Comparative analysis of ARIMA, RNN, LSTM, and GRU for multi-sector Indonesian stock price forecasting

Asep Muhidin (Department of Information Technology, Universitas Pelita Bangsa, Indonesia)
Agung Nugroho (Department of Information Technology, Universitas Pelita Bangsa, Indonesia)
Muhtajudin Danny (Department of Information Technology, Universitas Pelita Bangsa, Indonesia)

Journal of Soft Computing Exploration
Vol. 7 No. 3 (2026): September 2026

Forecasting stock prices remains a challenging task because financial time series are highly volatile, nonlinear, and often differ across sectors. In the Indonesian market, previous studies have mainly focused on single stocks or limited model comparisons, leaving cross-sector evidence relatively scarce. This study compared the forecasting performance of Autoregressive Integrated Moving Average (ARIMA) and three deep learning models, namely Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), using historical daily closing price data from 12 Indonesian listed companies across four sectors: banking, energy, consumer goods, and telecommunications. The data cover the period from 2010 to 2025 and comprise more than 47,000 observations. The historical datasets were preprocessed and divided into training and testing sets using an 80:20 holdout strategy. For the deep learning models, normalized closing prices were transformed into supervised sequences using a 30-day sliding window, and forecasting performance was evaluated under a rolling forecasting framework. Forecasting accuracy was measured using RMSE, MAE, and MAPE. The ARIMA baseline produced an average RMSE of 158.74 and an average MAPE of 3.30%. Among the deep learning models, RNN achieved the best overall performance, with the lowest mean RMSE (117.01) and mean MAPE (2.72%). In contrast, GRU obtained seven stock-level wins and the lowest median MAPE (1.97%), indicating greater consistency across individual stocks. LSTM showed weaker overall performance, with a mean RMSE of 146.96 and a mean MAPE of 3.83%. The Friedman test indicated significant differences among the compared models (χ² = 9.50, p = 0.0087). Pairwise Wilcoxon analysis showed a significant difference between LSTM and GRU (p = 0.0015), while no significant differences were found for RNN–LSTM and RNN–GRU. These findings suggest that simpler recurrent architectures remain highly effective for stock forecasting, and that model selection should consider both sector characteristics and evaluation criteria.