Abstract
Traditional deep learning models for stock forecasting often overlook the complex interdependencies between financial assets. This study investigates the efficacy of Spatio-Temporal Graph Neural Networks (ST-GNNs) for stock prediction on the Johannesburg Stock Exchange (JSE) by modeling the market as a dynamic graph. We evaluate ST-GNNs against sequence-based baselines like TCN and MLP on both point and trendline forecasting tasks. Our findings show that while sequence models excel at point prediction due to strong temporal autocorrelation, ST-GNNs demonstrate superior performance in trendline prediction by capturing directional market dynamics. Furthermore, network analysis of the learned graph structures identified systemically important equities and validated economically plausible inter-stock relationships. Simulated trading strategies based on trendline forecasts proved profitable, especially when coupled with risk management. We conclude that ST-GNNs offer a financially interpretable framework that is highly effective for understanding market structure and trend-based forecasting on the JSE.
Videos 1
Watch presentations, demos, and related content
Predicting the JSE with ST-GNNs! Discover how AI maps dynamic stock networks for smarter trend forecasting.
Like, comment, and subscribe on YouTube to support the creator!
Predicting the JSE with ST-GNNs! Discover how AI maps dynamic stock networks for smarter trend forecasting.
Gallery 1
Explore the visual story of this exhibit
Title Image