Abstract:Stock prediction is a challenging task due to multiple influencing factors and complex market dependencies. Traditional solutions are based on a single type of information. With the success of multi-source information in different fields, the combination of different types of information such as numerical and textual information has become a promising option. Although multi-source information provides rich multi-view information, how to mine and construct structured relationships from them is a difficult problem. Specifically, most existing methods usually extract features from commonly used multi-source information as predictive information sources, without further pre-constructing stock relationship graphs with dependencies using broader information. More importantly, they typically treat each stock as an isolated forecasting, or employ stock market correlations based on a fixed predefined graph structure, but current methods are not sensitive enough to aggregate the attribute features extracted from multi-source information and stock relationship graph, to obtain the dynamic update of market relations and relationship strength. The stock market is highly temporally, and the attributes of nodes are affected by the time perception of other attributes, which is not fully considered. To address these problems, we propose a novel dynamic attributes-driven graph attention networks incorporating sentiment (DGATS) information, transaction data, and text data. Inspired by behavioral finance, we separately extract sentiment information as a factor of technical indicators, and further realize the early fusion of technical indicators and textual data through Kronecker product-based tensor fusion. In particular, by LSTM and temporal attention network, the short-term and long-term transition features are gradually grasped from the local composition of the fused stock trading sequence. Furthermore, real-time intra-market dependencies and key attributes information are captured with graph networks, enabling dynamic updates of relationships and relationship strengths in predefined graphs. Experiments on the real datasets show that the architecture can outperform the previous methods in prediction performance.

GraphVAE: Unveiling Dynamic Stock Relationships with Variational Autoencoder-based Factor Modeling

FactorVAE: A Probabilistic Dynamic Factor Model Based on Variational Autoencoder for Predicting Cross-Sectional Stock Returns

RVRAE: A Dynamic Factor Model Based on Variational Recurrent Autoencoder for Stock Returns Prediction

HireVAE: An Online and Adaptive Factor Model Based on Hierarchical and Regime-Switch VAE

Dynamic CVaR portfolio construction with attention-powered generative factor learning

Understanding stock market instability via graph auto-encoders

Factor Investing with a Deep Multi-Factor Model

NeuralFactors: A Novel Factor Learning Approach to Generative Modeling of Equities

STORM: A Spatio-Temporal Factor Model Based on Dual Vector Quantized Variational Autoencoders for Financial Trading

Bayes-Factor-VAE: Hierarchical Bayesian Deep Auto-Encoder Models for Factor Disentanglement

ST-Trader: A Spatial-Temporal Deep Neural Network for Modeling Stock Market Movement

Stock Volume Forecasting with Advanced Information by Conditional Variational Auto-Encoder

Factor-GAN: Enhancing stock price prediction and factor investment with Generative Adversarial Networks

Graph-based Methods for Forecasting Realized Covariances

A Dynamic Attributes-driven Graph Attention Network Modeling on Behavioral Finance for Stock Prediction

Graphical Models for Financial Time Series and Portfolio Selection

Quantitative Stock Selection Model Using Graph Learning and a Spatial–Temporal Encoder

Global Stock Market Volatility Forecasting Incorporating Dynamic Graphs and All Trading Days

Constructing Dynamic Topic Models Based on Variational Autoencoder and Factor Graph

Dynamic Mean-Variance Portfolio Selection under Factor Models