Stock Price Forecasting with Technical Indicators: Comparing Windowed and Non-Windowed Deep Learning Models on the Saudi Stock Market
Abstract
Stock price forecasting increasingly uses recurrent neural networks with sliding-window sequences. However, whether explicit temporal windowing improves next-day prediction when technical indicators already encode recent market behaviour remains unclear, particularly in the Saudi stock market. This study aimed to determine how temporal input design affects LSTM, hybrid LSTM-GRU, and MLP performance. Daily OHLCV data from 2010 to 2023 were collected for 19 companies representing all TASI sectors. Twenty-two technical indicators produced a 27-variable feature space. Each model was evaluated without explicit windowing and with 5-, 10-, and 22-day windows, producing 228 neural model–company experiments. A naïve persistence baseline was evaluated separately for each company, yielding 19 additional evaluations. Performance was assessed using eight regression metrics and Wilcoxon signed-rank and Friedman tests. Non-windowed recurrent models outperformed all windowed variants. Non-windowed LSTM achieved the highest average R² of 0.9583, while non-windowed hybrid LSTM-GRU achieved the lowest average RMSE of 1.1585 and led 12 of 19 companies.