VRP Filter — Definitive Shuffle Test
PromotedVRP Filter — Definitive Shuffle Test
2026-05-19 · status: promoted · 106.6s
Hypothesis: vrp_filter@p20 lift of +13.8% (BB_EXTREME_2) and +37.7% (DONCHIAN_50) from Exp D1 is genuine VRP signal — confirmed if real-source z-score vs 10 shuffled-source backtests > 1.5.
Verdict: REAL SIGNAL — ADD TO MEGASWEEP — all 2 strategies z > 1.5 vs shuffled. Add vrp_filter to BB_EXTREME_2, DONCHIAN_50 COMPATIBLE_FILTERS in megasweep.py.
Key metrics
| metric | value |
|---|---|
| n_strategies | 2 |
| n_shuffles | 10 |
| n_signal | 2 |
| BB_EXTREME_2_real_lift_pct | +13.7545 |
| BB_EXTREME_2_z_score | +1.5679 |
| DONCHIAN_50_real_lift_pct | +37.7323 |
| DONCHIAN_50_z_score | +2.5304 |
Approach
Real-backtest with 10 shuffled vrp_4h.parquet copies (random permutation of vrp + vrp_pct_90d columns). Each backtest runs the full runner including re-routing. If lift is genuine VRP signal, real-source pnl should sit > 1.5σ above the shuffled distribution's mean.
Results
| strategy | baseline_pnl | real_pnl | shuf_mean | shuf_std | shuf_p5 | shuf_p95 | real_lift_pct | shuf_lift_pct_mean | z_score |
|---|---|---|---|---|---|---|---|---|---|
| BB_EXTREME_2 | 17.66 | 20.09 | 14.08 | 3.831 | 8.54 | 18.9 | +13.8% | -20.3% | 1.57 |
| DONCHIAN_50 | 5.52 | 7.6 | 5.13 | 0.979 | 3.35 | 5.96 | +37.7% | -7.1% | 2.53 |
Verdict
REAL SIGNAL — ADD TO MEGASWEEP — all 2 strategies z > 1.5 vs shuffled. Add vrp_filter to BB_EXTREME_2, DONCHIAN_50 COMPATIBLE_FILTERS in megasweep.py.