How Greystone Labs’ Machine Learning Algorithm Expedites Originations
- Nov 29, 2018
Three years ago, Greystone Labs Director Zac Rosenberg set out to find a tech-enabled solution to simplifying the firm’s agency lending process. By offering multifamily owners more comprehensive data faster, the firm is changing the way it serves its clients. A powerful machine learning algorithm is enabling these improvements, Rosenberg explained, with the potential to materially transform the way the firm underwrites loans.
MHN: When and why did Greystone decide to launch an in-house tech/innovation incubator?
Rosenberg: It came from humble beginnings around January 2015. I’ve always been the geek of the family. I have an engineering background and have worked at different tech startups. I came to Greystone and started applying my engineering background as an analyst. I realized things that were taking teams of analysts to complete could be done if you just wrote a quick script instead. It started from small hacks, which became scripts that we used to automate manual tasks. One of the first scripts ended up bringing in an additional $150 million in loan originations.
What are the goals of Greystone Labs?
Rosenberg: When Greystone Labs started, I don’t think anyone knew how far we’d run with this. In the beginning, the goals for Greystone were to find more opportunities to quickly adopt something to make someone’s life easier or to make the process more efficient. I was excited to work on something fun, something I’d be excited about at two o’ clock in the morning. It snowballed from there, and now our goals have become a lot more ambitious. We’re going to change the way you get a loan and access capital for multifamily and hopefully health care, too. I want to change the experience and the process of underwriting and risk assessment.
What is the concept behind the sizing app that you’re releasing?
Rosenberg: It’s a pretty (straightforward) concept. The idea is that if you’re an owner or borrower looking at a property you own or would like to own, you give us an address, and we’re able to look at that subject property as well as all the properties in the vicinity. You can see what income those properties are generating and what expenses they have. That gives us a sense of the cash flows for properties in that area. Then we look at the volatility—analyzing current capital markets by today’s rates, Libor, Treasuries, investor spreads—and try to assess the risk. From there, we come up with loan options that we think would work for that subject property today.
What’s unique about it is that, in the beginning, if you just give us an address, we can give you ballpark figures based on assumptions heavily based on data we have on (properties) in the vicinity. But if you already have some information—say, you already know what rents the property is (commanding)—you can give that to us, and we can be more accurate.
The app has been out in beta. We’ve used it internally and with a few of our clients. We’ve closed somewhere around $50 million in loans than directly came in through our sizing tool. We’ve been using it in some of our broker relationships. Even if one of our salespeople is sitting across from a client, instead of having the client send their financials and have one of the analysts go over it before sending it to the Fannie Mae and Freddie Mac teams to see what your options are, you can do that right from your phone.
So far, the response from users has been positive because it’s not such a revolutionary idea, it’s what should have been there. It allows people to see all the options across the board, instead of just one type of loan product. There’s even an educational aspect to it because you can see, for example, what types of features a Fannie Mae loan offers that CMBS doesn’t.
I envision this going much further. What’s happening in the background is we’re trying to get valuations on properties across the U.S. and see where (prices) are trending. I want to get to a point where we can actually start looking at forward rate locks. I want to have a rate that we’re so confident it that we’re willing to lock it in right there.
The platform is powered by machine learning. How does that work?
We used a clustering algorithm for the sizing app’s ultra-tailored marketing campaign.
What we wanted to be able to do, instead of marketing to people in a tone-deaf way, is to (provide useful information) to the different kinds of multifamily owners with various angles and incentives. What we wanted to do was to find a way to match information to the particular type of borrower and given them something that fits their portfolio strategy, activity level or the type of properties that interest them.
We’re starting in Seattle, where we combed over every multifamily transaction that has occurred there over the past five years, who owns what, who’s good at taking a B-class property and flipping into Class A, et cetera. We put that information about people into a clustering algorithm that (groups) owners into similar categories. Then we created a marketing campaign based on those clusters, such as owners that buy and flip properties and others like mom-and-pop shops that might own a single property for 15 years.
