Abstract This study presents a systematic comparison of linear, artificial intelligence (AI), and hybrid forecasting models for predicting daily returns of Pakistan’s KSE-100 Index, an emerging market characterized by volatility and data scarcity. Daily data spanning July 2019 to June 2024 is employed to evaluate traditional econometric approaches including autoregressive integrated moving average (ARIMA) and ordinary least squares (OLS) models, alongside deep learning architectures, including long short-term memory (LSTM) networks and convolutional neural networks (CNNs), as well as hybrid frameworks integrating signal decomposition and machine learning (Wavelet-LSTM and ARIMA-LSTM). The sample is segregated into 80% training and 20% testing sets to ensure rigorous out-of-sample validation. Forecast performance is assessed using root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Theil’s U statistic, and directional accuracy. Empirical evidence indicates that hybrid architectures consistently outperform both standalone AI and traditional models in out-of-sample forecasting. The Wavelet-LSTM specification achieves the lowest forecast errors and an out-of-sample directional accuracy of 89.26% which reflect substantial improvements relative to benchmark models. These findings highlight the importance of noise reduction and nonlinear feature extraction in modeling financial time series within emerging markets. However, statistical forecasting superiority does not inherently imply economic profitability in the absence of transaction cost and risk-adjusted performance analysis. Results further suggest caution in relying on MAPE in environments characterized by near-zero returns.
Rehman et al. (Thu,) studied this question.