A0497
Title: Product partitioned dynamic factor models: An application characterising Melbourne housing price co-movements
Authors: Zhuo Li - Monash Univeristy (Australia) [presenting]
Abstract: Bayesian nonparametric methods are used to accommodate the random partition issue surrounding block dynamic factor models (DFMs). The DFM with a block structure assumes that time series within the same block co-move with cluster-specific factors, where series from distinct blocks only communicate through global factors. The premise of this modelling framework is that the cluster assignment is given \textit{ex-ante}. In the presence of \textit{partition uncertainty}, three novel modelling frameworks are proposed that feature a random partitioning process in dynamic factor models to endogenously determine the partition size and cluster memberships, where each proposed model is designed for its own purpose, since no single model suits all applications. From the simulation experiment, the in-sample fit results show that allowing for the random partitioning process over the block DFM provides additional precision in prediction. Such precision gain is further confirmed in the empirical study of Melbourne housing price co-movement analysis, for both in-sample and out-of-sample exercises. In addition, it is suggested that the effect of economic drivers on housing price dynamics may not be fully captured if suburbs-level partition uncertainty is ignored while accommodating spatial dependence.
