STOCK MARKET PREDICTION VIA MULTI-SOURCE MULTIPLE INSTANCE LEARNING

Abstract

Stock market prediction plays a crucial role in financial decision-making, where accurate forecasting can help investors maximize profits and minimize risks. Traditional prediction models often rely on a single data source, primarily quantitative stock price data, which limits their ability to capture complex market dynamics. This project presents a multi-source multiple instance learning approach that integrates diverse data sources such as financial news, social media data, and historical stock prices to improve prediction accuracy. The system extracts meaningful features including structured events, sentiments, and vector representations from textual data using advanced techniques. Event extraction is performed using the HanLP algorithm to identify relationships within sentences, while sentiment analysis is conducted using the Latent Dirichlet Allocation (LDA) method. Additionally, Restricted Boltzmann Machines (RBM) are used to determine feature dimensions, and Sentence2Vec is applied to generate vector representations of textual data. These features are combined with quantitative stock data and used to train a Multi-Instance Support Vector Machine (SVM) model for predicting stock trends such as “rise” or “decline.” To further enhance performance, an extension using the XGBoost algorithm is implemented, leveraging multiple decision trees for optimized learning. Experimental results show that the proposed multi-source approach significantly improves prediction accuracy compared to traditional single-source models. The XGBoost-based extension achieves an accuracy of up to 94%, outperforming the basic SVM model. This project demonstrates the effectiveness of integrating multi-source data and advanced machine learning techniques for robust stock market prediction.