Wenting Luo's Dissertation Defense

Title: Transformer-Based Multimodal Frameworks for Stock Index Movement Prediction Using Financial News and Technical Indicators

To address this, this work develops a novel framework that combines summarized financial news from The Wall Street Journal with technical indicators derived from historical market data. A key contribution is a large language model–based summarization pipeline that extracts market-relevant information from long-form articles, reducing noise and enhancing signal quality for downstream modeling.

Building on this foundation, two Transformer-based architectures are proposed. DeepTransFuse employs an encoder–decoder structure with knowledge distillation to efficiently fuse textual and numerical data, achieving improved accuracy over traditional baselines. SwinStock, a multi-scale Transformer framework inspired by the Swin Transformer, captures market dynamics across multiple temporal horizons (from days to a year), leading to stronger predictive performance and better class balance.

Overall, this research demonstrates that combining filtered financial news with structured indicators through multimodal and multi-scale modeling provides a powerful and robust approach for stock index prediction.