How Our Match Prediction Model Works

A match prediction model turns team data into a probability for each outcome of a football match. Every percentage you see on TACTIX today, such as “Man City: 72% to win”, comes from one such model, and a calibrated 60% should come true in about 60 of 100 comparable matches. This article explains what goes in, how the model works and how it is tested.

What Data Goes Into the TACTIX Match Prediction Model?

• Recent form — results over each team’s last 10 matches; by default the input with the most weight
• Chance quality created and conceded — expected goals (xG), the quality of the chances a team makes and allows
• Shots on target, possession and goal difference — smaller shares of each team’s strength score
• Squad availability — known injuries, the smallest default share
• Elo ratings — a running team rating that rises and falls with every result, used by the match-result model
• Head-to-head and home advantage — the last 5 direct meetings give a small adjustment, alongside a home-advantage allowance, in the base model’s own match-result call
• The referee — the appointed referee’s card record adjusts the cards markets only

The published model does not use weather, confirmed line-ups or match stakes such as a cup tie or a European week. Late team news is therefore not reflected in the published probability, so it is worth checking the line-ups before you act on any prediction.

How Does the Model Turn Data Into Probabilities?

The six statistics combine into a strength score for each team. Each side’s attack is set against the other’s defence to give expected goals for the match, and a Poisson distribution, the standard way to model how often a rare event such as a goal happens, turns those two figures into a probability for every scoreline. A Dixon–Coles correction then adjusts the low scores, such as 0–0 and 1–1, that a plain Poisson model misprices. Goal markets are read from that grid; corners, cards, shots and player markets come from companion count models.

Two learned layers sit on top. A stacking model trained on settled results sets both teams to score and over/under 2.5 goals, and a three-way match-result model, which weighs the base model’s outputs with Elo ratings, form, rest days and rating trend, sets the headline home / draw / away probabilities. The full step-by-step is on the AI prediction models page.

How Is the Model Tested?

Every change is validated out of time: trained on earlier matches and tested only on matches that came after the training data, to prevent hindsight bias. A learned layer is used only where it beats the base model on matches neither has seen. Each market is then calibrated, its probabilities checked against how often those outcomes actually happened, and the curves are refitted every night from recently settled matches, so a stated probability keeps its meaning as the season moves on.

We publish a measured track record — settled hit-rates by market and confidence tier, plus a calibration curve — on the model scoreboard, the How accurate is TACTIX? page and at how accurate is TACTIX. Calibration is our headline measure, and every pick is fixed before kick-off in the public pick ledger.

How Is a Model Probability Different From Bookmaker Odds?

Bookmaker odds carry a margin, so the implied probabilities of every outcome in a market add up to more than 100%. The model’s probabilities carry no margin: they are its estimate of how often each outcome happens. Where the two disagree, the gap is an opinion to weigh, not a sure edge. Whether the model reads matches better than the market over time is a question for the public record, not for this article: the scoreboard grades every settled pick, losing ones included, so you can judge that for yourself.

18+ | BeGambleAware.org. Predictions are probabilities, not betting advice. Please gamble responsibly: free help is at BeGambleAware.org.