LLM Harness for Exact-Score Reranking in Football Forecasting

An arXiv paper (2608.05030) introduces an auditable LLM framework designed for exact-score reranking in football match predictions. This study integrates dynamic Poisson-family models with large language models (LLMs) to enhance score forecasting accuracy. It outlines four versions: V1 establishes a dynamic score-driven Dixon-Coles baseline; V2 incorporates LLM contextual ratings into expected-goal metrics; V3 substitutes scalar corrections with goal-by-goal simulations from a fixed score-candidate set; and V4 introduces shared judgments for first breakthroughs and post-goal cascades, along with time-aware stopping and deterministic tail candidates. The framework specifies input semantics, provides pre-match evidence, and seeks to ensure auditability in reasoning. It tackles the limitations of statistical models in understanding tactical matchups and motivation, while acknowledging that LLMs lack calibrated probability capabilities. The paper can be accessed on arXiv.