During some interview preparation, I learned about this type of modeling called “propensity modeling”. The premise is pretty cool. For companies trying to understand the “true” value of a customer, propensity modeling helps inform where attention should be focused. This is done in three often-connected ways, finding out who is most likely to become a customer, determining which of my customers hold most risk of churn, and determining which of my customers offer the most long term value.

General Model Overviews

Let us actually break down the different types of propensity models and why they are important.

Propensity to Buy Model

When companies search for potential customers, they generally want to find the targets that are willing to try out their product or service.

For example, Modelo Beer should never invest the calories to try and market my Dad. He does not drink alcohol and probably never will (except for when my sister got into college, still kind of hurt my college admission never got the same attention). They should instead target my Mom, or me, or my roommate who loves beer. But even amongst us three, we have different likelihoods to buy.

The main purpose of a propensity to buy model is to help companies determine which segment of customers is most likely to buy the offered product of service. This helps informs a company where they should be investing their outreach.

Propensity to Churn Model

If the goal is to maintain a healthy and growing company, selling is probably only half the battle. If your current customer base does not stay for long, then you will lose a lot of the revenue you worked so hard to obtain. As a business, it is critical to isolate these hotspots and determine the best course of action.

A propensity to churn model helps inform a company the likelihood in which its current (or future) customers are likely to leave. Why is that important?

Let me throw out another example. Suppose I am a software company that sells a communication platform primarily focused on messaging. One of my teams has built a propensity to churn model that has highlighted two hotspots: a group of folks that are moderately likely to churn, and a group of folks that are highly likely to churn. In addition, my model has shared that folks that are moderately likely to churn seem to love video-based communication, whilst folks that are highly likely to churn find product price to be the greatest concern.

As a business, I might use the model findings to decide that the group that is highly likely to churn requires too much effort to address and instead focus my attention to the group that I have a fair shot of keeping. I might decide to invest in adding video-chatting capabilities to my communication platform.

Customer Lifetime Value Model

Finally, something that is extremely valuable is having an understanding of a customer’s lifetime value (CLV). This is often useful when you run any form of subscription business, but can just as easily be relevant for something like Starbucks, where someone like me who has been converted as a customer drinks religiously.

The idea is to obtain a monetary value of a single customer over the course of his/her relationship with the company. In my Starbucks example, that might mean determining winning me over with that first cup of sweet, sweet caramel frappe will result in over $10k of lifetime spending. Once you adjust that amount using discounted cash flow analysis (money 10 years later is worth less than money today, have you seen the inflation rate?), subtract out the raw costs of my drink orders, and factor in my likelihood of suddenly converting to Peet’s Coffee (yeah right, will never happen), you might determine that my customer lifetime value is $5k.

From a ML standpoint, we might find it extremely useful of modeling out various cohorts of CLVs from my customer pool and factor that in when determining where I want to direct my finite marketing resources.

Putting it All Together: Example Use Case

Let us use an example to help illustrate how this type of modeling can be used. Pretend I was a company that sold a browser based design software. My primary revenue source is business software license subscriptions, but I currently have a large number of individual license users that are relatively young in age.

I decide moving forward I want to make a big investment into expanding my company’s footprint in educational institutions. The business case for it makes logical sense, as I would be introducing my product to people early on in their lives and it will hopefully stick with them as their go-to professional design tool when they transition into the workplace. After coming up with this decision, I want to make a data backed decision on what particular institution type I want to focus my marketing resources on. My choices range from high schools, undergraduate universities, specialized art schools, or post bachelors education.

Here we can leverage propensity modeling. First, we can leverage our pool of free license users and, with information we have on them such as demographics, preferences, and use cases (all of which we have internal data on or obtained through surveys they submitted whilst using our product), we can actually cluster our population via unsupervised learning to create general population archetypes. We might find from this that our users are primarily split into the following three buckets:

  1. Longtime interest users that transitioned from a free to paid plan
  2. Users that primarily signed up for a plan due to professional projects
  3. Users that tried out the product after recommendations from a friend

Using current customer data and past churn data, we can obtain churn rate estimates for our three cohorts. This merged with annual spend data allows us to obtain CLV estimates of our three cohorts. From the results, we decide that Cohort 3 has the highest CLV and is our ideal group to target.

We actually believe that Cohort 3 is mostly related to undergraduate universities, where the large student population and diverse pool of user provide an ideal environment for word-of-mouth product recommendations. However, we can easily pivot our strategy in this case if the results were different. Cohort 1 is most ideal for high school, where early exposure to the product might spark interest and getting many users to early on try out a free plan. Meanwhile, Cohort 2 might be tied to art schools or post-bachelors degree programs, as those tend to be more specialized and professional in nature.

Summing it Up

As shown and illustrated, propensity modeling can be a huge asset in helping plan out strategic business initiatives. As the need for efficiency increases, targeting the right customer population with data backed insights can be critical in ensuring companies make the most out of what they have.