A Big Data Analytics Framework for Decision-Making in Sports Performance Optimization

Abstract

Elite team sports now generate continuous, heterogeneous data streams from optical tracking, satellite-based wearables and manually annotated event feeds, yet most published analytics remain single-purpose models evaluated on proprietary corpora that cannot be reproduced. This paper proposes and empirically evaluates an end-to-end big data analytics framework that converts high-throughput spatio-temporal match data into decision support for coaching and performance staff. The framework comprises five layers: distributed ingestion, stream conditioning, a shared feature store, a multi-task learning engine, and an interpretable decision layer. It is instantiated on a fully public corpus of 1,517 matches and 5,321,459 events drawn from four European domestic seasons, and validated on three coupled tasks. First, an expected-goals module enriched with freeze-frame spatial descriptors reaches an AUC of 0.817 on 303 held-out matches, statistically indistinguishable from the commercial model shipped with the same data (0.819), while a distance-and-angle baseline reaches only 0.747. Second, a compact temporal convolutional network forecasts shot occurrence within a ten-second horizon at an AUC of 0.802 using 9,089 parameters, matching gradient boosting with a 50-fold smaller memory footprint and a median single-event latency of 0.034 ms. Third, a zone-based Expected Threat surface converges in 68 iterations and correlates with team-season goals at r = 0.859. An event-derived workload module shows that exponentially weighted load ratios are substantially more stable than conventional rolling ratios. The framework demonstrates that transparent, reproducible pipelines can match proprietary systems while remaining deployable at the edge.