Cuts

A cut fixes a body of data at one timepoint. The purpose of a cut in this repository is single: to provide a valuation that is fully reproducible. The body therefore holds everything the valuation reads, which includes the deals, the market data, the reference data, and the versions of the engine and of the code that computed the result. The kinds of cut name the occasion a cut is taken for, and the components record the two halves of the body.

A cut fixes a body of data at one timepoint. Three parts follow, and each answers a separate question.

• The label names the cut. A code, a reference date or a version serves as the handle that a report, a deal or a book carries.
• The timepoint fixes the cut. It is the instant held still, with its timezone.
• The object is the body of data the cut fixes. The rest of this note calls it the body. It holds the deals, the market data, the reference data, the engine and code versions, and the taking identity. The cut records the acts that took it, and the authorisation where its kind requires one, beside the body.

The label answers which cut is meant. The timepoint answers when it holds. The object answers what it holds. A statement that omits one of the three leaves the cut unidentified, and a claim about a cut carries the three together.

The sentence stating the definition constrains two things. Something is fixed, and one instant fixes it. Every further property of a cut is a fact about the kind: how often the instant recurs, which location owns it, and which process takes it. Those facts qualify a cut without entering the definition, and Kinds of cut works out what follows from that.

A kind names the occasion a cut is taken for. The purpose is single and fixes the body; the kind fixes the instant, the recurrence and the centre, and the use the result is put to. The classification, and the argument for it, is on Kinds of cut.

One kind takes its instant from the consumer rather than from a centre, a person or a process: a live cut is fixed at the instant the run that consumes it starts, and the run mints it. It is the kind a report selects when it says value this scope at whatever time you start, and it is why a run may bring a cut into being rather than only choose one.

Within ORE Studio the purpose of a cut is single: to provide a valuation that is fully reproducible. An official report, a verification of prices, and a view for a desk are uses a cut is put to, and every one of them is served by a valuation only if that valuation can be reproduced. The single purpose contains the others, which is why the body has one size and not several.

The size of the body follows from that purpose and from nothing else. The argument has three steps, and a public statement supports each of them.

The first step fixes what reproducibility is. A computation is reproduced when the same software runs on the same input data and returns the same result:

running the same software on the same input data and obtaining the same results

The sentence is Claerbout's definition, quoted by Plesser (Plesser, Reproducibility vs. Replicability). It places the software inside the condition rather than beside it.

The second step establishes that the software and its settings are inputs. A change to a program can change its output, and the change need not be large:

Even the slightest change to a computer program can have large intended or unintended consequences.

That sentence is Sandve et al.'s fourth rule (Sandve et al., Ten Simple Rules for Reproducible Computational Research). The same paper states what must therefore be recorded:

If the step is performed by a computer program, the critical details include the name and version of the program, as well as the exact parameters and inputs that were used.

and, for the version specifically:

In order to exactly reproduce a given result, it may be necessary to use programs in the exact versions used originally.

The third step establishes that the environment around the code enters as well. The Turing Way states the requirement in one sentence:

In order for research to be reproducible, the computational environment that it was conducted in must be captured in such a way that others can replicate it.

and gives the mechanism with an example in which the same script returns two answers:

Therefore this simple script returns different answers depending on the computational environment in which it is run.

The philosophical literature draws the same boundary around computation. The Stanford Encyclopedia of Philosophy describes computational reproducibility as

reproducing particular analysis outcomes from the same data set using the same code and software

adding that this is the most common form of it. Plesser, whose paper is the source of the Claerbout sentence, states the practical consequence in his own words:

I consider it to be essential to preserve artifacts such as software, scripts, and input data underlying computational science publications.

The three steps share one conclusion. Data, code, settings and environment belong to one class: each is an input whose variation changes the output. A cut holds the valuation's inputs, so it holds that class in full. The body therefore holds four inputs.

• The deals. The population the valuation values.
• The market data. Spots, curves, forward points, volatility surfaces, fixings, and the model parameters that are driven by the market.
• The reference data. The conventions, the calendars and holiday sets, the valuation and pricing settings, the curve configurations, and anything else the run reads to turn a deal and a quote into a number.
• The engine and the code. The version of the risk engine and the revision of the code that computed the result.
• The taking identity and the authorisation. Not inputs to the valuation: the cut records who took it, and, where its kind requires one, who authorised it and when, as the acts that made the body official. A kind whose occasion carries no signature records the taking act alone, and Kinds of cut states which occasions those are.

The list admits a test. Take a result the cut claims to support and ask which changes would make the number differ. Each of the four inputs answers yes. A field that answers no — the taking identity, the authorisation — is provenance about the run, and provenance is worth keeping, but the body is the set of things whose variation moves the number.

A cut that fixed the data and omitted the engine version would hold a body from which its own result does not follow. The omission survives casual review because the halves read as one variable and one constant: the market moves, and the software waits. The software waits only until the next deployment.

The component pages below take the two data halves of this body. The settings, the engine and the code have pages of their own, and this cluster holds them because the cut records them.

Within the body, the deals and the market data are the two halves that a process fixes by separate acts. The remaining parts, the reference data, the settings, the engine and the code, are recorded with the cut and are not the subject of a daily review, and the component pages below take only the two reviewed halves. The division between those halves has consequences for the process.

The components differ in owner, in review and in mode of failure. A deal is reviewed by reading it and comparing it to what the desk booked, and it fails by being absent or wrong. A rate is reviewed by comparing it to an external source, and it fails by being stale or unapproved. A corrected deal moves no rate, and an approved rate moves no deal. One act that fixed both would tie every trade correction to a fresh approval of the day's rates, and would therefore slow the correction the review exists to serve.

