Forecasting Monthly RMG Exports Demand in Bangladesh: with or without Change-Point Approach

Abstract

Time series models assume stationarity, but in reality, the mean and variance of time series often change over time. Though different techniques can stabilize variance, none can stabilize the mean. Hence, detecting change points in the mean is vital to segment the data and maintain stationarity in each segment for accurate modeling and predictions. This study explores change point techniques for identifying changes in the mean of time series data. Using the cumulative summation (CUSUM) test and Binary Segmentation algorithm, this study finds two significant change points in the mean of the monthly average export in million USD of specialized textiles in Bangladesh from July 2011 to July 2021. For forecasting, an ARMA(0,0) model with a non-zero mean is fitted for the data with the change points, while an ARIMA(2,0,0)(1,0,0)[12] model with a non-zero mean is estimated for the entire dataset without the change points. The findings of this research demonstrate that the accuracy of forecasting with the change points model is higher compared to forecasting without change points. Therefore, this study suggests that incorporating change point techniques in time series analysis can improve the forecasting process by considering the potential existence of a change point in the data.