AFL Model Accuracy: The Honest Numbers
We grade our AFL prediction model the hard way - walk-forward, out-of-sample. When we test on a season, the model has never seen a single game from it. No cherry-picking, no in-sample inflation.
Held-out accuracy by season
Each season below was fully excluded from training when the model predicted it. This is the only accuracy figure that means anything - a model graded on data it trained on will always look better than it is.
| Test season | Accuracy | Method |
|---|---|---|
| 2024 | 62.5% | Trained on 2019-2023 |
| 2025 | 72.7% | Trained on 2019-2024 |
| 2026 (season to date) | 74.1% | Trained on 2019-2025 |
| Weighted average | 69.0% | Across all test seasons |
2026 season, round by round (out-of-sample)
Every completed 2026 game predicted by a model trained only on 2019-2025 - so these are genuine, never-seen results, not the model marking its own homework.
| Round | Correct | Accuracy |
|---|---|---|
| Round 0 | 4/5 | 80% |
| Round 1 | 8/9 | 89% |
| Round 2 | 5/7 | 71% |
| Round 3 | 4/7 | 57% |
| Round 4 | 7/8 | 88% |
| Round 5 | 6/9 | 67% |
| Round 6 | 8/9 | 89% |
| Round 7 | 8/9 | 89% |
| Round 8 | 6/9 | 67% |
| Round 9 | 8/9 | 89% |
| Round 10 | 6/9 | 67% |
| Round 11 | 7/9 | 78% |
| Round 12 | 3/7 | 43% |
| Round 13 | 4/8 | 50% |
| Round 14 | 2/2 | 100% |
| 2026 total | 86/116 | 74.1% |
How the model works
An ensemble of XGBoost and Logistic Regression trained on 1,547 AFL matches back to 2019. It reads 53 features per game: recent form, head-to-head history, home-ground record, rest days, win/loss streaks, 5-game momentum patterns and points differential. It is the same architecture as our NRL model, retrained on AFL data with AFL-specific pattern rates.