From Score Matrices to Football Consciousness Simulation: An Auditable LLM Framework for Precise Score Re-ranking

Background and Context

Football score prediction has long been constrained by a fundamental dichotomy between statistical rigor and tactical nuance. Traditional dynamic Poisson-family models, such as the Dixon-Coles framework, provide a robust mathematical foundation for estimating team strength, expected goals, and coherent probability distributions based on historical data. However, these statistical cores are inherently blind to non-quantifiable factors: they cannot comprehend player roles, specific tactical matchups, shifting team motivations, or the dynamic behavioral changes that occur immediately after a goal is scored. Conversely, Large Language Models (LLMs) demonstrate exceptional reasoning capabilities regarding these high-context, unstructured tactical concepts. Yet, LLMs are not well-calibrated probabilistic engines; their outputs often lack the statistical reliability required for precise numerical prediction.

This research introduces an auditable hybrid architecture designed to bridge this gap by combining the probabilistic stability of statistical models with the contextual reasoning of LLMs. The core contribution lies in constructing a framework that defines clear input semantics, enforces pre-match evidence constraints, and mandates that the LLM follows a checkable reasoning path. By doing so, the study injects football-specific tactical consciousness into the statistical core without sacrificing mathematical integrity. This approach transforms the typically black-box nature of LLM inference into a traceable, auditable decision-making process, offering a novel methodological perspective for sports analytics that prioritizes both accuracy and explainability.

Deep Analysis

The technical evolution of this framework is documented through four distinct iterations, from V1 to V4, each refining the interaction between the statistical engine and the language model. V1 served as the baseline, utilizing a dynamic Dixon-Coles model to generate initial score probability distributions. V2 introduced a mapping mechanism that quantified the LLM’s assessment of contextual factors—such as tactical style and morale—and mapped these ratings back into the expected goals parameters, attempting to correct the statistical model’s inputs with LLM insights. V3 marked a structural shift, abandoning simple scalar corrections in favor of a per-ball simulation process. By freezing a set of score candidates, the LLM simulated the match ball-by-ball, evaluating the impact of each goal on the game state in real-time.

V4 represented the culmination of this iterative process, enhancing simulation granularity with several key mechanisms: a shared first-breakthrough judgment, cascade effect evaluation post-goal, time-aware stopping mechanisms, and deterministic tail candidate selection. This design ensured that LLM reasoning was not arbitrary text generation but a simulation strictly bound by mathematical constraints and football logic. Each step was designed to be auditable and verifiable, thereby increasing the credibility of the final predictions. The framework effectively forces the LLM to reason through the tactical implications of specific scorelines rather than merely predicting outcomes based on pattern recognition.

Experimental validation was conducted via chronological replay tests on the first 150 matches of the 2025-26 Premier League season. The results demonstrated steady performance gains across iterations. The V1 baseline achieved a Top-1 exact score accuracy of 10.0% and a Top-3 accuracy of 26.7%. V3 improved these figures to 12.0% and 30.0%, respectively. V4 further optimized performance, raising Top-1 accuracy to 14.7% and Top-3 to 30.7%. Candidate coverage also expanded from 77.3% in V1 to 84.7% in V4, indicating a broader consideration of potential outcomes. Baseline metrics for V1 included a 1X2 distribution argmax accuracy of 53.3%, a log loss of 0.9878, a Brier score of 0.5870, and a Ranked Probability Score (RPS) of 0.2095, providing a clear reference for subsequent improvements.

Industry Impact

The significance of this work extends beyond mere predictive accuracy, establishing a paradigm for auditable hybrid architectures in high-stakes domains. For the open-source community, the study provides a concrete design pattern for integrating the semantic understanding of LLMs with specialized statistical models. By constraining LLM reasoning paths to ensure interpretability, this framework offers valuable lessons for other sectors requiring both advanced reasoning and statistical reliability, such as finance and healthcare. The ability to trace why a specific scoreline was selected enhances trust and transparency, which are critical for user adoption in professional environments.

In industrial applications, particularly in sports betting and tactical analysis, this framework addresses the growing demand for decision transparency. Stakeholders can now understand the tactical rationale behind a prediction, rather than accepting a black-box output. This transparency is crucial for risk management and strategic planning. Furthermore, the study’s negative findings are equally impactful. The research revealed that while candidate coverage increased, no new tail candidates in V4 became exact Top-3 hits. This highlights a limitation in the current method’s ability to capture extreme long-tail events, suggesting that future developments must focus on better integrating low-probability, high-impact scenarios.

The study also candidly addresses the risks of memory contamination and time isolation, noting that the development slices were not entirely untouched benchmarks. This transparency calls for stricter blind testing mechanisms when evaluating LLM predictive capabilities. By acknowledging these limitations, the research pushes the field toward greater rigor in vertical AI applications. It serves as a cautionary tale against over-relying on LLMs for probabilistic tasks without rigorous calibration, while simultaneously demonstrating how hybrid approaches can mitigate these risks through structured, auditable workflows.

Outlook

Looking forward, the primary challenge lies in overcoming the limitations of football consciousness simulation in score selection. The current framework excels at modeling tactical interactions and behavioral shifts but struggles with the stochastic nature of extreme events. Future iterations must prioritize the development of more sophisticated mechanisms for handling long-tail probabilities. This may involve refining the time-aware stopping mechanisms or introducing new variables that better account for unpredictable match dynamics, such as referee decisions or sudden injuries, which are currently underrepresented in the simulation.

Another critical area for development is the mitigation of potential memory pollution in closed LLMs. As the study notes, results should be viewed as exploratory due to the possibility of the model having seen similar data during training. Implementing rigorous data isolation protocols and employing open-weight models with transparent training data could enhance the reliability of future predictions. Additionally, expanding the framework to other sports with different statistical profiles could test its generalizability and robustness.

Finally, the integration of real-time data streams could further enhance the framework’s utility. By feeding live match data into the LLM’s contextual assessment in real-time, the system could adapt its predictions dynamically, offering even greater value to analysts and bettors. This evolution from static pre-match analysis to dynamic, real-time tactical simulation represents the next frontier for explainable AI in sports analytics, promising to deepen our understanding of the game while providing more accurate and transparent predictions.