Multimodal multiscale dynamic graph convolution networks for stock price prediction

Predicting directional future stock price movements is very challenging due to the complex, stochastic, and evolving nature of the financial markets. Existing literature either neglects other timely and granular alternative data, such as media text data, or fails to extract and distill predictive multimodal features from the data. Moreover, the time-varying cross-sectional relations beyond sequential dependencies of stock prices are informative for forecasting price fluctuations, for which the modelling flexibility, however, is not adequate in most of the previous studies. In this paper, we propose a novel Multiscale Multimodal Dynamic Graph Convolution Network (Melody-GCN) to address these issues in stock price prediction. It contains three core modules: (1) multimodal fusing-diffusing blocks that effectively integrate and align the numerical and textual features; (2) a multiscale architecture that extracts and refines temporal features via a fine-to-coarse descending path and a coarse-to-fine ascending path progressively; and (3) dynamic spatio-temporal graph convolutional layers that learn the complex and evolving stock relations not only in between industries and individual companies but also across time horizons. Extensive experimental results and trading simulations on two real-world datasets demonstrate the superior performance of our proposed approach beyond other state-of-the-art models.