SP26 SOCCER PREDICTOR

ABOUT SOCCER PREDICTOR

Football, examined through data.

An independent football analytics site combining statistical forecasts, matchday analysis and a recorded comparison between model predictions and human decisions. The site is free; model content requires signup and login.

66,414
Historical matches in training datasets
755
Logged model predictions
605
Predictions with recorded results
4
Model families

Figures checked 2026-10-19. Training coverage counts distinct historical matches in the configured datasets across model families; shared datasets are counted once. Verified archived counts are used when older training files are not retained on the live server. Prediction counts cover all competitions in the model log, including fixtures where no bet qualifies. Settled counts require both recorded final scores. They are coverage figures, not a claim of predictive accuracy.

Four models, different contexts

  • Domestic leagues: team strength, attack and defence, recent results and expected-goal data where available. Major leagues use xG inputs and a Dixon–Coles low-score correction; leagues with less reliable xG use a different input mix.
  • Champions League: domestic club ratings combined with historical UCL results. The mechanism is calibrated separately; a correction that helps domestic leagues is not automatically applied here.
  • International tournaments: national-team ratings and form, with tournament and venue context.
  • Home/away internationals: shared national-team ratings with separate home-advantage treatment for qualifiers, Nations League and other home/away fixtures.

Goal estimates are converted into scoreline and outcome probabilities. Additional adjustments and filters are assessed for evidence and overlap, so the same information is not deliberately counted multiple times.

The human layer

Human judgments are recorded separately from the model. Comparing decisions with actual results, entry odds and the last captured odds strictly before kickoff helps assess both layers over time. A forecast and a qualifying betting edge are separate questions.

The model page includes the verdict comparison for signed-in users. Headline counts here describe model records, not the number of human bets.

How performance is evaluated

Winner accuracy is only one measure. Probability calibration and Brier scores assess the forecasts; closing-line value and settled yield assess recorded price decisions under their stated assumptions. Historical tests and logged results are evaluated separately.

The former 59.8% headline referred to one international classifier's historical holdout test. It was not an overall accuracy figure for the current four-model system. Combining that figure with other models' training accuracy would produce a misleading comparison.

Visit the track record and the model's performance panels for the relevant samples and metrics. Past results do not guarantee future performance.

Data sources

Core sources include football-data.co.uk for domestic results, The Odds API for market-covered fixtures and odds, Understat for major-league xG, and the international_results dataset for national-team results. Historical UCL results are curated from public datasets. Coverage, freshness and availability differ by competition.

Limits and ongoing research

Late lineup changes, rotation, tactical changes and unpredictable match events can materially affect results. Data feeds and recorded closing prices also have limits: the last quote we captured may differ from the final quote available at a bookmaker.

Current research includes finishing and goalkeeper performance beyond pre-shot xG, calibration across competitions, and whether recorded price advantages translate into returns as the sample grows.

All data, analytics and content are for individual, private use. Public or commercial use requires prior written permission. See the Terms and Privacy Policy.