Decision safety in football passing: A reproducible metric from probabilistic option sets

Abstract

Football analytics measures what happened on the pitch far more often than whether it was a good choice among the alternatives available. Using StatsBomb 360 freeze-frame data for all 64 matches of the 2022 FIFA World Cup (53,406 passes with a usable freeze frame and a recorded recipient), an XGBoost model trained on situational features alone predicts pass success and is applied to every visible teammate, giving each pass a probabilistic set of the passer's alternatives. The chosen pass is located within this option set by its estimated completion probability, in two normalizations of the same quantity: decision safety (DS), its position between the least and the most completable alternative, and opportunity-adjusted risk-taking (OART), the share of alternatives that were safer. The two are formulations of one construct (player-level r = −0.93) and are comparably reliable: the intraclass correlation coefficient (ICC), the share of variance attributable to differences between players, is 0.044 for DS and 0.035 for OART, and 2.9–3.7% once playing position is controlled. Passing safety is therefore mostly situational but modestly player-consistent, and the player signal sharpens with option-set heterogeneity, reaching an ICC of 9.8% for DS in the most heterogeneous decile. Safer passers create fewer chances (DS against expected assists per pass, r = −0.34; OART, r = +0.32); players of teams that reached the knockout rounds passed more safely within every position; and the apparent link between attacking-third risk-taking and possession expected goals is accounted for by the shot-assist pass itself.