A Comparative Study of Forecasting Models for Different Frequencies and Domains Using the M4 Competition Series

The purpose of this study is to compare the performance of statistical forecasting models such as the exponential smoothing model family, the ARIMA model, and the TBATS model in short-term point forecasting with univariable time series with multiple frequencies and domains. This study used 100,000 time series from the 4th Makridakis Forecasting (M4) Competition, which included annual, quarterly, monthly, weekly, daily, and hourly time series. with frequencies of 1, 4, 12, 1, 1, and 24, respectively, which are macro- and micro-level data, demographic, industry, financial, and other data are divided into training and test sets. Model performance is measured using three criteria: sMAPE, MASE, and processing time. The study found that 1) ARIMA and TBATS models produced lower sMAPE and MASE performance measures than ETS models for macro- and micro-level data. demographic, industry, and financial data; and 2) the ARIMA and TBATS models have the lowest error measurement and are not different from one another. for every frequency 3) The ARIMA and TBATS models have complex procedures for determining the best model. This results in processing times that are 2.3 and 4.4 times longer than the ETS model, respectively, and the ARIMA model is not appropriate for high-frequency data.