A1234
Title: Seasonal adjustment using Gaussian processes
Authors: Merlin Scherer - ETH (Switzerland) [presenting]
Abstract: A seasonal adjustment method is introduced for economic time series using Gaussian processes (GPs). Effective seasonal adjustment is crucial to accurately identify economic trends and turning points. While widely used, methods such as X-13ARIMA-SEATS and STL can struggle with structural breaks, outliers, and endpoint instability, which may limit their ability to capture evolving seasonal patterns and can lead to revisions when new data are introduced. Gaussian processes provide a non-parametric framework capable of addressing these limitations through kernels explicitly selected to reflect the characteristics of trend, seasonal, and irregular components. The proposed GP framework aims to effectively capture complex seasonal dynamics, including multiple seasonalities and non-linear patterns. Simulation studies show promising results, suggesting that GP-based approaches can significantly reduce historical revisions and handle structural breaks and outliers more effectively than conventional techniques. The contribution is in methodologically developing a GP-based seasonal adjustment approach designed explicitly to improve forecasting accuracy and enhance economic analysis.
