Home/Soccer Analytics/Outcome-Weighted Passing Network Analysis

Outcome-Weighted Passing Network Analysis

Passing MetricsLevel 3 — Advanced

What It Is

Building separate directed passing networks for each type of possession outcome — goals, shots on target, shots off target, and lost possession — then learning how much each outcome type predicts winning using logistic regression. Each player becomes a node; each pass creates a directed, weighted edge. The network is built by iterating through all passages of play, accumulating edge weights by outcome type. Player influence is then computed as network centrality (sum of incoming and outgoing edge weights) within each outcome-specific network, weighted by the learned importance of that outcome.

Correct Execution

(1) Iterate through all passages of play in the dataset. (2) For each passage, classify the terminal event: goal, shot on target, shot off target, or lost possession. (3) Build a directed graph: for each pass in the sequence, add weight to the edge between passer and receiver in the network corresponding to that outcome type. (4) After processing all passages, compute each player's centrality per outcome network: V(player, event) = sum of all adjacent edge weights (incoming + outgoing). (5) Learn outcome weights using logistic regression: train on per-match player centrality values as features, with match result (win/loss/draw) as target. The learned coefficients (alpha values) represent how much each outcome type contributes to winning. (6) Final player value = weighted sum across all outcome types: Σ(alpha_k × V(player, event_k)).

Progression Levels

Diagnostic Tree

Coaching Cues

  • "This tells you who drives YOUR team's setup, not who's the best player in the league."
  • "If Fernandinho's centrality drops when he's injured, and the team's results drop too — the metric caught something real."

Common Errors

  1. Using one network for all outcomes: A player who is central to goal-scoring passages and also central to lost-possession passages gets a conflated score. Separate networks per outcome type are essential.
  2. Not learning outcome weights: Treating all outcomes equally (goals = lost possession) produces meaningless values. The logistic regression step is what tells you how much each outcome matters for winning.
  3. Cross-team comparison without normalization: Teams with different possession styles produce incomparable raw values.

Sources

  • James, University of Southampton, StatsBomb Innovation in Football Conference, YouTube, 2019-10-30 — presented outcome-specific passing network construction with logistic regression outcome weighting; showed Man City/Arsenal player dominance; validated with Fernandinho injury performance correlation