Research
What we tested on live perpetual-futures data, what survived honest validation, and — mostly — what didn't. Every figure comes with what it was measured against and how many independent days stand behind it.
How we measure
Crypto perpetuals move together, so the unit of independent observation is the day, not the row: a hundred coins on the same red Tuesday are close to one observation. Every interval we publish is therefore bootstrapped by resampling days, and hit rates are compared against the same day's cross-sectional baseline — never against 50%, which on a one-sided day flatters a strategy that has no skill at all.
Model results are walk-forward out-of-sample with a purge, so no model is ever scored on days it learned from. Before reporting an improvement we run a noise control: the same inputs with randomly flipped signs. Anything smaller than what pure noise produces is not a small win — it is unmeasurable, and we say so.
Where a result contradicts what we ship, the result wins and the product changes. Several of these write-ups exist because a measurement killed something we had already built.