3 Tips for a Good Customer Lifecycle Model

The Art and Science of Grouping Customers

Cohort Analysis is a fancy name for analyzing customers by group. Cohort Analysis provides the most useful information about the stage of a customer’s lifecycle. A well-designed cohort will incorporate both timing and behavior. For example, a cohort might group customers with the same first purchase date or a similar time since their most recent purchase.

Since cohorts group customers by past behavior, it’s important to account for any marked change in business conditions when identifying and selecting groups to study. Examples include the impact of a holiday shopping season, a new incentive for the buyer or the launch of a competitive product line. In cases where a marketer doesn’t have much historical data, or where historical data is less relevant, it’s advisable to wait for new data to accumulate thus providing a more accurate picture of the world before defining cohorts to study.

To Find the Answer, We Must Start with the Question

The unspoken motivation behind Customer Lifecycle Modeling is that the marker can predict the behavior of a customer based on what stage they’re in. We use these predictions to anticipate the customer’s actions – enabling marketers to motivate progress through the lifecycle to the right stages, at the right times, and at the right cost. The key distinction between this perspective and the traditional marketing approach is that we aren’t looking for answers in the data. We are using our data to answer questions about reality, and the potential realities that lie ahead.

The key to keep in mind is that math is not magic, and data is not a crystal ball. Customer Lifecycle Models can be incredibly useful, but we must remember that they won’t show us what we don’t know. We do our best to anticipate what the future holds, but we will never know for sure. Since we are using these models to make business decisions, we want to be certain that the numbers are useful. We will discuss a technique to evaluate the usefulness of our models in the next section.

The More You Measure, the Smarter You’ll Be

We all know the saying: what’s measured gets managed. In order to improve performance of a metric, the first step is to get a sense of how we’re doing. Let’s start by seeing how well we can predict customer behavior.

Let’s Try It!

We will illustrate this technique with a simple example where we want to identify the lifecycle stage of 100 unique customers. In our example, a customer has 3 states: active, inactive and churning.

Active: Alice is a regular who has made at least one purchase in the last 12 months
Inactive: Where’s she been? Ina hasn’t bought a single thing in a year
Churning: Charlie’s been MIA for 11 months, but we’re not giving up on him yet

We will only consider active customers for this example. As stated, we use customer lifecycle models because customers behave differently at each stage of their lifecycle. In this case the difference is who is making the purchases. We would expect an active customer to make a purchase in the next 12 months. Let’s see how many customers we predicted correctly.

Rather than waiting a full year to see how we did, we can play a game with dates to get results now. Suppose today is 11/19/2018. We can start our measurement window on 11/19/2017 to get the most up to date picture of our model given our existing data.

How Did Our Model Perform?

Total Unique Customers: 100
Predicted Active Customers: 37
Actual Active Customers: 27

24 Customers: We Predicted Correctly!
We predicted 24 unique customers were active and they did indeed purchase something in the previous 12 months

3 Customers: We Missed Them!
We predicted these 3 individual customers would not purchase in 12 months but they ended up surprising us by making a purchase.

13 Customers: We Predicted Incorrectly.
We thought these 13 unique customers would be active, but they didn’t end up purchasing anything in 12 months.

We correctly identified 24 out of the 27 active customers. We were able to identify 89% of the active customers. Our model seems very good at identifying active customers. But only 65% of the customers we classified as active were actually active. Our model identifies most of the active customers correctly, but it incorrectly classifies some inactive or churning customers as active. Is this incorrectness a problem? We need a low effort baseline to put these measurements in context.

Establishing a Baseline

Since we’re predicting a binary outcome (purchase or not purchase) we can use a coin flip as a simple baseline. The results from our coin flip model would look like this:

Total Unique Customers: 100
Predicted Active Customers: 50
Actual Active Customers: 50

25 Customers: Right!
Predicted Active Who Are Active

25 Customers: Miss!
Predicted Not Active Who Are Active

25 Customers: Wrong.
Predicted Active Who Are Not Active

Our coin model correctly identified half the active customers and half of the customers it said were active are actually active. We are beating this baseline by solid margins in both areas, so we can be satisfied with this performance. Note that even though predicted actual active and actual active are both 50, the coin flip model is not picking the right active customers.

For our next step, we would want to dig into the customers our model is misidentifying as active and see whether they represent a key segment we would want to identify accurately.

There is much more to be said about evaluating the effectiveness of Lifecyle Models, but hopefully this gave you a basic understanding of how they can be used. Stay tuned for more information and examples in a future blog post!

‍