree, makes reporting slow to fix, and makes change ripple unpredictably. The long tail of exotics booked off-model is fragmentation's sharpest edge, and it is where the largest incidents are born.

Idea 6 — Real transformation moves cost or risk — including AI, governed

The durable wins are structural: consolidate the estate, converge on one canonical model, raise the STP rate correctly, master the reference data, and move to real-time cloud architecture. Generative AI genuinely helps — enrichment, anomaly detection, document extraction, break prediction — but only on clean data, under human control, augmenting the discipline rather than replacing its accountability.

Process Cheat Sheet — The Whole Flow on One Page

A single reference for the entire trade capture and booking flow. Keep it at the desk.

The flow, stage by stage

Stage Timing Core activity Key output Lead and systems
Capture Seconds Record executed economics in canonical model Skeleton trade record Front office; capture / e-trading, FIX
Enrichment Seconds Attach ref data by lookup Complete record Rules engine; EDM, reference masters
Validation Seconds Rules, tolerances, referential checks Record cleared to book Capture system
Verify & confirm Minutes–hours Four-eyes; match with counterparty Trusted, matched record Middle office; matching platform
Booking Seconds–minutes Fix representation, location, internal legs Booked, owned record Front / middle office; Murex, Calypso, store
Distribution Continuous Risk, finance, ops, reporting read record Positions, P&L, instructions, reports Downstream systems
Lifecycle Days–years Fixings, coupons, exercises, novations Versioned events Middle office; trade store
Settlement T+1 / T+2 Exchange cash and securities (DVP) Settled obligation Operations; SWIFT, ISO 20022

Key numbers and rules of thumb

Metric Rule of thumb
Volume distribution A few vanilla products are most trades by count; exotics are the risky long tail
Trade count vs deals Internal legs mean far more rows than real external trades
Position identity Position must always equal the sum of its trades
STP rate Higher is better — but only meaningful paired with the break rate
Capture timeliness Near real time, so intraday risk is right
Books of record Three (trade, position, accounting) that must all reconcile
Reporting timing Derivatives and their events reported, both sides, next day
Mandatory identifiers No valid LEI, no report; UTI pairs both sides; UPI names the product
Settlement principle Delivery versus payment — neither side transfers value without the other
Internal trades Must net to zero at group; correct internal-external tag is critical

The whole thing in one sentence

A trade capture and booking flow turns an executed agreement into one canonical, validated, enriched and correctly placed golden record — captured through many channels, defended by layered controls, maintained through its lifecycle, and read by risk, finance, operations and the regulator, none of whom keep any other copy of the truth.