China's Urban Housing Market: Applications of Machine Learning Methods to Price Prediction in Select Tier 1 and 2 Cities

Abstract

The focus of this thesis is on applying machine learning techniques to the investigation of China’s urban real estate boom. To date, much of the existing literature on Chinese housing markets has sought to construct an accurate house price index for the country’s major cities, usually through traditional approaches such as the repeat-sales method and the hedonic method. Unfortunately, existing methods of house price index construction have little to no predictive power. For example, although previous applications of hedonic price regression adequately control for factor inputs, they face severe limitations given their simple linear form. While such traditional approaches may prove useful in the absence of large and meaningful data sets, developments in machine learning combined with improvements in accessible data are showing promise. Thus, this thesis uses nonparametric machine learning algorithms to train predictive models of Chinese residential real estate prices. Boosting is applied to improve the performance of ordinary least squares regression, and its results are compared against those achieved by decision trees, random forests, and feedforward neural networks. These models are trained and tested on nearly one million secondary market home sale transactions across seven major Chinese cities, which include Beijing, Shanghai, Hangzhou, Shenzhen, Guangzhou, Chengdu, and Chongqing. Ultimately, random forest regression emerges as the leading performer, demonstrating impressive accuracy of price prediction and implying believable rates of historical growth.