Probabilistic Regression Trees Applied to Time Series

Taiane Prass (UFRGS)

Abstract: Probabilistic regression trees constitute a flexible class of nonparametric methods that replace deterministic partitions with smooth associations between observations and regions of the sample space. This formulation allows the reduction of instabilities associated with classical trees and broadens their applicability to problems with complex structures. In this talk, applications of probabilistic regression trees to time series will be discussed, with emphasis on predictive modeling, unsupervised learning, and treatment of missing data. Methodological and computational aspects of these models will be presented, as well as examples involving temporal dependence, exogenous covariates, and nonlinear structures.