Likely-to-Sell Recommendations for Real Estate

Foster Provost, Panos Ipeirotis, Eda Kaplan, and Nate Rentmeester

[Thanks to Compass’s NYC_AI and CRM teams!]

Hi! AI@Compass here.

In June, we completed the full launch of our machine-learning-driven “Likely to Sell” recommendations in the Compass Customer Relationship Management (CRM) system. This post will describe these recommendations: what they are, what they’re not, what we are trying to achieve, and why they are significant. We will provide a high-level introduction to the AI that creates the recommendations; we will dig into that more deeply in subsequent posts.

Why Likely-to-Sell recommendations in CRM?

Why are CRM recommendations relevant to the business? Experienced real estate agents often tell us that they generate the majority of their new business via their social and professional relationships. However, consider this seeming paradox: The National Association of Realtors reported that agents get only about 30% of their business from repeat customers and referrals. Also, 90% of homebuyers claim that they would use their agent again or recommend their agent to others. However, it turns out that only a small minority do! “One of the major reasons homeowners do not rehire their real estate agent is that they simply cannot remember their agent’s name.” What’s more, 75% of recent sellers contacted only one agent before finding the agent they worked with to sell their home.

The upshot is that being top-of-mind is vital. Maintained and nurtured agent-consumer relationships are more likely to generate new business. However, agents have hundreds or thousands of contacts in the CRM. Which should they reach out to today?

As described by a Compass Agent: “The hotter ones are buried in my past client lists. I have like 900 names in there. That’s pretty hard to go through. You have to be very disciplined.”

The goal of our Likely-to-Sell recommendations is to help Compass agents to engage systematically with just the right contacts — to provide both guidance and discipline. The recommendations focus attention on a small number of connections, those who are most likely to generate business soon. In particular, the Likely-to-Sell recommendations indicate those contacts who have the highest likelihood of selling their homes in the next 12 months.

As explained by Lance Pendleton, Head of Agent Development at Compass: “How many people a day do you reach out to directly in the hopes that they are thinking of selling their home? Chances are you don’t. Why? Because you aren’t sure who is thinking of selling, and it may seem desperate or random. Well, now you have a tool that takes that guesswork away. A tool that will give you direction to where and who you should be focusing on to increase the likelihood that you are connecting with the right people. Why work harder when you can work smarter!”

Of all the different AI-powered products and features that we could have built, why Likely-to-Sell CRM recommendations? In a future blog post, we will dig deeper into the general question of how we choose and prioritize AI products. For this specific project, here are some of the characteristics that pushed Likely-to-Sell contact recommendations to the top of the priority list:

1. Likely-to-sell recommendations directly address Compass’s strategic goal of helping our agents grow their business. Ideally, following our suggestion, the agent will reach out to the likely-to-sell contact; when the contact becomes ready to sell, they will not forget the relationship with our agent. And if the contact happens to be on the verge of wanting to sell right now, the agent may capture that sale immediately. (Read to the end for some success stories!)
2. Providing automatic recommendations via the Compass tech platform was straightforward because the recommendations could be integrated seamlessly with the Compass CRM. The tech platform integrates the data on homes and homeowners, the AI algorithms, the agents’ contacts, and the user interface to make outreach smooth and efficient. Agents receive these recommendations as part of their standard workflow.
3. We already had access to the data we needed to apply machine learning to “train” the AI models for estimating likelihood to sell, and (equally importantly) the data to evaluate the models.
4. The data science work done for this product has significant “option value” — it can be leveraged for future products: seller likelihood models can improve prospecting, farming, targeted marketing, and other tasks.
5. Individuals most likely to sell also may be among the most likely to refer other likely-to-sell customers (due to homophily — the social principle that birds of a feather do indeed flock together).
6. The Likely-to-Sell CRM recommendations can kick-start a program introducing a variety of business-enhancing recommendations for regular communication between agents and their contacts. How about identifying contacts that currently rent and are likely to buy? Those who visit the Compass website and seem to be interested in moving? Individuals in an agent’s sphere of influence who provide the best referrals?

So, what exactly are these “AI-driven” Likely-to-Sell recommendations?

We make a Likely-to-Sell recommendation when one of an agent’s contacts is estimated (by our AI system) to be among the most likely individuals in the region to sell their home.

Ok, first: what exactly are we talking about when we refer to the “AI” system. There are several parts to the AI:

1. the AI algorithm that estimates the likelihood-to-sell of any home, also known as “the model,”
2. the decision-making system that uses the likelihood-to-sell estimations to choose contacts for recommendations, and
3. the (separate) AI algorithm that creates the model from data — also known as “the machine learning system.”

In this blog, we will discuss #1 and #2: the likelihood estimation and how the system uses it to generate the recommendations. In our next blog post, we will discuss the machine learning of the model (and the model itself), for those who are interested in digging deeper into the AI.

