Football statistical models

Lens adapted to this topic: What established research and practice support

The mainstream view treats football models as useful probabilistic tools whose value depends on sound data, appropriate statistical assumptions, calibration, and out-of-sample validation. It favours interpretable and empirically tested models, while accepting that no model removes match uncertainty. Different methods can be appropriate for different tasks, from shot evaluation to league projection and tournament simulation.

• A practical modelling pipeline estimates attack and defence, incorporates factors such as home advantage, derives a scoreline distribution, and simulates remaining fixtures to produce probabilities.
• Interpretable models using FIFA ratings, team formation, player quality, and historical statistics can provide both predictive performance and information useful to coaches and players.
• A Bayesian mixed-effects xG model using seven shot and context variables achieved performance similar to the cited StatsBomb model, suggesting that simpler models can remain competitive.
• Bayesian dynamic goal models can allow attacking and defensive abilities to evolve over time, adapting more quickly when team performance changes substantially.
• Several model families, including Poisson, bivariate Poisson, Skellam, Student-t, machine-learning, and ensemble approaches, can be fitted and evaluated for out-of-sample prediction.

Lens adapted to this topic: Limits, failure modes, and model disagreement

The sceptical view argues that football models are often presented with more confidence than their data and validation justify. It emphasizes distribution shift between club and international football, small samples, changing team form, differing xG definitions, market comparisons, and the way modelling choices compound into divergent forecasts. This perspective does not reject modelling, but narrows the claims that should be made from it.

• Models trained largely on club football may transfer poorly to international tournaments because teams play fewer matches, squads change, and the data-generating environment differs.
• Tournament prediction is particularly vulnerable to small samples, unusual events, and distribution shift, so high-confidence forecasts can be misleading even when simulations are technically correct.
• There is no universally perfect xG model: accuracy, bias, context, provider definitions, and evaluation criteria can produce different conclusions about which model is best.
• Projection models can disagree substantially because they make different choices about team ratings, recent-form weighting, match distributions, home advantage, and simulation assumptions.
• A model may outperform a benchmark in a particular historical sample or betting exercise without establishing durable profitability or general superiority in other competitions.