Set Defense Attack Parameter Optimization

Using gradient boosting models with cluster-based interpretation to identify the optimal ball circulation parameters when attacking a set defense. Three key parameters emerge: speed of ball circulation, width of play (lateral pitch coverage), and time between consecutive passes. The analysis reveals sweet spots for each parameter and a critical context interaction: speed under pressure increases goal probability (pulling defenders out of position), but speed without pressure decreases it (unnecessary urgency reduces accuracy). Additionally, central play is more rewarding but riskier than wing play — a fundamental tradeoff that requires game-theoretic thinking.

(1) Label events as attacking against a set defense using the proxy framework. (2) Compute rolling features for each event: rolling speed (m/s), rolling width (lateral meters covered), rolling time between consecutive passes. (3) Train a gradient boosting model (XGBoost) to predict P(goal within 5 moves) using these features plus pressure state and distance from goal. (4) Interpret the black box: cluster events by distance-from-goal and feature values, then vary one feature at a time across the cluster centroid, predicting with and without pressure. This reveals local optima in each parameter.

Key findings:

• Speed sweet spot ~6 m/s: below this, the defense has time to adjust; above 8 m/s, accuracy drops and advantage is lost
• Speed under pressure helps: when pressed, increasing speed pulls defenders out of shape and creates gaps
• Speed without pressure hurts: when unpressed, increasing speed introduces unnecessary risk — the defense is already set, play patiently
• Width ~15m optimal: playing too narrow limits options; too wide stretches the team's own shape
• Inter-pass time ~1.5s minimum: faster than this is dangerous — below 1.5 seconds between passes, accuracy degrades
• Cutbacks disproportionately effective: against set defenses, passes into the 6-yard box come disproportionately from the byline; Man City exemplifies the half-space-to-byline-to-cutback pattern
• Central play more rewarding but riskier: the team that plays centrally against a set defense creates more but also concedes more when they lose it

1. Always playing fast against set defenses: Speed is only beneficial under pressure. Without pressure, patience and width are more effective.
2. Interpreting gradient boosting directly: The model is a black box. Cluster-centroid feature variation is needed to extract local optima; global feature importance alone is misleading.
3. Ignoring the central-vs-wing tradeoff: Neither central nor wide play dominates. The optimal strategy depends on risk tolerance and opponent-specific analysis.