Graph Neural Networks for Time Series Forecasting on Ghana Stock Exchange

Forecasting stock prices in emerging markets is challenging due to market inefficiencies, low liquidity, and complex interstock dependencies. While traditional models focus on temporal dynamics, Graph Neural Networks (GNNs) can capture both temporal patterns and relational structures among assets. We investigate spatio-temporal GNNs for multivariate stock price forecasting on the Ghana Stock Exchange (GSE), analyzing daily closing prices of ten stocks from September 2018 to March 2025. We construct correlation-based stock graphs and compare LSTM, LSTM-GNN, DCRNN, and StemGNN under single-step and multi-step forecasting. Results show that LSTM achieves the lowest errors (MAPE: 6.88%, RMSE: 0.767) for short windows (30 days), while DCRNN performs best for longer windows (90 days: MAPE: 11.39%, RMSE: 1.138), demonstrating superior spatio-temporal modeling. For multi-step forecasting, LSTM remains competitive, suggesting that architectural complexity may not always improve accuracy in low-liquidity markets. This work provides the first GNN-based analysis of GSE and establishes baseline benchmarks for emerging market forecasting.