A deep fusion model for stock market prediction with news headlines and time series data

Abstract

Time series forecasting models are essential decision support tools in real-world domains. Stock market is a remarkably complex domain, due to its quickly evolving temporal nature, as well as the multiple factors having an impact on stock prices. To date, a number of machine learning-based approaches have been proposed in the literature to tackle stock trend prediction. However, they typically tend to analyze a single data source or modality, or consider multiple modalities in isolation and rely on simple combination strategies, with a potential reduction in their modeling power. In this paper, we propose a multimodal deep fusion model to predict stock trends, leveraging daily stock prices, technical indicators, and sentiment in daily news headlines published by media outlets. The proposed architecture leverages a BERT-based model branch fine-tuned on financial news and a long short-term memory (LSTM) branch that captures relevant temporal patterns in multivariate data, including stock prices and technical indicators. Our experiments on 12 different stock datasets with prices and news headlines demonstrate that our proposed model is more effective than popular baseline approaches, both in terms of accuracy and trading performance in a portfolio analysis simulation, highlighting the positive impact of multimodal deep learning for stock trend prediction.