Function estimation with locally adaptive dynamic models (Q1424615)

This paper presents a nonparametric Bayesian method for fitting unsmooth and highly oscillating functions, which is based on locally adaptive hierarchical extension of standard dynamic or state space models. The proposed approach is fully Bayesian and used a two-stage prior for the unknown regression function. The first stage is first or second order random walk models as discretized Bayesian version of the common roughness penalty for smoothing splines. The second stage consists of analogous smoothing priors for varying variances of the random walk model errors used in the first stage. Inference is fully Bayesian and uses Markov Chain Monte Carlo techniques, combining Gibbs sampling with efficient block moves for functions and Metropolis-Hastings algorithm with conditional prior proposals. Performance of the proposed procedures is investigated by a simulation study, including nonsmooth examples constructed for wavelet shrinkage, and by application to real sales data.