The Open ModelForecast saved
Checking our work

Past predictions and results

Before kickoff, we publish the chance of a home win, draw and away win, then mark the highest number as the most likely result. After full time, we record whether that top pick was correct. Right and wrong results both stay visible.

70/102
Correct World Cup top picks
4/4
Final four teams identified early
Aug
Club-season record starts 2026-27

Club-season record begins in August 2026

Since the season hasn't kicked off, there are no club results to score yet. The pipeline is already running: every daily forecast snapshot locks the match probabilities that are public before kickoff, and finished results are scored against them automatically. This section fills in from the first completed match.

Where the model comes from — World Cup 2026

The same engine (Elo → Dixon-Coles → Monte Carlo) ran the 2026 World Cup as an open-source model: 70 correct top picks from 102 completed matches, with every prediction published before kickoff. When it picked a team and the match was drawn, that counted as incorrect.

The honest backtest: walk-forward on 913 real internationals

Before the tournament, the model was tested the strict way — out-of-sample: each match predicted using only data available before kickoff, then scored against the actual result with proper scoring rules, not just accuracy. 763 matches evaluated (150 burn-in). Reproduce it yourself with node backtest.mjs in the repository, or download the raw evaluation data.

MetricModelBaseline
Ranked Probability Score (↓)0.175coin-flip 0.241
Brier score (↓)0.520coin-flip 0.667
Log-loss (↓)0.886coin-flip 1.099
Expected Calibration Error (↓)2.3%<5% = well-calibrated
Correct result (win/draw/loss)61.9%always-home 49.1%
When a clear favourite (p ≥ 50%)69.0%481 matches
0%0%25%25%50%50%75%75%100%100%perfect calibrationPredicted 5.3% → happened 6.7% (225 outcomes)Predicted 15.0% → happened 13.4% (374 outcomes)Predicted 25.7% → happened 23.9% (804 outcomes)Predicted 34.8% → happened 32.2% (205 outcomes)Predicted 44.9% → happened 54.0% (200 outcomes)Predicted 54.9% → happened 55.7% (149 outcomes)Predicted 64.8% → happened 66.9% (136 outcomes)Predicted 74.6% → happened 75.8% (95 outcomes)Predicted 85.1% → happened 85.0% (100 outcomes)Predicted 91.1% → happened 100.0% (1 outcomes)Predicted probabilityObserved frequency
Backtest calibration: 763 internationals × 3 outcomes pooled into 10 bins. When the model said 30%, it happened about 30% of the time.
View as table
BinOutcomesAvg predictedObserved
010%2255.3%6.7%
1020%37415.0%13.4%
2030%80425.7%23.9%
3040%20534.8%32.2%
4050%20044.9%54.0%
5060%14954.9%55.7%
6070%13664.8%66.9%
7080%9574.6%75.8%
8090%10085.1%85.0%
90100%191.1%100.0%

Past results show how the model behaved; they do not guarantee future accuracy. New to probability? Read the beginner's guide before judging a single result — a prediction can be reasonable and still be wrong.