Measured, not claimed
How accurate are these projections?
Anyone can publish a prediction. This page publishes the marking. Every projection method this site uses was run over 1,707 finished matches from Europe's top five leagues in 2024/25 — forecasting each one from prices recorded two hours before kick-off, then scoring it against the result.
Skill vs base rates
+12.6%
Better than assuming every match is an average match. That baseline is harder to beat than it sounds — home teams win about 44% of the time, and a lot of published models never clear it.
- Brier score
- 0.572
- vs 0.654 for the baseline · lower is better
- Log loss
- 0.962
- punishes confident mistakes hardest
- Top pick right
- 53.8%
- the flattering number — see below
- Goals error
- 1.29
- average miss on total goals
The calibration curve
This is the chart that can embarrass us, which is why it's here. It asks: of all the times we said 60%, did it happen about 60% of the time? On the dashed line means the percentages can be taken literally. Above it means we're too cautious; below it means we're overselling.
Broken down by competition
A single pooled score can hide a lot. If the number only held up in one unusually predictable league you'd want to know, so here it is per competition.
| Competition | Matches | Skill | Brier | Goals error |
|---|---|---|---|---|
| La Liga | 380 | +13.2% | 0.562 | 1.22 |
| Serie A | 379 | +14.7% | 0.566 | 1.22 |
| Premier League | 334 | +12.1% | 0.577 | 1.32 |
| Bundesliga | 308 | +11.4% | 0.59 | 1.42 |
| Ligue 1 | 306 | +11.3% | 0.567 | 1.3 |
What this is, and what it isn't
Each forecast uses only what existed before kick-off: the de-vigged consensus of every bookmaker pricing that match, sampled two hours before it started, turned into a full distribution of scorelines. No result, no post-match statistic, nothing from the future.
We tested our own goal model against the market on exactly this sample, and the market won — clearly, in all five leagues. So we publish the market-anchored number rather than the one we built ourselves. Showing you a worse forecast because it was ours would be vanity at your expense.
"Top pick right" is the number people reach for and the least useful of the four. A model that always names the favourite scores well on it while being worth nothing to bet with. Skill and calibration are the ones that mean something.
Read the limits honestly: this is a backtest of the method over one season in five competitions, not a live track record. It says the projections are well calibrated. It does not say a bet at these prices makes money — that is a separate claim, and it lives on the record page with real picks and real results.