Sector-Level Benchmarking of Deep Learning Architectures for Predicting Saudi Stock Market

Accurate prediction of stock market behavior is one of the most common intractable problems in financial data science. In this study, we performed a large-scale comparison of three learning architectures, namely, Long Short-Term Memory (LSTM) Networks, gated recurrent units (GRUs), and a hybrid convolutional neural network with GRU (CNN-GRU). They were applied to the full historical data of the Saudi stock market (Tadawul), which covers the period from December 31, 2001, to August 3, 2025. The dataset contains over 910,000 rows for 279 listed companies and models were trained to predict the daily closing prices using a 30-day sliding window of key trading features such as opening price, high, low, traded volume, total traded value, and closing price. Model performances on the training and test data were analyzed using the root mean square error (RMSE) and the mean absolute error (MAE) and the results were aggregated by company and economic sector. Among the three architectures, GRU achieved the highest firm-level win rate (143 of 258 companies), while CNN–GRU recorded the lowest aggregate RMSE (2.465 vs. 2.575 for GRU), reflecting its stronger performance on high-volatility firms; LSTM trailed both with 38 firm-level wins. The sector-level analysis also revealed differences in model dominance, indicating that sectoral volatility and feature patterns may affect which model is the most appropriate. Altogether, this study provides an empirical basis for AI-driven decision support systems in Gulf financial markets and methodological guidance for the benchmarking of large-scale models.