Model accuracy over time

Walk-forward test: for each match day, the model - trained only on games before that day - predicts each game, then we score it against what actually happened. Because standings are decided by total games won, the headline metric is game-level accuracy: of all individual games, how often did the model's favored team win? (Series win-rate is shown too, but it matters less.)

Current season only
54.3%
game accuracy - 3204.0 games
All data (career prior)
53.7%
game accuracy
Does history help?
+5.0
match day 1 · full season -0.6
Finding: the career prior helps the cold start (+5.0 pts at match day 1, before this season has data) but slightly hurts (-0.6 pts) over the full season as current-season form takes over. History is a useful early-season anchor, not a season-long edge - which is why the live model fades it as games accumulate.

Honest A/B test of the two models, walk-forward. Current season only starts every team at the same rating; all data seeds each team from its roster's career skill from past seasons (a pre-season signal, so it's a fair test). The biggest difference should show up early in the season, before this year's results accumulate. "Series win-rate" matters less since standings reward total game wins (60.2% current).

Match dayGamesGame accuracyCumulative Series win-rate Brier
1400 49.2%49.2% 49.3%0.25
2400 53.8%51.5% 60.3%0.247
3400 51.7%51.6% 56.0%0.247
4400 57.2%53.0% 66.2%0.241
5400 61.3%54.6% 70.0%0.239
6400 49.8%53.8% 52.3%0.252
7400 55.2%54.0% 63.8%0.246
8400 56.8%54.4% 64.6%0.249
94 25.0%54.3% 0.0%0.264

Match day 1 sits near 50% because every team starts at the same rating, so early predictions are essentially coin flips; accuracy climbs as the ratings learn. These are genuinely high-variance games - the per-game edge is only a few points - but it compounds over a season into the playoff odds.


Data source: live (rscna.com) - standings & schedule refresh hourly.