Deep Learning Models for Predicting Stock Closing Prices in the Saudi Stock Market

Abstract

Financial time series data are inherently volatile and nonlinear, presenting considerable challenges in accurately forecasting stock market trends. This study aims to predict the closing prices of companies listed on the Saudi stock market by analyzing historical data spanning from 2013 to 2023. To this end, four deep learning models were employed-Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Units (GRU), and Convolutional Neural Networks (CNN). In addition, two hybrid models, GRU-LSTM and CNN-LSTM, were developed to enhance predictive performance. The models were evaluated using five key regression metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²). The findings indicate that the GRU and GRU-LSTM models demonstrated superior performance relative to the other models. The GRU model attained the lowest MSE and RMSE, demonstrating superior overall accuracy and robustness, while the GRU-LSTM model achieved the lowest MAE and MAPE, reflecting more precise pointwise and relative error estimates. Although the LSTM and CNN-LSTM models exhibited reasonable performance, the Bi-LSTM and CNN models were comparatively less effective. This study underscores the efficacy of deep learning, particularly GRU-based and hybrid architectures, in forecasting stock prices with high accuracy. The results offer valuable insights for investors and financial institutions operating within the Saudi stock market, demonstrating the practical implications of applying advanced deep learning methodologies. Moreover, the study contributes to the broader body of literature on financial forecasting and supports the development of more informed, data-driven investment strategies.