[Game data analysis] WAU model

Disclaimer: The data used in this blog is not real data, will transform the real data, focus on the game analysis of the idea of data.

Here is an analysis of the WAU model [article URL, demo URL] of the friend League, using a game (hereinafter called game a) data.

Role:

1. According to the transfer matrix, the future Wau can be predicted;

2. See "Wau user composition" To discover the game: too few new? Too much drain? Too few loyal users?

Concept Explanation:

The nth week mentioned here is the nth week of pushing forward, as shown in:

1. New User: The user who registered the time in the week.
2. Returning users this week : not logged into the game last week, there are users logged in to the game this week.
3. users who have been active for n weeks : have logged in for n weeks, but have not logged in for week n+1, for example: 3 consecutive weeks, that is, 3 weeks forward, but not logged in for the 4th week.
4. Loyal users: Active users for 5 consecutive weeks or more
5. Recent Churn users : N Weeks (1<=n<=4) users who have not logged in to the game (N+1 week), such as: 4 consecutive weeks have not logged in the game, but the 5th week has started the game.

According to the above definition, weekly active users (WAU) by: New users, 2 consecutive weeks active users, continuous active 3 weeks, continuous active 4 weeks, loyal users, composed. Each user type is treated as a user state, and the transition relationship between states is as follows:

Ideas:

1. Using the active user data of game A, the users are divided into each type according to the week, and the number of each type is obtained weekly.

2. Calculate the type transfer probabilities for the first 4 weeks of the week. The 3 transfer probability matrices are obtained, and then the average of each item in the matrix is obtained, and the following transfer probability matrix is used for the prediction of the later Wau.

Notice the last line:the sum of the probability of "recent churn users" converting to "this week's return users" and "recent churn" is not 1, which is not active in the first 4 weeks and the 5th week is active "recent churn", in the new week, if not continue to active, is not considered a "recent churn" in the new week.

3.WAU predictions

Multiply the number of users in the previous week by the corresponding transfer probabilities, and get the predicted number of users for the new week. Here's what you need to note : The "This week's return users" in the new week comes from two parts: 1. Last week's recent churn, 2. Users who have not logged in for 5 consecutive weeks or more last week. For the 1th part of the return users, directly with "Last week's recent loss of users" multiplied by the corresponding transfer probability. How do you calculate the return of the 2nd part? Through data discovery, (for a stable game) from "Last week's recent loss of users" conversion to "return users" accounted for the total return of users of the proportion is not small change, therefore, the calculation of "Last week lost users" conversion of "this week" users accounted for the first three weeks of the value, averaging. Then use: The recent loss of user * transfer probability/This ratio , you can get the predicted "return users this week." The forecast is as follows, and the error rate compared to the real data is 3.6%

Other conclusions:

1) from the above transfer matrix is known: The number of consecutive active users the more the probability of loss, the lower the probability of loss and then reflux is relatively higher;

2) by the following, various types of users in the Wau in the ratio can be derived: the new users in the Wau accounted for the most, and more than 2 weeks of continuous active people accounted for less, the user mainly from the game is the new users, and the loss of old users is more serious. Therefore, we should mainly focus on the conversion rate of new users and the retention of old users. The second aspect, "return users" accounted for more, but "return users" to the "2 consecutive weeks of active" probability is small, indicating that there are many users are not frequently logged in the game, more than a week to play, but after the game will be lost soon after play (7 days no login games defined as loss), Can be in-depth analysis of this part of the return of the user, whether reflux once and never come back? If this is the case, then you need to take a large percentage of the "back-up user" retention measures, such as login bonuses, items such as discount. Another reason is that users are already tired of gameplay and are reluctant to play games, leading to a low retention rate for old users, which requires considering whether to add new scenes or characters or props to attract users.

3) The number of loyal users in this 5-week overall downward trend (no specific data here), which requires the attention of operations, because these apps the loss of the best users of the app is the beginning of the scale of active users.

[Game data analysis] WAU model