• Unexpected by AI
  • /
  • End-to-end deep learning approach to property pricing prediction

End-to-end deep learning approach to property pricing prediction

Estimating the value of real estate is necessary for purchase and sale, investments, and financing. The market value of a property is determined by a number of characteristics, such as, its location, size, construction year, number of bedrooms, and others. An additional key element predicting the value of residential properties is what constitutes their visual aspects. Information deriving from the real estate interior and exterior photos can be of assistance when calculating value estimates. However, the interpretation of the property visual characteristics can vary significantly from person to person, and it is usually performed by appraisers or real estate agents, not by automated systems.

A framework allowing for automated value assessment based on both quantifiable characteristics of a property and its visual appeal could provide a great assistance in making better-informed decisions about the monetary worth of an estate. The approach adopted by Kellify envisioned the integration of multi-modal, heterogeneous data consisting of a list of property main characteristics (i.e. number of rooms, location, available surface etc.), textual description of the house style and character, interior, exterior and satellite pictures. The integration of this broad spectrum of data into a Deep Learning model allows for a highly accurate and unbiased estimation of the property value.

Currently, most of the services for real estate listings of flats and houses for sales use the comparative market analysis approach to calculate the possible value of a property. This model assumes that the price of a house can be appraised based on the selling prices of similar properties in the same area. This approach, even though can be useful in getting a grasp of the market trends, does not account for a number of factors, such as renovations done in the house, its visual design and style. Consequently, it carries the risk of under- or overestimation of the property value. The solution commonly adopted by the real estate listing sites is to provide a price range instead of accurately calculated value which, however, often results in poorly informative predictions.

Banks aiming at determining the house value for a mortgage or a home equity loan approval usually use a licensed appraiser’s assistance. They base their estimates on the basic home data, such as, the square footage, number of bedrooms and bathrooms, construction year, as much as on their subjective assessment of the property conditions. No matter how experienced and knowledgeable the expert is, their judgments are never free of personal bias and will always be exposed to the risk of miscalculation. These kinds of errors can lead to poor decision making and risk management.

Undertaking a scientific approach to the predictions of properties’ values leads to more accurate and unbiased estimates. Seeking for the most reliable solutions, we trained and tested our models on more than 3 million records of properties listed on one of the real estate digital platforms. Having analyzed the information available online, we have managed to obtain price estimation predictors with a mean accuracy of 85%. The employment of this innovative and automated valuation model holds the potential of shaking up and rebuilding the real estate, investments, and financing markets. It evolves into a cornerstone of properties’ marketability evaluation and improved risk management in the banking sector. We are joining forces with the leading real estate listing sites to help them create a more accurate reflection of the market reality and assisting world-class players of the financial industry in minimizing their risks and losses. 

Insights direct to your inbox​