Stock market price forecasting using evolving graph neural network

Abstract

Predicting stock prices is essential to inform investment decisions in the financial market. Analyzing financial market movements and stock price behavior is extremely complex due to the dynamic, nonlinear, non-stationary, non-parametric, and chaotic markets. Various approaches are used to analyze stocks for financial market forecasting purposes. Traditional methods based on time series information for one company's stocks do not consider the relationships between stocks of other companies, which can improve the efficiency of stock price forecasting. The use of graph neural networks for these purposes, in which the relationships of time series are represented as a relationship graph structure, and the variables are defined as graph nodes, significantly improves forecasting accuracy. Existing forecasting methods usually assume that the structure of the relationships graph, which is described by the relationships matrix and determines the aggregation method of the graph neural network, is fixed by definition. Therefore, they cannot effectively consider dynamic changes in relationship graphs. In this paper, an evolving graph neural network is proposed for forecasting stock prices in the stock market. To extract dynamic correlations between price movements in financial time series, a relationships graph is constructed in the form of clusters, the generation and the evolution of the structure and parameters of which are implemented using a dendritic artificial immune network (DaiNet). For each generated cluster of the relationships graph, the price encoding is performed using transformers to determine the price information. Then, the messages from the relational graph structure and the input time sequences are aggregated based on the use of the attention layer of the time graph. At the last GNN layer, the final prediction of the future price movement of each stock is performed using a multilayer perceptron to integrate the components.