The PELT-ensemble NARX Hybrid Method for Short-Term Forecasting Time Series Data with Change Points and Its Application

Abstract

This study proposes a new approach using the PELT-ensemble NARX hybrid method for short-term forecasting of time series data containing change points (CPs). The Pruned Exact Linear Time (PELT) method was employed to detect CPs and to determine cut-off points for the observation data, while the ensemble Nonlinear Autoregressive Neural Network with External Inputs (NARX) was used for forecasting. The method was applied to health data, specifically confirmed Covid-19 cases in the provinces of Java and Bali, Indonesia, from March 2020 to October 2021. Several Covid-19-related variables such as the number of recovered cases and deaths, and the weather variables, i.e., the maximum temperature, the average temperature, and humidity, are used as external inputs for the NARX model. The results indicate that the PELT method effectively identified the first day in the last segmentation for each province, and subsequent forecasting using the ensemble NARX method achieved high accuracy. Forecasting performance was evaluated using the Mean Absolute Percentage Error (MAPE) and several other metrics, showing that the PELT-ensemble NARX hybrid method significantly outperformed the single NARX approach and other forecasting methods, i.e. Exponential Smoothing, Autoregressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) model. These findings highlight the effectiveness of integrating change point detection with ensemble neural network modelling to substantially improve short-term forecasting accuracy for time series subject to structural change.