Customer Stories
August 9 2023

How to predict audience churn and save on user retention


Crypta specialists developed a model built on the AppMetrica infrastructure that predicts churn in mobile apps. Using the model, a service provider marketplace was able to more accurately calculate the number and nominal value of promo codes they should issue to bring back customers.


Here, we’ll use this example to demonstrate how the Crypta model helps apps increase LTV.

If you want to grow your mobile app, analyze user behavior and acquire top LTV traffic, it’s crucial to rely on accurate data for all your product and marketing decisions.

AppMetrica is an all-in-one mobile app analytics solution that allows you to take your app to the next level at an affordable price.

Get started for free now

Predictive churn model: what it is and how it will increase LTV

The predictive churn model is an ML model which predicts the likelihood that a user will stop using a service. Churn prediction can help apps that send push notifications to users or offer them promo codes.

To retain users in their apps, many companies offer discounts via promo codes

This approach may lead to more orders, but it reduces profits. That’s why some businesses build predictive models to avoid giving out promo codes to users who are already likely to stay on the app.

The service used its own predictive model, but it was suggested that they supplement it with Crypta's

In September 2022, Crypta specialists developed a model on the AppMetrica infrastructure that predicts user churn in mobile apps. They asked the service provider marketplace to compare models and decide which was more effective.

It was important for the team to understand whether the model would become a separate, boxed solution or an add-on to an already existing algorithm.

To build the predictive churn model, Crypta experts did the following:

  • Determined the periods over which users would return to the app, depending on its content
  • Identified signs that users were more likely to leave the app during a selected period based on their average return time
  • Based on behavioral patterns, they compared the user base with the resulting model and found the audience segments most likely to become inactive on the app (and which could be brought back using a promo code)

They checked whether they could more accurately predict churn with expanded data

Crypta experts assumed that the model would be more effective if it was based on more extensive data. To test this hypothesis, they launched a two-week A/B test where they compared the following:

  • Distribution of promo codes using the client's model
  • Distribution of promo codes using the Crypta model
  • Randomized distribution of promo codes

They tested these three approaches on the client’s app for the categories «Cleaning» and «Water delivery»

Sample sizes were all the same: 30% of the entire test audience. During the two weeks of the experiment, the only users considered were active app users who had ordered cleaning or water delivery at least once and who did not have an active current order.

The experiment showed that the predictive model built on the AppMetrica infrastructure perfectly complemented the client model’s churn prediction

There conclusions were reached regarding how more accurate forecasting affected business profits:

  • The Crypta model built on the AppMetrica infrastructure issued fewer promo codes, but the share of sales without a promo code didn’t drop: it actually rose from 7.7% to 15.5%. This showed that more complete data helps reduce the number of discounts issued, because the model more accurately determined potential user churn.
  • The number of sales without a promo code increased. The predictive model based on data from multiple apps (instead of just one) did not in any way harm the overall turnover of the business. In both models, the average nominal value of the promo codes did not differ by a statistically significant amount. This, considered in combination with the observed growth in sales without a promo code, may indicate that the Crypta model had a more targeted impact on user churn.

The predictive churn model is currently in closed beta testing. But you can already implement it as a test by filling out the AppMetrica feedback form. Get access.