Title: Distributional trees and forests for circular data
Authors: Moritz Nikolaus Lang - University of Innsbruck (Austria) [presenting]
Lisa Schlosser - Universitaet Innsbruck (Austria)
Torsten Hothorn - University of Zurich (Switzerland)
Georg Johann Mayr - University of Innsbruck (Austria)
Reto Stauffer - University of Innsbruck (Austria)
Achim Zeileis - University of Innsbruck (Austria)
Abstract: Circular data can be found in a variety of applications and subject areas, e.g., hourly crime rates in the social-economics, animal movement direction in ecology, and wind direction as one of the most important weather variables in meteorology. For probabilistic modeling of circular data the von Mises distribution is widely used. While most existing approaches are built on additive regression models, we propose an adaption of regression trees to circular data by employing distributional trees. In comparison to the more commonly-used additive models, the resulting distributional trees are easy to interpret, can detect non-additive effects, and automatically select covariates and their interactions. In addition, as a natural extension, ensembles or forests of such circular trees are introduced that can further improve the forecasts by regularizing and stabilizing the model. For illustration, short-term probabilistic wind direction nowcasts at different airports are obtained in order to direct airplanes to a safe landing. The predictive skill of the novel approaches is benchmarked with an additive regression model plus a persistency and a climatology model, employing the circular continuous ranked probability score. The proposed methods for circular distributional trees and forests are available in the R package 'disttree' from R-Forge.