Weekly EPA-OLS replay compared with recorded market spreads on the same completed games. Development and already-inspected confirmation remain separate. These are retrospective results, not frozen prospective forecasts.
| evidence | system | n | mae | rmse | bias |
|---|---|---|---|---|---|
| already_inspected_confirmation | epa_ols | 1022 | 12.761 | 15.968 | -0.089 |
| already_inspected_confirmation | market_recorded | 1022 | 11.982 | 15.142 | -0.871 |
| retrospective_development | epa_ols | 2826 | 13.213 | 16.768 | 0.136 |
| retrospective_development | market_recorded | 2826 | 12.244 | 15.581 | 0.137 |
MAE is average absolute point error. RMSE gives more weight to large misses. Bias is predicted minus actual Home Margin: positive means overstating the home side. A bias near zero can hide large errors that cancel.
Team cohorts include both venues but keep the home-side error sign. Samples below 30 common games are flagged sparse. Larger samples remain descriptive; games share teams and weeks. Error quantiles are not confidence intervals or a calibrated prediction distribution.
Coverage describes saved lined projections only. Games omitted before export cannot be counted here. Recorded spreads lack verified pre-kickoff capture times.
Home Margin alone cannot determine totals. Win and cover probabilities need a validated distribution or calibration mapping. No forecast mapping changed.