Steven Farber

This paper compares the use of several local regression models using residential property valuation as a case study. The dataset consists of 19,007 housing sales observations occurring between July 2000 and June 2001 within the City of Toronto. Presently, assessment offices rely on a vast number of structural variables in order to sufficiently model market values. Building on the findings of Farber and Yeates (2005), that local models using a small set of variables have a similar performance to the models used in industry, the aim of this paper is to compare the results of several localised regression models. In this paper, Global OLS models are compared to a variety of local models including spatially autoregressive techniques (SAR), geographically weighted regression (GWR), moving window regression (MWR), and a spatial model of error heterogeneity. The models are all calibrated using a small set of parsimonious and defendable variables. Spatial autocorrelation amongst the residuals, as measured with Moran's Index, is used as an indicator of spatial bias in the estimates. The results show that GWR produces the most accurate and least spatially biased estimates. MWR, the simpler alternative, produced results that rivalled those of GWR without incorporating the unknown effects of distance-decaying weights. This indicates that the additional estimation accuracy of GWR may not be worth the cost of the conceptual abstraction of a distance weighting scheme.