Comparison of ARIMA, LSTM, and Ensemble Averaging Models for Short-Term and Long- Term Forecasting of Non-Stationary Time Series Data

Abstract

⎯ This study aims to forecast the highest weekly selling rate of the Indonesian Rupiah (IDR) against the US Dollar (USD) and identify the most accurate model among ARIMA, LSTM, and Ensemble Averaging. The evaluation results indicate that ARIMA achieves an accuracy of 99.37%, demonstrating strong performance in short-term forecasting, while LSTM achieves an accuracy of 99.99%, excelling in capturing complex and dynamic patterns for long-term predictions. The Ensemble Averaging approach achieves an accuracy of 99.87%, proving to be the optimal solution by combining ARIMA's stability with LSTM's adaptability, resulting in relatively accurate and stable predictions. Although the Ensemble Averaging model has higher RMSE and MSE values compared to the individual models (ARIMA and LSTM), this approach remains quite effective in forecasting both short-term and long-term time series data. This shows that, despite larger prediction errors, Ensemble Averaging provides more stable and accurate results in the long term. The findings highlight that the ensemble approach is more effective than individual models, as it balances accuracy and prediction stability across various forecasting scenarios. This method serves as a reliable tool for addressing market volatility and contributes significantly to the advancement of more adaptive and accurate financial and economic forecasting techniques.