Method

How the football prediction model is calculated and tested

From completed league results to independent forecasts: what we use, what we have tested and what the results actually support.

See the experiments and results, or check the prediction record.

Model

Home and away goals are Poisson rates. Attack and defence ratings are estimated from finished matches only, with a one-year half-life so older games count less, and a penalty that pulls thinly observed teams toward the league average. Newly promoted clubs start from the typical rating of a promoted side instead, because promoted clubs have, on average, been weaker than established ones. Training uses Premier League results since 2005–06. A low-score correction, the Dixon–Coles term, stays positive. Setting that term to zero is the independent Poisson model.

Earlier results already influence the next forecast through these team ratings. Matches are not treated as identical coin tosses: scoring against a strong opponent contributes differently from scoring against a weak one. A winning streak does not receive an automatic extra bonus on top of those ratings.

Championship, La Liga, Bundesliga, Serie A and Ligue 1 models are fitted and validated separately using their own completed results since 2020–21. They exclude the Premier League promoted-club prior, FPL availability and playoffs. The four continental leagues use OpenFootball archive scores and exact football-data.org timestamps for recent seasons; archive records with no verified timezone become available conservatively three hours after the end of the match date in UTC. Explore league-specific studies.

European competitions: one rating scale

League models only compare clubs within a league, so the Champions League, Europa League and Conference League share a joint model. Every club gets an attack and a defence rating on a shared scale, fitted to domestic results from 22 leagues and to Champions League, Europa League and Conference League results. Each domestic league keeps its own goal rate and home advantage, and so does each European competition. A club's rating is pulled towards its league's level, and that level is learned only from matches between leagues: European ties, plus promotion and relegation. Clubs from countries without a domestic data file are rated from their European results, pulled towards a shared level for such countries. The same one-year half-life applies.

Before publication the model was tested walk-forward on 2,817 European matches from 2016-17, each forecast using only earlier results. Its log loss was 0.942, against 0.982 for the same model without league levels and 1.051 for historical outcome shares; the improvement over the model without league levels held in each of three periods. Averaging forecasts over the uncertainty in the ratings added nothing, so the published forecasts use the fitted ratings directly. Current Europa League and Conference League league-phase results come from fixturedownload.com; knockout scores are excluded until their 90-minute scope can be verified. The model has not been compared with bookmaker prices.

How we test an improvement

We judge a change by the quality of its probabilities on matches it has not seen. Log loss and Brier score reward probability assigned to the result that happened; lower is better. A change in log loss is not a percentage-point change in the chance of winning.

The offline experiments below refit team ratings every 28 days and split the last 3,800 evaluation matches per league into chronological tuning, calibration and test blocks of roughly 40%, 30% and 30%. Settings are chosen on tuning matches; probability confidence is calibrated on a separate block. The control and candidate each receive their own calibration. The final test block is reserved for evaluation. Historical results and match statistics are treated as available no earlier than 03:00 UTC the next day.

Our replacement rule requires log loss to fall by more than 0.005 and Brier score to fall in each of two chronological test halves. The upper bound of a 95% interval for the log-loss difference, estimated by resampling whole weeks, must also be below zero. Monthly production validation uses the same gate when testing alternative rating decay, rating shrinkage and probability calibration. A candidate stays out when the evidence does not meet the rule.

Monthly validation also excludes results that would still be unavailable at the transition from tuning to calibration or from calibration to testing, using the model's three-hour result-availability lag. This prevents overlapping fixtures at a boundary from influencing settings before their outcomes are known.

Experiment: recent form, winning streaks and losing streaks

Club season records

Club and club-season pages count unique completed league results with regular-time full-time scores; future fixtures, missing scores and duplicate fixtures are excluded, and the archive covers the loaded seasons rather than every season. Result points use three for a win and one for a draw, exclude administrative deductions and are not official standings.

