Model Accuracy
Every prediction we make is testable. Here is how our 7 models actually performed — measured honestly, with no look-ahead bias, across thousands of real matches.
47,691 predictions evaluated point-in-time: each match was predicted using only data available before kickoff.
Over 2.5 Goals — Combined model
When we say 54%, it happens ~54% of the time
Match Result (1X2) — Elo model
Our best-calibrated outcome model
Calibration by model & market
| Model | Over 1.5 | Over 2.5 | Over 3.5 | 1X2 | Exact Score |
|---|---|---|---|---|---|
| Combined ⭐ | 75.6% said 74.3% | 52.5% said 50.8% | 30.4% said 30.2% | 33.3% said 33.4% | 11.9% said 12.3% |
| Elo | 75.6% said 74.6% | 52.5% said 51.1% | 30.4% said 30.0% | 33.3% said 33.3% | 11.8% said 12.2% |
| Poisson | 75.6% said 74.6% | 52.5% said 51.1% | 30.4% said 30.0% | 33.3% said 33.3% | 11.8% said 12.2% |
| Dixon-Coles | 75.6% said 76.1% | 52.5% said 51.1% | 30.4% said 30.0% | 33.3% said 33.3% | 11.8% said 13.2% |
| Weighted Form | 75.6% said 74.9% | 52.5% said 51.3% | 30.4% said 30.0% | 33.3% said 33.3% | 12.0% said 12.0% |
| Normalized | 75.6% said 71.9% | 52.5% said 49.5% | 30.4% said 30.2% | 33.3% said 33.3% | 10.4% said 14.3% |
| Shots | 75.2% said 72.9% | 52.0% said 50.0% | 29.8% said 30.0% | 33.3% said 33.3% | 10.8% said 13.2% |
1X2 — how often the predicted winner was correct
How we test
We replay history match by match in chronological order. For each fixture, the models are fed only the data that existed before kickoff — rolling team form, home/away goal rates, shots and Elo ratings. The prediction is recorded, then compared to the real result. This eliminates look-ahead bias, the most common flaw in published accuracy claims. BTTS is excluded while we improve its calibration.
Last updated: 3 Jun 2026