Accurately establishing a home’s value is not just an important factor in real estate transactions and mortgage lending, it is integral to household wealth creation, as homeownership continues to be one of the primary pathways for Americans to build wealth. Inaccurate appraisals and valuations can impact the type of mortgage financing a potential buyer receives, whether a seller commands a fair value when selling their home, or whether a homeowner has opportunities to tap into home equity.
Traditionally, a property valuation is conducted by a certified appraiser — a practice that appraisal industry leaders have recently acknowledged may contribute to bias and discrimination for certain borrowers and markets. One result has been the historic undervaluing of homes in majority-Black neighborhoods. So, what can be done to mitigate bias and inequality in the valuation process? The answer lies in evolving valuation technology.
First, we must recognize that traditional appraisals are a unique blend of art and science. The “science” is driven by hard data and analytics: previous home sales, combined with information about the property and location. The art is derived from the expertise of the appraiser: their professional assessment of the property condition and how it compares to other homes sold. Unfortunately, this art form is inherently subject to human bias, whether conscious or unconscious.
Consider two types of bias that can impact appraisal values: neighborhood and individual. Neighborhood bias would be akin to an appraiser redlining and adjusting values for an entire neighborhood because the area is home to a particular demographic group. This type of bias impacts the value of all homes in that area.
Meanwhile, individual bias suggests a prejudice based on an individual’s race, culture, gender, sexual orientation or other characteristic. We have all read the stories of minority homeowners hiding their family photos and eliminating cultural clues from the home in hopes of ensuring a fair appraisal value. This type of bias and discrimination could happen in any neighborhood.
However, the home valuation process has begun to evolve, presenting new opportunities to remove bias from the picture. In the last year, limitations of in-person valuations during the pandemic, coupled with technological advances, have accelerated wider acceptance of alternative valuation products, including automated valuation models (AVMs).
The benefit of these models is their ability to help minimize the potential for bias by focusing more on the “science” and less on the “art.” As the industry relies more and more on model-based solutions like AVMs, the promise of accurate and impartial valuations is on the horizon.
Predictive models themselves, unlike humans, lack emotion and therefore inherently lack the associated biases. Yet, it is important to understand that models are only as good as the data they are fed (or not fed). Most data for property valuations is collected in person by the same person determining the final value, making those values susceptible to human bias. And so, even when applying technology-based alternatives to in-person appraisals, biased data can impact values if not recognized.
For instance, an appraiser walks through a home and unconsciously notes cultural elements. The appraisal comes in $40,000 below market value, and, as a result, the home sells for a lower price. Now that sale is part of the data set that will inform future appraisals and home sales in the neighborhood. An AVM will use this sale record as part of the formula to estimate value. Thus, bias can still be perpetuated in data even when the human component is seemingly removed from the process.
Fortunately, with advances in modeling we now have the technology and tools within our power to begin correcting these issues. The solution lies in training our models to identify and eliminate bias in the historical data set and feeding fresh data into the modeling tool that is free from human bias:
- Data augmentation — Sometimes the data available for modeling is insufficient in breadth or depth. These “thin” data sets may deliver inaccurate results. By supplementing this data with additional ancillary information or breaking data into constituent parts, models can sometimes increase accuracy. For example, instead of using a concept such as “neighborhood” as a single input, we can treat individual variables, such as the age of the home or the distance from other structures, as separate data inputs that can produce a more tailored valuation and convert more information from limited data.
- Machine and deep learning — These sophisticated algorithms not only analyze data and look for patterns, but also correct and refine their conclusions based on new and changing data. In this way, machines can learn to separate inaccuracies or “noise” from the data and focus instead on the most relevant information that consistently delivers the most accurate valuations without bias. Neural networks are a terrific example of how modern analytics can mimic human behavior, all the while potentially weeding out biases. And while explainability, or the ability to justify results, is necessary in models, explainability of the differences between two appraisers’ results can often seem like a black box too.
- Artificial intelligence — Perhaps the most exciting approach to reducing bias is through the growth of artificial intelligence. One promising discipline known as Computer Vision analyzes images to assess a home with virtually no human input. By not telling the model the demographics of the homeowner or of the neighborhood, we can see the home, just like the visiting appraiser, but exclude unhelpful, prejudicial data. In this way, we can begin to eliminate unintentional human bias and replace it with objectivity.
Accurate property valuations are essential for both lenders and borrowers, but they also have a key role to play in making homeownership accessible, establishing housing equity and building generational wealth. While AVMs have the potential to deliver objective valuations without human bias, they rely on data collected by human beings, rendering them susceptible to bias.
Advanced technology can help overcome those errors by augmenting data and applying processes that result in more equitable and accurate conclusions.
Artificial Intelligence and automated valuation solutions may not eradicate human bias in the broader housing industry, or eliminate all model-borne biases in valuations, but they put us a step closer to creating a fair system for all Americans to benefit from their dream of homeownership. Sharing a goal of reducing bias is important, and we should embrace models trained to minimize or eliminate bias to achieve this mission.
This column does not necessarily reflect the opinion of HousingWire’s editorial department and its owners.
To contact the author of this story:
Steve Gaenzler at steve.gaenzler@radian.com.
To contact the editor responsible for this story:
Sarah Wheeler at swheeler@housingwire.com