Over 2.5 goals means three or more total goals; both teams scored means each club scored at least once. These are historical frequencies, not next-match probabilities. Form trends describe the loaded results without adjusting for opponents or home advantage and use actual goals, not shot-based expected goals. Head-to-head records cover league meetings only, from the first season shown; cup games and other competitions are excluded.

Across 6,846 test matches in six leagues, no recent-form or winning-streak addition met our adoption rule against separately calibrated goals-only controls. Small overall score improvements were insufficiently stable across both test halves. This does not show that successive results are independent.

Test samples and dates

Premier League: 1,140 test matches, 2023-08-11 to 2026-05-24; Championship: 1,142 test matches, 2024-04-20 to 2026-05-02; La Liga: 1,140 test matches, 2023-08-11 to 2026-05-24; Bundesliga: 1,142 test matches, 2022-10-15 to 2026-05-16; Serie A: 1,140 test matches, 2023-08-19 to 2026-05-24; Ligue 1: 1,142 test matches, 2022-12-29 to 2026-05-17.

Retrospective, overlapping historical samples. The form and referee variants share a 54-comparison uncertainty adjustment; their results are not independent replications. Experiment form-referee-v1. Stored report · Link to this experiment.

Test design and full results

Across 6,846 test matches in six leagues, we compared recent points, goal difference and points above the model's pre-match expectation over five and ten games, consecutive wins and losses, and a combination of these signals. Points above expectation help distinguish a surprising run from one the model already expected against weaker opposition.

Test log-loss changes against the calibrated goals-only control; negative is better
LeagueMatchesGoal difference, 5 gamesGoal difference, 10 gamesWin / loss streaks
Premier League1,140−0.001659−0.001574−0.000051
Championship1,142−0.000688−0.000679−0.000809
La Liga1,140−0.000745−0.000599−0.000774
Bundesliga1,142−0.000739−0.000729−0.000037
Serie A1,140−0.001701−0.002098+0.000336
Ligue 11,142−0.000730−0.000597+0.000853

Decision: keep recent form as a research candidate, without changing published probabilities.Recent points and goal difference gave small improvements in every league's overall test score. Points above expectation added less. Streaks were inconsistent and worsened the overall score in Serie A and Ligue 1. Combining all form signals also worsened the Ligue 1 score. Some improvements reversed in the later test half, so an overall improvement alone is insufficient.

Forecast timing, samples and uncertainty

These forecasts use information available 24 hours before kickoff. Results become usable no earlier than 03:00 UTC the following day, and unavailable outcomes are removed at both stage boundaries. Both controls are independently calibrated. All 54 league-and-variant comparisons, including the referee experiments below, are accounted for using a Bonferroni adjustment to weekly block-bootstrap intervals with 20,000 resamples. Every adjusted interval includes zero, and no candidate passes the existing replacement rule. This is not evidence of a reliable improvement in live prediction accuracy.

  • Premier League: 2023-08-11 to 2026-05-24.
  • Championship: 2024-04-20 to 2026-05-02.
  • La Liga: 2023-08-11 to 2026-05-24.
  • Bundesliga: 2022-10-15 to 2026-05-16.
  • Serie A: 2023-08-19 to 2026-05-24.
  • Ligue 1: 2022-12-29 to 2026-05-17.

Experiment form-referee-v1, model dc-v4. The control uses fixed score-model settings and excludes player-availability adjustments. These samples overlap earlier experiments.

Experiment: does knowing the referee help?

Across 2,282 Premier League and Championship test matches, referee identity, prior card tendencies and combined form/referee additions did not meet our adoption rule. No referee adjustment was adopted. This does not establish that refereeing has no effect on matches.

Test samples and dates

Premier League: 1,140 test matches, 2023-08-11 to 2026-05-24; Championship: 1,142 test matches, 2024-04-20 to 2026-05-02.

The archive records the final referee but cannot prove the appointment was available 24 hours before kickoff. The target is match-result probabilities, not card-count prediction or causal refereeing effects. Experiment form-referee-v1. Stored report · Link to this experiment.

