📊 Model performance
Here we honestly measure whether our model gets it right. Two different things: the VALIDATION over thousands of already-played matches (does it predict the past well?) and the LIVE RESULTS of the tips we publish (which build up match by match).
🧪 Historical model validation
Backtest over already-played matches: for each one we recompute ratings using ONLY prior matches (no lookahead) and compare with the real result. "Favourite hit-rate" = how often the model's most likely team actually won.
| Sport | Favourite hit-rate | Lift over chance | Total error | Matches |
|---|---|---|---|---|
| ⚽ Football | 62% | +10.6% | ±1.43 goals | 281 |
| 🏀 Basketball | 59% | +2.2% | ±17.66 pts | 122 |
| 🤾 Handball | 72% | +6.8% | ±6.42 goals | 26 |
| 🎾 Tennis | 61% | +7.2% | – | 237 |
The "lift over chance" compares the model error (Brier) against always betting the base rate: a positive number means the model adds real information. It is not a profit promise: it measures model accuracy, not return vs the odds.
📡 Live results of our tips
Real hit-rate of the published tips, settled when each match ends. It fills up over time on its own (the more matches played, the more reliable).
⏳ Collecting data: no published tips have settled yet. Check back in a few days.
Methodology: opponent-adjusted Elo engine + attack/defence for totals. Predictions are anchored to the market to avoid overconfidence. Betting carries risk; no statistic guarantees results. Gamble responsibly.