The Curious Case of the Trade Secrets that Weren’t

By Curt Levey
September 17, 2019

“Amrock built MyAVM using a publicly available machine learning algorithm that is not anyone’s trade secret. Unlike a quarter of a century ago, when I built AVMs, knowledge of how to build them is widespread and publicly available today. ”

https://depositphotos.com/61369901/stock-photo-perspective-word-on-a-puzzle.htmlNote: the author has no relationship to either party in this case; the following analysis represents his opinion based on publicly available facts.

The normally staid world of intellectual property law was buzzing last year about one of the biggest trade secret cases and largest punitive damages awards in American history. The case involves automated valuation models (AVMs), which are computer models typically generated by machine learning—a form of artificial intelligence—and used to estimate property values by analyzing the property’s attributes, comparable properties, and the like.

Jaws dropped last March when a Texas jury awarded HouseCanary, a Silicon Valley company specializing in residential real estate data and analytics, more than $700 million in compensatory and punitive damages after accepting its claims that it possessed AVM-related trade secrets that were allegedly misappropriated by Amrock (formerly Title Source), one of the nation’s largest appraisal and title service companies.

The jury’s verdict might lead you to believe that Amrock is guilty of one of the most blatant and outrageous intellectual property thefts in history. But when you look closer, it is the jury’s verdict that is outrageous and nearly impossible to justify. I say that not only as a lawyer, but also as someone who built AVMs much like those at issue here before attending law school.

Of Whistleblowers and Reality Checks

Problems with the verdict became clear just a day after it was handed down, when Amrock’s CEO received an email from a whistleblower, a former HouseCanary executive, whose conscience was bothering him. He confessed that “your recent loss was based on fallacies and spin. … HouseCanary never had any proprietary anything. It’s all a lie.” Three more whistleblowers, also former HouseCanary executives, eventually came forward. The four testified under oath that HouseCanary did not have “any IP to steal,” let alone trade secrets worth hundreds of millions of dollars. Amrock is appealing the verdict so that this and other information can be properly considered.

Before the relationship between the two companies soured, Amrock agreed to license a mobile app for human appraisers that HouseCanary was developing. HouseCanary claims that during this time, it gave Amrock information containing trade secrets, which it misappropriated. But it’s telling that HouseCanary made no effort to protect that information from public disclosure during the trial, which is not how one treats valuable secrets. Only later did HouseCanary try to seal the relevant documents in response to this very argument.

Even if HouseCanary had intellectual property worth stealing, the jury’s verdict makes no sense for several reasons. For one thing, it’s hard to see how Amrock could have misappropriated anything. After all, it was undisputed at trial that HouseCanary never provided Amrock with its source code or algorithms for any of the purported trade secrets.

Moreover, even if we were to take all of HouseCanary’s allegations at face value, the size of the judgment bears no relationship to economic reality. The judgment was based on the jury’s wholesale acceptance of HouseCanary’s assertion that the trade secrets in question were worth $200 million. That’s quite a claim given that HouseCanary was a struggling start-up company, founded in 2013, with little history of success in the AVM market.

The $200 million figure is more than 100 times HouseCanary’s total revenues—not just profits—for all product sales during the period in question. HouseCanary might have been dreaming of a future in which its struggles morphed into $200 million worth of success, but speculative damage awards are not permitted under Texas law.

What makes this verdict most shocking to me is not my knowledge of the law but my prior experience, 25 years ago, building AVMs much like those created by HouseCanary and Amrock. At the time, I was working as a scientist at an artificial intelligence startup company in San Diego. There I built machine learning models—including AVMs for homes in California—and invented and patented pioneering technology for providing explanations and confidence measures for the decisions made by such models. You can understand my surprise when I heard that HouseCanary’s lawyers had convinced a jury that AVM technology virtually indistinguishable from the models I built 25 years ago was invented by a company founded just six years ago.

To better understand both the technology at issue and my skepticism, let’s take a look at HouseCanary’s specific trade secret allegations and why they defy credulity.

[[Advertisement]]

“MyAVM”

First and foremost, HouseCanary claims that Amrock’s in-house AVM, named “MyAVM”, was based on technology misappropriated from HouseCanary and was “virtually identical” to its AVM. Right away, this claim is called into question by HouseCanary’s own expert witness, Dr. Vernon Rhyne, who testified after examining the source code for MyAVM that there was no evidence of “any fingerprints, any clues, any reference to any HouseCanary technology.”

The truth is that Amrock built MyAVM using a publicly available machine learning algorithm that is not anyone’s trade secret. Unlike a quarter of a century ago, when I built AVMs, knowledge of how to build them is widespread and publicly available today. The result is that it’s difficult to make much money selling AVM models and free ones are available, including online. The idea that AVM technology, whether stolen or not, is worth many millions of dollars is laughable.

