Backtesting Methodology
Before any strategy goes live it must survive a rigorous out-of-sample validation process. We believe in publishing methodology so you can hold us accountable.
Parameters are never optimised on the full history. We use a rolling walk-forward procedure: calibrate on an in-sample window, then evaluate on the immediately following out-of-sample window. The out-of-sample windows are stitched together to produce the published performance curve. This mimics how a live system encounters unseen data.
The in-sample period is used solely for parameter search (lookback, dispersion threshold, gate thresholds). The out-of-sample period is locked before the search begins. We publish only the OOS results. In-sample numbers are shown for transparency but are never the headline metric.
In-sample Sharpe ratios are almost always inflated by curve-fitting. We committed early to publishing only out-of-sample figures so that live performance comparisons are meaningful. If the strategy works, the OOS Sharpe should be close to the IS Sharpe - a meaningful divergence flags over-fitting.
All backtests assume: (1) entry at the open of the bar following the signal, (2) Bybit taker fees of 0.055 % per side, (3) funding settled every 8 h at the prevailing rate, (4) 3× slippage on the 24 h volume to model market impact. No look-ahead bias - only data available at signal time is used.
Backtests are reported under three BTC market regimes (Risk On / Cautious / Risk Off) and also unconditionally. A strategy must show positive OOS Calmar ratio in at least two of the three regimes before deployment. This prevents strategies that only work in one perpetually bullish environment.
Every deployed strategy is re-evaluated at the end of each month. If the rolling 30-day OOS Calmar ratio falls below the pre-specified retire threshold, the strategy is suspended pending investigation. This keeps the signal catalogue fresh and honest.