« How the models work

Logit

column: Logit votes in H:D:A

The one model on the site that is learned rather than hand-set: a multinomial logistic regression trained on six past seasons of results, fed only things known before kick-off. It reads each side's rolling form and outputs home/draw/away probabilities.

How it works
  1. For every team and every game, build a running average over the team's last 25 games (home and away mixed) of: goals for and against, expected goals for and against, and points per game. Five numbers per side; last season's games seed the window in August, and a short history is blended towards the league mean so two games cannot pass for form. (Shots and corners were dropped once xG was available - on the holdout, goals + xG + points beat the shot-based set, and adding xG on top of shots was worse than either.)
  2. A fixture's feature row is the home side's nine numbers, the away side's nine, and which division it is. Nothing from the game itself - never its own shots or half-time score.
  3. Train a multinomial logistic regression (three classes: H, D, A) on every played game from the previous six seasons across all five divisions, about 8,000 games, skipping each team's first three games of a season. Features are mean-imputed (xG exists from 2024-25 for the EFL and Premier League, from 2025-26 for the National League) and standardised; light L2 regularisation.
  4. Predict: the model returns three probabilities for each fixture, shown as home:draw:away percentages. It is retrained from scratch every morning as part of the pipeline.
Worked example
Home side, last 251.8 scored (1.7 xG), 0.9 conceded (1.0 xG against), 2.0 pts/game
Away side, last 251.1 scored (1.2 xG), 1.5 conceded (1.6 xG against), 1.1 pts/game
Modelweights those against each other plus the division's baseline home advantage
Output58:24:18 → home win
How BetChair calls it

The largest of the three percentages is the vote. Held out on the whole of 2025-26 (trained on the five seasons before it), the favourite was right 47.7% of the time with a log-loss of 1.053 - better than always backing the home side (~43%), short of the bookmakers' favourite (~50%). Honest, not magic.

Where it shines
  • Learned from data rather than thresholds someone picked; calibrated probabilities.
  • Uses expected goals as well as goals - a side creating chances but not scoring shows up here before it shows in results.
  • Carries form across seasons and divisions, so promoted and relegated clubs are not blank slates.
Where it falls down
  • Rolling averages are a blunt instrument; it knows nothing about opponents' strength beyond what shows in the numbers.
  • National League xG only exists from 2025-26, so its predictions lean more on goals.
  • Retrained daily on the same history, so it will not change much within a season - the features move, the weights barely do.

See it on the fixtures table (tick the column under Columns → Predictions) and on every tips slip. Related: Poisson, Market & Value.