2018HIT
Russia
Saudi Arabia
- Predicted
- 2–1(Russia)
- Actual
- 5–0(Russia)
A backtest of past World Cups. We re-ran the model over the 2018 and 2022 tournaments — all 128 matches — and graded every call against what really happened. The misses are shown right alongside the hits, no cherry-picking.
This is a retrospective test. For each match, the model was given only the data available before that match — no future results — then it predicted the score, and only afterwards did we check it against the real result. These were never locked in ahead of time for real, and never money. It is evidence the method is reasonable, not a promise about the future.
For the live, going-forward record — predictions we lock before kickoff and grade in public as matches finish — see the Results on today's page. That is the real-time test; this page is the historical one.
Outcome accuracy
55%
Outcome accuracy
55%
Outcome accuracy
55%
Outcome accuracy = how often the model's top win/draw/loss call was right. RPS and Brier score the full probabilities — lower is better, 0 is perfect. Exact score = how often the predicted scoreline was spot on.
Showing 128 matches — 70 correct, 58 missed
Retrospective backtest, not live picks -- these were never locked in for real money. The model parameters (K, home advantage, supremacy, rho) were tuned on >=2023 data, so 2018 and 2022 are out-of-sample for the parameters as well as for the ratings. Metrics: outcomeAccuracy is top-pick (1X2) hit rate; RPS is the ranked probability score for ordered Home>Draw>Away (lower is better, 0 perfect); Brier is multiclass (lower is better); exactScoreRate is how often the rounded predicted scoreline equalled the real scoreline. Football is high-variance and a single tournament is only 64 matches -- treat these as credibility evidence, not a guarantee.
These predictions are statistical estimates for context and fun — not advice, and not a betting product. A single tournament is high-variance; treat the backtest as credibility evidence, never a guarantee.