Big data in real estate: from manual appraisal to automated valuation

We developed a practical application of “big data” in combination with sophisticated modeling techniques, providing an automated, machine-based valuation model for the commercial real estate sector. Automated valuation models (AVM) are clearly beating traditional appraisals — the absolute error of the automated model now stands at 9 percent, which compares favourably against the accuracy of traditional appraisals, while the model can produce an instant value at every moment in time, at a very low cost. Read our full article in the Journal of Portfolio Management.

Real estate is the largest asset class in the world, received its own global industry classification (GICs) in August 2016, and makes up, on average, 5.1% of any institutional real estate portfolio. But determining the value of commercial real estate remains elusively hard, with a workforce of 74,000 appraisers in the U.S. alone still manually assessing the value of assets sometimes worth billions of dollars.

Appraisals are typically based on the capitalization of the net income of an asset, using a yield (or: cap rate) that has been inferred from neighbouring transactions, “comps.” However, transactions of comparable properties are never truly comparable, neither in time, nor in building characteristics, and appraisers are typically anchored on previous valuations or the previous transaction price of a building. The result is that property appraisals typically lag the market and provide “smoothed” approximations of true market values, with values that are artificially low in bull markets and high in bear markets.

The precision of appraisals has been topic of both popular debate and academic study for multiple decades. We developed a practical application of “big data” in combination with sophisticated modeling techniques, to develop an automated, machine-based valuation model for the commercial real estate sector. To build the model, we exploit a wide set of both standard demographic and economic measures and more modern, “hyperlocal” metrics, such as proximity to music events, bars and restaurants, green space, and local crime incidence. Rather than traditional hedonic models, which are limited both statistically and by a researcher’s predisposition towards “standard” explanatory variables, we apply assisted machine learning models that rely on (stochastic) decision trees.

Automated valuation models (AVM) are clearly beating traditional appraisals — the absolute error of the automated model is 9 percent, which compares favourably against the accuracy of traditional appraisals, while the model can produce an instant value at every moment in time, at a very low cost. The models also show the importance of using “hyperlocal” information on the location of an asset. Our automated valuation models are directly applicable for real estate lenders and investors, and have important implications for the traditional appraisal industry.

For underwriting and refinancing purposes, automated valuation models can provide an instant indication of property value, which saves both portfolio investors and lenders, as well as those interested in a single property, significant time and resources. This is especially beneficial on the lending side, where the debt service coverage ratio is a leading indicator, with the LTV as an important, but secondary input in the underwriting process. Automated appraisals can provide banks, insurance companies, pension funds, and other institutional investors and lenders with an accurate, instant revaluation of the assets on the balance sheet, obviating the need for an annual (or quarterly) expensive and lengthy revaluation process which regulators increasingly require. Such instant assessment of the market value of the book is especially useful in times of market volatility.

The timeliness of automated valuations also allows for the development of financial trading strategies and innovations in the underwriting process. For example, instant assessment of the value of a commercial real estate asset can be used for new investment models, comparable to what Opendoor has developed in the single-family market. Instant and accurate assessment of the value of a portfolio of assets could be used for arbitrage trading on real estate investment trusts (REITs), not dissimilar to the quant trading strategies by some hedge funds. And banks can resort to automated origination models that obviate the need for (costly) underwriting processes.

For the industry to accept an automated valuation over a traditional, manual appraisal will take significant education, as well as market adoption by leading lenders and investors. Exciting times, although perhaps unnerving to some…

Read our full article in the Journal of Portfolio Management.