The three properties of a working model

Why the return number comes last

It is tempting to judge a model by its return number, but the return is the last thing to look at, not the first. A spectacular figure tells you nothing if the rule behind it cannot be run twice the same way, if the losers were quietly removed before the number was printed, or if the call would have come out differently on a nervous afternoon. The three properties below are the prior questions — the ones that decide whether a return number is evidence at all. They are not a wish-list of nice-to-haves; they are load-bearing, and they hold each other up. An untestable model cannot be measured, an unmeasured model cannot be trusted, and an unrepeatable model cannot be tested in the first place.

Read them as a short sequence. Testability gives a model a record to be judged by. Measurability insists that record be counted in full, losers and drawdown included, so it tells the truth rather than the flattering half. Repeatability is the quiet property underneath both: a record only means something if the same inputs reliably produced the same call. Each lesson below takes one property apart and shows how a set of measured, A-to-D graded mean-reversion models satisfies it — not as proof of profit, but as proof that the question of whether the approach works can actually be answered.

One warning before you start: a model can satisfy all three properties and still be a losing model. None of this is a promise of profit. A perfectly testable, fully measured, rigorously repeatable rule can simply have no edge — in which case the same three properties will tell you so, clearly and early, instead of letting you discover it with real money. That is the point. The properties do not make a model good; they make a model honest, so that its quality — good or bad — becomes visible. A model that fails them is not necessarily worse at trading; it is worse at letting you find out, which on a long enough timeline is the more expensive flaw.