Plain-English reference

Football prediction and model glossary

The exact meaning of the terms used in our forecasts, testing and public record—without hiding behind model jargon.

12 definitions

From 1X2 to walk-forward testing

1X2 probability: A three-outcome forecast for a football match: 1 is a home win, X is a draw and 2 is an away win. The three probabilities should add to 100%.
As-of join: A data join that only uses information known at the prediction timestamp. It prevents later results, ratings or revisions from leaking into a historical forecast.
Append-only prediction ledger: A record in which a published prediction is inserted once and protected from update or deletion. Later results and corrections are added as separate events instead of rewriting the original call.
Brier score: A probability-quality score based on squared error between predicted probabilities and the result. For this site's three-class 1X2 implementation, the class errors are averaged so lower is better on a 0–1 scale.
Calibration: The agreement between stated probability and observed frequency. Across a large enough sample, outcomes forecast at 60% should occur about six times in ten.
Data leakage: Any use of information that would not have been available when a prediction was made. Leakage can make a backtest look strong while producing an unusable live model.
Elo rating: A continuously updated team-strength rating. Match results move each team's rating, while home advantage and season-to-season regression can be modelled explicitly.
Fair odds: The decimal price implied by a model probability before bookmaker margin: fair odds equal 1 divided by probability. A 50% probability therefore implies fair odds of 2.00.
Isotonic calibration: A non-parametric method that adjusts raw model scores into probabilities while preserving their order. It is fitted only on past validation data, never on the matches being evaluated.
Lead time: The interval between publication and scheduled kick-off. Football Proof AI records it so readers can distinguish genuinely pre-match predictions from late or retrospective claims.
Settlement: The automated act of connecting a locked prediction to a verified final result, then deriving its hit or miss and Brier contribution without changing the original probabilities.
Walk-forward backtest: A time-ordered evaluation in which a model trains on the past and predicts a later period, then moves forward and repeats. It better resembles live use than a random train-test split.