A Local Gaussian Process Regression Approach to Mass Appraisal of Residential Properties

Mass appraisal of single-family homes is now possible using scalable versions of Gaussian process regression. However, it is here shown that the valuation accuracy of such models tends to suffer for higher priced properties where samples are thin. To remedy this, we turn to the industry standard practice of identifying small sets of comparable properties (comps) using rules loosely based on assessor methods. By using a real-world empirical dataset built on a decade’s worth of Assessor database backups, it is shown that this combination of domain expertise with machine learning improves predicted appraisals in a significant way. As part of this analysis, we also introduce and discuss a novel metric, average comp quality, for evaluating the predictive effectiveness of alternative comp sets.