Multimodal Stock Price Direction Prediction
Dr. Khaled Solaiman, Department of Computer Science and Electrical Engineering.
This project, Multimodal Stock Price Direction Prediction Using Historical Prices, News, and Sentiment, aims to predict short-term stock movements up, down, or stable by integrating numerical and textual financial data. The approach combines historical stock price features with sentiment analysis of financial news to capture both market trends and investor sentiment. Historical price data are collected from Yahoo Finance, while relevant news articles and headlines are sourced from the News API or GDELT. Sentiment scores are derived using FinBERT, a transformer-based model optimized for financial text, and aggregated on a daily basis. These sentiment features are then merged with technical indicators such as moving averages, relative strength index (RSI), and multi-day returns. The resulting multimodal dataset is labeled according to future price direction and used to train classification models including Random Forest, XGBoost, and Multilayer Perceptrons. Model performance is evaluated using accuracy, F1 score, and confusion matrices to compare the predictive power of price-only, sentiment-only, and combined feature sets. The study explores both early and late fusion approaches to integrate modalities and applies feature importance analysis (e.g., SHAP) to interpret model behavior. The findings aim to demonstrate how combining quantitative and qualitative signals can improve stock trend forecasting and provide insights into the interplay between market sentiment and price dynamics.