An intellectual property court case that was decided this spring has elevated the topic of Automated Valuation Models, or AVMs, into the public discourse. While normally confined to real estate industry circles, the subject has received increased discussion as media outlets report on the case that led to a $706 million jury verdict.
The case involves a housing data analytics firm called HouseCanary, which claims that Amrock – another real estate valuation company then known as Title Source – misappropriated its trade secrets by allegedly incorporating components of its AVM into Amrock’s own model.
This intellectual property dispute over AVMs and its remarkable payout warrant a discussion on the nature of the technology and the feasibility of classifying one’s own AVM as a “trade secret.”
First, AVMs are the general name for computer-based models that predict what price—or price range—that an asset would likely sell for, given current market conditions and the particular attributes of the asset.
They are commonly deployed by the real estate industry, especially in the residential sector, to provide quick and easy spot valuations of individual dwellings. Lenders, insurers and investors are the most common consumers of these valuations.
However, even individual homeowners and buyers can access them via online services, such as Zillow, by simply typing in their property’s address and instantly obtaining an estimated price. Essentially, the models work by analyzing recent sales of properties that are similar to the subject property, and then arriving at a reasonable prediction of current value.
The two primary inputs that are needed to create an AVM are historical data on sales of comparable assets (e.g. homes), and computer software that is capable of mathematical and statistical modeling; especially regression modeling. Although there are advanced software packages that are capable of very sophisticated models (e.g. SAS, STATA, R), it’s possible to build a viable model in a basic spreadsheet software such as Excel.
The sales data is typically from a public or semi-public source. A public source would either be a particular county’s or city’s tax roll, which contains every property’s most recent sales price in addition to its assessed value that is used for tax purposes.
A typical semi-public source would be the local Multiple Listing Service, which is a database accessible to local realtors and brokers that contains all sales and listings in a given geography. It is also common to obtain this data from third party vendors that specialize in collecting, compiling, cleaning and re-selling this data.
However, a key point here is that both the software and data are widely available to anyone or any entity that wants to obtain it. Moreover, this has been the case for decades, since these models and their use date back to the 1980s.
As such, there is both a very high and very specific standard for any one party who develops these models to make claims of proprietary knowledge or intellectual property theft against any other parties who also develop and use these models.
Under what circumstances could a developer of such models feasibly make such a claim? Since the use of these models are so widespread, and both the data and the documentation of how to build these models are essentially in the public domain, it would be very challenging.
In essence, the aggrieved party would have to demonstrate that the accused party is using a model that is either an exact duplicate of the aggrieved party’s or is extremely close to it – perhaps by incorporating the source code.
Establishing this stringent condition is incredibly challenging. The reason being is that both the data used to estimate these models and the research showing how to develop these models are widely available.
There is a large, longstanding and well-established literature on the factors that determine house prices. These include the physical attributes of the structure, the geographic attributes of the house’s location, and the public services the owner of the property is entitled to and the price they pay for these services.
A typical AVM is simply a regression of a dwelling’s sales price on these factors, plus whatever other factors that the model’s developers think affects house prices in their particular market of interest.
As such, most AVM models are very similar to each other. What differences do exist are fairly minor: developers may combine multiple variables into a single variable (called “interactions”), or nonlinear transformations of the variables may be performed (taking the natural logs or squares of a variable is typical), or location-specific variables that are relevant in one location but not another (central air is a critical amenity in Florida but not Maine).
As such, the internal similarity of this general class of models creates a high burden of proof for an aggrieved party to claim intellectual property theft.
As an analogy, consider a restaurant owner whose chef quits and then opens up a new restaurant. This new restaurant has the same name, design and menu as the original restaurant. In such an instance, it would seem that the original owner would have legitimate claims of intellectual property theft.
Now consider an example at the opposite extreme: the chef who quits opens a new restaurant that is completely different from the original, with a different menu, name and design.
Clearly, it would not seem as if the original owner has a claim. It would be ludicrous for the original owner to claim that he has a monopoly on the concepts of “dining out” or “cooking food,” since that is the only thing the two restaurants have in common.
But, consider an example between these two extremes: the new restaurant has both a design and name that is different from the original, but not that distinctly different.
And there are a number of items on the menu that are new, but some that are similar to items on the menu in the original. Yet, even the menu items that are similar are clearly not identical copies.
For example, both restaurants serve grilled steak and baked potatoes, but the new restaurant uses different cuts of beef and different toppings on its potatoes. Does the original restaurant owner have a case? The standard is still extremely difficult for him or her to meet – especially in a world of countless competitors with overlapping similarities.