Several regulatory agencies on June 1st issued a Notice of Proposed Rulemaking (NPRM) and request for public comment regarding Automated Valuation Models (AVMs) used in residential lending — by mortgage originators and secondary market issuers. The NPRM is specific to real estate transactions for a consumer’s principal dwelling.
Before jumping into the NPRM, we’ll remind readers about how AVMs are a bit different from an appraisal.
AVMs tend to be widely-available online or from a vendor, some for free, some for payment, but they all utilize some sort of mathematical model to home valuation data in a large database in order to arrive at a valuation for a specific property. Over the years, AVMs have become quite sophisticated, often times utilizing multiple statistical models, and most recently machine learning technology. For example, the model might be trained on which sales comparable to use across which neighborhoods. (Hold that thought, as it’s key to the NPRM). Whereas an appraisal is conducted by a licensed residential real estate appraiser, and typically involves an on-site visit to the property to see the inside and outside of a property. The appraiser will then select sales comparable and ‘adjust’ property valuation for a number of elements using a well-trained and experienced opinion. It goes without saying that AVMs are usually less expensive, and depending on the residential mortgage product and program, an AVM might be used by the lender as a substitute for an appraisal.
The NRPM shows us that regulators want to require financial institutions that utilize AVMs to have an AVM risk management program in place comprised of policies, procedures, and control systems (such as testing) centered around the integrity of the AVMs. Given the focus over the past few years on Appraisal Bias, one can assume that avoiding discrimination against borrowers is at the forefront of this NPRM.
What does this mean to you? Risk and Compliance Officers will have to completely understand the models being used by the AVM vendors. What are the datasets? How is the model weighting one factor over another? How is the model selecting the comparables and from which neighborhoods? How is it doing adjustment calculations? Asking your vendor for a copy of their model validation testing is a good start, but you’ll need to periodically test and validate the model as well. This testing will likely involve reperformance of the valuation to determine if there are any gaps in the value per the AVM and per traditional appraisal methods and human opinion. This is not to imply that the latter is correct. You’re just identifying gaps for research.
The comment period is open on the NPRM; might be a good time to provide your feedback to the regulators on this important topic.