Time Series Forecasting of KSE-100 Index Using a Hybrid ESN-LSTM Model

The volatility and complexity of stock markets, particularly in emerging economies like Pakistan, pose significant challenges for accurate forecasting. This paper aims to improve the predictive accuracy of the Karachi Stock Exchange index (KSE-100) by comparing and combining advanced deep learning models. The research investigates the effectiveness of Echo State Networks (ESN), Long Short-Term Memory (LSTM) networks, and a hybrid ESN-LSTM model in capturing both short-term fluctuations and long-term trends. Daily KSE-100 index prices from September 1, 2019, to September 1, 2024, were used, with scaled closing price, 50- and 200-day moving averages as input features. The models' performances were evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R-squared and Directional Accuracy (DA). Results indicate that the ESN-LSTM hybrid model outperformed both individual ESN and LSTM models, achieving the lowest MAE of 513.10 and RMSE of 650.59, along with the highest R-squared of 0.975 and DA of 94.12%. The superior performance of the hybrid model is attributed to its ability to use the efficient temporal feature extraction of ESNs and the long-term memory capabilities of LSTMs. These findings suggest that hybrid approaches can significantly improve stock market forecasting accuracy in emerging markets. Future research should explore the integration of sentiment analysis and macroeconomic indicators to further enhance predictive capabilities.