Expected goals
xG°An estimate from a model we built. Useful, but not a fact — every modeled stat's page says where it's weak.
Definition
Every shot is scored by a model that reads the chance's geometry and context — distance, angle, body part, assist type, defensive pressure — and returns the probability that an average NWSL finisher converts it. A player's xG is the sum of those probabilities.
How it is computed
Fitted on the full NWSL event record. Each shot's features are mapped to a conversion probability; the season figure is the sum over that player's shots.
How much is enough
xG per shot is meaningful almost immediately; a player's season xG total is meaningful once they have taken roughly 30 shots. Below that, the total is mostly a count of chances.
Where it is weak
Weakest on unusual chances the training set is thin on: deflected shots, scrambles, and shots taken while falling. Penalties are modeled at a flat rate and will not reflect a specific taker's skill.
What it refuses to measure
xG does not measure finishing skill. It deliberately describes the chance, not the player who took it — the difference between a player's goals and their xG is the finishing signal, and it is noisy over a single season.
What would prove this wrong
If, over two consecutive data releases, summed xG diverges from actual non-own-goal league goals by more than 10%, the calibration is broken and this model tier is voided.
Model vs reality — shown, not asserted
League conversion per season: the dim line is what actually happened (goals per shot); the dark line is what the model predicted (xG° per shot). The falsifier for this metric lives in this gap — currently 5.1% in 2026.