Near-high signals in crypto futures

A pre-registered search for simple signals that rank crypto perpetual futures against each other, and a strategy based on the ones that held up.

The data is about three years of hourly prices from 55 perpetual futures markets on Binance, from August 2023 to July 2026. The question was narrow: do simple, rule-based measures say anything about which of these coins will do better than the others over the next day or two?

Backtests charged 8 basis points per side on actual turnover, plus funding payments, with no leverage.

Fourteen candidate signals were written down first, each with a formula and a pass/fail rule, before any testing. Four passed in-sample. The two strongest measure how close a coin trades to its 20-day high and to its running high, the highest price since the data begins. The other two were a rule for trading breakouts after unusually quiet periods and a filter that drops the least liquid fifth of coins. Plain 20-day momentum failed. That contrast was the main finding: where a coin sits relative to its highs carried information that its recent return did not.

A strategy used all four. It held an equal-weight basket of the top-ranked coins only while bitcoin was above its 20-day average; the breakout trades ran either way. It passed its one-time out-of-sample test, as measured before the September fixes to the backtest engine described in Auditing my own backtests, and then lost money the following month.

An attribution of that loss showed the breakout rule had picked the right direction only 49% of the time on held-out data. I cut it and registered the rest as a new strategy. Because the out-of-sample period had already been looked at, the only clean test left for the new version is paper trading, which started in August 2026.

That version and a dollar-neutral variant now run as paper trades. Neither has passed my validation checks yet: one does not have enough data behind it, and the other’s results move too much when its settings are changed slightly.

A separate low-volatility signal did rank coins: a later retest of low-volatility signals on non-overlapping windows still found a positive rank correlation with returns over the following days. But it could not be turned into a profitable long/short portfolio: one that included it lost money to costs, funding payments, and sharp rallies against its short positions.