Training two post-shot xG models — one without goalkeeper location (estimating xG against average positioning) and one with it — and using the difference to isolate the positioning contribution from the shot-stopping contribution. Enables counterfactual "what if the keeper stood here" analysis by sweeping goalkeeper position across the goal and running inference. The optimal position surface emerges as a property of the model, not a separate optimization.
Model A: post-shot xG without GK position → estimates shot danger against "average" positioning. Model B: post-shot xG with GK position → estimates shot danger given actual positioning. Positioning value = Model A output - Model B output. Positive = good positioning reduced danger. Negative = bad positioning increased danger.
Standard GSAA (Goals Saved Above Average) lumps positioning quality and shot-stopping reflexes into one number. A goalkeeper with elite reflexes and poor positioning can produce the same GSAA as one with elite positioning and average reflexes — but the training prescriptions are opposite, the sustainability is different (reflexes decline with age, positioning improves), and the team-building implications are different.