Are the probabilities honest?
This is the real test. Group every prediction by the chance we gave it, then check how often that group actually happened. A well-calibrated model sits close to the diagonal: 30% tips land about 30% of the time, 70% tips about 70%.
| We said | Predictions | Expected | Actually happened | Calibration |
|---|---|---|---|---|
| Under 20% | 1640 | 15.0% | 13.8% | -1.1 pts |
| 20–30% | 3624 | 25.5% | 23.9% | -1.5 pts |
| 30–40% | 2649 | 35.1% | 34.0% | -1.1 pts |
| 40–50% | 3551 | 45.2% | 44.7% | -0.5 pts |
| 50–60% | 3189 | 54.7% | 56.7% | +2.0 pts |
| 60–70% | 1748 | 64.3% | 66.0% | +1.6 pts |
| 70–80% | 1321 | 75.1% | 76.5% | +1.5 pts |
| Over 80% | 762 | 84.2% | 85.3% | +1.2 pts |
The marker on each bar is what we predicted; the fill is what happened. Within about four points is good; a persistent gap in one direction means the model is biased there.
By market
Draws deserve special attention. A draw tip at 30% that wins three times in ten is a correct prediction, even though it lost seven times. Judge it on the gap between expected and actual, and on whether backing it at our quoted price made or lost money.
| Market | Settled | Expected | Happened | Brier | Profit at best price |
|---|---|---|---|---|---|
| Home wins | 2078 | 44.5% | 44.9% +0.4 pts | 0.210 | −104.19 ROI -5.0% |
| Draws | 2078 | 24.8% | 24.5% -0.3 pts | 0.182 | −60.04 ROI -2.9% |
| Away wins | 2078 | 30.6% | 30.6% 0.0 pts | 0.184 | −159.22 ROI -7.7% |
| Over 2.5 goals | 2066 | 55.0% | 59.8% +4.9 pts | 0.234 | +129.45 ROI 6.3% |
| Under 2.5 goals | 2066 | 45.0% | 40.2% -4.9 pts | 0.234 | −270.23 ROI -13.1% |
| Both teams to score | 2035 | 55.4% | 59.6% +4.2 pts | 0.244 | +77.69 ROI 3.8% |
How to read this page
Hit rate is not accuracy. If we tip a draw and give it a 30% chance, we are saying it will lose about seven times in ten. Losing seven times in ten is then the correct outcome, not a failure. The only way to judge a probability is to collect many of them and check whether reality matched the number.
Calibration is that check. Take every prediction we rated between 30% and 40%; if roughly 35% of them happened, the numbers are honest. A gap of more than about eight points, repeated across hundreds of predictions, means something is wrong — not bad luck.
The Brier score squares the error on every prediction and averages it. Zero is perfect, 0.25 is a coin flip, and lower is better. It punishes confident mistakes far more than cautious ones, which is exactly what you want from a forecaster.
Profit at best price answers the practical question: if you had backed every one of these at the price we showed, staking one unit each time, where would you be? This is the number that matters most, and it is usually negative for favourites and closer to break-even on draws, because the market systematically overprices short odds.
Predictions are logged before kick-off and never edited afterwards. Past performance is not a guide to future outcomes. 18+.