Combining Property Price Predictions from Repeat Sales and Spatially Enhanced Hedonic Regressions

Hedonic regression and repeat sales are commonly used methods in real estate analysis. While the merits of combining these models when constructing house price indices are well documented, research on the utility of adopting the same approach for residential property valuation has not been conducted to date. Specifically, house value estimates were obtained by combining predictions from repeat sales and various hedonic regression specifications, which were enhanced to account for spatial effects. Three of these enhancements—regression kriging, mixed regressive-spatial autoregressive, and geographically weighted regression—are widely utilized spatial econometric models. However, a fourth augmentation, which addresses systematic residual patterns in regressions with district indicator variables and the presence of outliers in housing data, was also proposed. The resulting models were applied to a dataset containing 16,417 real estate transactions in Oslo, Norway, revealing that when the repeat sales approach is included, it reduces the median absolute percentage error of solely hedonic models by 6.8–9.5%, where greater improvements are associated with less accurate spatial enhancements. These improvements can be attributed to the inclusion of both spatial and non-spatial information inherent in previous sales prices. While the former has limited utility for well-specified spatial models, the non-spatial information that is implicit in previous sales prices likely captures otherwise difficult to observe phenomena, potentially making its contribution highly valuable in automated valuation models. © 2019. This is the authors' manuscript to the article.