Multiple seasonal time series forecast using decomposition, ARIMA model and discrete Fourier transform: a preliminary study. Kong Hoong Lem, Yi Xian Yap
Universiti Tunku Abdul Rahman
Abstract
Multiple seasonalities often appear in time series observed at high frequency. For example, an hourly observed data may exhibit multiple seasonal patterns due to combination of daily, weekly, monthly or even yearly periodicity. In this study, we first decomposed the data into components using MSTL. For the seasonal components, we leveraged the properties of discrete Fourier transform to serve as a regressor, whereas the trend and the remainder components underwent an ARIMA model. Experiments were done on two datasets and compared with the TBATS approach. The proposed method yielded superior forecast performance.
Keywords: multiple seasonal time series, discrete fourier transform, ARIMA, MSTL
