Price Prediction and Computer Vision in the Real Estate Marketplace

Abstract

House price valuation plays a critical role in real estate decisions. Home buyers and sellers use valuation estimates to judge the fairness of an asking price, and lenders use valuation estimates to assess risk. This thesis will explore potential house price prediction models and describe possible predictors that drive these models. We will present to our models a triad of data types: descriptive, spatial, and visual. Hedonic characteristic spaces describe a house in terms of a set of features. Location information is an integral component of house price prediction. Current advances in the field of computer vision will help us find predictors in images of a house for sale. Using this triad, we will explore the impact of each data type on house price prediction. We will show that each type of data adds value to predictive price models. An important contribution of this thesis is presented with respect to visual data: we will provide a mechanism by which abstract data can be extracted from images and successfully used to improve a house price prediction algorithm.