Football Intelligently: GNN Expected Goals

Classic expected-goals (xG) models treat a shot as a handful of scalar features: distance, angle, body part. But whether a shot goes in depends on where every other player is standing. I modeled each shot as a graph, players as nodes and their spatial interactions as edges, so the model can reason about defensive shape and goalkeeper positioning as relational structure.

The production model is a HybridGAT+T: a GATv2 graph-attention network over the StatsBomb 360 freeze frame, fused with 27 dimensions of shot metadata (post-shot xG placement, goalkeeper geometry, defensive blocking, foot preference), with per-competition temperature scaling for calibration. It scores AUC 0.760 and a Brier of 0.148 on 8,013 shots across seven competitions.

The model reaches 95.7% of StatsBomb's proprietary xG AUC using only free, open data, validated across men's and women's competitions (World Cups, Euros, Bundesliga). Permutation importance across 12 feature groups showed goalkeeper distance dominates (+0.223 AUC when shuffled), ahead of shot distance and the header flag.

The dashboard turns the model into something an analyst can use: shot maps coloured by model xG, a shot inspector with gradient-saliency and attention overlays showing which players the model attended to, calibration and reliability diagrams, full match reports with cumulative xG timelines, and a "surprise goals" view that surfaces sub-15% xG strikes.

Five novel contributions are confirmed against the literature, and a paper targeting the MIT Sloan Sports Analytics Conference is in progress. It is the side of me that takes an idea from raw open data to a calibrated, evaluated model with an interface, not just a notebook.