Like-for-like model map
| Family | Typical target | Main strength | Main risk |
|---|---|---|---|
| Independent or adjusted Poisson | Home and away goals | Transparent score grid | Goal dependence and changing rates |
| Elo-style rating | Relative team strength | Compact sequential update | Parameters and draw mapping need validation |
| Logistic or ordinal regression | Match outcome | Interpretable coefficients | Misspecified relationships |
| Tree ensembles | Outcome or goals | Non-linear interactions | Leakage, tuning, unstable probabilities |
| Neural networks | Flexible targets | High-capacity representation | Data demand and opaque failure modes |
Dixon and Coles provide a canonical football score-model example. Peer-reviewed model comparisons and newer benchmark work show why conclusions belong to a stated dataset and evaluation design.
Define the target first
A 1X2 classifier, exact-score model and over-2.5 probability model solve different tasks. Specify settlement period, cancelled fixtures, promoted teams and whether probabilities or labels are required. Accuracy is unsuitable when the decision depends on probability quality; calibration guidance explains how forecast probabilities require reliability checks across cases.
Fair comparison protocol
scikit-learn's TimeSeriesSplit guidance supports the chronological validation rule used in the next step.
- Freeze raw data and feature availability timestamps.
- Create one chronological train, validation, and final test design.
- Fit a naive baseline and market baseline where available.
- Tune every candidate without viewing the final test period.
- Score the same fixtures using log loss, Brier score, calibration and coverage.
- Report uncertainty and failure slices by league, season and probability band.
TimeSeriesSplit documents time-ordered validation. Calibration guidance explains reliability checks for probability outputs.
Reproducibility record
Store code version, provider schema, data cutoff, exclusions, trained parameters, random seeds and prediction timestamp. A result that cannot be recreated cannot support a model-comparison claim.
Baselines and model choice
Peer-reviewed football model-evaluation research supports the baseline, time-order, and reporting protocol used below.
Choose the simplest baseline that represents the real task: league home-draw-away rates, a rolling goal average, a basic Poisson model or a market probability captured at the same cutoff. The baseline must use no information unavailable to the candidate model.
| Candidate | Useful when | Extra audit burden |
|---|---|---|
| Poisson regression | Goal counts and interpretable rate effects | Dependence and dispersion checks |
| Ordinal or multinomial model | Direct match-result classes | Calibration across all outcomes |
| Rating model | Sequential relative team strength | Update rule and competition transfer |
| Tree ensemble | Non-linear tabular relationships | Leakage, drift and explanation stability |
| Neural model | Large structured or sequence inputs | Data scale, tuning and reproducibility |
No row is a universal winner. Compare models on identical rolling cutoffs and the same scoring rules.
Release record
Peer-reviewed football model-evaluation research supports the baseline, time-order, and reporting protocol used below.
For the selected model, preserve the feature list, transformations, code version, training window, hyperparameters, calibration method and untouched final test period. Report proper scoring rules as well as decision-specific metrics. Slice results by season, league and probability band.
Peer-reviewed football model-evaluation research supports the baseline, time-order, and reporting protocol used below.
If a complex model wins by a very small or unstable amount, the operational cost may outweigh the score difference. Record latency, missing-input behaviour and retraining requirements alongside accuracy so the comparison reflects the system readers will actually use. Football model-evaluation research supports reporting operational and performance limits beside model comparisons.
Related resources
Build the simplest score model with Poisson, or continue to machine learning for supervised-learning mechanics.
Continue learning
- Next guide: Football Shot Statistics
- Related guide: Expected Threat (xT)
Assumptions and limitations
The table describes model families, not guaranteed rankings. Performance changes with target, data quality, sample and implementation. Market prices are useful baselines but must be timestamped and de-margined consistently.
GambleAware
GamCare
Gambling Therapy
GAMSTOP