Evaluation
How good was the forecast?
- Backtest: 15 past seasons, each predicted from earlier ones. Used to pick the model.
- 2026: the real test — the forecast was fixed before the season.
2026 game log loss
0.6845
coin flip 0.6917
2026 games called
55.3%
favourite won
Calibration error
0.6%
stated vs. observed win rate
Win-total RMSE
8.4
wins; correlation 0.63
1Model comparison, season by season
- Elo beat Team Marcel in 14 of 15 backtest seasons — narrowly.
- Both crush the naive baselines (last season’s record, coin flip).
- Elo’s edge: it sees games in order, so late-season form counts more.
EloTeam MarcelLast seasonCoin flip
EloTeam MarcelLast seasonCoin flip
2Calibration
- When it said 60%, teams won about 60% — off by 0.6% on average.
- Deliberately modest: the most lopsided game was 72%. A preseason rating can’t know injuries or starters.
- Win ranges were honest too: 26 of 30 teams landed in their 80% range.
3All numbers
| Method | Sample | Games | Log loss | Brier | Accuracy | Wins MAE | Wins RMSE |
|---|---|---|---|---|---|---|---|
| Elo, carried over | Backtest | 36,444 | 0.6832 | 0.2451 | 55.9% | 7.95 | 9.92 |
| 2026 | 2,429 | 0.6845 | 0.2458 | 55.3% | 6.62 | 8.41 | |
| Team Marcel (3-yr Pythagorean) | Backtest | 36,444 | 0.6843 | 0.2456 | 55.7% | 8.23 | 10.28 |
| 2026 | 2,429 | 0.6848 | 0.2459 | 55.4% | 6.89 | 8.48 | |
| Last season's W% | Backtest | 36,444 | 0.6903 | 0.2483 | 55.4% | 9.67 | 12.01 |
| 2026 | 2,429 | 0.6927 | 0.2496 | 54.4% | 8.95 | 11.08 | |
| Coin flip + home field | Backtest | 36,444 | 0.6909 | 0.2489 | 53.4% | 10.06 | 12.30 |
| 2026 | 2,429 | 0.6917 | 0.2493 | 52.9% | 8.50 | 10.85 |
Per-season backtest table
| Season | Elo log loss | Team Marcel log loss | Last season log loss | Coin flip log loss | Elo RMSE | Team Marcel RMSE | Last season RMSE | Coin flip RMSE |
|---|---|---|---|---|---|---|---|---|
| 2010 | 0.6787 | 0.6798 | 0.6829 | 0.6868 | 7.5 | 8.3 | 9.3 | 10.8 |
| 2011 | 0.6878 | 0.6884 | 0.6965 | 0.6925 | 9.9 | 9.9 | 12.4 | 11.2 |
| 2012 | 0.6848 | 0.6866 | 0.6901 | 0.6912 | 9.9 | 10.4 | 11.5 | 11.7 |
| 2013 | 0.6853 | 0.6834 | 0.6881 | 0.6903 | 10.5 | 9.8 | 11.3 | 12.0 |
| 2014 | 0.6900 | 0.6916 | 0.7022 | 0.6916 | 8.9 | 9.5 | 12.6 | 9.4 |
| 2015 | 0.6893 | 0.6910 | 0.6962 | 0.6896 | 10.2 | 10.8 | 12.4 | 10.3 |
| 2016 | 0.6848 | 0.6865 | 0.6906 | 0.6915 | 8.1 | 8.7 | 10.2 | 10.5 |
| 2017 | 0.6872 | 0.6877 | 0.6962 | 0.6900 | 10.4 | 10.6 | 13.0 | 11.3 |
| 2018 | 0.6795 | 0.6821 | 0.6831 | 0.6919 | 10.9 | 11.7 | 12.0 | 14.4 |
| 2019 | 0.6709 | 0.6723 | 0.6740 | 0.6916 | 9.9 | 10.3 | 10.8 | 15.6 |
| 2021 | 0.6819 | 0.6826 | 0.6909 | 0.6902 | 12.0 | 12.2 | 14.2 | 14.2 |
| 2022 | 0.6799 | 0.6814 | 0.6904 | 0.6911 | 11.4 | 11.9 | 14.0 | 14.4 |
| 2023 | 0.6855 | 0.6873 | 0.6961 | 0.6929 | 10.9 | 11.4 | 13.6 | 12.9 |
| 2024 | 0.6832 | 0.6849 | 0.6929 | 0.6927 | 9.3 | 9.8 | 12.1 | 12.3 |
| 2025 | 0.6792 | 0.6794 | 0.6841 | 0.6895 | 7.6 | 7.8 | 9.4 | 11.5 |