SETAR-Tree: a novel and accurate tree algorithm for global time series forecasting

Abstract Threshold Autoregressive (TAR) models have been widely used by statisticians for non-linear time series forecasting during the past few decades, due to their simplicity and mathematical properties. On other hand, in community, general-purpose tree-based regression algorithms (forests, gradient-boosting) become popular recently ease of use accuracy. In this paper, we explore close connections between TAR trees. These enable us rich methodology from literature on define a hierarchical model as tree that trains globally across series, which call SETAR-Tree. contrast do not primarily focus forecasting, calculate averages at leaf nodes, introduce new forecasting-specific algorithm global Pooled Regression (PR) leaves allowing learn cross-series information also uses some time-series-specific splitting stopping procedures. The depth is controlled conducting statistical linearity test commonly employed models, well measuring error reduction percentage each node split. Thus, proposed requires minimal external hyperparameter tuning provides competitive results under its default configuration. We develop forest where forecasts provided collection diverse SETAR-Trees are combined process. our evaluation eight publicly available datasets, able achieve significantly higher accuracy than set state-of-the-art benchmarks four metrics.