Bayesian Models for Sports Data

Small samples are the norm. A player has a handful of games, a team has a season, and a specific matchup has almost no history. Frequentist point estimates from tiny samples are extremely noisy, and the usual fix, regression toward the mean, is exactly what a Bayesian hierarchical model does by construction.

Structure is nested. Players sit within teams, teams within competitions, matches within seasons. Hierarchical models represent that directly, letting each level inform the others. A player with twenty minutes played is estimated mostly from the team and league distribution; a player with three seasons is estimated mostly from their own record.

Uncertainty is the answer. Many sports questions are really about how confident we should be. A posterior distribution answers that natively, where a point estimate plus a standard error often does not communicate it usefully.

Prior information exists. Analysts genuinely know things before seeing data: scoring rates fall in known ranges, home advantage is positive but modest, and player ability does not change wildly between weeks. Encoding that is legitimate and often improves estimates substantially.

A typical model. Team attack and defence parameters drawn from league-level distributions, a home advantage term, and a likelihood appropriate to the sport's scoring process. That structure, with partial pooling across teams, is the standard starting point in published football modelling work.

Partial pooling in practice

Probabilistic programming. Stan is the most widely used engine for this work, with interfaces from R and Python. In R, brms provides a formula-based interface that compiles to Stan and covers most hierarchical models without writing Stan code directly, while rstanarm covers common cases. PyMC fills a similar role in Python. Sampling is usually done with Hamiltonian Monte Carlo variants.

A workable sequence.

1. Write the model, starting simpler than you think you need.
2. Simulate data from the priors alone and check that it looks plausible for the sport. Implausible prior predictive output means the priors are wrong, and this check costs minutes.
3. Fit to real data.
4. Check diagnostics before interpreting anything: divergent transitions, effective sample size, and convergence statistics across chains.
5. Run posterior predictive checks, comparing simulated results against the observed distribution of scores and margins.
6. Compare against a simpler baseline using out-of-sample predictive performance.

Cost. Fitting is slower than fitting a regression, sometimes by orders of magnitude. For large datasets and simple questions, penalised regression or gradient boosting may deliver similar predictions far faster, and the Bayesian machinery earns its cost mainly when uncertainty and hierarchy matter.

Priors are assumptions. State them, and refit with reasonable alternatives to see how much the conclusions move. A result that depends heavily on a prior nobody can justify is not a finding.

Diagnostics are not optional. Divergences and poor convergence mean the sampler did not explore the posterior properly, so the numbers are not trustworthy regardless of how sensible they look.

Posterior predictive checks. Simulate matches from the fitted model and compare distributions of goals, margins, and totals against observed data. Mismatches in shape, particularly clustering on common scores and the size of the tails, show up clearly here.

Out-of-sample evaluation, ordered by time. Fit on earlier seasons, evaluate on later ones. Random splits leak future information in a sport with strong time structure.

Calibration. Group predicted probabilities and compare against observed frequencies. Bayesian models can be well specified and still poorly calibrated if the likelihood is wrong.

Compare against the market. For any model producing match probabilities, closing prices are the benchmark worth beating. Systematic disagreement in one direction usually indicates a missing term, not an edge.

Version everything. Model code, priors, data snapshot, and sampler settings. Posterior summaries are not reproducible without them.

This page describes data and method and is not betting advice.