Title: Model confidence sets for nonparametric time series models
Authors: Rolf Tschernig - Universitaet Regensburg (Germany) [presenting]
Christoph Rust - University of Regensburg (Germany)
Abstract: Nonparametric estimation of nonlinear autoregressive models requires an appropriate bandwidth choice and lag selection. This can be achieved by using cross-validation. Alternatively, plug-in-bandwidth can be combined with lag selection criteria well suited for nonparametric time series models. We propose another approach: Associate for a given set of lags a specific bandwidth choice with a specific model. Then different bandwidths imply different models and bandwidth selection corresponds to model selection. This allows us to use model confidence sets. Such sets estimate the set of superior models that contains all those models that exhibit identical and lowest risk w.r.t. a user specified loss function such as quadratic loss. In practice model confidence sets typically contain several models. We suggest applying all models in the model confidence set for the modeling purpose, e.g. prediction. In a next step one can combine the model confidence sets for each lag combination to a new set and estimate a new model confidence set such that lag selection is included as well. In a Monte Carlo study we compare our proposal to existing methods.