Everything on Beat the Press is built from match-level event data with our own models - not stats copied from elsewhere. This page covers what goes in, how each model works, and how we check it. The working for every model, with its numbers, is on the Math Models page.
The foundation is event data for every match: each pass, shot, tackle, aerial and substitution with its pitch coordinates and timing, from the top five leagues, the next tier of European leagues, the English Football League, MLS, the Brasileirão and the major European and domestic cups, from 2024/25 on. Player season tables, line-ups and formations come from the same source, and market values and listed positions from Transfermarkt.
From the raw events we build our own per-player, per-match numbers - progressive passes and carries, box touches, regains, errors that led to shots, aerials by pitch area and dozens more - so every stat on the site is computed the same way in every league.
Bookmaker odds are collected only as a benchmark to score the predictions against. They never go into any model.
xG. A logistic regression fitted on every shot we have gives each shot's chance of being scored, from what's known before it's struck: distance, angle, body part, and how the chance came about (cross, through ball, counter, set piece, rebound). Nothing about where the shot went is used.
P(goal) = 1 / (1 + e^-(b0 + b·features))
Expected points. Each shot is treated as a coin flip with its xG as the chance of heads, so a team's goal count follows a Poisson-binomial distribution, computed exactly. Win and draw chances - and expected points - follow from the two teams' distributions.
Post-shot xG and finishing. For shots on target, a second model adds where in the goal the ball was placed. Finishing = post-shot xG - xG over a player's shots: a much steadier skill signal than goals - xG, which is mostly luck over a season.
Expected threat (xT). The pitch is split into zones and, from every open-play pass, carry and shot, we learn how likely possession in each zone is to lead to a goal. Each completed pass or carry is credited with the threat it adds: xT(end) - xT(start).
Players are ranked within a position pool - strikers, wingers, No. 10s, central and defensive midfielders, centre-backs, full-backs, keepers. Pools come from where a player actually lines up in matches (the share of his minutes in each slot), with Transfermarkt's listed position as a fallback.
Each role (Poacher, Target Man, Ball-Playing CB, Sweeper Keeper...) is a set of per-90 stats. Each stat becomes a percentile in the pool, and the percentiles are combined into the role score. All Round sums a player's roles for his position.
Where a role has a measurable outcome, a regression decides which stats carry weight. Stats that add nothing once the others are known are shown on profiles but not scored.
Stats are scaled by the strength of the competition, so a number in the Premier League counts for more than the same number in League Two. Discipline is exempt.
Rates are pulled toward the pool average until a player has enough minutes, so a hot five-game spell doesn't top a list.
Every team gets an attack and a defence rating, fitted on every match at once - league, cups and Europe - so they're opponent-adjusted and European and cup ties link the leagues together.
home goals ~ Poisson(exp(c + home_adv + att[home] - def[away])) away goals ~ Poisson(exp(c + att[away] - def[home]))
0.7 × our xG + 0.3 × goals, not raw goals - goals alone are noisy.
Matches fade with age: a 365-day half-life in the top five leagues and Europe, 180 days elsewhere, with an extra discount for last season's matches since squads change in summer.
Each team is pulled towards a prior from its squad's market value relative to its league, so a promoted or rebuilt side starts where its squad says rather than at the league average.
Match probabilities use the Dixon-Coles correction for 0-0, 1-0, 0-1 and 1-1, and a stretch factor that restores the gap between strong and weak sides.
The Match Simulator turns two teams' ratings into expected goals and a full score grid, giving win, draw and loss chances and fair odds. The League Simulator plays the rest of each season 10,000 times with each league's real format - playoffs, the Scottish split, MLS conferences, the Champions League phase - keeping results already played. Each simulated season first nudges every team's ratings by their uncertainty, so a team we know less about has a wider range of outcomes.
Settings (how fast old matches fade, how hard ratings are pulled to league level) were chosen on 2025/26 and then scored untouched on 2026/27. On the bookmakers' prices, the honest answer: the market is sharper in the top leagues, and our edge, where there is one, is small and in lower-league totals.
What happens to a team's season when one player changes? We learn it from 530 real squad changes - every team that stayed in its league from one season to the next: how much its xG and xG against moved, against how its line-up changed.
Each player's opponent-adjusted npxG + xA per 90 (shrunk with his minutes), his RAPM attack, and whether he's unproven - players with little previous league data come in below average.
His team's opponent-adjusted xG against while he's on the pitch, aerial defending relative to his league, errors that led to shots, and his RAPM defence.
Goals prevented (post-shot xG faced - goals conceded) per 90.
RAPM (regularised adjusted plus-minus) splits every match into stints between substitutions, goals and red cards - about 155,000 of them - and fits every player at once, so each one's effect on xG for and against is separated from his team-mates and opponents. It was the one addition that clearly improved the defensive model.
The change in xG feeds the same ratings and simulator as the League Simulator, so a swap shows up in points, finish and relegation odds. Checked out of sample, the predictions are the right size (actual-on-predicted slope 0.96-0.98). By this data a top centre-back is worth roughly 2-3 goals a season over a weak one - one player in eleven moves a team less than it looks from the stands.
For passes attempted, shots and fouls, each likely starter's expected count is his recent rate per 90 (recent appearances weighted most, pulled toward his position's average), times his expected minutes, times context from a Poisson regression: his team's level, what the opponent allows, home or away and, for fouls, the referee.
expected = rate per 90 × minutes / 90 × exp(b · context)
The count around that is a negative binomial, so any line gives a chance of More or Less, then calibrated on held-out matches (passes were overconfident and are pulled in). Keepers have their own passes model, half from their season average, since a keeper's passing follows his team's build-up.
Tested on matches since July 2026 against lines near 50/50 at each player's season average. Real sportsbook lines are sharper, so the logbook on the Props page is the true test.
Simulators are tested walk-forward: ratings refitted on only the matches before each week, predicting the next. The Team Builder uses cross-validation that keeps each team's seasons together, and ranking weights come from regressions against real outcomes.
Models never take market prices as an input; odds are only a yardstick.
Ideas that don't beat the simpler model on unseen data are dropped - and recorded. Rejected so far include game-state adjustment, priors on RAPM, aerial win %, dribbled-past counts, midfield progression stats in the Team Builder, fatigue and style matchups in the simulators.
Fits include an intercept for season-to-season drift, so a model can't look better just by soaking up a league-wide trend.
Every Tuesday the whole pipeline reruns: