Stock Movement Prediction and Portfolio Management via Multimodal Learning with Transformer

Abstract

This paper introduces a novel high performing multimodal deep learning architecture(Trans-DiCE) for stock movement prediction utilizing financial indicators and news data. Our multimodal architecture uses dilated causal convolutions and Transformer blocks for feature extraction from both data sources. The masked multi-head self-attention layers inside Transformers preserve causality and improve features based on contextual information. To integrate the derived multimodal model representations, we use stacked Transformer blocks. We show empirically that our model performs best compared to state-of-the-art baseline methods for S&P 500 index and individual stock prediction and provides a significant 3.45% improvement from 74.29% to 77.74%. We also demonstrate our model’s utility for the Portfolio Management task. We propose a Deep Reinforcement Learning Framework utilizing Trans-DiCE for Portfolio Optimization, providing noticeable gain on Sharpe Ratio and 7.9% increase in Portfolio Value over the existing state of the art Models.