Enhanced Stock Market Prediction Using Multi-Source Multiple Instance Learning

Abstract

The stock market is influenced by a multitude of dynamic and often interdependent factors, making accurate prediction a complex task. Traditional machine learning models typically rely on single-source data and assume precise instance-level labels, which may not effectively capture the multifaceted nature of financial markets. In this project, we propose a novel approach for stock market prediction using Multi-Source Multiple Instance Learning (MS-MIL), which enables the model to learn from grouped data instances (bags) derived from various heterogeneous sources such as historical stock prices, financial news, social media sentiment, and macroeconomic indicators. By treating each source as a distinct set of instances and aggregating them under the multiple instance learning framework, the model can better handle weak supervision and uncertainty inherent in financial data. Our MS-MIL framework integrates both numerical and textual data, applying advanced feature extraction and attention mechanisms to learn discriminative representations. Experimental results demonstrate that the proposed method achieves superior performance in predicting stock movement direction and market trends compared to traditional learning models. This approach offers enhanced robustness, adaptability, and interpretability, making it a promising tool for investors and analysts in making informed decisions.