Enhancing Deep Learning Forecasts with Wavelet Decomposition: Evidence from the Ghana Stock Exchange

Forecasting stock market returns in emerging economies remains challenging due to market volatility, structural irregularities, and limited data availability. This study investigates whether discrete wavelet transformation can enhance the predictive performance of deep learning models when applied to financial time series from emerging markets. Using daily returns of the Ghana Stock Exchange Composite Index (GSE-CI) spanning 2011 to 2022, we evaluate three widely used deep learning architectures—Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN)—in both their standard form and with preprocessing based on the Daubechies-4 (db4) discrete wavelet transform. The empirical results indicate that wavelet preprocessing consistently reduced forecasting errors across all three deep learning architectures, highlighting its effectiveness as a multiscale feature extraction and noise reduction technique for financial time series. Among the models considered, the Wavelet-LSTM achieved the lowest forecasting error, while the wavelet-enhanced variants consistently outperformed their corresponding baseline models. These findings suggest that the benefits of wavelet decomposition extend beyond a specific neural network architecture by providing a richer representation of nonlinear temporal dynamics in volatile and data-constrained financial environments. As one of the first studies to systematically evaluate wavelet-augmented deep learning models for stock market forecasting in an African equity market, this work contributes to the growing literature on hybrid forecasting frameworks and provides practical insights for researchers, analysts, and investors interested in forecasting emerging financial markets.