When Real Is Too Real

What training data can you use in your model?

In October of 2022, the journalism outfit ProPublica published a report about software developed by a company called RealPage. The software used an algorithm that users (i.e. landlords) noted was driving significant increases in the rents they could charge in their market(s).

This week, the Department of Justice along with several Attorney Generals sued RealPage.

The AI Hook

Training AI models to make predictions is AI 101. In fact, it is arguably machine learning and not really the kind of AI that is now commonly known via Large Language Models (LLMs). But, semantics.

Machine learning models that predict a future value are quite easy to create and refine. I was for years a teacher in various data analysis boot camps and we regularly taught tech novices during that six month intensive how to build such models and deploy them in applications.

Those models work pretty simply, but they need data. Without data, no predictive machine learning model can be created. Here’s how they work.

For example, to create a machine learning model that predicts the future price of a stock one might add in five years of historical stock price data. Combine that with relevant historical data that could impact that stock. For example, other related stocks, inflation rates, even weather patterns for companies whose products or services are affected by weather. (Orange futures for example). Combining all of that data provides variables that were present for each relevant period when that stock, historically, had a value of X. 80% of that data is then used to train a model. That is, computer code is written creating a model of what features from that data are most predictive of the stock price as reflected in that historical data. Then, once that model is created, it is tested. 20% of that historical data, never before seen by the model, but whose results are certain since it is historical data, is then presented to the model minus the actual stock price information. The model is then asked to predict the stock price based upon the data it is given and using its developed “knowledge” from the training it underwent on that 80% of data where it was given the stock price.

In this way, the prediction that the model creates when fed the testing data can be precisely scored because you, the user, know the stock price value that you have held back from the model during this testing.

If your model is sufficiently accurate about its predictions, let’s say more than 80% accurate, you then set about using the model to predict the stock price - in the future. If others are convinced of the value of your model, they may well make financial decisions based upon the predictions the model provides.

Does Every Company Deserve To Maximize Profit?

When companies offering goods or services decide on pricing, their goal is to maximize profit. Pricing an offering too low leaves money on the table - lost revenue. Pricing it too high leaves sales on the table - insufficient potential customers are willing to pay for the offering. So, finding that sweet spot is an age old task for anyone selling anything.

Whatever job you have now or had in the past, you did the same thing. You tried to determine how much you should negotiate as your pay for the job. The employer, meanwhile, was trying to figure out how low a salary they could offer and still attract and retain the desired quality of worker.

In short, capitalism thrives on the basic process of companies offering the best product or service at the best price. But, how to determine that price? Before the availability of competitor data, companies offering products and services had to use trial and error. Sometimes they left money on the table and other times they priced themselves out of business. But, more data means everyone is closer to the ideal price. That is, the price which maximizes revenue while customers perceive maximizes the value of what they are purchasing.

Into that world comes AI algorithms. They are ideal tools to use data and then predict the price of something. So far nothing sounds all that groundbreaking. Why is RealPage facing a federal lawsuit alleging violations of what is commonly known as price-fixing?

Algorithms Fix Everything (But Not The Way Everyone Expects)

RealPage’s rent-setting software dominated the market. RealPage, according to the report, “acknowledged…feeding clients’ internal rent data into its pricing software, giving landlords an aggregated, anonymous look at what their competitors nearby are charging.” In the pre-AI, machine learning model days, that looks a lot like price-fixing. But, is it?

You are a landlord trying to figure out what you can charge as rent for your units. You can guess. You can use your experience. You can drive by and notice signs on neighborhood buildings and decide based upon that. You can ask your other landlord or property management friends what they are charging. None of that is illegal. Having information is not price-fixing. In fact, that is what capitalism yields in the market - fierce competition and that competition is most often reflected in pricing decisions.

Even the federal government gets in on the act. The government recently passed legislation to cap insulin prices at $35.00. The way it arrived at the $35.00 price was, you guessed it, analyzing data from various states which had already capped insulin prices at slightly different prices, some higher and some lower. The analysis compared how little could be charged to enable insulin makers to still justify producing insulin and how much could be charged enabling typical patients to afford their insulin. Cap the price too low, manufacturers stop making the product. Cap it too high, patients in need still cannot afford it.

To train a model predicting what rent price in a given market will yield the most customers, i.e. the most renters agreeing to pay that rent, the model needs rent price data. There is no other way to train that model. There is no useful model that can be trained on synthetic data or something other than the actual data regarding rents, square footage, neighborhood crime stats, etc. That is what RealPage did.

It gathered data (just like a competitive business-minded landlord did 50 years ago) and used it to train a model enabling it to predict what rent price was ideal in a given market. The “how” of its data gathering, however, has been highlighted as one of the issues fueling the DOJ lawsuit.

Part of the DOJ analysis was the increase in rents seen over the past ten years or so. But, as with anything in a diverse economy, it is impossible to peg an increase or decrease in prices to a single factor.

Landlords turned to RealPage because they became aware that setting prices too low effectively dragged down all local rents. As ProPublica summarized a property management executives impression of their problems, “[i]f you have idiots undervaluing, it costs the whole system.”

Private Versus Public Information

Doing a search for available rental units online yields a mix of buildings advertising prices and those that do not. What the DOJ is alleging as RealPage’s main anti-competitive feature is the use of and training of their algorithm on private information from landlords. The mixing of that data, although anonymized according to RealPage, trained the algorithm to help landlords set prices by essentially sharing pricing data between - competitors.

However, let’s consider a somewhat manual tool, let’s call it RentPriceScrape. That is a tool that regularly scrapes rental prices, square footage, addresses, amenities, etc. from the websites of apartment buildings across the nation. It then aggregates that data, trains a model and offers the model’s predictions as a service to landlords. It is hard to see how this could amount to price-fixing given the data it is using is publicly available.

There are many data aggregation services that rely entirely on publicly available data on the internet that users are willing to pay for in exchange for not having to perform the comprehensive web scraping of that data themselves. This is not price fixing or even collusion given it is merely organizing public data.

How Can RealPage Navigate This in the AI World?

Like so many lawsuits in the era of widespread AI, the core issue is not about the AI. AI is not a product, it is a tool. The core issue is whether creating a model using information the public cannot access constitutes an illegal price-fixing scheme. The solution for RealPage seems simple. Communicate to its customers to publicize their rental information on their apartment websites in a consistent format that RealPage can then scrape, aggregate, train a model (the same predictive model is has now) and continue charging landlords for access to the model’s predictions.

It is hard to see how price-fixing allegations could survive the use by landlords of a tool that aggregates the data that any other member of the public could aggregate if they wanted to take the time and effort to do so.

Now, to protect against competitors, RealPage may convince landlords to publish rules on web scraping their sites, or ban IP addresses other than those that pay for access. That gets closer to collusion, however, if it turns out that only RealPage is willing to pay the price for the right to scrape the data.

The Future

I don’t think the outcome of the lawsuit will be that no company can train a model to predict the optimal rental price. After all, no matter what the model predicts, if too few renters pay the price, the rent is going to decrease no matter what a fancy AI model says is the ideal price. Likewise, if enough renters pay new, higher rental prices, then it seems the model is simply predicting the price that the market will tolerate. If that price is 10% or even 50% higher than the current price, that is a hardship for current renters to be sure, but that hardship does not, in itself, mean the price is inflated or unfair. One of the harsh realities of capitalism is that the price for a good or service is never set. It is always the price that the provider is willing to accept and the price the purchaser is willing to pay. AI cannot change that fundamental.