Read these only after you have written your own three sentences. Two of the paragraphs are good, one is not. The question is not which seat is right — the lesson does not settle that, and the unit returns to it at the end — but what a paragraph has to do to count as a verdict at all. Hume’s seat is not modelled here; if you took it, the checklist at the bottom applies to you unchanged, and item 3 asks you for the forecasting method you would call rational.

Model A — the practitioner’s seat

The modellers were rational, in the only sense of “rational” a bank can run on: the best-available-methods standard, not the philosopher’s demand for an inference immune to Hume. By that standard, their inference was an enumerative induction from twenty years of price moves, and the failure was local — the auxiliary assumption that asset classes would not collapse together was never stress-tested, though the data to imagine it existed. The price I pay is that my prescription has to bite in 2007, not in hindsight: I must say which correlated-collapse scenario a competent stress test would have run that year, and why the model that ran it would have refused the ABN AMRO purchase; if I cannot name it, “better stress-testing” is a wish, not a verdict.

What it does: names the seat; fixes one sense of “rational” and holds it; locates the reason at a specific inference and a specific auxiliary premise, not at the collapse; names the price and admits it is a debt.

Model B — Taleb’s seat

The modellers were playing the wrong game, and “rational” here has to be judged relative to the domain: in a fat-tailed market the long run is decided by exactly the events the twenty-year window excludes, so every confirmed day did not make the model safer, it built the position the tail would destroy. The reason is not that VaR was unfalsifiable — it was falsifiable — but that falsifiability inside the window says nothing about what lies outside it, and the model licensed a €71 billion purchase on the strength of a sample that could not, in principle, contain its own refutation. The price I pay is Monday morning: the bank still has to size its positions, and my answer is that it caps exposure by what it can survive losing, not by what the distribution says it will lose — a rule that uses no forecast of the tail at all — which is why this is not Hume’s scepticism in a trader’s jacket but a different question put to the same data.

What it does: the same three moves, on another seat. Notice that both good paragraphs agree on the inference (induction from inside the window) and disagree on what standard of rationality applies to it. That is where the case lives.

Model C — a paragraph that does not count

The modellers were obviously irrational, because as Hume showed no amount of past data can guarantee the future, and 2008 proved it — the bank lost £45.5 billion of public money. On the other hand they used the best methods available at the time, so you cannot really blame them. In the end the whole idea of modelling risk is flawed and banks should stop pretending they can predict the market.

What goes wrong:

  • The sense of “rational” switches mid-paragraph: “irrational” by the philosopher’s standard in the first sentence, “cannot blame them” by the practitioner’s in the second. That is not nuance; it is two verdicts wearing one coat.
  • The reason is an outcome — “2008 proved it”, the £45.5 billion — not an inference. A model can be irrational and lucky, or rational and ruined; the collapse by itself settles nothing about the reasoning that preceded it.
  • “Hume showed” is used as if Hume had shown the models unreliable. He showed induction cannot be justified without circularity — which is true of the models and equally true of every method the writer would have used instead.
  • No price. “Banks should stop pretending” is Taleb’s seat without its Monday morning: it does not say what the bank does instead, so it costs the writer nothing — and that is exactly the exit the lesson warns is not a seat.

The checklist

A paragraph counts as a verdict when it:

  1. names a seat;
  2. fixes one sense of “rational” — the philosopher’s, the practitioner’s, or the domain-relative one — and holds it to the last word;
  3. gives a reason that names an inference (the induction from inside the window; a rescue with or without a new prediction), not the outcome, and names the price the seat pays — Hume’s method, the practitioner’s 2007 fix, or Taleb’s Monday morning;
  4. answers, or at least states, the strongest objection from another seat.

Model C fails 2, 3 and 4. Go back to your own paragraph and mark it 1–4.