Frame2Threat — Possession-Danger Prediction

Independent research · StatsBomb open data

What this is

Predicts possession danger from partial event sequences with an XGBoost + GRU ensemble, a graph neural network (SAGEConv) and a SHAP explanation layer. Executed notebooks and 49 committed figures.

Metrics

Ensemble

0.965 AUC (n=2,475)

Validation

Match-level split, so no possession spans train and test

Components

PossessionGRU 0.9524 · XGBoost 0.9505

Pass-level

0.882 AUC (n=7,344)

Honesty

360 context adds only +0.001 AUC

Stack

XGBoostPyTorch GNNGRUSHAPStatsBomb

What I would not claim

Not a production tracking-data model: StatsBomb open data with partial 360 coverage.