When we’re doing the underwriting, what’s very important is looking at the performance of properties in the vicinity to compare to the subject property. What’s key is picking a good set of representative properties. There might be a ton of properties in the area, but they don’t all correlate with the subject property. Up until today, people have used models to do this heuristically, asking if they were built around the same time or have the same number of units. We use machine learning to further that (analysis).
Let’s say we look at one block with a few multifamily buildings. Over time, we can look at the impact of rents changing in one property on the others in the vicinity. If you see rents increasing on one (asset) and not on another, then renters (probably) don’t view those as comparable properties. We feed that in the model, which tries to predict which properties are most similar. And because we have the historical data, we’re able to tell if we were right about (our predictions). As we repeat this with thousands of transactions, the algorithm learns how to spot perfect comps.
Tell me about the pilot that launched in Seattle.
We wanted to see if there was a better way to reach out to and resonate with owners in a market. Instead of going in with a best guess of what kind of loan they may want, you might consider what their current loan is. For example, they may have a Fannie Mae loan already, so we might pitch (the same type of financing) to them since they’re familiar with it. What we wanted to do this time was to take a more data-driven approach (to that process).
We wanted to consider what each owner’s (focus areas) were. If you’re good at flipping a property from Class C to Class A, for instance, then we’ll consider which Greystone product could help you do that more effectively. From there, we’ve been starting to send out a digital campaign tailored to each individual owner. For each owner, we created an individual landing page where you would see one of the properties you own, on a map that shows similar properties, influential businesses in the area, how your building compares to the competition on the block and then we also give you real financing options. We don’t give you a range but a specific rate based on where we think your property would size out to today. And even if you don’t do a loan with us, we’re hoping the information we send out is still valuable.
What is the benefit of this tool to the end user?
What I’d like to be able to give to any multifamily owner is a way to view their property from the lender’s standpoint. Lenders look at properties differently than investors do, as far as how we underwrite risk. Often, when we first speak to owners, there is a little bit of shock when they’re making a certain kind of cash flow and we have to use a lesser amount because we can’t include things like amenity fees. The shock comes from the owner learning they can’t get the type of financing they thought they could.
I think the sizing tool that we’re releasing allows you to see how Fannie Mae, Freddie Mac or FHA would look at your property. This enables an owner to plan ahead for when they refinance the loan and assess how (capital improvements) might impact how lenders view the property.
How does machine learning technology enable better solutions for commercial real estate investors?
Machine learning helps us see trends that we may not have caught ourselves. In the end, when you’re trying to predict the cash flows of a property, it’s not just the building’s historical operations that matter. It’s also important to think about the other businesses in the area, cuisines that are nearby, job growth, (upcoming public works projects). There’s so much that factors into the desirability of the building. As humans, we can only analyze maybe 10 things at once, whereas machine learning algorithms can analyze thousands of data points to give us greater ability to predict. We want to be able to walk the entire block, not just look at a single parcel.
How do you see technology disrupting multifamily lending?
I think the lending space will change because the way multifamily properties are built will become very different, due to methods like modular construction. You’re going to have more standardization. For the borrower, the loan process will be faster, which means the lender needs to move faster, too. I don’t see the 60-day closing process continuing. We’re going to have to be very quick in locking in terms and delivering capital.
I don’t think there’s an area of the multifamily industry that won’t be impacted be technology. Take smart homes, for example. We might be able to tap into these technologies and extract data on how a property is operating and refinance a loan as soon as rates drop without any additional paperwork because I’m plugged into your property and I’m plugged into the capital markets.
What do you think the multifamily lending process or landscape will look like in five years?
Five years from now, we’ll be able plug into property management software and in real time evaluate risk level and see how the property compares to its peers, every day. As long as we have direct access to information, I think executing the loan will be the easy part. I’d like to think we’d be able to finance a transaction within a few hours because you’ll have very accurate underwriting up front. From a technology standpoint, we could definitely get there.