Figure 2: The Likely-to-Sell (LTS) AI model takes an instance of a Home & Homeowner as input and produces as output an estimation of that home’s likelihood to sell in the next 12 months. The recommendation procedure takes as input the estimate and decides whether or not to generate an outreach recommendation for an agent.

The function of the likely-to-sell algorithm or “model” is shown in Figure 2, which helps to clarify one of the key flows in the recommendation process. The model embodies a statistical estimation procedure, which takes as input an instance of a home (and homeowner) and produces as output an estimate of the home’s likelihood of selling within the next 12 months. When we say “an instance of a home,” that means “a carefully selected set of data about the home and homeowner.”

What is that carefully selected set of data? Here is a partial list of factors that the likely seller modeling currently considers (we add new features continually):

• Details about the property (bedrooms, bathrooms, square footage, etc.)
• Time since the last sale, and frequency of past transactions for the property
• Home value appreciation; home value compared to others in the neighborhood
• Mortgage status and estimated equity held in the home
• People movement data (percent of owners, renters, how often they move)

The modeling can combine these factors in complex ways. For example, a particular neighborhood may have a high overall moving rate. Small homes, owned for four years, may have unusually elevated likelihoods of sale — for example, because this is a “starter neighborhood” for young families who later move to a larger home. We use this example to illustrate a salient point for any application using machine-learned models: the AI does not “think” like a person. A person might say, “well, I’d have to know more about this particular family…maybe they don’t have kids at all!” However, the AI deals in probabilities estimated from extensive collections of diverse evidence. Even though it does not “know” that this family has kids, it might still be able to predict accurately that this home is one of the most likely to sell. Importantly, the artificial intelligence is complementary to expert human intelligence. Thus the AI predictions plus the knowledge and expertise of an experienced agent create a powerful combination.

One agent’s reaction for a high likely-to-sell recommendation was that it was “very interesting because they are empty-nesters, recent empty-nesters. And this house is worth probably five times as much. This house probably worth two and a half million, and they are recently empty nesters. So this is an excellent suggestion.” (See below for more agent reactions.)

The output estimate of the likelihood-to-sell model is just what it sounds like — for example, the probability of some particular home selling within the next year might be 13%. For some other homes, the score might be just 2%.

The key for helping agents to increase their business to be able to focus on the most likely to sell. As one agent told us, “If somebody tells me, this person is going to sell 12 months from now… there is a 10% chance or a 5% chance, that’s the first person I want to reach out to.” In this agent’s region, the majority of homes have less than 2% likelihood of selling each year.

It is also important to realize that not every contact recommended today will want to sell right now. Some of the contacts may decide to sell eight months from now. And some may not be ready to sell in the next year at all — we are talking likelihoods here, not certainties. Nonetheless, as we saw with the agent above, a CRM contact with a selling probability of 10% may be one of the most likely to sell homeowners in the area: out of such connections, 9 out of 10 will not list their houses in the next 12 months. The seemingly low precision of the recommendations doesn’t mean reaching out to them is not valuable. Many of these properties are more likely to be ready in, say, 18 months (rather than within 12) than others. They also are expected to know other people willing to sell (see the comment above about birds of a feather).

How are likelihood predictions turned into recommendations?

What is happening to create the recommendations?

The recommendation system analyzes each contact in each agent’s CRM database, and checks to see that an address is present (so it can identify the home). It then applies the AI model to each of these homes/homeowners, resulting in a likelihood of selling score for each such contact.

What is a high-enough likelihood of selling to warrant a recommendation?

This decision is made based on several factors:

• Is the home one of the most-likely-to-sell in the local area?

Our recommendations for the highest likelihood-to-sell homes are those in the top 10% (the top “decile”); we will focus on those for the rest of this post. We also recommend “medium” likelihood to sell homes — those in the top quartile (top 25% most likely to sell).

• What is the “base rate” of selling in the area?

The base rate is a statistical concept that’s pretty easy to understand in context: how many homes sell every year in the focal area? If “9% of homes sell every year”, then a 10% likelihood is less impressive than if “2% of the homes sell every year”. In the latter case, 10% is a fantastic “lift”: We went from 1 out of 50, when selecting a home at random, to 1 out of 10, for a highly likely-to-sell contact.

• Are the top-decile homes much more likely to sell than the average home?

We insist that the recommended homes, in the aggregate, sell at least twice the base rate — in data science, that’s called “a lift of at least 2x”. Note that this is a bare minimum criterion: in almost all of our regions, the top decile lift from even our very first deployed AI model was significantly larger than 2x.

But come on, what sorts of likelihoods-to-sell are we looking at here?

These can be very different in different geographic areas, but Figure 3 shows an example from one of our early AI models in one geographic region. On the horizontal axis, we see the probability of selling across all homes in the area. The graph shows the distribution of probabilities for the homes.

