How to evaluate football prediction accuracy

Hit rate (how often the top pick wins) is easy to read but ignores confidence: a 51% call and a 90% call count the same. Brier score (Brier, 1950) is the mean squared error between the probabilities and what happened. For a 1X2 forecast, sum the squared errors over the three outcomes. Illustrative example (fictional): forecast 50/27/23 and the home team wins → (1−0.5)² + (0−0.27)² + (0−0.23)² = 0.25 + 0.0729 + 0.0529 = 0.3758. Lower is better. Log loss is −ln(probability given to the actual outcome): here −ln(0.50) ≈ 0.693. It punishes confident misses heavily. Always compare with a baseline such as the bookmaker closing favourite. In WinSight's registered backtest of 16,322 matches the model picked the winner 50.2% of the time against 52.4% for the Bet365 closing favourite — we publish that gap. Keep each model version separate; pooling versions hides weak periods.