The Basics of Churn Analysis in Social Games

Posted by Dmitri Williams

Find me on:

Sep 23, 2013 4:30:00 PM

What is churn?

Churn is quitting, plain and simple. It’s bad, but it’s also inevitable. No one lives forever, no one plays forever, and no game lasts forever. In practical terms, this means that working with churn is all about minimizing it rather than eliminating it completely.

Churn of course hits your bottom line, but it’s different depending on your business model. If you’re running a subscription model it’s very straightforward. You lose X dollars/month if the player drops. If you are running a free-to-play or microtransaction-based title, you are losing the player’s particular spending. And, as most people know, player spending is highly skewed, with a small minority spending the most, and the vast majority spending nothing. As a general rule, you care the most about the “whales” and want to prevent their churn. Social ripple effects (below) are an exception to that thinking.

And, if you are playing on a new mobile bingo app, you're joining many iPhone, Android and tablet players worldwide. Online bingo sites and casino sites are applying a mobile-first approach for its users, so smartphone gamers' experiences are now prioritised over desktop.

How should you define it?

In a subscription-based model, it’s pretty obvious, but still has a little bit of variation. There are those who let their subscription lapse, those who cease playing but keep paying (zombies, see below), and those who do a hard-stop cancellation where they proactively cancel.

In a free-to-play or console title, you have a conundrum since playing patterns differ widely. Take two players, A and B. Player A tends to play daily, and Player B plays weekly. Neither has played in 3 days. Who has churned? Right there you can see that a one-size-fits-all definition is going to have errors. A truly sophisticated approach will be custom to each player, determining when each has exceeded their own personal pattern and has likely churned. Machine-learning models can handle this, when driven by an expert.

The bottom line is that unless you have a hard-stop cancellation, you will always have some error in your estimations. It’s going to be a question of whether you want to put up with false positives (it looks like they quit, but they’re coming back) or false negatives (it looks like they are staying, but they’re already gone).

Time-based metrics

Few firms can handle this kind of sophisticated modeling. Much more typical is to look across the player base and use aggregated metrics based on retention from the day of installation. These are important because they focus on the early days (and minutes) of game play, where most player churn typically occurs. With these, firms use 1-day, 2-day, 3-day, 7-day, 30-day, 60-day, 1-year, etc. metrics. These are saying “for those who started on day X, this many came back within 24 hours” or 48, etc. These need to be rolling, letting you know what the 2-day retention is for March 5, March 6, and every other day. Those numbers then become a relative baseline you can graph out and look in for patterns.

One further wrinkle. You probably aren’t going to be able to do a truly bespoke solution based on the exact minute a player joins, and then starting their unique 24-hour clock. If you can, kudos. You’re most likely going to need to batch things on fixed 24-hour windows. If you do that, we recommend using midnight GMT. Although that means that North American developers are 6-8 hours off, it’s not too bad. They will get “end-of-day” numbers in the afternoon. The reality check is that games are truly global anyway, and there has to be a standard.

Social ripple effects

When a player leaves, it’s sad. When a player leaves and takes their friends with them, it’s a disaster. This is why you cannot look at your players in isolation. Most players are connected to others and they impact each other.

Social Network Graph

If you can see a social network graph like this one, you’ll recognize that some players are off and solo, some are connected to one another, and some are highly connected. Obviously, those who are heavily connected are going to have a larger potential ripple effect. These graphs aren’t practical, so you’ll need some network statistic to reflect who is more connected. We developed Social Value to do this as intelligently as possible, adding the degree of impact in addition to just the number of connections.

Opportunity cost thinking

If you have a lifetime value model or our Social Value approach, you’ll know that players are not at all equal. Some are worth more, and some have larger ripple effects. However you get to those numbers, you know they aren’t the same. Let’s say you know that Player C is worth $10 to you and Player D is worth $20. Theoretically, it’s worth $9 of effort to keep Player C and $19 to keep Player D.

In practice, you won’t spend anything like this, but you get the idea. You should spend more to retain more valuable players.

Combining with predictive analytics

Now let’s add the ability to know the likelihood of the player churning out. Let’s say with players C and D above, that they are both 50% likely to churn out next month. You should multiply their value by this churn rate to get their expected value. Player C is essentially “worth” $10 x 50% = $5, and Player D is essentially “worth” $20 x 50% = $10. Now you should change your calculus about how many resources you should spend to retain this player. Player C is worth spending up to $4 to retain. In the long run, if your churn models and value models are good, you can use these for intelligent cost-benefit decisions and planning.


In a subscription-based title, some players continue to stay in your system but cease playing. I call them zombies. On the one hand, these are the world’s most ideal customers since they take none of your cycles yet they pay full price. On the other hand, it’s arguably unethical to charge people for a service they aren’t using. Will you point this out? Most don’t. If you do decide to prompt them, be careful what your goals are. I recall the case of a developer who wanted players to “come back” and enjoy the title, but ended up simply reminding the zombie players that they were still signed up and paying for nothing. Those players promptly quit en masse. Ethical, yes. On purpose, no... prediction-the-future-of-game-analytics


Topics: Churn Analysis, Social Ripple Effect, Social Games, Zombies

Spread the word


Spread the Word


Subscribe to Email Updates

Ninja Metrics - Prediction The Future of Game Analytics
Ninja Metrics Katana Demo

Ready for a Demo of Katana®?

If you have a social game or application, we can help you learn who matters most in your network and what features are affecting the health of your community