Poisson Regression for Football Predictions

Fixed rate versus regression

A basic Poisson calculation starts with a supplied rate, lambda. Poisson regression estimates that rate from features. OpenStax's Poisson distribution guide defines the distribution and its rate parameter:

log(lambda) = beta0 + beta1 x home + beta2 x team strength + ...

The exponential transformation keeps lambda positive. OpenStax's Poisson guide provides the underlying count distribution; regression adds a feature-to-rate model.

Define the inputs

Illustrative rate calculation

Suppose the fitted linear predictor for a home team is log(1.40). Exponentiating gives lambda = exp(log(1.40)) = 1.40 expected goals under the model. That rate can feed a one-team goal distribution. Building an independent home-away score grid adds another assumption about dependence.

Dixon and Coles provide a football-specific score-model reference and a low-score dependence adjustment. Their model is evidence for a method, not a universal parameter set.

Fit and validation sequence

1. Freeze fixture and feature availability timestamps.
2. Estimate parameters on earlier matches.
3. Tune decay, pooling and regularisation on later validation blocks.
4. Inspect residuals, mean-variance relationships and low-score cells.
5. Lock the model before the final future period.
6. Compare goal and outcome probabilities with simple baselines.

TimeSeriesSplit documents ordered validation. Football outcome-model research supports comparing alternatives under a stated sample.

Assumption audit

Keep the basic model in every comparison. Football benchmark research supports reporting the target, data and evaluation protocol beside model scores.

Derive an illustrative rate safely

Suppose an illustrative log-rate model is:

log(lambda_home) = intercept + home term + attack strength − opponent defence

If the linear predictor equals log(1.60), the expected home-goal rate is exp(log(1.60)) = 1.60. That rate can feed a Poisson probability calculation. It is not a statement that the team will score 1.6 goals or that the fitted covariates are causal. OpenStax's Poisson guide supports interpreting lambda as a distribution rate rather than a deterministic outcome.

OpenStax's Poisson guide defines the rate and probability mass function. Dixon and Coles provide a football-specific score-modelling method with explicit assumptions and a low-score adjustment.

Diagnose the fitted model

Compare observed and predicted score frequencies, especially 0-0, 1-0, 0-1 and draws. Check whether variance materially exceeds the mean, whether residuals cluster by team or season, and whether home and away goals remain dependent after the covariates. Evaluate scoreline probabilities as well as aggregated match-result and totals probabilities.

Refit through historical cutoffs and preserve the team-strength estimates available at each forecast. Using a season-end attack estimate for an early-season fixture leaks later results. Promoted teams need a declared prior or pooling rule, not a retrospectively convenient starting value.

Continue the workflow

Use the football Poisson distribution guide when the rates are already supplied and the task is to build and audit the score grid.

Continue learning

• Next guide: SportSignals Prediction Methodology and Attribution
• Related guide: Training vs Testing Football Prediction Models

Assumptions and limitations

The rate example is illustrative. Football goals can be dependent and rates can change within a match. Regression coefficients describe a fitted sample and do not establish causal effects or betting value.