World Cup Forecasting
Can squad value and historical form improve pre-match win, draw, and loss probabilities?
Research question
How accurately can international football outcomes be forecast when every prediction uses only information available before the match?
Method
- Estimated Elo–Davidson and Dixon–Coles models with venue-aware home advantage and exponential time decay.
- Added squad-strength features derived from player valuations and ratings.
- Refit models yearly using prior matches only, then evaluated log loss, Brier score, and expected calibration error.
Result
Across 11,879 walk-forward test matches, the primary Dixon–Coles model achieved
0.8976 log loss versus 1.0518 for the base-rate benchmark, with 0.0078
expected calibration error.
Boundary
The published table records a historical evaluation run. Reproducing the exact numbers requires the original frozen data snapshot because the upstream match dataset changes over time.