Model track record

Outcome accuracy

The share of matches where the model's highest-probability outcome (home win, draw, or away win) was the actual result. A coin flip would score ~33% across three outcomes; the model aims to beat that with calibrated forecasts.

Brier score

The average squared error between the forecast probability vector [P(home), P(draw), P(away)] and the actual outcome (encoded as [1,0,0], [0,1,0], or [0,0,1]). Ranges from 0 (perfect) to 2 (maximally wrong). Lower is always better. Unlike accuracy, it rewards confident correct calls and penalises confident wrong ones.

Skill vs naive

How much the model improves on a naive baseline that just uses the observed home/draw/away base rates for each match. Calculated as 1 − (model Brier ÷ baseline Brier). A positive number means the model beats the baseline; zero means it matches it; negative means the baseline was better on that slice. A good forecasting model consistently posts positive skill.

Log loss

The mean negative log-likelihood of the correct outcome under the forecast distribution. Similar to Brier but penalises overconfident wrong predictions more severely — assigning 1% probability to an outcome that actually happens costs far more than assigning 40%. Lower is better. Shown for completeness; Brier and skill are the primary metrics here.

Backtest vs live

Historical data (Jan 2010 – Jul 2026) uses walk-forward backtesting: each match is forecast using only results that existed before it, with no look-ahead. Live forecasts (Aug 2026 onward) are genuine pre-match snapshots logged in real time. Both are included together in this record.

Sorting the tables

Click any column header to sort by that metric. Click again to toggle ascending/descending. Brier and log loss sort ascending by default (lower is better); matches, accuracy and skill sort descending (higher is better).