Entry-Filter Testing
Candidate entry filters tested against trade R: Spearman rank correlation + chronological 70/30 train/test split. "held up" = filter direction consistent out-of-sample.
| Metric | n | ρ | p | Sig? | Train (filter vs other) | Test (filter vs other) | Held up? |
|---|---|---|---|---|---|---|---|
| Deviation (x ATR) | 200000 | 0.032 | 0.0 | p<0.05 | +0.39R vs +0.31R | +0.40R vs +0.30R | held up |
| Direction (long=1) | 200000 | -0.032 | 0.0 | p<0.05 | +0.30R vs +0.40R | +0.30R vs +0.41R | held up |
| Hour of day (UTC) | 200000 | -0.076 | 0.0 | p<0.05 | +0.27R vs +0.43R | +0.28R vs +0.43R | held up |
Read: with very large n, small ρ is statistically significant but practically negligible. Focus on filters that both reach significance AND show a meaningful train/test gap that holds direction.