A0517
Title: Efficient divide-and-conquer approach for spatio-temporal modeling of real estate data
Authors: Kapil Gupta - Indian Institute of Management, Bangalore (India) [presenting]
Soudeep Deb - Indian Institute of Management Bangalore (India)
Abstract: Statistical research in real estate markets has recently garnered attention from the perspective of understanding the spatiotemporal dynamics of house prices. Markov chain Monte Carlo (MCMC) is generally used for Bayesian inference in spatio-temporal modelling. However, standard techniques of MCMC are usually slow for large datasets such as real estate data due to the requirement of multiple passes through the entire data in each iteration. A divide-and-conquer spatiotemporal modelling approach is proposed to tackle this problem. The method involves partitioning the data into multiple subsets of sufficient locations and utilizing an appropriate Gaussian process model for each subset in parallel. The parameters corresponding to each subset are then combined to obtain the global parameters for the original problem. The proposed methodology allows us to assess the spatially varying impact of various factors on the house-price dynamics. It is also proved to be much faster than a conventional Bayesian approach. As a real life application of the proposed model, we analyze house price data from London from January 2011 to October 2019, covering 53 bi-monthly time points and 906 middle layer super output areas (MSOAs). The results furnish insightful analysis and render good predictive accuracy, as demonstrated by a cross-validation study.
