From Score Matrices to Football-Aware Match-State Simulation: An Auditable LLM Harness for Exact-Score Reranking explores a hybrid architecture designed to improve football score forecasting by combining statistical models with the contextual reasoning of Large Language Models (LLMs). The research addresses the difficulty of predicting exact scores, where statistical models excel at identifying historical patterns but struggle to interpret tactical nuances, such as how a specific player absence or a first goal might alter a match's trajectory.
A Hybrid Approach to Forecasting
The system uses a "harness" to bridge two distinct components. The first is a dynamic Poisson-family model (V1), which provides a stable probabilistic backbone by calculating team strength and expected goals. The second is an LLM that performs causal match simulation. Instead of replacing the statistical model, the LLM is constrained by the harness to interpret pre-match evidence—such as formations, rest, and tactical matchups—and construct goal-by-goal paths. This ensures the LLM’s reasoning remains auditable and subordinate to the statistical prior, preventing it from inventing arbitrary scores or ignoring numerical probabilities.
Evolution of the System
The research documents four iterations of the system, moving from simple scalar adjustments to complex path simulations:
V1: A pure statistical baseline using a dynamic Dixon-Coles model.
V2: A "sandwich" architecture where an LLM provided a scalar residual to adjust the statistical model’s parameters. This proved unstable, as a single scalar could not capture the conditional nature of football events.
V3: The LLM was tasked with building three explicit goal-by-goal paths from a 0–0 starting point, reranking a frozen set of score candidates.
V4: Added "shared-root" and "cascade" judgments, allowing the model to simulate whether a match would open up or close down after the first goal, while also incorporating deterministic tail candidates.
Performance and Results
Testing on the first 150 matches of the 2025–26 English Premier League showed that the V4 iteration achieved 14.7% Top-1 and 30.7% Top-3 exact-score accuracy, compared to 10.0% and 26.7% for the V1 baseline. While the model improved at selecting among central low-score outcomes, it struggled to identify 0–0 results or promote "tail" candidates (high-scoring or blowout matches). Although V4 increased the number of available score candidates, it did not consistently learn how to rank those tail outcomes correctly.
Limitations and Considerations
The author notes that these results are exploratory rather than a claim of state-of-the-art performance. Because the development of the model was informed by the 2025–26 season data, the benchmark is not an untouched test set. Furthermore, the system relies on a closed-weight LLM, meaning the research cannot rule out the possibility that the model "memorized" match outcomes during its pre-training. The study concludes that while the LLM adds value by providing a structured, causal story for a score prediction, it does not yet function as a calibrated probability engine, and its semantic judgments—such as whether a game will "open up"—remain difficult to calibrate.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!