Analyzing real estate transaction and pricing data via statistical machine learning

点击翻译按钮获取中文摘要

Real-estate property price estimation and prediction is key to Canadian home buyers, investors and policy makers. To date, most real-estate market assessments accessible by Canadian users are based on big trend analysis, qualitative prediction or traditional real-estate market indices, leading to a limited accuracy. The research literature has presented a large number of both parametric and non-parametric models to assess the market value of an individual property based on the sold prices of other properties transacted in the same region. However, even with non-parametric models and the latest kernel-based regression, existing approaches have limited accuracy and are unable to handle irregular regions with interior uninhabitated areas or to capture price jumps across heterogeneous areas. To overcome these difficulties, in this project, we will propose and extensively study a number of new statistical learning tools for applications to real-estate valuation, price estimation and prediction, including finite element analysis, spatial spline regression, roughness and fused-LASSO-type regularization, low-rank matrix completion and state-space models with penalized Kalman filters. With carefully designed structures, our proposed models can be decomposed and solved using convex optimization. We will also study the implementation issues for faster computation in commercial real-estate market assessment applications with a large amount of data.