The clock nobody hears ticking
A risk modeller builds a probability-of-default model on ten years of bureau history. The model works. Then a benchmark rate is reformed, a regulator issues new guidance on what counts as forbearance, or a bureau changes how it flags a missed payment during a relief scheme, and the model keeps working exactly as before — producing the same confident numbers on a relationship that no longer holds. Nothing crashed. Nothing threw an error. The string "arrears" still parses. What broke is the index between that string and the event it used to mean.
This is terminology churn, and credit risk is an unusually good field for watching it happen, because the vocabulary is regulatory, contractual and market-driven all at once, and each of those three sources rewrites terms on its own schedule, out of sync with the other two.
Two positions worth taking seriously
Set two defensible claims against each other and see which survives contact with the sector's own history.
Position one: a lookup table is enough
The strong pragmatic position says the problem is smaller than it looks. Attach a current terminology service to any model — the latest ISDA benchmark fallback definitions, the current IFRS 9 application guidance, this quarter's bureau data dictionary — and a model trained years ago reads today's contracts correctly. You do not need a system that watches every stream continuously; you need a refreshed reference table and a model that queries it before it answers. This is retrieval, and retrieval is cheap.
Nomenclature drift in credit is documented and centralised. The transition from LIBOR to SONIA and SOFR was published, dated and phased over years with an official fallback protocol. Any modeller who missed it wasn't failed by their tools; they weren't reading the register that existed for exactly this purpose.
Position two: the drift that keeps its spelling
The second position says the lookup table only catches the churn that got formalised. Most of the damage in credit risk comes from terms that never changed their spelling at all. "Significant increase in credit risk" is the operative phrase in IFRS 9, adopted in 2018, and it still reads today exactly as it read then. What changed, repeatedly and without any renaming event, was what triggered it. When the UK's payment holiday scheme launched in 2020, roughly a million and a half mortgage accounts entered forbearance in the first weeks. Some banks initially treated enrolment itself as evidence of significant increase in credit risk and moved those accounts to Stage 2 provisioning. Regulators then issued guidance — the EBA's statement, the Bank of England's supervisory steer — saying that participation in a general relief scheme should not, on its own, be read that way. The referent of "significant increase in credit risk" moved twice inside a single year, and no lookup table has an entry for "the current common-sense reading of a phrase everyone already knows."
Where credit risk actually breaks
The failure mode named for this domain is precise: a portfolio scored on a relationship that broke with the last rate move. Trace how that happens.
A benchmark reform changes what a floating rate resets against. The string "the reference rate" survives the transition from LIBOR to SONIA intact in every covenant that used it as shorthand, but the compounding convention underneath is different — SONIA is typically applied in arrears with a compounding period, where LIBOR was set in advance. A model that learned the historical correlation between rate resets and delinquency spikes learned it against LIBOR's timing. After the 2021–2023 transition, the same phrase in the same covenant clause now describes a different cash-flow event. The model does not know its own premise moved. It keeps scoring the relationship as if the clock still ticked the old way.
Bureau behaviour compounds this. During relief schemes, some bureaus mark an account "up to date" if a missed instalment was formally deferred, rather than flagging it as missed. The field name — "current," "arrears," "days past due" — is unchanged. Its meaning, for that cohort, is not. A model trained to treat "current" as evidence against default risk absorbs a cohort where "current" was a policy label, not a payment fact, and the correlation it learns for that period is contaminated in a way no error message will surface.
Sector reclassification adds a third front. Rating agencies and index providers revise industry taxonomies periodically — GICS sub-industry definitions were revised in 2018 and again in 2023 — and a borrower's sector code can migrate between revisions without its business changing at all. A model that reads sector code as a stable feature for concentration risk is reading noise dressed as signal every time the taxonomy underneath it moves.
The objection that doesn't fully die
The retrieval position is not wrong; it is under-scoped. It is exactly right for the curated cases: benchmark fallback rates, IFRS 9 formal amendments, official bureau schema releases. Where a body has published a register and dated it, attaching that register is cheaper and more auditable than any continuous-intake architecture. Building a system that re-derives the SONIA compounding convention from raw market chatter, when ISDA already published the fallback protocol with an effective date, would be wasted engineering.
The gap is everything that changes meaning without changing spelling, and in credit risk that is most of the damage. Nobody issues a bulletin announcing that "significant increase in credit risk" will be read more strictly this quarter. Nobody publishes a register entry for the moment a bureau's internal treatment of a payment holiday shifts. Those changes surface as usage — in supervisory statements, in servicer guidance, in the loan-level detail of how a handful of large lenders actually apply the phrase — and detecting that shift requires watching the stream of usage over time, not looking anything up. A lookup table answers "what is the current definition." It cannot answer "has the working definition drifted since I last checked," because nobody wrote the drift down as a definition change. It has to be inferred from a moving pattern of application, which is continuous intake by another name, whether or not anyone builds it that way.
There is a second objection worth taking seriously, and it cuts the other way.
A model that ingests every servicer memo, every supervisory speech, every forum post using "cure rate" loosely will absorb noise faster than signal. Reproducibility suffers: a portfolio scored today and rescored next month on the same inputs gives different answers, because the model's working vocabulary moved underneath it. A frozen model is at least auditable. You know exactly what it believed "arrears" meant, because it never changed its mind without your permission.
This is the strongest challenge, and it lands. Unfiltered recency is a real failure mode, not a hypothetical one. A model that treats every unofficial usage as equally authoritative will chase transient market slang and abandon it faster than a modelling team can document the change, and audit trails genuinely suffer when the ground moves without a paper record. The honest answer is not to freeze the vocabulary back down to protect reproducibility. It is to date every term-to-referent binding and keep the provenance attached to it — this model treated "significant increase in credit risk" as triggered-by-forbearance-enrolment as of March 2020, and as not-automatically-triggered as of the EBA statement in June 2020, and here is the citation for each. That is a record you can pin for an audit and still update for use. A frozen corpus cannot do either half of that; it has one belief, undated, indistinguishable from certainty.
What narrows, not what falls
| reads | cannot see | |
|---|---|---|
| Large Language Model | the corpus at cutoff | any regulatory guidance, rate reform or bureau policy change issued after it |
| Large World Model | the current loan book, the current transaction | which current phrase the servicer or regulator now means by it |
| Large Universe Model | the ongoing stream of guidance, servicer practice and rate-reform documentation, dated and sourced | nothing structurally — only coverage, latency and trust remain as open questions |
The claim survives, narrowed. Terminology churn in credit risk is not a general argument that language changes and therefore all trained knowledge is worthless; the vast majority of the sector's vocabulary — "default," "collateral," "amortisation" — has held its meaning for decades and a good analyst infers the rest from context exactly as a human specialist does. The specific claim is that churn concentrates in the phrases regulation and benchmark reform touch most often, its failures are silent because the string never breaks, and its rate is set by bodies external to any model — the EBA, ISDA, the bureaus themselves — not by how carefully the model was trained. A frozen lexicon cannot detect that it has been overtaken by its own subject matter. Retrieval fixes the part of that problem which was ever going to get written down. The rest requires watching the stream, dating what you believe about it, and being willing to say when you last checked.