Non-Gaussian score-driven models with non-linear unobserved components combinations

A commom technique in time series modeling is to decompose the time series into it's trend and seasonal components. In Score-Driven Models class, this decomposition usually takes an additive form, so that the series is expressed as the sum of its trend and seasonal components. However, it is not unusual that, even with the seasonal component being considered into the model, the residuals still show signs of seasonal dependency that were not captured by the model. With that said, the main objective of this article is to study if a non-linear combination of those components is able to improve forecast accuracy in score driven models. Our results, albeit limited, indicate that this non-linear combination can be more effective in explaining the variability of seasonal time series.