House Price Estimation with Alternative Data and Machine Learning : A Case Study of Residential House Prices in Surrey, Canada

Increasing availability and acceptance of Alternative Data in recent years have created challenges and opportunities for experimental exploration from unorthodox sources. In the meantime, rising popularity of Machine Learning algorithms has probably accelerated interest for more data ingestion, at the cost of demand for higher compute power. In this study, we attempt to propose and apply an empirical framework for house value attribution that incorporates Macroeconomic, Neighborhood-specific, and property-specific data, compatible with spatially adjusted econometric approaches and ML algorithms. Our case analysis of all residential properties in the City of Surrey, Canada, uses Random Forest method with over 40 variables on all 3 levels. The empirical results show impressive modeling accuracy with 76% Adjusted R-square for continuous price estimation and 10% misclassification rate for market segment analysis. Across price spectrum of market segments, property interior size and Mortgage Rate seem to possess high importance for all sub-markets. Certain Neighborhood variables seem to exhibit more segment disparity than others, a sign of differential customer preferences. Take Safety and Mobility for example. While Safety seems to be ubiquitously important for all market segments, Mobility is attributed with more value almost exclusively for High-end. This case study shows strong potentials for further analysis and interpretations through Economic and Social lens