This paper examines the predictive properties of the structural Theta model, which represents an extension of exponential smoothing models. These models are widely used in forecasting, particularly for their ability to handle a large number of time series. The Theta model forecasts are evaluated using the M4 database, which is commonly employed for comparing short-term and medium-term forecasting techniques, and contains time-series data at multiple frequencies spanning various fields including macroeconomics, finance, demography, and industry.
The results show that the proposed model provides particularly accurate forecasts in the presence of highly persistent time series (such as the inflation rate) or non-stationary series (such as GDP). Moreover, the model produces more precise point forecasts than the (various) alternative models here considered, though they are computationally more complex.
