How to reduce user churn through effective operational activities?

The “Double 12” promotion kicked off the year-end sales rush. The upcoming Christmas and New Year holidays provide various opportunities for APP to increase sales at the end of the year. In such a market environment, if operations want to successfully increase sales, they must combine "loosening" and "blocking". Operations must not only stand out from the crowd of activities and gimmicks, attract users and seize users' time, but also accurately prevent user loss.

How can operations “precisely” prevent user churn ? As user operators, we can regard the problem of user churn as the water reservoir problem in elementary school mathematics Olympiads. First of all, it is impossible for us to prevent user churn. So , what we need to do is to accurately predict user churn, retain users with a higher probability of churn through effective operational activities, and make the user churn rate as lower as possible than the user growth rate. This will ensure the upward growth of the number of users. With the quantity, there is a foundation for conversion.

So how can operations accurately predict user churn? This requires solving three major problems: concepts, data, and tools:

Question 1: How to define lost users?

Question 2: What kind of data can be used to predict user churn?

Question 3: What kind of tools can be used to improve prediction accuracy?

1. How to define lost users?

Different products have different definitions of user churn. If you use a unified standard to define it, there will be problems. For example, the formula for calculating user churn rate is the ratio of the number of users who churn to the total number of users who use/consume the product (or service).

In actual operation, if we only calculate according to the literal definition, the user churn rate will become lower and lower as the total number of users increases. In other words, nothing was done in terms of user operations, but the KPIs were getting better and better. This will create the illusion that users are loving our products more and more. However, that was not the result.

When operations are trying to accurately prevent user churn, the first step is to clearly define churned users. Operations need to define the concept of lost users based on the type, tone and user portrait of their own products. If it is user churn for a specific activity, the concept of user churn needs to be defined based on the purpose and significance of the activity.

For example, the value of social apps lies in solving communication problems, and lost users are usually defined by the length of time since the last login. If a user does not perform any operation for one or two months, it can be considered that the user has been lost. One thing to note here is that QQ and WeChat are highly social software. Even if we don’t use them, they will still be installed on our mobile phones.

For example, e-commerce apps make profits through user purchases, especially on special days such as Double Eleven and Double Twelve when sales are critical. Lost users are usually defined by the level of purchasing activity. If a user only browses but does not buy, the e-commerce company may lose the user.

Only when the definition of lost users is clear can we formulate good judgment criteria for user churn prediction.

2. What kind of data is used to predict user churn?

What is the likelihood that a user will churn next? Mathematically speaking, we can use the Bayesian formula to estimate the probability of user churn. This mathematical formula contains a simple truth:

When you cannot accurately know the essence of a thing, you can judge the probability of its essential attributes based on the number of events related to the specific essence of the thing.

This method of churn prediction has a bit of statistics + psychology flavor. Taking e-commerce operations as an example, if you see that a user browses a lot but buys little on Double Eleven, then there is a high probability that this user will not shop on Double Twenty. However, such predictions are still not accurate enough.

With the development of big data technology, more accurate predictions are made through data analysis, model algorithms and deep learning techniques to predict user behavior. Before making behavior predictions, what user data do operations need to consider that can help us predict user churn? This is a crucial step in building a computational model.

From a data perspective, at least detailed user portrait data and behavior data are required, namely:

• User portrait information: ID, gender, age, region, member type, user source...
• User behavior data: login days, online time, login frequency, registration days...
• User consumption data: the number of days since the last order, cumulative orders, cumulative consumption amount, average order value, etc.

It should be noted here that the consideration criteria for each small dimension are different in different APPs. For frequently used apps, such as social apps, video apps, taxi apps, and music apps, the login frequency should be appropriately increased; for apps that value user time, such as reading apps and information apps, the online time should be appropriately increased; e-commerce apps pay more attention to conversions, and operations can use visual embedding technology to accurately count conversion data such as purchase pages and payment pages.

3. What kind of tools can be used to improve prediction accuracy?

The data mentioned above alone are not enough, because there are many external factors that restrict the accuracy of the data.

First of all, different environments and geographical locations will lead to different user behaviors and interest preferences. As the user's geographical location changes, he moves from first- and second-tier cities to third- and fourth-tier cities, and the user's APP usage will also change, which cannot be reflected in the APP's own data.

Secondly, when making churn predictions, the volume of the APP’s own data is seriously insufficient. If users are silently leaving and rarely open the APP to use it, how can we still generate enough data?

Furthermore, the APP’s own data has limitations and cannot tell operators about changes in user interests. Users who are no longer interested will definitely leave, so there is no need to retain them.

Therefore, at this time, operations need to rely on external forces to improve the accuracy of predictions. At present, a more feasible approach is to cooperate with third-party big data service providers, find effective data for churn prediction through data combing, and then integrate data from two or even three parties to expand the data volume and dimension, and finally complete accurate behavior prediction.

At present, a few companies in the data field have launched behavior prediction products. A company that is leading the way internationally is Google. In China, Getui is one of the earliest data companies in the industry to develop behavioral predictions, and has opened a corresponding function in its application statistics product "Number", which can provide predictions of key behaviors such as churn and uninstalls for APP operations. In addition, “Number” can also provide a visual tracking tool to realize the statistics of custom events, and perform data analysis while statistics, and provide behavioral predictions for custom events such as purchases and sharing.

With the help of big data behavior prediction, operations can gain insight into user churn behavior in advance, intervene early, and retain users who are about to churn through appropriate operational means, truly achieving the effect of "blocking".

In short , whether it is the year-end rush month, the mid-year promotion, or various activities and festivals, operations must have the operational concept of "combining unblocking and blocking". Especially today when the traffic ceiling has already appeared, predicting and preventing user loss will become more important. This requires operations to not only have careful data thinking and understanding of cutting-edge data technologies, but also to find good data partners to jointly explore the deep value of operational data, starting from user needs, retaining users with services, and promoting conversion with experience.

Source: Shrimp Operation