Personalized Likely-to-Sell: Elevating Real Estate Recommendations with “Likely-to-Win”
[Thanks to Compass’s AI team and to Yin Zhang, Panos Ipeirotis, Foster Provost!]
In the dynamic world of real estate, staying ahead of the curve is crucial. Compass, a pioneer in leveraging technology to improve the agent and consumer experience in real estate, introduced an innovative feature in its Customer Relationship Management (CRM) system in June 2020: “Likely-to-Sell” (LTS) recommendations powered by machine learning. The success of LTS, which contributed over $150MM in incremental revenue in 2021, paved the way for further advancements. We’re now taking a closer look at the LTS’s evolution and the introduction of the “Likely-to-Win” model, which offers more tailored recommendations for real estate agents in line with their specific business goals.
A recommendation’s effectiveness is measured by the property’s likelihood to sell and the real estate agent’s ability to list and sell that property successfully. It’s important to select properties based on their selling prospects and consider the agent’s likelihood of winning the listing. For instance, our existing “Area of Agent” model (see details in our blog post: Estimating the Geographic Area of a Real Estate Agent) provides geo-matching between the property and the agent. The more central the contact’s property is located in the area where the agent usually works, the higher the chance of winning the listing.
Another important aspect is the agent’s sphere of influence. The stronger the agent’s connection with the contact, the greater their likelihood of winning the listing.
Hence, in addition to the Likely-to-Sell model that predicts the sale probability of a property, we developed a new model called the “Likely-to-Win’’ model. This model assesses the probability of an agent representing the contact at the property’s sale. This article delves into personalized LTS recommendations: why we need them, what they are, where they can be applied, how they are created, and how to evaluate them. We offer an overview of the motivation behind these recommendations and outline the basic principles of the machine learning approach used in their development.
Why do we need personalized Likely-to-Sell recommendations?
Consider an agent, Bob, presented with two properties that are deemed likely to sell: Property A, within his familiar working area, and Property B, in a distant location he’s unfamiliar with. Even if Property B has a higher selling likelihood, Bob might prefer Property A due to his familiarity with the area. This scenario underlines the necessity of personalized LTS recommendations. Recommending a distant property like Property B could diminish Bob’s trust and lead to resource inefficiency, showing the importance of aligning recommendations with each agent’s specific expertise and operational zones.
Figure 1. The Area of Agent for Bob who works primarily in Queens and Midtown Manhattan, but also does business in lower Manhattan and western Brooklyn
Recommendations more aligned with agents’ business
Our goal with LTS CRM Contact Recommendations is to boost agents’ business. A successful recommendation isn’t just one that leads to a sale, but one where the agent who got the recommendation makes the sale. Even with a perfect LTS model, i.e. the LTS model’s score equals the probability of the property being sold in the next 12 months, recommendations generated from it can still be improved by adding a new dimension of information. While our current LTS model effectively predicts property sales within the next year, it lacks insights on whether an agent is likely to represent the seller. By developing a model that estimates an agent’s likelihood of winning the listing, we can refine our recommendations to more directly support agents’ business goals.
More trust on our product
Since the roll-out of the Likely-to-Sell recommendation system, Compass agents have shown higher success rates with recommended properties compared to similar, non-recommended ones. This suggests that agents trust these recommendations, often investing time to nurture them, which leads to higher win rates. Agents tend to prioritize recommendations that align with their expertise and knowledge of the area, enhancing their success. However, an excess of misaligned recommendations could decrease their trust and usage of the system.
What is Likely-to-Win?
As previously discussed, the goal of our recommendation system is to help Compass agents to increase their business — in particular, to represent more sellers. This goal is quantitated by the agent’s (recommended) transaction win rate and can be computed as:
Agent’s recommended transaction win rate = Recommended property’s probability of sale * Agent’s probability of winning the listing given sale
The first part, the property’s probability of sale, is estimated by the Likely-to-Sell model. To assess the second part, we’ve developed the “Likely-to-Win” (LTW) model, which estimates the agent’s probability of winning the listing given that it sells. With the LTW model, recommendations are ranked based on this combined transaction win rate, prioritizing properties that are not only likely to sell but also have a strong possibility of being listed by the agent. This approach aims to increase the number of successful listings for the agent and Compass, and in turn, foster greater trust in our system.
Where to use Likely-to-Win?
The Likely-to-Win (LTW) model is primarily used in CRM contact recommendations, enhancing the Likely-to-Sell (LTS) score by combining it with the LTW score for better queuing, prioritizing, and ranking of property recommendations. Beyond CRM contact recommendations, the LTW model has several other applications:
Orphan matching to the right agent
Agents’ CRM can only cover a very small portion of the properties eligible for recommendation. Those properties not belonging to any agent’s contacts are called orphan properties. There will be a lot of opportunities from orphan properties and we can help agents to extend their business by matching those orphan properties with the right agents.
