In use

Predicting risk, not direction

With the same data and the same method, price direction can't be predicted and volatility can — but turning that into an exposure rule didn't work.

The question: If the project can predict when turbulence is coming, why not cut position then?

What was found

The volatility-regime model is right in 13 of 13 independent yearly validations, across all four horizons. It is by a distance the most robust result in the whole project. The natural consequence seems obvious: if we know when a storm is coming, reef the sail. It was tested, with two allocator designs declared BEFORE running them, and it didn't work — neither in exploration nor in the holdout. The reason isn't statistical but economic, and it's interesting in itself: in Bitcoin, the periods of highest drawdown risk COINCIDE with those of highest expected return. Cutting exposure in them cuts the upside too. The model has informational value — a warning — but not allocation value.

Try it yourself

AUC is the metric this model is judged by, and it has a very concrete literal reading: pick a stormy day at random and a calm day at random; AUC is the probability the model scores the stormy one higher.

Out of every 100 (stormy day, calm day) pairs, it gets the order right in…

Model AUC0.737

pairs out of 100

How far above pure chance:

Try 0.576: it's the low end of the drawdown model's confidence interval. A "significant" result whose interval reaches there licenses no talk of certainty.

Illustrative example numbers for practice — not real data.

How it was tested, step by step

The model is trained only on data before each evaluated slice, with a 30-day purge between train and test: the target looks 30 days ahead, so without a purge the last 30 training rows leak into the test.
TWO allocator designs are declared before seeing them run: a linear one (exposure falls continuously as the probability rises) and an asymmetric one (it only cuts in the top risk decile).
They're compared against a MATCHED-AVERAGE-EXPOSURE baseline, not against always being 100% invested. Comparing an allocator that is sometimes out of the market against one that is always in measures beta, not skill.

Allocator	Window	Sharpe	Matched base	p
Linear	Exploration	0.822	0.816	0.720
Asymmetric	Exploration	0.648	0.816	0.569
Linear	Holdout	0.968	0.928	0.762
Asymmetric	Holdout	0.599	0.928	0.720

None comes close to significance, and the asymmetric one is clearly worse than doing nothing. The model knows something; the portfolio doesn't know what to do with it.

What this does NOT say. A null result on the allocator doesn't invalidate the model; it invalidates one use of it. And the reverse must be said: the "probability of a drawdown over 20% within 30 days" target DOES pass its permutation test (AUC 0.655 in exploration, p=0.015; 0.737 in the holdout, p=0.0495), but it rests on 68 events and its 90% confidence interval reaches down to 0.576 — the low end of the interval nearly grazes chance. It's positive and it's clean, and it's marginal. Those aren't the same thing.

Tracking

This finding does have a real number that gets published and kept updated.

Data last updated on 2026-09-09.

← All research