Yan, Wei QiNguyen, MinhYang, GuangLiang2025-05-202025-05-202025http://hdl.handle.net/10292/19242In this thesis, we explore the adaptation of large language models (LLMs) for structured time-series forecasting, focusing on predicting table tennis serve landing points. Traditional time-series models rely on specialized architectures, while LLMs are inherently designed for textual data processing, posing challenges in numerical sequence modeling. To address this, we introduce ChatPPG, a multi-modal framework that integrates time-series data into LLMs through structured embeddings, cross-modal attention, and parameter-efficient fine-tuning (i.e., LoRA). Our findings demonstrate that alignment-based approaches significantly enhance forecasting accuracy compared to prompting-based methods, with DeepSeek-R1-Distill-Qwen-1.5B achieving the lowest MSE (0.432) and MAE (0.441). However, our study also highlights a trade-off between accuracy and inference efficiency, as prompting-based methods introduce excessive latency, making them impractical for real-time applications. Ablation experiments further validate the importance of multi-modal feature alignment, interleaved embedding fusion (IEF), and domain-informed prompting, showing that their removal leads to substantial performance degradation. In this thesis, we extend the application of foundation models beyond natural language processing, establishing a scalable and computationally efficient framework for integrating LLMs into structured forecasting tasks. Our future research directions include the development of a fully end-to-end multi-modal sports analytics system, leveraging real-time vision models for spatiotemporal reasoning, as well as the exploration of generative models like stable diffusion for stochastic time-series forecasting. These advancements aim to enhance automated match analysis and intelligent coaching applications, further bridging AI, computer vision, and predictive modeling in sports analytics.enThesisOpenAccess