How the forecasts work
Methodology
Definitions, assumptions and practical limits of the ratings and season forecasts.
What a rating means
The 0–100 rating is expected points share against a fixed portfolio of 100 reference opponents on neutral ground. It summarizes attack and defense; it is not a league standing or a probability of beating every lower-ranked club. Matchups and season simulations use the full attack/defense pair.
How the model learns
A Dixon–Coles goal model estimates each club’s attack and defense and each competition’s home advantage. Results decay at 0.002 per day: roughly a one-year half-life. Where matched Understat xG is available, the actual score contributes 40% of a match’s weight and an xG observation contributes 60%. Squad values provide a soft within-league prior. No historical squad-value snapshots are available, so retrospective fits exclude that signal.
What connects leagues
Continental matches connect domestic leagues. Connections between confederations are sparse and heavily affected by time decay. A decade of archived matches does not equal a decade of equally strong evidence. Small differences between clubs from different confederations should not be read as precise measurements. Weighted match counts describe evidence volume; they are not statistical confidence intervals.
Season uncertainty
Each simulated season draws persistent attack and defense perturbations for every club. Domestic scorelines include the Dixon–Coles correction even when ratings are perturbed; UEFA knockout rounds retain the same strength perturbations as their league phase. The current noise scale is 0.12. Corrected preseason and midseason checks show mixed benefits: preseason bottom-position accuracy improves, while title/top-N evidence is inconclusive and midseason distributions may be too wide. This setting remains provisional pending forward evidence.
Competition rules and limits
Each competition page lists its scope and simplifications. Qualification-place columns represent specified league finishing positions; they exclude cup cascades, titleholder routes, performance spots and qualification rounds. Relegation playoffs use a fixed survival estimate. Head-to-head mini-tables are included where configured, with remaining subgroup and special deciding-match rules approximated. MLS projections concern the regular season and playoff entry, not the MLS Cup winner. The Championship uses its expanded 2026–27 playoff field. Domestic cups retain simplifying assumptions about undrawn rounds and lower-division strength. Completed first-leg scores are retained; extra time uses proportional scoring rates and penalties are a 50/50 split.
What-if scenarios
Matchday scenarios select simulated seasons in which a fixture ends in a home win, draw or away win, then summarize those same seasons. They condition on the outcome without choosing a scoreline or refitting the ratings. Rare outcomes with fewer than 100 matching simulations are withheld. Impact is the largest probability difference across these outcomes for either club’s listed season events; small swings may be Monte Carlo noise.
Validation and the archive
Match backtests freeze ratings at quarterly cutoffs and compare forecasts with available bookmaker odds, closing where available. They therefore differ from the live hourly update schedule. Season backtests use exact preseason and midseason cutoffs and exclude incomplete round-robin seasons. Reconstructed finishing-position events are proxies, not official relegation or qualification outcomes. Historical windows have been studied before and are not a newly untouched holdout. The forward forecast archive retains the first prediction recorded before kickoff and scores it when a regulation-time result becomes available.
Freshness and changes
Hourly refresh checks update available feeds; a generation timestamp is not proof that every source is current. The data coverage page lists included observations and unresolved fixtures. Postponed fixtures stay in the remaining season unless a replacement supersedes them. A release is generated and checked in staging before replacing public files. Movers only compare snapshots sufficiently close to the selected time window and never compare different rating versions. A recorded result is context for movement, not proof of its sole cause.
Sources and local preferences
Football-data.co.uk supplies historical results and available odds; ESPN supplies continental matches and current fixtures/results; Understat supplies xG; Transfermarkt supplies squad values. API-Football currently provides comparison data and squads rather than an automatic results fallback. Theme choices and aggregate interaction counters stay in your browser; this implementation sends no behavioral analytics to a server.
Validation results · Data coverage · Forward forecast archive