Abstract
The paper designs an automated valuation model to predict the price of residential property in Coventry, United Kingdom, and achieves this by means of geostatistical Kriging, a popularly employed distance-based learning method. Unlike traditional applications of distance-based learning, this papers implements non-Euclidean distance metrics by approximating road distance, travel time and a linear combination of both, which this paper hypothesizes to be more related to house prices than straight-line (Euclidean) distance. Given that - to undertake Kriging - a valid variogram must be produced, this paper exploits the conforming properties of the Minkowski distance function to approximate a road distance and travel time metric. A least squares approach is put forth for variogram parameter selection and an ordinary Kriging predictor is implemented for interpolation. The predictor is then validated with 10-fold cross-validation and a spatially aware checkerboard hold out method against the almost exclusively employed, Euclidean metric. Given a comparison of results for each distance metric, this paper witnesses a goodness of fit (
) result of 0.6901 ± 0.18 SD for real estate price prediction compared to the traditional (Euclidean) approach obtaining a suboptimal
value of 0.66 ± 0.21 SD.