Author(s): LIU, H., YANG, J., CUI, Y., Institution: BEIJING SPORT UNIVERSITY, Country: CHINA, Abstract-ID: 984

INTRODUCTION:
Corner kicks have transitioned from routine set-pieces into high-value attacking opportunities as modern defenses increasingly compress open-play. Despite their tactical importance, traditional performance analyses primarily rely on basic event statistics and overlook the rich spatiotemporal information embedded in player and ball tracking data. This study addresses the gap by developing machine learning model to predict corner kick threat, defined as the likelihood of a subsequent shot, and examining key spatiotemporal determinants that distinguish Shot and No-Shot outcomes.
METHODS:
Utilizing StatsBomb event and tracking data, a total of 3,025 corner kick sequences from 780 matches across the Premier League (2022-2024), Bundesliga (2023/24), World Cup (2022), and UEFA Euro (2020/2024) were extracted, incorporating tracking frames at the moments of delivery and reception. Samples were categorized into Shot and No-Shot groups based on the sequence outcome. Twenty-one metrics were constructed across four dimensions: (i) delivery quality, (ii) collective tactical formation, (iii) recipient opportunity, and (iv) goalkeeper positioning. A two-stage analytical framework was employed: first, Mann-Whitney U tests identified significant feature differences between outcome groups; second, five machine learning models were trained, optimized using Optuna Bayesian optimization and refined via threshold-moving techniques.
RESULTS:
Significant spatiotemporal differences were observed between Shot and No-Shot sequences (effect sizes reported as non-parametric r). The Shot group exhibited significantly higher recipient space quality (45.99±41.81 vs 32.81±49.66, p<0.001,r=0.34), despite facing higher defense density (5.92±2.83 vs 5.10±3.08, p<0.001, r=0.15) and closer proximity to the nearest defender (2.10±2.97 vs 2.45±4.41, p<0.001, r=0.17).Shot corners featured longer pass lengths (37.39±10.70 vs 35.02±13.15, p<0.001, r=0.12) against a more compact defensive shape, indicated by lower defensive average pairwise distance (7.57±1.70 vs 8.05±2.09, p<0.001, r=0.12). Among all models tested, the optimized LightGBM achieved the highest predictive performance (AUC=0.80, Accuracy=76.1%, Precision=0.84), with recipient space quality and tactical combination strategy emerging as the most influential predictors.
CONCLUSION:
The findings indicate that recipient space quality, defensive intensity, and tactical combination strategies are pivotal determinants of corner kick shots. High-threat corners are not merely those with open space, but those that create high-quality space under congestion. Practitioners should therefore emphasize strategies that maximize spatial advantage in tight defensive zones and refine delivery precision to enhance shot conversion rates.