Limited Data Forecasting of Financial Time-Series Using Graph-Based Class Dynamics

Rohan T. Money, Joshin Krishnan, Baltasar Beferull-Lozano

Abstract: This paper addresses the issue of limited data in financial time series forecasting, a challenge regularly faced in the context of newly listed companies. While deep learning models excel at learning relevant representations for forecasting, their efficacy is hindered by the need for substantial data. To over-come this limitation, we propose a forecasting model that takes advantage of transfer learning using graph neural networks. Our model is constructed based on the assumption that a company's stock price is affected not only by the pertinent economic factors specific to the company but also by the shared class-specific factors within its business sector. The model comprises three neural network modules: <tex>$i$</tex>) a module that captures the individual dynamics of a company, ii) a graph neural network (GNN) that captures class-specific dynamics in which the graph is learned from the data, and iii) a module that integrates both the individual and the class-specific dynamics. We propose a novel transfer learning approach to train the GNN, enhancing its efficiency in forecasting time series with limited historical data. Experimental results on real financial time series demonstrate improved forecasting accuracy when incorporating information from class-specific time series.