Spatio-temporal stability of housing submarkets. Tracking spatial location of clusters of geographically weighted regression estimates of price determinants.

• Novel universal methodology of spatio-temporal comparison of stability of spatial data-driven clusters. • Integrates spatial econometrics, spatial statistics and unsupervised learning. • K-means clusters of GWR model coefficients are rasterised and the mode cluster id labels given raster cell, Rand index and Jaccard Similarity test if pairs of pairs of raster cells belong to the same clusters in both periods. • Clusters of GWR coefficients indicate spatio-temoral stability of submarkets, while clusters of standard errors of GWR coefficients detect stability of local spatial heteroscedasticity. • Methodology can be applied to any spatial phenomena, including housing submarkets. This paper fills the gap in rich housing literature by testing the spatio-temporal stability of real estate submarkets. We start with standard Geographically Weighted Regression (GWR) estimation of the hedonic model on point data, and we cluster model coefficients to detect housing submarkets. We check spatio-temporal stability - we add novelty by comparing if clusters move over space or stay in the same place. We rasterise surface and apply the Rand Index and Jaccard Similarity to check if clusters assigned to raster cells yield stable spatial structure. This approach allows for quantitative assessments of how much determinants of price are stable over time and space. The same mechanism applied to standard errors of GWR coefficients is a good test of the spatio-temporal stability of local heteroscedasticity. A Case study of apartments' transactions in Warsaw-Poland for the 2006–2015 period, evidences relatively high spatio-temporal stability. [ABSTRACT FROM AUTHOR]