Hedonic Pricing of Real Estate using Computer Vision

Automated real estate valuation models (AVMs) play an increasingly important role in the industry (bank valuations, index calculations, etc.). These models are based on hedonic price models and rely on observed prices and value-influencing variables. However, only variables that are fully and uniformly recorded can be taken into account in such models. Due to increasing digitization, real estate listings are nearly always accompanied by images of the property, which contain a substantial amount of information and therefore represent a great potential for these models. While image analysis is already a fixed component in areas such as medicine, facial recognition, etc., it has so far only been applied to a limited extent in the real estate industry and real estate valuations. To address this gap, this project aims to automatically extract information from real estate images and to integrate this information into hedonic pricing models.

Methodology and objectives

The aim of this research project is to integrate automated image recognition into hedonic pricing models of real estate. The focus is on (i) the analysis of features that can be automatically extracted from the images (e.g., the classification of the condition of the property) and (ii) the prediction accuracy of the hedonic price model. In particular, the condition or quality of a property represents a significant factor influencing value. However, people involved in property valuation may perceive characteristics relating to amenity, quality, or condition differently. The automated extraction of these characteristics from images can therefore create a uniform standard and thus improve the regression results. The potential and the possibilities of image recognition are investigated on different modeling levels. The regression analyses range from analyses based on individual image features to a completely image-based approach without further information.

The methodology of the project is based on a two-part approach: (1) Image analysis techniques are used to extract value-influencing features from real estate images (exterior/interior images as well as aerial/satellite images of properties). For this purpose, multiple neural networks (Convolutional Neural Networks/CNNs) are constructed and trained using either image patches or the entire segmented building contours as input. The classification of the images is then performed based on the extracted features. After the network has been trained, only one input image is required for classification and all steps up to the classification result are fully automatic. (2) The features extracted from the images are then implemented within hedonic pricing models. Here, it is investigated whether information extracted from the images improves the price estimation and which features have the greatest influence on the price in this context.