How our football predictions work

1. Data: where the numbers come from

• Historical results come from openly licensed data (public domain), at least four seasons per league.
• Fixtures and kickoff times come from official league and club sites, checked against a second major source.
• Results need the same 90-minute score on the official site and one major outlet. If they disagree, the match is flagged and not graded until it is resolved.
• Team news comes from club statements, press conferences and reputable outlets. We keep the fact and the link, never copied text.
• Point in time: when we fit the model for a match, we only use results that were in our system before the data cutoff, so a forecast never sees the future.
• No betting odds in version 1. Free odds feeds do not allow public or commercial use.

Every source and its licence

2. The statistical model: Elo and Dixon-Coles

Two well-tested models each give a probability for home win, draw and away win. We blend them and calibrate the result, so that our 60% picks win about 60% of the time.

log λH = b + h + attH − defA
log λA = b + attA − defH
P(x, y) = τ(x, y) · Pois(x; λH) · Pois(y; λA)

• Recency: older matches count less (a half-life of about a year in our backtests), and small samples are pulled towards the league average. Promoted teams start from a league prior and carry a “small sample” label. Clubs moving between the Premier League and the Championship keep their rating instead, adjusted for the gap between the two leagues.
• Elo updates both ratings after every match, with a goal-difference factor, and converts the rating gap into home, draw and away probabilities with a draw parameter.
• Blending: p_stat ∝ (p_DC^w · p_Elo^1−w)^1/T. The weight w is chosen on validation seasons. The temperature T is set for each league on that league’s own validation seasons and pulled towards a shared value when the league has little history. The current values belong to the model version listed below.
• One grid for every market: the score grid gives the most likely scores, Over/Under 2.5 and both teams to score, so they always agree with the result probabilities.

3. What the AI analysts do

Claude and Codex read the same file for every match: sourced team news, the model’s expected goals and its probabilities. Then they:

• write the analysis on the match page, where every claim needs a source, and an editor checks it before it is published;
• give their own home, draw and away probabilities;
• may propose an adjustment of up to ±15% to a team’s expected goals, only for sourced news the model cannot see: a missing key player, rotation before a cup game, a team with nothing left to play for.

In version 1 all of this runs in shadow: the AI view is locked and graded like everything else, but it does not change the published probability. AI never edits a locked forecast.

4. When the AI layer can count

Today the published probability is the statistical model alone. The AI layer gets a weight (at most 0.5) only after at least 300 graded matches and 6–8 weeks of live running, and only if all of these hold on the same matches:

• better log loss and better RPS than the statistical model;
• the 95% interval of the difference, bootstrapped by round, excludes zero;
• the improvement holds in at least two thirds of the leagues.

If it passes, the model version goes up and the change is listed below. If not, it stays in shadow.

5. How we choose a model

• Time order only: fit on earlier seasons, choose settings on the next one, and report on a final test set that was never used for choices.
• Benchmarks: league base rate, always the home team, independent Poisson and Elo on its own. A candidate that cannot beat them is not used.
• Scores: log loss and RPS out of sample come first; Brier score and calibration are checks.
• Fair comparison: same matches, same cutoff, same 90-minute target, with uncertainty from a bootstrap by round.
• Simple wins ties: if an improvement is not stable, we keep the simpler model.

Backtest results

6. Locking

• Every forecast locks 24 hours before kickoff, or straight away if the match enters our tracking later than that; the match page shows the actual lock time, and every forecast locked before kickoff is graded. The file gets a SHA-256 hash and is committed to our public ledger straight away; the hash and time appear on the match page.
• A locked forecast is never edited. Mistakes are fixed with a correction record, listed below.
• If a match is moved, the old forecast is voided but kept, and a new revision is locked before the new kickoff.
• A forecast made after kickoff is labelled late and not graded.

How to verify a forecast

7. Grading

• We grade the 90-minute score, stoppage time included. Extra time and penalties are recorded separately.
• Postponed, abandoned, cancelled and awarded matches are not graded, and neither are late forecasts.
• If a source corrects a result, we grade again under a new revision with the reason, source and time, and keep the old record.

RPS      = ½ · Σk=1..2 (Σj≤k pj − Σj≤k oj)²     order: home, draw, away
Log loss = −(1/N) · Σ ln p(actual result)
Brier    = (1/N) · Σi Σk (pik − oik)²
Coverage = published eligible forecasts / eligible fixtures

Track record

8. What we don’t claim

• This is not betting advice. We make no profit claims and link to no bookmaker. 18+.
• Probabilities are not certainties: a 70% pick still loses about 3 times in 10.
• We make no accuracy claims until we have 6–8 weeks of live results; until then numbers are marked as an early sample.
• Models miss things: news after the lock, weather, motivation, a red card in the 5th minute.

Model version history

Record changes

No corrections so far. Every change to a graded result will be listed here, newest first, with the reason and the source.

FAQ