Our working convention follows from that. The two components are fixed by separate acts, the deals first and the market data second. The deal cut is retaken until the population is clean, because a person may still correct a deal in production. The market data cut follows, since a provisional market data set is needed to value the deals and confirm they are right before the day's rates are committed.

The classification comes first. The purpose fixes the body, and the kind contributes the occasion, which is the recurrence, the centre and the instant. The component pages then take the two halves of the body, and the last page takes the definition into the expiry sense of the word.

The purpose of a cut, and the size of the body that follows from it, rest on the reproducible-research literature quoted in The purpose and the body: Claerbout's definition as reported by Plesser, Sandve et al.'s rules on programs, versions and parameters, the Turing Way's account of the computational environment, and the Stanford Encyclopedia's definition of computational reproducibility. Each quotation is reproduced exactly and carries its URL.

Three further claims in the cluster have a public source, and each is recorded on the page that makes it. The admissible sources of an independent price verification rest on a prudential standard (EBA RTS (EU) 2016/101, Article 3), and the stated open and close of a centre's business day rests on one settlement system's published hours (Fedwire Funds Service); both appear in Kinds of cut. That an official rate is published once per working day at a stated time rests on the central bank reference rates (ECB euro foreign exchange reference rates), which appears in Market data cut.

Everything else is a definition of a term or our working convention. The order of the two component acts, the existence of the flash and start of day kinds, the numbered end of day sequence, one cut per book, the after-hours rule, the arbitrary instant a process may name, and the times that cut codes name have no public source we could verify. They are written here as the convention we work to. Practice at a particular firm is a matter for that firm's own record.

Term	Definition
Cut	A body of data fixed at one timepoint
Label	The name of a cut: a code, a reference date, a version
Timepoint	The instant a cut fixes, with its timezone
Object	The body of data a cut fixes
Kind	The occasion a cut is taken for, which fixes its instant, recurrence and centre
Authorisation	The sign-off a kind requires, which makes a cut official and immutable; a kind whose occasion needs none records the taking act alone
Trade cut	The component of a cut that fixes the deal population
Market data cut	The component of a cut that fixes the market data
Cut code	A label for the time of day at which an option expires
Cut time	The time of day a cut code names, with its timezone

The pricing settings themselves, the engine's configuration documents, and the reads served over stored market data belong to other clusters. This cluster states what a cut must hold and why, and leaves the meaning of each setting to the page that owns it.

• Knowledge — the index this structure note hangs from.
• Pricing Configuration — states the reproducibility contract from the settings side, which this cluster holds in the body and leaves the meaning of to that page.
• A report binds to a published curve, not to a name — the binding that names a published set and leaves the deal population unbound, which is the dependency this cluster answers.
• Model market data cuts as proper ORE Studio reference data — the story that designs the cut time as reference data.
• Book — the entity a cut would be assigned to.
• Book groups and rates centres — why every book that revalues together shares one centre's market data.
• Business Centre — the coded trading location a book carries in place of a cut.
• Zettelkasten — what a structure note is and why this order is authored rather than computed.
• Naming the cut — why the general concept keeps the word cut, what each component is called, and what a change of term would cost.
• The cut is the body a run prices from — the facts a cut records, the states those facts derive, the permissions and the authorisation, the derived label, and the second binding that lets a run name a body.
• Cuts journeys — taking a cut, pricing a report from one, and rerunning and auditing one, as the person works them.
• The cut prototype — the list, the record, the take flow, the picker, the run and the named cut times, drawn before they are built.

1. Commission Delegated Regulation (EU) 2016/101 on prudent valuation, Chapter I, Article 3. Verified: the independent price verification process is named, and consensus service data, tradable broker quotes, exchange prices and counterparty collateral valuations are accepted as sources.
2. European Central Bank, euro foreign exchange reference rates. Verified: one reference rate per pair is published around 16:00 CET each working day, after a concertation procedure among central banks.
3. Board of Governors of the Federal Reserve System, Fedwire Funds Services. Verified: the business day starts at 21:00 ET on the preceding calendar day, ends at 19:00 ET, and the deadline for third-party transfers is 18:45 ET.
4. Hans E. Plesser, Reproducibility vs. Replicability: A Brief History of a Confused Terminology, Frontiers in Neuroinformatics 11:76 (2018). Verified: it quotes Claerbout's definition of reproducing as "running the same software on the same input data and obtaining the same results", and states that preserving the software, the scripts and the input data is essential.
5. Geir Kjetil Sandve, Anton Nekrutenko, James Taylor and Eivind Hovig, Ten Simple Rules for Reproducible Computational Research, PLOS Computational Biology 9(10): e1003285 (2013). Verified: rule 1 requires "the name and version of the program, as well as the exact parameters and inputs that were used"; rule 3 requires the exact versions of the programs used; rule 4 states that the slightest change to a program can have large consequences.
6. Stanford Encyclopedia of Philosophy, Reproducibility of Scientific Results. Verified: computational reproducibility is described as "reproducing particular analysis outcomes from the same data set using the same code and software", and as the most common form of reproducibility in computational disciplines.
7. The Turing Way, Reproducible Environments. Verified: the computational environment must be captured for research to be reproducible, the environment comprises the operating system, the software and its versions and configurations, and the same script returns different answers in different environments.