STOCK MARKET PREDICTION VIA MULTI- SOURCE MULTIPLEINSTANCE LEARNING(ML)

Abstract

Stock market prediction is a complex task influenced by multiple dynamic factors such as financial indicators, company performance, global economic conditions, and sentiment from news or social media. Traditional machine learning models often rely on single-source data, limiting their prediction accuracy and making them sensitive to noise. This paper proposes a Multi-Source Multiple Instance Learning (MS-MIL) framework for stock market trend prediction. MS-MIL treats each day as a “bag” of heterogeneous instances that include technical indicators, historical market patterns, macroeconomic parameters, and text-based sentiment features. Unlike conventional supervised learning, MIL processes bags rather than individual instances, making it suitable for handling incomplete labels, uncertain data sources, and non-uniform feature contributions. This approach improves robustness against noise and captures hidden relationships across data sources. The proposed system integrates feature extraction, normalization, bag construction, and multiple-instance classifiers such as MIL-Boost, mi-SVM, and neural MIL networks. Experimental results demonstrate that MS-MIL outperforms traditional single-source models in predicting stock price movement, achieving higher accuracy, precision, and stability under volatile market conditions. The system provides a scalable and intelligent solution for modern financial forecasting.