Model performance
Active experimental model. Frozen M0 parameters transported from NRL; not yet validated on Super League.
Probability quality is measured by Brier score and log loss — lower is better for both. Winner accuracy is shown as a secondary measure because it ignores how confident each forecast was.
Scored sample — 3 predictions: 3 prospective, 0 backtest.
0.377
Lower is better · n = 3
0.960
Lower is better · n = 3
0.0%
Secondary · 3 decided matches (draws excluded)
3
Completed matches with a scored forecast
Cumulative & rolling Brier score
Rolling window: last 10 predictions. Reference line at 0.25 = always forecasting 50%.
Calibration
Average forecast home probability vs observed home win rate per 10% bucket. Dot size = sample size.
Probability distribution
How often the model issued each home win probability.
Winner accuracy over time
Cumulative, draws excluded. Secondary to Brier and log loss.
Model comparison
Compare models only on the same matches for a fair comparison.
| Model | Scored | Of which backtest | Brier | Log loss | Winner accuracy |
|---|---|---|---|---|---|
| M0 v0.1.0 | 3 | 0 | 0.377 | 0.960 | 0.0% (3) |