Methodology · Premier League
How the Premier League model works
Title, top-four (Champions League), and relegation odds for all 20 clubs.
- Version
- v1 (independent-Poisson, regression 0.5)
- Updated
- Weekly (nightly in season)
- Data source
- openfootball (history) + football-data.org (live) — free, no quota
- Calibration
- Reproduces recent real tables; regression tuned by a 3-season backtest
The question
Where all 20 clubs finish — who lifts the title, who takes the four Champions League places, and which three go down.
The data
Free and open: openfootball's historical match results seed the ratings, and football-data.org supplies live fixtures and scores (10 requests/minute, no monthly cap, cached nightly).
How strength is measured
Each club gets an attack rating and a defense rating from goals scored and conceded, expressed relative to the league average and shrunk toward the mean so a small early sample can't run away.
Recency & regression
For the preseason projection each club is regressed halfway (50%) to the league mean. A backtest — train on season T-1, predict season T, across three seasons — showed heavier weighting on last season (e.g. 70%) was overconfident on Brier and log-loss, so 50% is the tuned default.
Injuries & roster
Big transfers and news enter as a documented strength overlay (a dated delta with a reason) rather than player-by-player modeling. Promoted clubs, with no top-flight data, are priced off the relegation-zone prior.
Home advantage
A home scoring bump is baked into the Poisson goal rates, so home sides both score a little more and concede a little less.
The simulation
An independent-Poisson match model turns two clubs' attack/defense into a full home/draw/away distribution (draws are ~25% of matches and modeled explicitly), scored 3-1-0. A Monte Carlo plays all 380 fixtures thousands of times, producing each club's finishing-position distribution — the table as a cloud, not a single line.
Postseason format
None — the Premier League is decided by the 38-game table. The stakes are the title race, the top-four Champions League cutoff, and the bottom-three relegation line.
Backtesting
The data pipeline reproduces recent real final tables, and the regression strength was chosen by the T-1 → T backtest above rather than by eye.
Known weaknesses
No player-level or lineup modeling; transfers are approximated as club-level deltas; and newly-promoted sides are the hardest to price because there's no recent top-flight signal.
Every published number is timestamped to the Ledger the moment it's made and graded against results — nothing here is backdated. Percentages are rounded for readability.