Given the lack of “any fingerprints” and the fact that HouseCanary never gave its algorithms to Amrock, HouseCanary was forced to claim that Amrock must have reverse engineered HouseCanary’s AVM from the Value Reports – printout examples of the AVM’s operation – that it shared with Amrock. This claim is so vague and speculative that HouseCanary’s attorneys could only describe it to the jury as feeding the Value Reports data into a “magic machine.”

Building an AVM using machine learning requires a “training set” of sample properties for which the actual value of the property (the “target”) is known from a sale or human appraiser. What HouseCanary’s reverse-engineering allegation amounts to is a claim that Amrock built MyAVM using a training set in which the targets were the outputs of HouseCanary’s AVM – that is, machine-generated estimates of various properties’ values – rather than the actual values of the properties.

Not only is that unnecessary—Amrock had its own database with actual values for millions of properties—but it would also be scientifically foolish. Optimal AVM performance requires that the models be trained with accurate targets, i.e., actual property values, rather than machine-generated estimates.

As an example, consider the ability of some cell phones to recognize the user’s fingerprint. That ability is achieved through machine learning in which the training set targets are images taken of the user’s actual fingerprint. Using targets which are, instead, estimates of what the user’s fingerprint looks like would surely degrade performance.

Optimal AVM performance also requires maximizing the number of training samples. It would make utterly no sense for Amrock to train MyAVM on the thousands of Value Reports HouseCanary provided instead of what it actually did – train MyAVM on the millions of properties in its own database.

Show Me the Data

Like Amrock, HouseCanary had a database consisting of millions of sample properties, including their attributes (location, square footage, and the like) and comparables (similar pieces of property with known values). HouseCanary accuses Amrock of “improperly using and taking [this] data.” However, despite devoting more than 30 pages to detailing their claims, HouseCanary never gets any more specific about what data it’s talking about beyond the Value Reports it gave to Amrock.

What is clear is that Amrock had plenty of its own data and thus no need to acquire HouseCanary’s data. That’s why, when HouseCanary offered to license its database to Amrock, Amrock declined.

What’s also clear is that there was nothing proprietary about HouseCanary’s database. Like most such databases, it consisted of data acquired from public sources such as county or city tax rolls and semi-public sources such as multiple listing services. That severely undercuts HouseCanary’s claim that its data was a trade secret because, under the law, information cannot be a trade secret if it is generally known or readily ascertainable through proper means.

Nonetheless, HouseCanary tried to spin their database into a trade secret by claiming the selection of which attributes or variables to include in the data was proprietary. While there can surely be some differences in the variables included in different companies’ AVM databases, HouseCanary’s database was just standard fare, cobbled together from public and semi-public sources and containing the typical variables well known to be potential determinants of property values.

No doubt, HouseCanary eliminated or combined some of the standard variables when building its AVM and those choices, if novel enough, could constitute a trade secret. But HouseCanary never shared those choices with Amrock, and Amrock had no way of divining them.

HouseCanary claims that, in addition to stealing its data, Amrock also misappropriated its “data dictionary,” a descriptive index of the variables or categories in the HouseCanary database. But an index is an obvious and inseparable part of any database. Since there were no trade secrets in HouseCanary’s database, there cannot be any in the accompanying index.

Similarity and Complexity Scores

HouseCanary also alleges that Amrock stole its “similarity score,” a concept HouseCanary describes as a measure of the “similarity of comparable properties relative to the subject property.” To be sure, a similarity score for comparables was part of the Value Reports HouseCanary provided. But Amrock never produced its own similarity score and, thus, nothing was misappropriated.

Even if we put that aside, there is nothing proprietary about the idea of measuring the similarity of properties. Real estate appraisal, whether done by a human or machine, has always relied heavily on the value of comparables—that is, similar properties. That, in turn, requires a way of measuring similarity.

In theory, one could invent a novel way of measuring similarity, which might make the specific method proprietary. But that’s not what HouseCanary did. To the contrary, HouseCanary says its similarity score was computed based on geographic information and important property attributes. There’s nothing novel about that. As Amrock has pointed out, “Any real estate agent, mortgage lender, appraiser, or indeed average American knows that property characteristics and location drive property value.”

There’s no need to steal the obvious. But in any case, Amrock could not have stolen HouseCanary’s similarity score for the simple reason that Amrock was given only background information about the score. It was never given the details. That background information itself cannot legally be a trade secret because it was lifted from the description of HouseCanary’s similarity score on its website. Information cannot be a trade secret unless the holder of the information uses reasonable measures to protect its secrecy.

