← All projects

Forecasting / Applied ML

FIFA World Cup Forecasting

A two-stage probabilistic forecasting system for the 2026 World Cup, combining a Dixon-Coles match model with a LightGBM calibrator and a Monte Carlo tournament simulation.

View repository

Problem

International football outcomes are hard to forecast: teams play few matches a year, strength shifts between tournaments, and draws happen more often than a naive scoring model predicts. This project builds a forecasting system for the 2026 FIFA World Cup designed to be evaluated on held-out matches, not judged by how a narrative reads after the fact.

Contribution

Personal project, designed and built solo: data pipeline, statistical model, calibration layer, and tournament simulation.

Approach

A two-stage pipeline:

  1. Dixon-Coles bivariate Poisson model, fit by maximum likelihood on roughly 32,000 international matches (1990 onward), with a context-dependent correction for low-scoring outcomes and a fitted home-advantage term.
  2. LightGBM calibrator layered on top, using squad market values (Transfermarkt) and a confederation-aware Elo rating to correct for team-strength shifts the core model is slow to pick up.
  3. Monte Carlo tournament simulation (100,000 runs) over the official 48-team bracket — groups, third-place qualification, extra time, and penalties — to produce round-by-round advancement probabilities.

Supported results

Evaluated on a held-out test set of international matches from 2022 onward (n = 4,552) that no part of the model saw during fitting, scored by mean per-match log-loss against a uniform three-outcome baseline (log-loss 1.0986):

Model Test log-loss vs. uniform baseline
Dixon-Coles alone 0.9261 +15.7%
Dixon-Coles + LightGBM calibrator 0.8711 +20.7%

The improvement from adding the calibrator was statistically significant under a paired bootstrap test on per-match log-loss.

Limitations