Different from the CRM contact recommendation where we rank properties for each agent, in orphan matching, we will compute the expected value of the recommendation for each (agent, orphan property) pair and rank the expected value across all those pairs, followed by an optimization process to assign properties in a way that maximizes overall value and balances agent workloads.
Agent referral
Besides predicting the probability of the agent winning the listing, the LTW score can also serve as a measure of the match between the agent and the property. This is particularly useful when an agent needs to refer a seller to another agent due to geographical or time constraints. The model finds the most suitable agent for referral based on the LTW score.
Seller-side lead routing
Lead routing is to source, verify, nurture and deliver leads to agents. To do that, we need to first identify high quality properties (sellers) for Compass to nurture and then identify good-matching Compass agents to deliver those leads. Our LTW model can apply for identifying both properties and agents.
In order to identify properties with high quality for further verification and nurturing, we will use both the LTS score and the LTW score. We will require the property having a decent LTS score and there are a fair amount of Compass agents in the area having a decent LTW score if paired with the property. By doing that, those identified properties are actually likely to be sold and there are many Compass agents having a high chance to win the listing. Then, we can spend time nurturing those leads. After we have verified qualified leads, the LTW score could be used to route them to the optimal agents, and increase the probability of converting the listing.
How to model Likely-to-Win?
Recommend, or not recommend, that is a causal question.
As previously demonstrated, the existing LTS recommendations have a positive impact on the win rates of Compass agents. We posit that recommendations, along with the agents’ efforts in nurturing them, influence their likelihood of winning listings, while having a limited effect on the probability of properties getting sold. This implies the necessity for a Likely-to-Win (LTW) model that specifically addresses the causal impact of the recommendation process.
Agents can win listings in two ways: by following up on our recommendations or through organic engagement with non-recommended properties. To represent both scenarios, we employ two distinct versions of the LTW model:
Recommendation Model:
- This model evaluates the probability of an agent winning a listing following a recommendation. It specifically addresses the question: “what’s the agent’s chance of winning the property if we send the recommendation and it gets listed?”
Non-Recommendation Model:
- In contrast, this model focuses on situations where agents win listings without the influence of recommendations, answering: “What are the chances of an agent winning a listing organically?”
Using different data for two models
The LTW model is to estimate the probability of an agent winning the listing given the property gets sold. Unlike the LTS model that relies on property data, the LTW model uses data comprising (agent, property) pairs, tracking sales within a specified timeframe. Positive instances in this dataset are where the agent successfully sells the property, while negative instances are where sales are completed by other agents. For the recommendation model, data includes properties recommended to the agent before listing, while the non-recommendation model uses data from properties not recommended to the agent.
Key features in the model
The LTW model considers three main feature categories, with continuous updates and additions:
Agent-Property Match:
- Proximity to the agent’s previous transactions.
- Property’s estimated price range relative to the agent’s typical pricing.
Agent-Homeowner Connection:
- Relationship strength between agent and homeowner.
- Frequency and recency of communication.
- Past client history with the homeowner.
Agent’s Profile:
- Historical engagement with recommendations.
- Win rate on previous recommendations.
- Basic information like performance history, experience, motivation, tenure, and role.
Engagement analysis through online experiment
In March 2023, we introduced recommendations based on the Likely-to-Win (LTW) model. To evaluate its effectiveness, we conducted an A/B testing experiment for a month, dividing agents randomly into control and treatment groups. Agents in the treatment group received property recommendations prioritized according to LTW scores.
The key objective of this experiment was to assess and compare agent engagement levels between the two groups. We primarily focused on the outreach rate as our metric for engagement, defined as the ratio of the number of properties an agent contacted within a day of viewing the recommendations to the total number of recommendations viewed. The findings were significant: agents receiving new recommendations ranked by the LTW model showed a 39% higher outreach rate compared to those in the control group. Furthermore, for existing recommendations re-ranked according to the LTW model, there was also a 16% increase in outreach rate compared to the control group. This outcome highlights the enhanced engagement driven by the personalized recommendations.
What’s next?
Having explored the transformative journey from Likely-to-Sell (LTS) to the groundbreaking Likely-to-Win (LTW) model, Compass is set to expand the reach of the model beyond CRM contact recommendations, applying it to various real estate activities like lead routing and agent referrals. Insights from recent A/B tests will guide improvements, focusing on agent win rates to better align the model with their workflows.
Ongoing enhancements to the LTW model will include adding new features for a deeper understanding of agent, homeowner, and property dynamics. Emphasizing user trust and engagement, Compass plans to integrate agent feedback continuously into the system, ensuring it adapts to the changing needs of the real estate sector.
Disclaimer: Compass is proud to help everyone find their place in the world. Essential to this mission is providing equal housing opportunities to all individuals in accordance with fair housing laws. This means that Compass and its agents cannot and do not discriminate or assist in discrimination against potential renters, owners or buyers.