Modelling Rainfall Time Series: The Case for Score-Driven Models with Structural Components

Understanding rainfall dynamics is essential for economic planning and daily decision-making. While physical models remain important, the literature and industry also rely on machine learning methods that generate point forecasts for rainfall time series. However, such point estimates fail to capture uncertainty and distributional features. Some authors such as Harvey and Ito [2019] suggest addressing this uncertainty through score-driven (GAS) models.

Despite recent advances, significant gaps remain. There is limited empirical comparison between score-driven and machine learning approaches. Furthermore, rainfall series are rarely decomposed into trend and seasonal components. Finally, evidence across different climates and temporal aggregations, particularly monthly data, is scarce. These limitations are addressed in this article by proposing an integrated framework for rainfall modeling, applying the findings to different climates, and combining score-driven models with structural components by Harvey and Peters [1990].