Single backtests lie
Pick any 12-month window and any indicator and you can build a 100% backtest. The signal-to-noise ratio in crypto is low and the parameter space is huge. Curve-fitting happens by accident, not by malice.
That's why we don't trust any sleeve until it survives walk-forward — and even then, we publish every revert.
How Phase 37 actually works
- Roll a 180-day train → 30-day test window across 5 years of history (~77 windows).
- At the start of each window, re-rank the entire strategy catalog by train-period Sharpe.
- Pick the top-5 sleeves, with greedy decorrelation (max pairwise corr 0.7).
- Allocate Sharpe-weighted, vol-target 1.6% daily, James-Stein shrinkage λ=0.25.
- Honour the per-class quota: trend 90%, carry 10%.
- Concatenate only the out-of-sample test periods into the curve.
Result on the May 25, 2026 snapshot: +104.58% last-12m at 10 bps, Sharpe 2.39, MaxDD -10.51%.
What we tell new users
Backtest performance is not a guarantee. Live realistic target is +110-140% over 18 months at 5-8 bps actual slippage on a Tier-1 exchange. Some windows in the backtest underperformed. Some live months will underperform too.
The reason we still trust Phase 37 is that the walk-forward survives the slippage stress test (0/5/10/20 bps), survives sleeve drops, survives the 2022 bear market, and survives weekly re-snapshots.