Read Achilles markRead Achilles

Risk Management

The Models That Fail Quietly

In more than twenty years of bank risk management, I have never been hurt by a model that failed loudly. Loud failures are merciful: the number is obviously wrong, everyone stops, there is a post-mortem, someone is…

♥ 0 Sign in to like 1 comment

This post’s retention & record terms are stated in full in the colophon below.

In more than twenty years of bank risk management, I have never been hurt by a model that failed loudly. Loud failures are merciful: the number is obviously wrong, everyone stops, there is a post-mortem, someone is reassigned. The models that did the damage failed quietly — returning smooth, plausible numbers, month after month, while the world underneath them had already changed. [author assertion]

A quiet failure has a particular cruelty: it trains trust. A model that is slightly wrong every day does not trigger scrutiny; it accumulates a track record. Committees learn to wave its output through. New staff learn that the number is simply what the number is. By the time the model is badly wrong, the institution has organised itself around believing it — limits, bonuses, capital plans, all downstream of a figure nobody has genuinely questioned in years. [author assertion]

Validation, as usually practised, does not catch this. Validators test a model against history, and history is precisely the wrong witness: the risk lives in the regime change that the sample does not contain. A model can pass every back-test ever devised and still be a machine for being surprised, because back-tests measure how well it describes the world that produced the data — the one world we know it will not meet again. [author assertion]

Correlation is the classic quiet failure, and I have watched it from inside. Portfolios that are diversified on paper — different names, different sectors, different regions — concentrate, in the one state of the world where diversification is actually needed, into a single exposure. The spreadsheet says the risks are independent. The spreadsheet was calibrated in fair weather, and independence is fair weather's most generous assumption. [author assertion]

Underneath all of this is a governance fact that risk managers are reluctant to say aloud: a model is an opinion with an org chart. Challenging the model means challenging the people whose budgets, headcount and status the model justifies. So the challenges that do occur are technical — a parameter here, a window there — because technical challenges are safe. The challenge that matters, whether the whole framing still fits the world, has no natural owner. [author assertion]

I no longer ask whether a model is right. I have never once seen that question answered in a way that survived contact with the following year. I ask instead: what would we see, and how soon would we see it, if this model stopped being right? A risk function should be built around tripwires rather than verdicts — observable signs, agreed in advance, that the quiet failure may have begun. [author assertion]

Model risk, in the end, is not a modelling problem. It is an attention problem. The number on the page is the least dangerous part of any model; the exposure is the confidence arranged around it. Institutions that understand this hold their numbers the way sailors hold a forecast — gratefully, and with a hand on the rail. [author assertion]

Julia Zhao

All posts by Julia Zhao → · Follow Julia Zhao by RSS

The Board 20 essays 0 corrections Agent citations: —— — citation attached ≠ claim validated As of 16 November 2026 Full board →

Provenance

About this version

How this post was verified, labelled and shared — in plain language first. The complete machine-readable record is kept below, word for word, for anyone who wants to check it.

Version
1 · in good standing
Author
Julia Zhao · verified at publication
Published
6 Nov 2026
Claims
7 labelled — 7 author assertion
Citations
No external citations
Distribution
Shared with the author’s consent — including to agents acting for readers, under the author’s per-post opt-in
Content fingerprint
3ffe2d56…932dfc
Full verification record

The verbatim graded record of this version: identifiers, claim classes, citation targets, hashes and standing, exactly as issued.

This record is also published in structured text for agents and tools — the same fields, word for word. View structured text

House editorial — Read AchillesAuthor verification: verifiedStanding: governingHash: 3ffe2d569fc4d43619547352da45d361755f9162e5dc63c80aea3fb291932dfc

Version & standing

Post post-the-models-that-fail-quietly-df7a63 · version 1 (post-the-models-that-fail-quietly-df7a63-v1)

@julia-zhao · verified state at publish: verified

Published: 2026-11-06T23:43:31.244Z · Standing: governing

Claims & citations

  • c-1: author_assertion
  • c-2: author_assertion
  • c-3: author_assertion
  • c-4: author_assertion
  • c-5: author_assertion
  • c-6: author_assertion
  • c-7: author_assertion

    Disclosure & trust

    When your words reach the end of their retention period they are destroyed, on the schedule above. What the platform keeps after that — indefinitely, with no expiry — is a skeleton of acts: that an act occurred, its class and time, the bundle and section it touched, its change class, a hash that identifies the record without containing it, and the keys that link a correction to what it corrects. The skeleton is not your words and cannot reproduce them. Your elections to shorten a period, to ask ephemerally, or to skeletonise now govern your words; they do not shorten, and cannot remove, the skeleton — it is the record that lets the platform prove what it did with your work after the words are gone. The platform will never produce your skeleton by choice against you; like any record it can be reached by a court order the platform cannot refuse, and if that happens you will be told wherever the law allows and the production will state, on its face, that it is a skeleton and that the words expired on this schedule.

    When your words reach the end of their retention period they are destroyed, on the schedule above. What the platform keeps after that — indefinitely, with no expiry — is a skeleton of acts: that an act occurred, its class and time, the bundle and section it touched, its change class, a hash that identifies the record without containing it, and the keys that link a correction to what it corrects. The skeleton is not your words and cannot reproduce them. Your elections to shorten a period, to ask ephemerally, or to skeletonise now govern your words; they do not shorten, and cannot remove, the skeleton — it is the record that lets the platform prove what it did with your work after the words are gone. The platform will never produce your skeleton by choice against you; like any record it can be reached by a court order the platform cannot refuse, and if that happens you will be told wherever the law allows and the production will state, on its face, that it is a skeleton and that the words expired on this schedule.

    Integrity & distribution

    Content hash: 3ffe2d569fc4d43619547352da45d361755f9162e5dc63c80aea3fb291932dfc

    Distribution: opted_in

    Canonical presentation: https://blog.readachilles.com/post-the-models-that-fail-quietly-df7a63/

    What we retain, and why — the retention disclosure

    The retention disclosure for this version is the disclosure line of record, stated in full in the colophon fields above.