LLM Harness Improves Football Score Forecasting Accuracy

Football score forecasting traditionally relies on statistical models like dynamic Poisson-family models to estimate team strength and goal probabilities. While effective for core statistics, these models lack the ability to understand nuanced contextual factors such as player roles, tactical matchups, or how a goal changes game dynamics. To bridge this gap, researchers developed an auditable information harness that integrates Large Language Models (LLMs) with statistical forecasting. The LLM provides reasoning about contextual elements, while the statistical core handles probability calibration. This hybrid approach allows for a more comprehensive prediction. Through four iterations, the system evolved from a basic statistical baseline to V4, which incorporates sophisticated features like shared first-breakthrough and post-goal cascade judgments, along with time-aware stopping. On a chronological replay of 150 English Premier League matches, V4 achieved a 14.7% Top-1 exact-score accuracy, a notable improvement over the 10.0% of the statistical baseline. This demonstrates the potential of combining LLM reasoning with statistical rigor for complex predictive tasks, though the results are exploratory and require further validation.