How Novanta Lab works

Novanta Lab publishes its own football models. We don't resell anyone's data — we build models from publicly available match records and publish what they produce, with the error bars shown. Every model here can be checked, and nothing we publish can be taken away from us.

TrueGoals — our expected goals model

TrueGoals estimates how many goals a team's performance in a match typically produces, based on shots, shots on target, corners, home advantage, and opponent strength. It's trained on over 250,000 team-matches across 20+ years and 21 leagues.

It is a team-level model built from match statistics — it does not use shot-location data, so it won't match shot-by-shot models exactly, and we don't pretend otherwise.

How well calibrated is it?

A good model's predictions should match reality on average — when it predicts 2 goals, teams should score about 2.

TrueGoals calibration curve — Each point is a group of predictions; the closer to the diagonal, the better calibrated. Validated on 75,677 held-out matches the model never saw during training.

The honest numbers: TrueGoals beats a naive shots-on-target baseline on both Poisson deviance (0.840 vs 0.870) and mean absolute error (0.758 vs 0.771). It is slightly conservative on very high-scoring performances — a known limitation we're refining.

xPoints and Luck

xPoints simulates how many league points a team's expected goals should have earned. The Luck Index is actual points minus xPoints — a positive number means a team is over-performing its underlying numbers (and may regress), a negative number means the reverse.

Our data

Match results, shots, corners, and other statistics are compiled from publicly available records going back over 20 years. We mirror this data ourselves. Our models and their outputs are our own.