This piece originally appeared in the July 2022 edition of MReport magazine, online now.
Well-trained associates at home improvement stores will often ask customers how they will be using a tool or product to narrow their recommendations and match the tool to the customer’s needs.
They know the needs of a first-time homeowner are different than those of an avid do-it-yourselfer, and certainly different from those of a professional contractor. Bottom line: the right tool for one customer might not be best suited for another. Now, a similar approach is informing the development of the next generation of Automated Valuation Models (AVMs).
Everybody Knows What an AVM Is … or Do They?
Mortgage professionals understand that AVMs are generally used in purchase and refinance transactions to confirm that an appraised value is within a certain band of confidence for the underwriter. In some low loan-to-value (LTV) refinances, AVMs may even be used as the value indicator for the property.
Additionally, AVMs are widely used by servicers and investors to make collateral-based decisions. The number of use cases for AVMs outside of lending continues to grow as data sources, technology, computing power, and modeling techniques evolve and become more sophisticated.
The real estate and mortgage industries, marketing firms, and consumers all have very different needs, and modelers now can provide an AVM that is best suited for each purpose.
Why Are There So Many Types of AVMs?
More than 25 AVMs have come to market in the past 25 years. Why so many? It’s because the various modelers are using different data and modeling techniques to test if their assumptions can produce more accurate valuations, a higher hit rate (in which the model can deliver a score on a higher percentage of properties), or increased confidence scores (the score within a certain tolerance band, e.g., 95% assurance that the score would be within 5% of the top and bottom model score). In many cases, AVM developers use back testing, the method in which machine-learning techniques are used to “teach” the model, to prove their efficacy against closed loan data, list price, sale price, and property characteristics.
Generally, there are three grades of AVMs commonly available: marketing, consumer facing, and lender grade. Most effective AVMs use multiple sets of data and sub-models, running on a base of MLS listing data and public record assessor data. What distinguishes one model from another is how the data, and how much of the data, is used. For example, if you’re using an AVM for marketing purposes, you may want a model that has the highest hit rate, so that every property shows at least a rough value. To attain a higher hit rate, marketing AVMs may generate valuations using less data and fewer comparables (comps) and include fewer property characteristics. However, a lender-grade model may not produce a value for that same property due to a lack of available information needed to generate a value with the desired confidence score.
This category of AVMs provides values on 100% of properties at an extremely low cost per property, delivering valuations with a high hit rate and an acceptable level of accuracy.
However, these models often sacrifice some degree of accuracy to attain the extremely high hit rate, as they use fewer comps and characteristics than higher-grade AVMs. Lenders often use a marketing model AVM to identify prospects for home equity lending based on an estimate of the current loan-to-value of a property.
Similarly, these models might be used by a variety of home improvement companies targeting potential prospects for direct marketing offers to replace or repair the property’s existing roof, pool, HVAC, or landscaping or to offer other improvements.
Consumer-facing models are used extensively in online real estate and lending portals. They offer a quick reference tool to help consumers understand what a home is worth and what potential financing options might be for a certain property. Online lenders, proptech firms, and iBuyers use these types of “white-labeled” AVMs to engage consumers for prospecting, initial screening, and to instantly triage deals and channel prospects to the appropriate product. These AVMs also provide a high hit rate, but with looser tolerances since no underwriting requirements are built into the model. They also can include owner-stated property conditions, upgrades, and past appraisal results.
Lender-grade AVMs play a critical role in the determination as to whether to underwrite a loan. They include multiple data sources to produce a valuation with a high-confidence score.
These more sophisticated AVMs frequently layer in multiple sub-models, including regression analysis, appraisal emulation, data mining, current and historical market performance in the area, and the latest machine-learning modeling techniques.
The sub-model results are then blended, weighted, and reconciled by a final model to produce the property valuation. AVMs are using new sources of data that are harder to acquire.
These can include geospatial characteristics, true neighborhood boundary information, and the “holy grail” of AVM modeling—the current condition and quality of the property versus similar properties in the comparable geographic area. Many valuation experts believe the availability of property condition and quality data will produce a paradigm shift in the lending and investing markets.
Lastly, the convergence of cloud computing and new data extraction techniques with access to property condition and quality data is enabling AVM performance to reach new heights.
Many observers believe that this new generation of AVMs will be realized in the not-too-distant future.
When this next generation of AVMs becomes the norm for the industry, the continuous feedback loop will drive AVM modelers, researchers, and users to understand how these models are performing in “real life.”