CFE 2019: Start Registration
View Submission - CMStatistics
Title: Sparse change-point VAR models Authors:  Arnaud Dufays - Namur University (Belgium) [presenting]
Jeroen Rombouts - ESSEC Business School (France)
Yong Song - University of Melbourne (Australia)
Zhuo Li - University of Melbourne (Australia)
Abstract: Change-point (CP) VAR models face a dimensionality curse due to the proliferation of parameters that arises when new segments are detected. To handle large data set, we introduce the Sparse CP VAR process that determines which parameters truly vary when a break is detected. By doing so, the number of new parameters to estimate at each segment is drastically reduced and the CP dynamic is easier to interpret. The Sparse CP VAR model disentangles the dynamics of the mean parameters and the covariance matrix. The latter is driven by an infinite hidden Markov framework while the former stands for a CP dynamic with shrinkage prior distributions on the first-difference parameters. We argue that limiting the number of possible breaks in the mean parameters has several theoretical and empirical advantages over the standard practice of time-varying parameter (TVP) models. An exhaustive Monte Carlo study highlights that the framework operates for detecting the correct number of breaks per model parameter in small and medium-sized dimensional settings. The Sparse CP VAR model also takes advantage of common breaks in the cross-sectional dimension to more precisely estimate them. Our applications on 7 and 25 macroeconomic systems highlight that the Sparse CP VAR model help to interpret the detected breaks. In addition, the Sparse CP VAR model outperforms several recent TVP and CP VAR models in terms of log predictive densities.