The Potential of Machine Learning Real Estate Valuation Models (5 mins)

Property valuation is a necessary task for parties across the real estate industry. Development, investment, lending, and brokerage all rely on determining the value of property by either using external valuations and appraisals or by constructing internal valuation models, typically on ARGUS or Excel. Underwriting is core to the real estate industry as it is used to determine both transaction prices and lending limits. Value is typically estimated by combining the output of cash flow models that are driven by cap rates with a relative valuation based on comparable recent sales. Both cap rates and comps are largely imperfect measures as neither can be objectively adjusted for a property’s unique location or characteristics, while both are lagging as they are based on past transactions. Though yet untapped, machine learning and predictive analytics tools have the potential to upend the valuation and appraisal process in real estate.

Machine Learning in Commercial Real Estate

It’s a buzzword that we’ve all heard, but what does “machine learning” really mean? According to the SAS (Statistical Analysis System) Institute, “machine learning is a method of data analysis that automates analytical model building. It is a branch of artificial intelligence based on the idea that systems can learn from data, identify patterns and make decisions with minimal human intervention.”[i] The combination of increasing data availability and advancements in machine learning has lead to transformative uses across industries.

Though machine-learning based appraisal models are now prevalent in the single-family market, particularly with the uptake of Zillow’s “Zestimate” tool, this technology has yet to widely penetrate the commercial real estate industry. In June of 2017, a group of researchers in the Netherlands developed a test automated valuation model (AVM), based on machine learning, for US multifamily assets in three states.[ii] Instead of relying on comps or cap rates, this AVM utilized decision tree models to divide the data into subsets, then hedonic models to apply a regression on each subset of data. The model was able to determine property values within a 9% error (comparing value estimates to actual transaction prices), compared to a 12% industry average between 1984 and 2010 as reported by the National Council of Real Estate Investment Fiduciaries (NCREIF). In addition to a lower error level, the model runs at no cost and can produce valuations instantly, contrasting the weeks-long process typically required to have a property appraised. This test model demonstrated the potential of AVMs, but automated valuation technology based on machine learning is still in its infancy.

Actors in the real estate industry assess the value of a property through their own lens, which often leads to different determinations of value on the same property. An optimistic developer, for example, may price some of the potential value creation, whereas the value to a pessimistic lender would be driven by historic certainties. The best use-case for AVMs as they stand, given the data and technology available today, is for the models to augment or replace appraisals. Appraisals are relied upon by a wide variety of users, from lenders who are seeking an estimate of value, to accountants who might be looking to mark property to market, to a range of parties seeking to comply with regulatory requirements that require an independent determination of value. Instead of relying on appraisals, which can take time and are often costly, these parties could instead begin to rely on automated models. Though the diligence that appraisers uphold extends beyond what a model could undertake, for example appraisers will often walk a property as part of their appraisal process, models may be able to statistically determine value more accurately, especially when self-learning is involved. AVMs could also uncover arbitrage opportunities in many real estate arenas. For example, AVMs could estimate the value of a REIT’s underlying assets instantly, which can be compared to the price of the security.[iii]

Developers and investors rarely rely on appraisals alone, as they instead want to reflect their own assumptions about the market and their strategic plan for the property in an estimate of value. Machine learning models determine value by comparing attributes of properties transacted in the past, and market conditions at the time of those transactions, to the attributes and timing of the target. Unlike today’s development and investment valuations, they are not based on cash flow models. The value proposition of AVMs to developers and investors is tangent to a simple appraisal; in addition to a value as-is today, these parties should be able to change assumptions and estimates to determine potential values given different value-add strategies and market conditions. With further development, AVMs may even be able to direct strategy, as opposed to just output the fruits of it. If realized, this could disrupt the transaction process, as each bidder would arrive at the same determination of value. By removing pricing inefficiencies, intangible factors such as relationships may determine which bidder acquires a property.

Barriers to Machine Learning Models

The commercial real estate industry has yet to see an uptake of machine learning automated valuation models. One available product, Enodo, is a software that uses machine learning and statistical modeling to determine NOI and operating expenses for the multifamily market. This software uses a model similar to that of the Dutch researchers cited above, but is focused on the asset’s characteristics (for example, level of finishes), in addition to locational characteristics.[iv] It can be argued that predictive modelling is simpler for multifamily assets due to the relatively homogeneous nature of the product in comparison with office, retail, and industrial space. Commercial leases are assessed based on some quantitative factors that can be modeled with machine learning, such as lease length and tenant credit quality, along with qualitative judgements that are much more difficult to assess automatically, such as a tenant’s likelihood to renew. With the lack of product offering across asset classes, firms are left to either wait for products to become available or attempt to develop AVMs in-house, forcing them to stray from their core competency. Competitive forces have yet to incentivise CRE firms to innovate as nothing has challenged the status quo, regardless of the potential for cost savings and improved decision making that AVMs present.

Another barrier to AVM development and propagation may be conflicting interests of large real estate services firms. These services firms have the data and the resources to innovate, but would be directly cannibalizing major lines of business. The third-party appraisal market is estimated to gross $90bn annually – it’s a cash-cow that CRE service firms don’t want to disturb. With accurate automated valuations, the appraisal industry, which is a service these firms provide, may become less relevant. Low-cost automated valuations would push pricing down the point where the cost of traditional appraisals may not be exceeded by the market price for these services. Brokerage, to a lesser extent, would be threatened as well, as there may be little room for negotiation if buyers and sellers have the same determination of value.

With an explosion in data availability and continuing development in machine learning software, automated valuations are likely to grow in prominence. The market size presents a lucrative opportunity for new entrants, and the real estate industry has yet to face enough competitive pressure to lead this innovation internally. As with disruptors in other industries, we may see a team of data scientists take advantage of the opportunity that machine learning in real estate presents by developing models that they then sell as a service to the industry. Though AVMs have yet to be developed to a degree of sophistication with which they can be relied upon for decision making in commercial real estate, with time these automated models could pose a threat to appraisals and could serve as a valuable tool to actors across the industry.

[i] Predictive Analytics and Machine Learning

[ii] Big Data in Real Estate? From Manual Appraisal to Automated Valuation (Kok, Koponen, Martínez-Barbosa, 2017)

[iii] Big Data in Real Estate? From Manual Appraisal to Automated Valuation (Kok, Koponen, Martínez-Barbosa, 2017)

[iv] Enodo: Automated Multifamily Underwriting

Print Friendly, PDF & Email