Here I introduce some of the standard prediction models for free-to-play and in-app-purchased mobile games. Predict churn event and purchase behavior of players are the most common models that are easy to apply.
This post is written with the help of my University teacher with theory and materials, and discussions with the CEO and CTO of a Mobile Game Company, Hyperkani. I was lucky to be able to use Hyperkani as a case study for my research about prediction models for Mobile Game.
Hyperkani is a game company offering several in-app purchase mobile games. The company is located in Tampere City, Finland. One of the most popular games of Hyperkani is Bomber Friends. For further info about this company: http://hyperkani.com/
After reading several materials, I have summarised several popular prediction models used for the mobile game that could apply to the game Bomber Friends or many other similar mobile games of Hyperkani.
In this discussion, I will mention two models among other standard prediction models for the multiplayer mobile game with in-app purchase, as follow:
1. Prediction churn event of all players (or high-value players) (unsubscription decision)
It is used to predict how long a player stays in the game, whether they will leave in the near future, how many of them will stop playing and which one will stop, etc.
Why do we need to predict this? Because it directly affects the revenue of the game developers.
Purposes of this model:
- Prevent the player from quitting the game, or to extend the players’ lifetime in the game by adjusting strategy, for example, giving incentivization on time to encourage the players to stick to the game.
- Increase the enjoyment, the playing experiences of the game (i.e., their characters advance to new levels too fast or too easy that can reduce their playing experiences in general for the game, other players’ cheating) or changing design, content, lengthen the game life-cycle
The expecting outcomes: to predict whether a player can be retained with some adjustment before the churn event: by changing their playing experience, incentivization, and communication. Is it also possible to ignite a new life in another game (cross-linking or cross-sold)?
2. Prediction player purchase behavior in the game
This model is used to predict the buying pattern of players. The Common patterns that we can use are the time of purchases, most likely of product to be purchased, who make the repeated purchase, and generate a continual revenue stream and minimalize maintenance cost.
Why? Highly frequent advertisements likely lead to advertisement fatigue. It makes players no longer liking or buying goods from the ads. Also, advertisements wear out, which will lead players to ignore them so that these advertisements are no longer effective.
Purposes of this model:
- Help to decide which users should be focused on in terms of revenue.
- Predict the level of loyalty of future users.
- Predict the readiness of purchasing items, such as when it will happen? For example, after finishing a match or another time? “Advertisement timing” increases the likelihood of making purchases to improve revenue. This prediction model will suggest players purchase in-app items only in the time they would really buy them. It allows dynamic advertisement placements for efficiency.
- Understand the potential users, and to create a long-term perspective of the potential relationship.
- Help to tailor strategies to deal with different segments, build the connection between game enjoyment and purchase behavior in free-to-play games.
Some necessary data required for building these prediction models (depending on game type):
- User logins data and in-app actions, action frequency
- Number of matches: online, tournament and friend games
- Distribution of match types: number of each game type the player played: match types according to his skill, friend games through some social platform.
- Distribution of levels played: the number of games played at each level.
- Performance in matches: the number of wins in different game types, the number of wins per level match properties
- Number of trophies
- Bonus: daily reward
- Rank-ups: when the player has just unlocked the level
- Time play: from log in to end of the game, time series (day/night/ hour/minute/day/month)
- Inventory: number of the purchase, number of bought coins and package, the amount spent, number of users with repeated purchases and the average number of purchases per user, time series of purchase
- Game hours: the average daily playtime, average daily session count, overall subscription time
For both models, there are many methods combined to analyze the data and usually employed two methods/models to compare the results. The methods are often used in those including Kaplan-Meier estimator, classification, Random Forest, linear SVM and decision tree, hidden Markov model, neural network, A/B testing, beta-geometric negative binomial distribution model.
Note: For this churn event prediction model, it is crucial to choose the cut-off observation to build the model, depending on the type of game.