Enhancing Fantasy Premier League Strategies through Machine Learning and Large Language Models

Fantasy Premier League (FPL) is the official fantasy football game of the English Premier League, where millions of participants worldwide assemble virtual teams of real football players each season. Players must balance their budget, form, schedule, and injuries each week when selecting their starting lineup and substitutes, earning points based on the players performance on the pitch. The combination of in-depth statistical analysis, strategic team building, and continuous decision-making has made FPL one of the most engaging fantasy sports. In this thesis, an AI-powered digital assistant with explainable capabilities will be introduced to help FPL users make data-driven decisions to maximize their points each round. By constructing both basic and advanced variables, the models are trained on FPL's scoring components. Two modeling methods were used to achieve the predictions: linear/logistic regression and XGBoost. Both gaining similar results in accuracy and actual fantasy points in the simulation which was run from gameweek 1 through 21 in the 2024/25 season. The AI assistant manager's workflow identifies the optimal starting squad and weekly substitutions during the simulation, with the best results being the linear models, yielding 1293 points which would place the algorithm in the top 12\% of all FPL managers. The predictions were fed into a LLM integrated layer which generate personalized recommendations via an LLM. By translating prediction metadata into clear, understandable explanations, the "black box" problem is addressed. Evaluation of the full-stack solution shows that the natural language explanations increased AI assistant users trust and understanding compared to just raw score predictions. By integrating predictive analytics, optimization constraints, explainable AI and conversational interfaces, this work offers transparent, human-centric decision support in fantasy sports as well as a proof-of-concept for similar applications in other data-rich settings.