Identifying and correcting the goal-mouth-angle inversion artifact in xG models: at very low angles (near the byline), apparent xG increases because unintended events (clearances, long passes) are only labeled as "shots" when they accidentally result in goals. This creates a selection bias where the training data at extreme angles is contaminated by non-shot events. The fix: extrapolate the expected monotonic angle-xG relationship and downsample over-represented "goal" observations in peripheral regions.
Plot xG vs. goal-mouth angle. If xG increases at very low angles (where it should monotonically decrease), the model has learned from contaminated data. Fix by: (1) identifying the angle threshold where the inversion begins, (2) extrapolating the monotonic curve, (3) downsampling or reweighting peripheral-angle goals.
At very low goal-mouth angles, xG increases counterintuitively because events from extreme angles only enter training data as "shots" when they accidentally result in goals. Selection bias toward goals contaminates the peripheral-angle training data.