Start-up uses AI to zero in on US property deals

Team including former cyber intelligence agents builds model to target real estate transactions

A US real estate investment firm is counting on artificial intelligence for an edge in scoping out suitable property purchases, but it isn’t planning on surrendering complete control to the machines.

Skyline AI, based in New York, scrapes data from public and less conventional sources to feed its predictive model as it seeks signals in target assets. The firm’s co-founder and chief technology officer, Or Hiltch, likens the process to a medical procedure.

“Our artificial intelligence and data methods are just the same as when we need both a human doctor and technology like X-ray to look through the matter we study. And the precision is all we care about,” Hiltch says.

The recent start-up, founded in 2017, partners with private equity firms and fund houses such as Deutsche Bank’s asset management unit DWS, by offering its artificial intelligence for investing in real estate.

The firm’s investment scope is multi-occupancy residential buildings in the US. The sector historically has high tenure rates for rentals, providing investors with a predictable revenue stream. Skyline’s artificial intelligence model has two main functions: it aims to determine an accurate price for new investments, and it sniffs out potential purchases before the asset has even come to market.

Property pricing is heavily dependent on historical data, but rapid changes in market environment can quickly render transaction data obsolete. An alternative marker for pricing is the income that the property is expected to generate, known as the capitalisation rate. Hiltch says the firm uses an algorithm to combine these two inputs.

“We try to predict the discount or premium in capitalisation rate terms that the buyer and seller would agree upon, given the property’s economic attributes. The cap rate is then compared to recent similar deals to translate it into a forecast selling price. The value computed with the algorithm will probably be very different from calculating with the most recent historical cap rate,” says Hiltch, who works out of Tel Aviv in Israel.

Even though the firm’s statistical model focuses only on the US multi-occupancy lettings sector, the vagaries of the market mean the model requires constant modification depending on a property’s physical location or other fundamental factors. It’s an example of where the human touch is required to guide the machine.

Data sources also need attention. For instance, Fannie Mae and Freddie Mac provide public data which the firm funnels into its model. However, the data does not always come in a usable format. Freddie Mac data is provided in a PDF document, requiring computer technology to extract and convert it into a table to be loaded into the database.

“Setting up a systematic strategy is not just about building the algorithm from scratch. This step follows some very laborious processes like extracting and cleaning the data,” Hiltch says.

Other sources of structured data include historical transaction data and rental figures. Away from structured sources, Skyline AI taps into alternative data for crucial nuggets of information, ranging from FBI crime figures to local weather data, to mobile device usage. The firm even tracks the number of Whole Foods organic stores in a neighbourhood as a likely predictor of affluence.

Or Hiltch, Skyline AI

Hiltch explains that information gleaned from mobile devices can help the algorithm interpret a high-level picture of a neighbourhood’s demographic, without having to rely on public data which may be out of date or insufficiently accurate.

Hiltch says: “If a device is parked somewhere overnight for more than two weeks, our model would interpret that the owner of that device lives there. The device location also tells the identity of the device owner. For example, we can assume that the device owner is a student if the device commutes every day to the university.

“This method allows us to calculate the percentage of students in the area and to improve the accuracy in estimating the need for leasing property.” The dataset enables the model to stay roughly two years ahead of the US government’s official demographic figures, Hiltch claims.

The model’s second purpose is to help the firm look for potential deals that are expected to come on to the market in the foreseeable future. The algorithm detects leading indicators that a vendor is positioning the property in advance of sale.

Hiltch explains: “It happens when the potential seller inflates the occupancy by providing a concession in a combination of an interest-only period for assets coming to the end of their mortgage. This is a signal that the property is going to sell even though the asset has not gone to the market yet. We will be able to use that to source market opportunities.”

Data crunchers

Hiltch’s background in multi-occupancy real estate investment helped convince him that AI could find a useful application in this sector. “We figured out the numbers provided by the real estate operator can be random,” he says. “But it was impossible to verify the numbers because there were so many variables that could impact things like rent growth.”

Skyline AI received initial backing from venture capitalists including Sequoia Capital, TLV Partners and Nyca Partners. The firm won’t put a number on its assets under management because its funds are for private investment and wealth clients.

Hiltch leads a research and development team of 15 staff. The team carries a heavy engineering and data science background, with some members previously serving the cyber intelligence security unit for the Israel Defense Forces.

The skills relevant to such a military background can be transferred to investment analysis, Hiltch believes, in particular the ability to assimilate, verify and interpret data. Some technical elements of each role are similar too, such as satellite image analysis, mobile data analysis and natural language processing.

Others on the team have a more traditional physics and computer science background to PhD level, enabling the team to “go beyond the standard open-source deep learning library and integrate some proprietary elements into the models”, Hiltch says.

The firm’s external investment partners have differing objectives for their funds, with some targeting high rates of return and others seeking stable, more modest yields. Typically, the partners list out the investment scope as well as the benchmark used for performance comparison.

None of the funds have yet completed the full investment lifecycle, so Skyline AI is unable to provide internal rate of return figures. Back-testing results show that the use of Skyline AI’s machine learning algorithms can lift the internal rate of return for one sample vehicle to 21.87% from 15.60%, according to the firm.

“We have already made cash distributions to investors from some of the vehicles, which exceeded the targeted cash-on-cash returns,” Hiltch says.

In the near term, Skyline AI does not intend to widen its investment scope beyond the familiar territory of the US. “The data quality elsewhere is hardly on par with the country and the market is broad enough for us to focus on for now,” Hiltch says.

Despite a heavy involvement of machine learning and big data techniques in investment analysis, the firm has no plans to replace the work of human analysts.

“We do not allow the technology to override human decisions,” Hiltch says. “Eventually, the decision is made on the investment committee level. So, you could think about the AI as an additional member of the committee. After all, we still need humans to bring in creativity to design the strategy from scratch.”

Editing by Alex Krohn