Methodology
How ProbaBall Predictions Work
Overview
Every ProbaBall forecast starts from the same place: a statistical model that gives each team an attack rating and a defence rating, estimated from two seasons of results. Those ratings produce a full probability distribution over exact scorelines for every fixture — the win/draw/loss percentages and expected scores you see on the site all follow from it.
Season forecasts then come from playing out the rest of the season thousands of times in simulation and counting how often each team lands in each final position.
Data Sources
Match Results
Scores from the current and previous two seasons, weighted so recent matches count for more.
Fixtures & Standings
The remaining schedule and current table, which the season simulation plays out to a finish.
Player Statistics
Goals, assists, minutes and match ratings behind the player tables. These inform the player pages only — the match model works purely from team results.
Nothing is published about a season's players until matches are played, so before kickoff these tables show last season's figures for the players registered in this season's squads, labelled as such. The Projected column is separate: expected minutes and goals for the coming season, from each player's share of his team's output combined with what the simulation expects that team to score. Goalkeeper clean sheets come straight from the match model. Ratings and assists are not projected — they carry too little from one season to the next to be worth publishing.
Bookmaker Odds
Used only as a benchmark when evaluating the model, never as an input to it.
The Match Model
Ratings
ProbaBall uses a Dixon-Coles model, a well-established approach from the football-modelling literature. Each team's expected goals in a match come from its attack rating against the opponent's defence rating, plus a league-wide scoring baseline and a home advantage. Goals are then modelled as Poisson counts, with a correction that gives low-scoring draws (0-0, 1-1) the extra weight they have in real football.
- Time decay: every match is weighted by its age, with a half-life of just under a year — recent form counts for more, but old evidence never vanishes abruptly
- Priors: ratings are pulled toward a prior so a few surprising results don't swing a team's rating wildly; after roughly twenty matches the current season dominates
Starting a New Season
Forecasts are available from matchday 1 because each season starts from what the model already knows:
- Returning teams carry over last season's ratings, shrunk toward the league average — squads change over the summer, and some of last season was luck
- Promoted teams start at replacement level: about as strong as the sides that just went down, which is where promoted teams historically finish
- As results arrive, the carried-over knowledge gradually washes out in favour of this season's evidence
Season Simulation
The remaining fixtures are played out thousands of times, with a real scoreline drawn for every match. Points, goal difference and goals scored accumulate exactly as in a real league table, so tiebreakers behave properly. Position, qualification and relegation probabilities are simply how often each outcome happened across all simulated seasons.
The simulation deliberately tracks two kinds of uncertainty:
- Match randomness — even the better team loses on the day; every simulated match samples a scoreline rather than assuming the most likely one
- Rating uncertainty — how good each team really is. The simulation runs across many "rating worlds", and teams the model has watched less (promoted sides especially) get appropriately wider ranges
How We Check It
The model is evaluated by walk-forward backtesting: replaying past seasons matchday by matchday, refitting the ratings using only the results that were available at the time, and scoring the predictions against what actually happened.
- Scored with proper scoring rules (log loss, Brier score, ranked probability score) that reward honest probabilities, not just picking the favourite
- Compared against baselines, including de-vigged bookmaker odds — the strongest public benchmark available
- Season projections are checked for calibration: when the model gives a team an 80% range for its final points, the team should land inside it about 80% of the time — and in backtests over 14 league-seasons, it does
Limitations
Honest forecasts require honest caveats:
- Ratings are fixed within a simulated season, while real teams drift — so the true spread of outcomes is slightly wider than shown
- Transfers and injuries only register through results. A team that rebuilds heavily over the summer will be mispriced until its matches say otherwise
- No fixture context: European fatigue, derbies and manager changes are not modelled
- Head-to-head tiebreakers (used by Serie A) are not modelled; ties are broken on goal difference
Our predictions are calibrated probabilities, not certainties — a 70% favourite should still lose about three times in ten.