Methodology

1. 01

   Team strength

   Every team carries an Elo rating updated after each match. Margin of victory scales the update, so a 4-0 moves the rating more than a 1-0. Home advantage is a fixed offset per sport.

2. 02

   Match probabilities

   For football we fit a Dixon-Coles Poisson goal model on top of the ratings to get home, draw and away probabilities. Other sports use the Elo win expectancy directly, so there is no draw bucket.

3. 03

   Market blend

   Where bookmaker odds exist we remove the vig and blend the market-implied probability with the model. Backtests showed the market alone beats the model on match winner, so the blend leans on the market and the model adds value on totals and handicaps.

4. 04

   Publish gate

   A pick is only published when it clears a confidence threshold, and never against the market favourite unless the model is above 60%. Weak contrarian picks lost money in backtests, so we do not show them.

5. 05

   AI panel

   Independently of the statistical model, several large language models are asked to predict the same fixtures before kick-off. Their picks are timestamped and graded exactly like ours, and shown side by side on the accuracy page.

6. 06

   Grading

   Once a result is final each pick is marked hit or miss. Postponed or abandoned matches are not graded. Predictions created after kick-off are excluded from every statistic.

Full benchmark notes, calibration curves and data sources are published on Scorebase, which runs the models.