Helsinki, Finland

Game Analytics – Prediction models (P.2)

Gamer’s playing hours

From game operator’s perspective, they want to know:

  • How many players will join the game?
  • How long will they stay in the game? (The length of time since player first joined the game to the time of his/her last login.)
  • Gameplay hours & patterns are embedded in their game hours.
  • Will the players leave in the near future (unsubscription decision, how many, and which one?)

Predicting Unsubscription decision of players is important because it affects directly to revenue of game developers. Prediction of how long players will stay in the game can be interfered by external behavior, for example, how quickly/difficult their characters advance to new levels and how long they spend in the game each day.

Method: using the Kaplan-Meier estimator.

  • use short-term behavior to predict long-term behavior (Game hours as the input, the average session time, average daily session count, average daily playtime as a factor for short-term behavior variables: average length of ON period, average season length and the overall subscription time to indicate long-term behavior)
  • need to know whether there is temporal dependence between namely days, weeks, ON period (a group of consecutive days during a player play every day) and seasons.

The benefit of this model is to prevent players from quitting a game or at least minimize the number of quitting players, when players are dissatisfied with design, content, or other players’ cheating.

Predicting Players’ transactions

Predicting user behavior in an in-app purchase game by predicting purchase events. This model uses the variable “time of purchases” but not “product to be purchased”.

Many players invest in a game by purchasing the goods to improve their skills and level in the game.

A problem that many game developers often make is running too many advertisements in their game. Frequent advertisements do not lead to increase revenue because of advertisement fatigue and advertisements wear out. Ads fatigue makes customer no longer liking or buying goods from the game because it bothers them too much. Ads wear out means that customers will ignore the ads, so it makes no effect.

There is a solution for this problem: “Ads timing”. We allow advertisements appear only when the player is likely to make purchases.  Ad timing makes the likelihood of converting through the ad is increased, which leads to raising the revenue.

The predictive model is based on in-game signals: players success, curiosity, social interaction, playtime lengths and action frequencies. Predicting players’ readiness to buy is a cue after finishing a match. predicting the buy of a cue for gold coins for real currency.

There are two stages:

  • 1st stage: predict that the buyer would buy a cue of any kind, then decided to only predict the gold cue buys.
  • 2nd stage: use a segment of paying users as a subset of all users

Goal: improve the monetization of online games, using microtransactions by reducing the amount of advertisement sent to the user. suggest user purchase in-app items only in the times he would really buy them

Method: machine learning and data mining, matchmaking, user segmentation, cheater detection.

The study also gives recommendation systems in a game: recommend in-game item timely which allows dynamic ads placements

Data collection: D1: a data set of all users and all of their actions generated by clients and server in a period of time 1 month. They divide this set into subset: newly registered/non-registered, play 1 game/10 games. They find out a group of newly registered and bought some items.

D2: a set of all users that registered and bought some gold cue in the same period of time. the method used Random Forest, linear SVM, and decision tree

Log data: user logins, in-game actions including features:

  • number of matches: matched 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 (in relation to the bet by coins, higher bet higher level)
  • match properties: time spent in game
  • number of trophies
  • bonus coins: daily reward
  • rank-ups: when the player has just unlocked the level, shown by a dialog
  • time play: from login to end of the game
  • inventory: number of already bought cues and gold

Divided into 2 parts: training set 20% and testing set 80%, use 5-fold cross-validation

Information about features importance can be helpful to the game designers in order to improve the game and where to focus their design goals.

Churn prediction for high-value players in casual social games


  • Implement a hidden Markov model to address temporal dynamics.
  • Using a neural network to best predict performance in terms of area under the curve.
  • A/B testing to assess the business value of churn prediction.

Result: contact players shortly before the predicted churn event substantially improves the effectiveness of communication with players. Giving out free in-game currency doesn’t significantly impact the churn rate or monetization of players.

Conclusion: players can only be retained by remarkably changing their playing experience before the churn event. Cross-linking to other games in the company’s portfolio may be the more effective measure to deal with churning players.

Assess the business impact, the effect on communication with players, churn rate and revenues

Benefits of the prediction when a user will leave a game lead to the opportunity to adjust the game/playing experience to extend the lifetime of a user in a game, incentivized to stick with a game, bring potential monetization and possible to ignite a new lifetime in another game (cross-linking or cross-sold) or even lengthen game life-cycle.

Prediction action:

Churn prediction model for high-value players based on players’ gaming activity data and test incentivization as a method to deal with churning high-value players.

They define high-value player segment. They formulate churn prediction as a binary classification (churn or no churn) problem. They use AUC, the area under the receiver operating characteristic (ROC) curve to compare performance because it allows them to compare models across all possible classification thresholds.

The raw dataset includes 3 categories of data:

  • First: in-game activity tracking data (time series of logins per day)
  • Second: revenue-related tracking data (a time series of revenue generated by players)
  • Third: player profile data (how long the player has been playing the game, which country the player is from)

They also measure the business impact by applying the churn prediction model on a small scale by implemented A/B testing into 3 group of players.

  • Group 1: sent out incentives to high-value players after the churn event happened (14 days of inactivity).
  • Group 2: sent out incentives to players that about to churn according to the prediction model.
  • Group 3: no taken

The measurement is used by means of key performance indicators (KPIs): churn rate, revenue as well as email campaign click-through rate and facebook notification click-to-impression rate to measure the effectiveness of communication with players.

Definition 4. Churn Rate (CR) CR = 1 – (# active high-value players) / (# high-value players)

Definition 5. Daily Revenues (DR) DR = total revenues from players in a group during a day

Definition 6. Email Click Through Rate (CTR) CTR = (# gifts claimed by email) / (# emails delivered) Definition 7. Facebook Click To Impression Rate (CTI) CTI = (# gifts claimed by notification) / (# notifications seen by players)

Results show that sending substantial amounts of free in-game currency (monetary value approximately 10 USD) to churning and churned high-value players does not affect the churn rate remarkably. This indicates that highly engaged players cannot be retained at the end of their lifetime by simple incentives.

Compared to a reactive churn management policy, the one leveraging the prediction model improved the communication channel effectiveness by factor four to five. For email campaigns, the click-through-rate was increased from 2.4% to 12%. For Facebook notifications, the click-to impression rate was increased from 8.7% to 31.5%.

A possible extension of our research is an investigation of cross-linking to other (ideally similar) games in the company’s portfolio. This appears to be a more promising approach to deal with player churn than incentivization. Another option that we discussed and that would be valuable to explore is to touch the deeper gameplay mechanics and change players’ game experience in a way that keeps them interested in the game.



Leave a Reply

Your email address will not be published.