You can see that most homes (the fat part of the distribution) have relatively low probabilities of selling within the next 12 months. However, the distribution has a long “tail” extending to the right. These are the homes with high scores. The green vertical dotted line shows the base rate: in this population of households, about 5% of homes sell each year. The red dotted line shows the 90th percentile of likelihoods to sell, which falls at just below 10%. The distribution to the right of the red line is zoomed in on in the bottom panel. This represents the homes that have the highest probabilities to sell in this particular market. You’ll see that the probabilities of sale span a wide range, with the bulk being between 10% and 20%, but with probabilities all the way up to 80%. (The data scientists should investigate that unusual bump at the right end.)

Figure 3. The top graph shows the distribution of (estimated) probabilities of sale for homes in a particular geographic market. It shows that most homes have a low likelihood to sell, but the distribution has a long “tail” to the right, where homes are much more likely to sell. The green vertical line shows the “base rate” of sale for this market — the overall sales rate. The red line shows the 90th percentile (“top decile”) boundary; homes to the right would be candidates for high likelihood recommendations. The bottom graph zooms in on the top decile, showing that the probabilities of sale span a wide range.

Note that the AI model continually improves. The recommendations will be better next week, next month, next year. At this point, don’t think of a fancy AI system learning as it works … think of a team with a ton of data science expertise continually experimenting with additional data and improved algorithms and models, some of which are significantly better. (How would you know that? Stay tuned for a future post!)

Do these recommendations work?

Understanding the impact of likely-to-sell recommendations is a long game. Any individual recommended homeowner may not be ready to sell immediately, but starting work on strengthening the relationship with them now may pay off down the line. Furthermore, our agent may receive other sorts of value from the outreach — such as new referrals.

Nevertheless, we would hope that we would start to see evidence of impact soon after introducing these recommendations. And we have!

Roy Towse — Team Towse (Seattle)

Roy logged in and saw his very first recommendation. Because they were friends, Roy seldom brought up real estate business when they spoke. After seeing the recommendation, Roy gave his friend a call, to advise him that his property had increased in value significantly since they closed their deal several years back. Much to Roy’s surprise, his friend actually was in the market to sell — and now Roy is listing his home.

“The client was a close friend who purchased a condo from me about 8 years ago. He’s excelled in his career, recently got married, and is having his first child. We’ve communicated off and on for years, but I never thought to ask him if he’s given thought to selling. I got the notification and it made me do a little research about what his home was worth, and then immediately called him. Turns out it was perfect timing and they wanted to buy a bigger home for their growing family. I would have never thought to ask if I didn’t get the recommendation!”

Other agents also have already had positive interactions with recommended contacts. For example, here are a few examples:

• “I was in the CRM to look up a few clients to contact them when I noticed the likely to sell feature. I contacted a client that I was pretty sure didn’t need to sell her home. Just checking in with her since it had been awhile. Turns out, she does need to sell it! We will be listing her home. It’s a starter home and will be in demand when we come to market. We have several buyers looking for a home like this one. This is a great feature to help focus our efforts. Love it!” — Kristine Halverson
• “I LOVE this feature. Just this past week, I reached out to one of the people that I was notified may be ready to sell. I reached out to her and will likely be listing her house in the coming weeks. I see no reason why this won’t help me win more business in the coming years!” — Todd Brunsvold
• “I was a bit skeptical about this feature at first. The system shares some names with me and some make sense, while others, well, I thought — how can this be — I just sold them a house 3 years ago —so I did not really act upon it. And a couple of days later, low and behold this client calls asking to see a house. I was SHOCKED!!! We are now actively looking to find them a larger home and then sell their current home. I have no clue why the Compass System thought they were likely to sell, but now I am a believer!” — Sally Marcelli

These real cases illustrate the power of these recommendations, but let’s make sure we’re being realistic. Most recommendations won’t lead to an immediate win — they are part of a strategy for systematically nurturing relationships to increase business in the long run.

What’s more, the success and continuous improvements of the recommendations depend critically on the agents embracing them. So, if you’re an agent and you have suggestions for how to make the recommendations work better for you, please let us know. We are always trying to improve them. You can reach out to your Agent Experience Manager or just track us down.

Stay tuned for more!

We have a series of blog posts — some out already, some in the works — that will dig into our Likely-to-Sell recommendations in more detail and depth, for those of you who are interested in more of the AI/data science details. For example, the next post in the series presents the machine learning system that creates the AI algorithms (models) that estimate seller likelihood. We also will explain the comprehensive evaluation platform that allows us to improve the predictive models continuously and have confidence that the new-and-improved models will improve the recommendations. (For you AI aficionados: we will get into stuff that you won’t get in your typical machine learning class! For example, the machine learning post shows why different homes were estimated as being likely to sell.)

Finally, we will continue to tell the broader story of AI@Compass: product, science, engineering, and strategy. Stay tuned!