Beyond accuracy: why we measure P&L in units of volatility

"68% hit rate." Sounds good. But an isolated percentage can lie in the worst way: an engine can win often and small, then lose rarely and big. On balance it loses money while showing a nice accuracy. That's the classic trading trap, and we refused to fall into it without measuring it.

The real judge: R in σ

A 2% move on gold and a 2% move on the VIX don't mean the same thing: the VIX naturally moves far more. To compare bets across different assets, we scale each gain or loss to the asset's own volatility over the signal's horizon. We call it R in σ: how many "typical moves" of the asset did the bet capture, in the right direction?

An R of +1σ is a normal move of the asset in our favour. An R of −1σ is the same move against us. The honest question isn't "how often are we right?" but "when we're right, do we win as much as we lose when we're wrong?".

What we found

The analysis was pre-registered: the definitions were written before seeing the numbers, and we publish the result whatever it is (that's the rule of our public protocol).

First result: the payoff is nearly symmetric. When a bet wins, it returns on average +0.95σ; when it loses, it costs −1.03 to −1.06σ. The working hypothesis holds: our expectation per bet is consistent with our hit rate.

Second, more unexpected result: we compared several confidence floors (emitting from 57, 58 or 60). In pure accuracy, raising the floor to 60 looks better (72.7% vs 68%). But in R per σ, the advantage disappears entirely: the highest-confidence bets lose bigger when they lose (−1.26σ instead of −1.06σ). Once risk is accounted for, all three floors are equivalent — the confidence intervals all overlap.

Cumulatively over the year, it's even the lowest floor (57) that dominates: it captures more total move because it emits more decisions, without degrading per-bet quality. The practical conclusion: we keep the emission floor at 57, and the TRADE badge (reserved for an even more selective subset) remains risk management, not a claimed superior edge.

Why we publish this

Because a figure without its risk is marketing. From now on, next to every hit rate on the Truth Table, you'll find the σ-normalised P&L per bet, with its confidence interval and the number of observations. Where the high floor doesn't dominate, it's written — even when it contradicted our starting intuition.

The full technical report, with the tables, backtest runs and calculation method, is public and verifiable here.