Hierarchical Bayesian Fuzzy Clustering Approach for High Dimensional Linear Time-Series

This paper develops a computational approach to improve fuzzy clustering and forecasting performance when dealing with endogeneity issues and misspecified dynamics in high dimensional dynamic data. Hierarchical Bayesian methods are used to structure linear time variations, reduce dimensionality, and compute a distance function capturing the most probable set of clusters among univariate and multivariate time-series. Nonlinearities involved in the procedure look like permanent shifts and are replaced by coefficient changes. Monte Carlo implementations are also addressed to compute exact posterior probabilities for each cluster chosen and then minimize the increasing probability of outliers plaguing traditional clustering time-series techniques. An empirical example highlights the strengths and limitations of the estimating procedure. Discussions with related works are also displayed.