epl mid-dog v2: filter stack vs v1, holdout + generalization
← lab · · ran 7/7/2026
What this is: A one-off study — its numbers below are exactly what the run reported.
results
| n | hit | roi | label |
|---|
| 540 | 0.35 | 0.1796 | EPL v1 (all 21 seasons) |
| 232 | 0.392 | 0.346 | EPL v2 (filters, SAME data — inflated by construction) |
| 104 | 0.394 | 0.3903 | EPL v2 on 2005-15 (derivation era) |
| 315 | 0.343 | 0.1504 | EPL v1 on 2016-26 (holdout) |
| 128 | 0.391 | 0.3101 | EPL v2 on 2016-26 (HOLDOUT — the number that matters) |
| 1,734 | 0.289 | -0.0416 | Other 6 leagues v1 |
| 768 | 0.279 | -0.067 | Other 6 leagues v2 (do the patterns travel?) |
Reading the columnswhat each number actually means
| AUC | predictability: 0.50 = coin flip, ~0.70 = ceiling for sports |
| Importance | how much the model leans on this factor (permutation importance) |
| Direction | sign of the raw correlation with the outcome |
| Survives all eras | effect points the same way in every historical era |
Spec · the reproducible recipe
{
"kind": "v2_filters",
"filters": {
"away_rank_max": 14,
"home_form_min": 4,
"home_rank_max": 17,
"exclude_months": [
8,
9,
10
]
}
}