Time Series Forecasting for Double Seasonal Event: A Simulation Study Approach

The objective of this study is to compute the model for forecasting the data which contain a non-multiplicative double seasonal pattern. This study analyzes the food material demand simulated data which contains two types of seasonal, i.e. a weekly Gregorian calendar with 7 days in each cycle and a weekly Javanese calendar with 5 days in each cycle. Two methods on the simulation dataset are presented. In the first method, the time series regression is combined with the seasonal ARIMA model. The second method applied the two-stage seasonal ARIMA with the different orders of seasonal. These two methods are aimed to remove the different seasonal cycles successively. As the result, the time series regression as preprocessing combined with seasonal ARIMA provides better accuracy compared to the double seasonal ARIMA model, based on the value of RMSE and MAE. This implied that time series regression is able to conceive the pattern of the different seasonal cycles. In conclusion, the time series regression-ARIMA is able to capture the pattern of a non-multiplicative double seasonal pattern, especially for forecasting the simulation data relating to the food material demand.