Test design and full results

We tested referee identity and lagged yellow/red-card tendencies on the Premier League and Championship, where our test samples have complete referee names. Card tendencies are adjusted for prior team discipline and league home/away card rates. Small samples are pulled toward the league average, and an unseen referee receives no identity adjustment.

Test log-loss changes for win / draw / loss forecasts; positive is worse
LeagueReferee identityPast card tendenciesForm + referee signals
Premier League+0.003832+0.000082+0.002079
Championship+0.002221−0.000044+0.001137

Decision: no referee adjustment to published forecasts. Referee identity worsened the overall test score in both leagues; card tendencies produced changes close to zero with inconsistent results across test halves. The combined form-and-referee model was selected on tuning data in both leagues, then worsened on the final test. This illustrates why a good fit to earlier data is not enough.

The archive records the final referee, without proving that the appointment was known 24 hours before kickoff. These tests assume it was known; a live feature would need dated appointment records and replacement checks. Penalty events are unavailable in these inputs and were not tested. The experiment scores match-result probabilities, not the ability to predict card counts or individual refereeing decisions.

Experiment: do past shots on target help?

Across 6,846 test matches in six leagues, adding each club's lagged shots on target did not meet the adoption rule. These are earlier match totals, not shot-location xG; no current-match statistics were used as pre-match inputs.

Test samples and dates

Premier League: 1,140 test matches, 2023-08-11 to 2026-05-24; Championship: 1,142 test matches, 2024-04-20 to 2026-05-02; La Liga: 1,140 test matches, 2023-08-11 to 2026-05-24; Bundesliga: 1,142 test matches, 2022-10-15 to 2026-05-16; Serie A: 1,140 test matches, 2023-08-19 to 2026-05-24; Ligue 1: 1,142 test matches, 2022-12-29 to 2026-05-17.

The current production results feed lacks the required shots-on-target input. Retrospective statistical association would not by itself establish live availability. Experiment lagged-sot-offset-v1. Stored report · Link to this experiment.

Test design and full results

We added a feature based on each club's last ten observed matches with shots-on-target data, pulling small samples toward the league average. These are match totals, not shot-location expected goals (xG). The comparison uses archived football-data.co.uk matches and a goals-only Dixon–Coles control.

Held-out log loss by league; lower is better
LeagueTest matchesGoals onlyWith past shotsChange
Premier League1,1400.9836600.979413−0.004247
Championship1,1421.0492861.047739−0.001547
La Liga1,1400.9705830.968920−0.001664
Bundesliga1,1420.9859200.984081−0.001839
Serie A1,1400.9839410.981045−0.002895
Ligue 11,1421.0021431.000394−0.001749

Decision: keep the current model. All six leagues showed a small reduction in log loss, with negative weekly 95% intervals. None passed the full replacement rule: the improvement was not greater than 0.005 in both test halves. Our production results feed also lacks the shots-on-target input this feature would need. It remains an offline finding, rather than an input to published forecasts.

Test dates and uncertainty
  • Premier League: 2023-08-11 to 2026-05-24; 95% interval for the log-loss change: −0.006023 to −0.002204.
  • Championship: 2024-04-20 to 2026-05-02; 95% interval for the log-loss change: −0.003034 to −0.000204.
  • La Liga: 2023-08-11 to 2026-05-24; 95% interval for the log-loss change: −0.002959 to −0.000664.
  • Bundesliga: 2022-10-15 to 2026-05-16; 95% interval for the log-loss change: −0.002735 to −0.001009.
  • Serie A: 2023-08-19 to 2026-05-24; 95% interval for the log-loss change: −0.004758 to −0.000909.
  • Ligue 1: 2022-12-29 to 2026-05-17; 95% interval for the log-loss change: −0.002940 to −0.000472.

Experiment lagged-sot-offset-v1. Both controls use the same fixed 365-day decay and rating penalty of 4 before independent calibration.