Even more difficult to believe is HouseCanary’s claim that Amrock stole its “complexity score,” which HouseCanary intended to be a measure of how reliably a property’s value can be appraised. But that sounds like an ordinary confidence measure. There is nothing new about computing confidence measures to deal with the uncertainty surrounding estimates of a quantity or value, be it real estate values or anything else.

As with a similarity score, one could imagine turning this standard fare into a trade secret by inventing a novel way of measuring “complexity.” However, that certainly didn’t happen here because HouseCanary admits that it never developed its idea for a complexity score into a detailed model or formula. As a result, Amrock was never given anything more than a generalized description of the concept in the form of a PowerPoint presentation.

Nor did Amrock ever develop or produce a complexity score. That the jury saw fit to award damages for misappropriation of a technology that never existed is baffling.

Indeed, almost everything about the verdict in this case is baffling. None of HouseCanary’s claims about stolen trade secrets hold up to even a modest amount of scrutiny. They all involve technology or information that either falls far short of being a trade secret or is something Amrock had no opportunity or need to steal. In many instances, both are true. Add an astronomical judgment that bears no relationship to economic reality and you have an injustice that cries out for overturning.

HouseCanary Responds

In January 2019, Judge David A. Canales denied Amrock’s request for a new trial, thus upholding the jury verdict. At the time, HouseCanary issued a statement saying that “this was precisely the right decision” and that Amrock had “produced no new information but used the Court’s time to parade out so-called ‘whistleblowers’ who only testified to advance business interests with Amrock and its affiliate Quicken Loans.”

“Amrock cannot distract from the fact that it stole HouseCanary’s proprietary technology,” said Max Tribble, Partner at Susman Godfrey LLP and counsel to HouseCanary. “The jury saw all of the facts and rendered a careful verdict after a seven-week trial, and the attempts by Amrock to nullify the jury’s decision were entirely dismissed by the Court after another four days of hearings.”

 

 

The Author

Curt Levey

Curt Levey is a constitutional law expert and President of the Committee For Justice, a nonprofit legal and policy organization that focuses on the judiciary, law and technology, regulatory reform, and individual liberty. Before attending Harvard Law School, Curt earned an M.S. and B.A. in computer science from Brown University, where he studied artificial intelligence and cognitive science. Subsequently, he worked for five years as a staff scientist at an AI startup company, where he designed and built numerous AI systems, published peer-reviewed articles about machine learning, and invented and patented pioneering neural network technology.

Warning & Disclaimer: The pages, articles and comments on IPWatchdog.com do not constitute legal advice, nor do they create any attorney-client relationship. The articles published express the personal opinion and views of the author and should not be attributed to the author’s employer, clients or the sponsors of IPWatchdog.com. Read more.

Discuss this

There are currently 2 Comments comments. Join the discussion.

  1. Mike Pellegrino September 17, 2019 12:03 pm

    Like you, I was a former professional software developer (and a periodic one in a data warehouse company I own currently) and understand how it can appear that items that originate in the public domain are trivial. However, having valued scores of software programs over the years, including many based on commercial-off-the-shelf technologies for both capital formation and in litigation, it is clear that the mere presence of knowing “what works” can be valuable even in view of free public data. I have valued numerous software applications for companies based on this very fact pattern (my own patent data warehouse being a prime example).

    There is a more nuance to trade secret damages issues than what you present in your article. While I have not had a chance to review the record in this matter, I feel it is important to highlight that damages and value–especially trade secret value–does not manifest exclusively in revenues and profits. Trade secret value can manifest in other forms such as risk reduction, market entry acceleration, the step up in value at various inflection points for a company (e.g., higher value in A round versus seed round), and others. These value calculations can be much greater than nominal value measurements like revenues and profits. It is a primary reason why early-stage, high-risk companies can sell for remarkable sums despite lacking economic fundamentals.

    Whether the value benefits that accrued to the prevailing party are worth $200 million is subject to fact and expert opinion and I am sure those are well-developed in the record. That a $200 million damages opinion survives the customary motions in limine is certainly interesting. That this matter even went to a verdict highlights a likely suboptimal economic decision making process by the defendant. An optimal solution would have likely been to dispose of this matter quickly instead of spending what are likely millions to defend a purportedly worthless trade secret accusation.

  2. Eric September 24, 2019 2:11 pm

    Who benefits from the apparent agenda to destroy patent protections for less-than-billionaire interests, while simultaneously raising those of so-called “trade secrets” to a similar, heretofore-unknown legal and economic status?

Post a Comment

Respectfully add to the discussion.

Name *
Email *
Website