A neural network that predicts, for any passing moment, the expected ball speed given the spatial context — where attackers, defenders, and the ball are positioned. The architecture is adapted from Fernandez & Bornn's soccer_map: a CNN that takes spatial input surfaces and produces a prediction surface for the entire pitch, but evaluates loss only at the actual pass destination. The key modification: replace the sigmoid activation (pass success/failure) with a linear activation to predict a continuous output (ball speed in m/s). The model captures that passes into crowded areas require higher speed, passes to open teammates can be slower, and certain pitch zones have characteristic speed profiles.
Architecture: 4 spatial input channels, each a continuous surface over the pitch (1 cell per square meter):
Train with linear activation on the output layer; loss = absolute difference between actual and predicted ball speed. ~1200 trainable parameters (much smaller than soccer_map's ~400K). Train on ~50 games of 360 data.
Alternative approach (physics-based): iterate Spearman's pitch control across a range of ball speeds (8-20 m/s); compute pitch control along the pass trajectory at each speed; the minimum speed where pitch control stays >50% for the entire trajectory is the "minimum viable ball speed." Strengths: physics-based, accounts for player velocity. Weaknesses: slow, fails with missing off-screen players in 360 data.
A pass played at 20 m/s when the expected speed for that context is 12 m/s reveals urgent intent — the player was trying to execute quickly, regardless of whether the pass completed. Conversely, a pass at 8 m/s when expected was 12 m/s suggests hesitation or a deliberate tempo change. The actual-minus-expected speed delta is an intent signal that event data doesn't capture, because event data only records what happened, not how urgently the player tried to make it happen.