Home/Soccer Analytics/Spatial xT via CNN and 360 Spatial Maps

Spatial xT via CNN and 360 Spatial Maps

Expected Value ModelsLevel 4 — Expert

What It Is

Extending the xT model by replacing zone-average transition probabilities with a CNN trained on five 360-derived spatial maps (teammate distribution, pitch control, interception likelihood, xT surface, historical xT transitions) to compute context-sensitive ball-movement probabilities. The resulting "Threat Above Expected" (TAx) metric measures how much more or less threatening a situation is than the zone average, given actual player positioning.

Correct Execution

For each action, generate 5 spatial input surfaces from 360 data. Feed through CNN to predict context-specific transition probabilities. Compare actual xT to the CNN's prediction of expected xT for that zone → TAx = (actual - expected) / expected. Positive TAx = better positioning than average; negative = worse.

Diagnostic Tree

Sources

  • Gregory Everett, University of Southampton / Sentient Sports, StatsBomb Conference 2022, YouTube, 2022-10-03