Hybrid Ensemble Method for Residual-Based Forecast of Time Series Data with Interventions

Abstract

The modeling of time series data is an increasingly prominent concept in recent decades. The focus has extended beyond conventional statistical methods to incorporating Machine Learning (ML) technologies. Conventional statistical methods are effective when data have a linear pattern and meet assumptions such as normality of errors and white noise. Meanwhile, ML models are better suited for non-linear data patterns that do not rely on assumptions. It was observed that both methods were not effective when data exhibited both linear and non-linear patterns simultaneously, showing the need for more adaptive modeling to address time series data with mixed patterns. Therefore, this study aimed to apply a hybrid model, specifically an ensemble method for residual forecast (ERF), to predict the import volume of Indonesian fabrics woven from cotton. The results showed that ERF outperformed the conventional Autoregressive Integrated Moving Average (ARIMA) Intervention by providing more accurate forecasts for data with both linear and non-linear patterns. The ERF model outperforms other hybrid models, namely ARIMA-MLP and ARIMA-SVM, across all forecast periods by delivering better accuracy through its ensemble-based architecture that effectively captures time series patterns with intervention effects.