Methodology

What the models use, what a model score means, and what it does not mean.

Statistical inputs

Models are built on observed football data rather than opinion. The inputs include team scoring and conceding rates split by home and away, recent form, expected goals (xG) for and against, clean-sheet and both-teams-to-score rates, corner averages, and league standing. Data is drawn from a commercial football data provider and refreshed on a schedule.

Market context

Model output is read against the prices available for the fixture. The market aggregates a great deal of information and is a strong baseline; Punter Insight treats it as context to compare against, not as a benchmark it claims to beat.

What a model score means

A model score or probability is the model's estimate of how likely an outcome is, given the data it was given. Some markets additionally carry a calibrated probability, which adjusts the raw score using how that model's past estimates actually resolved.

A model score or probability is not a guaranteed outcome and not a guaranteed return. A 70% estimate is expected to be wrong roughly three times in ten. A run of losses at high model scores is normal and does not by itself indicate the model is broken.

Different model families

Punter Insight does not run one model. Different markets use different approaches — rate-based models, Poisson-style goal models, form and signal models, and machine-learning classifiers — because the structure of each market differs. Each prediction is stored with the version of the model that produced it.

Pre-match by design

Every prediction is produced and published before kickoff. Nothing is generated or adjusted in-play, and nothing is added retrospectively. The price shown is the price recorded when the prediction was published.

Sample size and uncertainty

Football has high outcome variance. Short-run results are dominated by noise: a 60% win rate over 20 selections is entirely consistent with a model that is no better than the market. Punter Insight therefore publishes the sample size next to every figure, including for individual AI Traders, and treats small samples as uninformative rather than promotable.

Percentages on this site are historical records of settled selections. They are not forecasts, and past performance does not indicate future results.

What is not published

Punter Insight publishes its record, not its implementation. Model thresholds, feature weights, selection cut-offs, trader configurations and internal strategy parameters are not disclosed. This is deliberate: the evidence is verifiable without the implementation being public.

See also how it works and how we verify results.