In this lesson, you’ll learn how to forecast the likelihood of user churn and how to proactively communicate with those users to keep them engaged before they churn.
No matter the app size, category, or business model, retaining app users is a big problem...but also an opportunity. The Apple App and Google Play Stores each offer over 1.5 million apps vying for the same audiences. These stores are buyer’s markets filled with products that often have high substitutability. Just search for “photo editor” or “subway maps” and brace yourself for literally hundreds of options. The market for app downloads is additionally characterized by cut-rate switching costs -- free-to-download apps and ever-decreasing download times are among the most prominent drivers of shrinking barriers to install.
It goes without saying that user churn is a major concern. And even with a robust, data-driven marketing solution, churn can be a particularly tricky problem to diagnose and treat because it trumps the logic of descriptive analytics: once you’ve observed users to churn, it’s already too late to save them. Wouldn't it be nice to have the ability to forecast which users are likely to churn and preemptively reengage them before they do?
The following steps will walk you through creating and analyzing predictive insights to identify users at risk of churn and their associated behaviors and attributes.
Navigate to Predictions under the 'User Insights' section of your dashboard. Once there, click on the green 'plus' icon to bring up the "Create a New Prediction" module. There are several levers at your disposal to define the behavior you want to predict:
- Churn or Conversion
- If Churn, not performing any Events OR a subset of Events
- If Churn, the number of consecutive days that the Events will not be performed
- The specific Events you want to associate with churn or conversion
The first option you have in defining a new Prediction is deciding if it will be a churn or conversion Prediction. Generally churn Predictions will forecast whether users will not behave in some way while conversion Predictions will forecast whether users will behave in some way. New Predictions are set to be churn Predictions by default. To alter this field, simply click the dropdown and select 'churn' or 'conversion' to specify the Prediction type you want to create.
Next, you'll want to define what Event(s) to associate with the new churn or conversion Prediction you're creating. For churn Predictions, click the dropdown and select whether you want churn to be defined as not performing either:
- Any Events - to forecast the likelihood that users will not perform any Events at all
- The following Events - to forecast the likelihood that users will not perform specific Events that you select
For both churn and conversion Predictions, specific Events can be added or removed to detail the behavior you want to forecast for your users. In addition, AND/OR operators can be uniformly applied across multiple Events in a Prediction.
For conversion Predictions only, you may only forecast whether users will perform specific Events.
For churn Predictions only, you can define the number of consecutive days that the Event(s) will not occur in order for a user to churn. You can do this by changing the number in the 'consecutive days' field. The number of consecutive days for churn Predictions can range from 7 to 90 days.
We recommend naming your Prediction by describing the type of user behavior you intend to forecast. This will help avoid confusion later on, and will allow you and your team to more easily interpret and use new Predictions as more are created. For example, a good name for the churn Prediction in the screenshot in Step 1 above is "Churn - no Events in 30 days."
To save your new Prediction, simply click 'Save & Close'. Once you save your Prediction, you will return to the main Predictions screen in the Dashboard, where a card for your newest Prediction will appear with the disclaimer:
All new Predictions will first be generated at midnight ET of the day it was created. The Prediction will subsequently be updated weekly from the time that it was first created. For example, if you create a new Prediction on Tuesday at 2pm ET, it will first become available on Wednesday at 12am ET, and subsequently be updated weekly every Tuesday at 2pm ET.
We are frequently asked why we update Predictions weekly. The predictive modeling techniques we use are carefully tuned to not overreact to new user data being generated. We make forecasts based on an analysis of longer-term trends. As such, the incremental data required to move users from one predictive segment to another is generated over the course of days.
Once a new Prediction is generated it will appear in the Predictions section of the Dashboard as a Predictions summary card. Each Prediction summary card will include four key pieces of information:
- Name and type (Churn or Conversion) of the Prediction
- Number of users in the high churn or low conversion likelihood segment
- Top Related Behavior
- Top Related User Attribute
Click on the Prediction summary card to explore detailed insights: Prediction 'Type', 'Criteria Definition', 'Next Update', and 'Baseline Churn Rate'.
Likelihoods: total 'Number of Users' and 'Percentage (%) of Active Users' that are predicted to have a High, Medium, and Low likelihood of churn or conversion.
Related Behaviors: Related Behaviors are the usage patterns that serve as lead indicators of future behavior like churn or conversion. Facebook’s former head of growth discussed how important it was for them to discover their keystone metric: Getting any individual user to add 7+ friends in their first 10 days.
- These behaviors describe the ‘aha!’ moments in your app when users discover core product value, and we observe the greatest shift in retention or conversion as a result of that discovery. We comb through all of the Event and attribute data sent from your app to find these inflection points over all time, as well as within certain key time bins like the first 1, 3, 7, 14, and 28 days of your users’ lifetime.
- With these insights, you can better target retention campaigns towards users that exhibit early behavior indicative of churn. For example, if I know that users who view 5+ articles with no video in their first 7 days are 20% less likely to churn, I can now communicate with users who have viewed fewer than 5 articles and encourage them to move past that key gateway to product adoption.
- By default, Related Behaviors are both ranked by z-score and grouped by Event. To view Related Behaviors in attribute-level detail, uncheck “Group by Event.” To cherry pick the most valuable insights for your app, use search, sort and filter functionality:
Related user attributes: User attributes most related to the predictive target behavior. For each attribute, you will see the proportion of active users observed to have each attribute, and the relative difference in churn or conversion between users observed to have that attribute vs. all active users.
- Predictive analytics are a valuable tool for identifying users that are at a high risk of churn, and offering them an incentive, such as access to a new feature, to continue their relationship with your brand and your app.
- Use Predictions to discover the key behaviors that are early indicators of user retention; like Facebook's growth team, which found that users who add 7 friends in their first 10 days are more likely to be retained.