Accounting for locational, temporal, and physical similarity of residential sales in mass appraisal modeling: the development and application of geographically, temporally, and characteristically weighted regression

Geographically weighted regression (GWR) has been recognized in the assessment community as a viable automated valuation model (AVM) to help overcome, at least in part, modeling hurdles associated with location, such as spatial heterogeneity and spatial autocorrelation of error terms. Although previous researchers have adjusted the GWR weights matrix to also weight by time of sale or by structural similarity of properties in AVMs, the research described in this paper is the first that has done so by all three dimensions (i.e., location, structural similarity, and time of sale) simultaneously. Using 24 years of single-family residential sales in Fairfax, Virginia, we created a new locally weighted regression (LWR) AVM called geographically, temporally, and characteristically weighted regression (GTCWR) and compared it with GWR-based models with fewer weighting dimensions.GTCWR was the only model to achieve IAAO-accepted levels of the coefficient of dispersion (COD), price-related differential (PRD), price-related bias (PRB), and median assessment-to-sale price ratio in both the training and testing samples, although it did not fully correct the existence of heteroscedasticity. With lower PRD and PRB levels, the application of temporal weighting to this data set did appear to help reduce indicators of vertical inequity. Along with an equitable, uniform, and defensible methodology that mirrors the sales comparison, GTCWR presents a new AVM that demonstrates an ability to value over 24 years of sales at IAAO standard levels, without the creation and implementation of time-based variables, the trimming of outliers, and time-intensive model specification and calibration