How the Bundesliga model actually performs
A model is only worth its calibration. Every number below is an out-of-sample prediction for a 25/26 Bundesliga fixture, made before kick-off from earlier data only, then checked against what actually happened. No cherry-picking, no hindsight.
Read this first. This page measures two things: whether our probabilities are honest — do the events we call 30% happen about 30% of the time? — and whether the prices we flag as value beat the closing line. It is not a profit record: calibration proves the numbers are sound, closing line value proves they beat the market, and neither is a promise of profit. We grade everything against a baseline so a confident-looking model can't hide.
Every market — does the model call it straight, and beat guessing?
Two honest questions per market. First, are our percentages true — when we say something happens 30% of the time, does it? The bar shows it: grey is what we said, green is what actually happened; the closer they sit, the straighter the call. Second, is the model genuinely sharper than just predicting the season average for everyone? That's the badge.
Show the full numbers — predictions, Brier score, baseline
Brier score measures per-prediction accuracy (lower is better); the baseline is what you'd score by predicting the same base rate for everyone. Skill is how far the model beats that baseline — the maths behind the badge above.
| Market | Predictions | Model avg | Actual | Brier | Baseline | Skill |
|---|---|---|---|---|---|---|
| To be carded | 7,923 | 14% | 14% | 0.117 | 0.117 | -0.1% |
| 1+ shot on target | 7,923 | 23% | 26% | 0.164 | 0.191 | +14.3% |
| 2+ shots on target | 7,923 | 7% | 7% | 0.056 | 0.063 | +11.1% |
| 1+ shot | 7,923 | 45% | 52% | 0.197 | 0.250 | +21.0% |
| 2+ shots | 7,923 | 23% | 25% | 0.146 | 0.188 | +22.5% |
| 3+ shots | 7,923 | 12% | 11% | 0.082 | 0.101 | +18.5% |
| 1+ foul committed | 7,923 | 48% | 49% | 0.214 | 0.250 | +14.2% |
| 2+ fouls committed | 7,923 | 20% | 19% | 0.139 | 0.153 | +9.2% |
| 3+ fouls committed | 7,923 | 8% | 6% | 0.053 | 0.055 | +3.0% |
| 1+ foul won | 7,923 | 44% | 47% | 0.213 | 0.249 | +14.4% |
| 2+ fouls won | 7,923 | 18% | 18% | 0.131 | 0.149 | +12.2% |
| 3+ fouls won | 7,923 | 8% | 6% | 0.051 | 0.054 | +5.3% |
| 1+ tackle | 7,923 | 54% | 59% | 0.186 | 0.242 | +23.1% |
| 2+ tackles | 7,923 | 29% | 31% | 0.179 | 0.213 | +16.2% |
| 1+ outside-box shot on target | 7,889 | 8% | 8% | 0.069 | 0.073 | +5.5% |
| 3+ saves | 618 | 53% | 56% | 0.246 | 0.246 | -0.1% |
| 4+ saves | 618 | 35% | 35% | 0.221 | 0.227 | +2.9% |
| 5+ saves | 618 | 22% | 21% | 0.162 | 0.167 | +3.0% |
| Team & match markets | ||||||
| Match Over 2.5 goals | 272 | 55% | 65% | 0.241 | 0.228 | -5.5% |
| Match Over 3.5 cards | 306 | 58% | 53% | 0.257 | 0.249 | -3.2% |
Calibration, up close
The aggregate can look right while the details are wrong, so here's the finer view. We sort every prediction into groups by how confident the model was — all the ~20% calls, all the ~30% calls, and so on — then check how often each group actually happened. The model is well calibrated when the two line up: when the things it calls 30% happen about 30% of the time. In each row, grey is what the model said and green is what actually happened — the closer they are, the better. Tap any market to open it.
To be carded
1+ shot on target
2+ shots on target
1+ shot
2+ shots
3+ shots
1+ foul committed
2+ fouls committed
3+ fouls committed
1+ foul won
2+ fouls won
3+ fouls won
1+ tackle
2+ tackles
1+ outside-box shot on target
3+ saves
4+ saves
5+ saves
Match Over 2.5 goals
Match Over 3.5 cards
Where the model earns its keep
The shots, fouls and tackles markets are the bulk of what the tool prices, and it beats a base-rate baseline by up to 23% on them — scored using only matches played before each game, so nothing is fitted with hindsight. That's most of the board, and the part the tool is built to price.
We show the ones that don’t, too. To be carded, 3+ saves, Match Over 2.5 goals and Match Over 3.5 cards came out no better than guessing the league average this season — the noisiest markets on the board, where even a sharp model has little to add. We leave them on the page because a track record you only show when it flatters you isn't one.
Closing line value
Whether our prices beat the market's own closing view is measured across every league at once, not per competition — the Championship has the most legs of any single league and still sits below the sample a closing-line read needs. Pooled, the record covers 3,186 legs.
See the full closing line value record
Method: player markets scored as-of each kick-off using only prior matches (the recency-weighted rate never sees the future); team markets fit on earlier seasons and evaluated on 25/26. Regenerated from source data on 2026-10-12. Full method on the methodology page.
The scoreboard above is free, and it always will be. The paid tool is that same model, pointed at this week's prices — the edge, and the confidence, on every leg.