Performance Analysis of ARIMA, LSTM, and Hybrid ARIMA-LSTM in Forecasting the Composite Stock Price Index

Abstract

This study evaluates the performance of ARIMA, LSTM, and hybrid ARIMA-LSTM models in predicting the closing and opening prices of the Indonesia Stock Exchange Composite Index (IHSG) over various periods (2007-2020, 2007-2022, and 2007-2024). For the LSTM model, a lag of 1 was chosen based on MAPE analysis, showing strong dependence on the previous day’s price. Different learning rates (0.01, 0.001, 0.0001) and batch sizes (16, 32) were tested on various network architectures. Results indicate that while ARIMA effectively captures linear patterns, LSTM consistently outperforms with lower MAPE values—2.27% for closing and 2.02% for opening prices—especially with a simple (1-50-1) architecture and a learning rate of 0.001. The hybrid ARIMA(0,1,1)-LSTM(1-50-1) model showed competitive results, achieving MAPE of 2.00% for closing and 1.74% for opening prices using batch size 16. However, its success depends on ARIMA’s ability to model linear components. Key findings emphasize LSTM’s dominance in accuracy, the importance of parameter tuning, and the effectiveness of simple network structures. The hybrid approach holds promise when linear and nonlinear data components are clearly separable. This research offers methodological insights for optimizing stock price prediction models and practical guidance for model configuration, contributing to the advancement of financial market forecasting.