How the model works

Model v1.1

The current public production model is v1.1. It is based on a Poisson goal framework with a Dixon-Coles adjustment for low-scoring football outcomes.

The model produces expected-goal rates for the home and away teams and converts those rates into match-outcome probabilities.

Markets we publish

• 1X2: home win, draw and away win probabilities
• Over 2.5 goals probability
• Both Teams To Score — Yes probability

Pre-match inputs

Features are built only from matches that occurred before the target match. The model uses recent historical information such as scoring and xG form, longer recent windows and schedule/rest context.

Where xG is used, SmartMatchData relies on stored full-quality xG data. npxG is not used by the current public model.

• Recent form windows such as last 5 and last 10 matches
• Recent xG-based attacking and defensive form where full-quality xG is available
• Rest and scheduling context
• No post-kickoff information is used to create a public pre-match prediction

Locked predictions

When a public prediction is created, its probabilities and feature snapshot are saved before kickoff. The prediction remains immutable after kickoff.

This means a later result cannot be used to rewrite the original public probability.

Forward evaluation

Once a locked match finishes, SmartMatchData settles the prediction against the final result and updates public forward-performance metrics.

Accuracy measures classification correctness. Log loss measures probability quality and penalizes confident wrong predictions more heavily.

• 1X2 accuracy and log loss
• Over 2.5 accuracy and log loss
• BTTS accuracy and log loss
• League-by-league forward tracking

Data-quality safeguards

• Administrative results are excluded from historical modelling workflows where appropriate
• Cancelled matches are not treated as normal completed matches
• Public probability records are separated from later settlement
• Missing statistical fields are not silently invented