Experiment: react faster to changing teams and seasons?

Across 6,846 test matches in six leagues, none of the three additions—season scoring, 180-day team-rating decay, or both—met the adoption rule. The existing model was retained in every league.

Test samples and dates

Premier League: 1,140 test matches, 2023-08-11 to 2026-05-24; Championship: 1,142 test matches, 2024-04-20 to 2026-05-02; La Liga: 1,140 test matches, 2023-08-11 to 2026-05-24; Bundesliga: 1,142 test matches, 2022-10-15 to 2026-05-16; Serie A: 1,140 test matches, 2023-08-19 to 2026-05-24; Ligue 1: 1,142 test matches, 2022-12-29 to 2026-05-17.

Settings were selected on tuning data; all variants are shown as retrospective ablations. Coach effects and a historical player-availability layer were not tested, and these samples overlap the other studies. Experiment hierarchy-ablation-v1. Stored report · Link to this experiment.

Test design and full results

We tested three additions separately: an adjustment for the current season's league scoring level, a faster team-strength estimate with a 180-day half-life, and a combination of both. Blend weights were chosen using tuning matches only, with a penalty against unnecessary adjustments.

Decision: keep the current model in every league. None of the three additions passed the replacement rule. For the Premier League, tuning assigned zero weight to every addition, reproducing the control. Elsewhere, differences were very small or did not improve consistently across both test halves and both scoring measures.

Log-loss changes for all three additions
Change against the separately calibrated control; negative is better
League / sampleSeason scoringFaster teamsBoth
Premier League1,140 matches · 2023-08-11 to 2026-05-240.0000000.0000000.000000
Championship1,142 matches · 2024-04-20 to 2026-05-020.000000−0.000084−0.000084
La Liga1,140 matches · 2023-08-11 to 2026-05-240.000000+0.000015+0.000015
Bundesliga1,142 matches · 2022-10-15 to 2026-05-16+0.000000+0.000000+0.000001
Serie A1,140 matches · 2023-08-19 to 2026-05-24+0.0000040.000000+0.000004
Ligue 11,142 matches · 2022-12-29 to 2026-05-170.000000−0.000058−0.000058

Experiment hierarchy-ablation-v1, model dc-v4. Values are rounded to six decimal places. Showing every variant does not mean choosing the best one on the test set.

Weather, rest days and consecutive away games

We also tested precipitation, wind, temperature and humidity in the Premier League, and league-only rest and away sequences across six leagues. None of the 24 weather comparisons or 36 schedule comparisons met the adoption rule. The signals are not used to adjust published probabilities.

These retrospective samples cannot establish that rain or fatigue has no effect. Severe-weather samples are small; cup matches and player workload are missing, and historical input availability is not fully auditable. Read the full weather and fixture congestion study, with charts, samples and downloadable results.

What these experiments can establish

These are retrospective comparisons on historical seasons already studied during earlier model research. The final blocks are held out within each experiment; they are not fresh, prospective confirmation. The controls use fixed model settings, rather than reconstructing every historical production policy. The overlapping samples should not be counted as independent replications.

Coach effects were not tested, and the team-and-season experiment did not have enough archived player availability to test that layer. Our trials of AI-assisted injury-news extraction measure source coverage, player identification and freshness. They do not establish a gain in prediction accuracy and have not replaced the production availability feed. The model record separately reports held-out accuracy against simple baselines and predictions frozen before kickoff.

Player availability

Before kickoff, expected goals are adjusted for players the Fantasy Premier League feed lists as injured, suspended or doubtful. Each player counts by their share of the club's league minutes this season, and a doubtful player by their chance of missing out. The adjustment sizes were estimated on 1,140 matches from 2022 to 2025. Coaches and confirmed lineups are not used. If two sources disagree, the page says availability is unclear. Missing fields stay empty.

Record

A prediction row is inserted, never replaced. Backtests and predictions frozen before kickoff are separate. See the record. Plain-English explanations are in the guides.