Multi-Branch Transformers for Stock Market Prediction using Previous Market Data and News Articles

Abstract

The stock market is a complex and dynamic system influenced by numerous factors, including technical indicators, financial news, and long-term historical price movements, among others. Understanding and accurately forecasting market behavior requires integrating diverse data sources and identifying underlying trends across multiple modalities. The primary objective of this research is to demonstrate the advantages of employing multi-branch Transformer architectures for managing multimodal financial data and to evaluate the model's effectiveness in predicting stock market trends over short, medium, and long term horizons. In this study, we investigate the implementation of a multi-branch Transformer model designed to forecast stock market prices by integrating multiple data sources, such as news articles and historical market data, over extended periods. The proposed architecture comprises two main branches: the first is a BERT based Transformer that processes textual information related to daily stock performance, while the second is an LSTM based neural network that analyzes long term historical price data. After the feature extraction and processing stages, the outputs from both branches are fused through dedicated layers to enable highly accurate and efficient stock price predictions. Leveraging advanced artificial intelligence, particularly deep Transformer architectures, the proposed multi-branch model processes heterogeneous financial data simultaneously, significantly improving forecasting accuracy and predictive capability. The model achieves a mean square error (MSE) of 0.0006, demonstrating its strong performance and minimal loss value. This study underscores the potential of multi-branch Transformer architectures to seamlessly integrate textual and numerical financial information, offering a robust and advanced framework for stock market prediction and trend analysis. The proposed approach relies on large scale datasets, which pose challenges related to data quality, accessibility, and processing efficiency. Furthermore, the model's substantial computational requirements may limit its practicality for small organizations or institutions with constrained resources.