Skip to content

How churn prediction can maximize your retargeting ROI

Unravel the enigma of churn prediction and its power to skyrocket your app retargeting ROI. At MAU Vegas 2023, Adikteev’s COO Kate Lovejoy teamed up with Reggie Singh from Adjust to navigate the landscape of user retention in today’s competitive app industry. Discover effective strategies to keep your users, prevent churn before it happens, and make the most out of your marketing investment.


What you’ll learn: 


Watch the full video here:




My name's Kate Lovejoy. I'm the COO of the retargeting business at Adikteev, a mobile growth engine. As a retargeting DSP, we help you grow your user base and make the most of your existing users. We do everything after the installs. Today I'm joined by Reggie from Adjust. Reggie, I'll let you introduce yourself a bit.


Thank you for having me, Kate. My name is Reggie Singh. I'm the Director of Partnerships at Adjust. For those of you who are not familiar with it, Adjust is a mobile measurement platform, essentially an SDK, to measure where your users are coming from and where they're going afterwards. I've been with Adjust for about six and a half years now, and previous to this I was on the UA front at a company called Smule.

For all the marketers out there, I totally understand what you're going through and hopefully we can address a couple of the pain points you have, at least as it relates to churn.


To kick things off, Reggie's going to talk about the problem of churn, which I think all of us in the mobile space have been facing and dealing with and trying to address. He'll share some of the latest figures in the industry and then I'll talk about a solution that Adikteev has for churn prediction: moving from a rule based methodology for defining when a user will churn and starting to predict with machine learning when a user will churn. So, Reggie, take it away.


The problem with user churn


Cool. Thank you. I'm not going to go through each of these slides here, specifically bullet point by bullet point. Essentially, what we’re looking at are retention rates data from 2021 and 2022, and you’ll notice a common thread: retention rates have been pretty stable between the two years, and that's a good thing. For us marketers, current or former ones, having such stability and predictability with retention is going to be really key in coming up with a proper strategy as it relates to churn. 

As you really think about churn, and creating a strategy,  one really, really important point as you go through these analyses is the data that you're utilizing and where you are getting this data. When you think about user acquisition, you realize that you really do have a leaky bucket problem. Too many of us are spending tons of time, tons of money, acquiring lots and lots of new users. But sometimes we forget about the leaky bucket problem that we all deal with, whether you like it or not. Maybe that's related to product or aggressive monetization strategies. Whatever it is, you're spending a ton of money acquiring users. You want them to stick around, especially in this economy. You want your money to go as far as it can.


Absolutely. We see that time and time again that it's more efficient to keep a valuable user who you know has opened their wallet before, who's making purchases, than it is to go acquire a bunch of new fresh installers.


Adding to your point, with MMPs like Adjust, you get a lot of data insights as it relates to retention and churn. And a couple of those data points, at least that we're looking at from our side at Adjust, is looking at session data— retention, dormancy periods, how long have people gone inactive. We consider a user dormant typically after seven days and that's when they're available for re-engagement.

In addition to that, we also measure things like uninstall. So many of us download so many apps where we're engaged for the first couple of months and then we leave it and not use it again. Perhaps, once a year you go through your phone and delete a bunch of apps. Really acquiring or re acquiring and reengaging with those users is the best bang for your buck.


What we're trying to do is get an ad in front of the user at that moment when they're thinking about leaving your app, when they're thinking maybe they're not as interested anymore, rather than showing them an ad when it's too late, when they've already uninstalled.


Yeah, that's a great point. Utilizing a partner such as your MMP or Adjust and another partner like Adikteev to complement your strategies is really, really key. Why wait for that user to leave just to get them back again? Why not maintain that user, keep them engaged on your platform, rather than waiting for them to leave? You probably need to spend a lot more money to get them right back. I think using some of the products at Adikteev, and Kate's going to talk about them here, is really important to complement your strategy, not only utilizing the data points that your MMP is providing and also the tools when we're talking about audience and audience segmentation. So if that user has churned, you can create audiences and provide them to your partner, such as Adikteev, for them to utilize in their systems— regarding retargeting, reengaging and other tools are available out there, such as push notifications and whatnot. So it’s important to really work very closely with your partner, such as Adikteev, with the tools that you currently have and your data sets to really come up with a cohesive strategy as it relates to churn. That being said, I think it'd be really interesting to see what Adikteev has on the churn prediction front.


Adikteev's user churn prediction technology


Yeah, absolutely. As I mentioned, we are all about predicting when a user will churn. I wanted to illustrate what this problem actually looks like. 

What are the pain points of churn? One is a blind spot for UA channel quality. If I ask most marketers here to score the vendors or partners that they're working with on churn likelihood of the users that they're bringing in, that would be a hard metric to really measure. 

Another real pain point that churn prediction solves for is moving from a rule-based methodology of targeting a user once they have lapsed for 7, 10, 15 days and then get an ad in front of them. Instead of using this rule-based methodology, we want to move to a predictive methodology: getting them right when they're thinking about churning rather than waiting a set number of days for every single user. The problem that we see with waiting those X number of days is it's late, right? It's after the churn has already happened. We want to address it before it happens.

When we're talking about user churn, it's not just churn that we're talking about. We're talking about revenue. And so, a cool thing about our solution is that we can actually attach the revenue to the churn that you are addressing. We want to be able to tell you by getting in front of these users, there is a set dollar amount that you're going to be getting back by keeping them in your app ecosystem.

Last thing that we're really solving for when we talk about this problem of churn is, as Reggie mentioned, in today's economy, efficiency, efficiency, efficiency is really key. We're all seeing that UA costs are rising. It becomes all the more important to invest in keeping those users that you've spent this money to acquire.


Indeed. I would say 2023 is the year of retention and re-engagement.


Tips on creating churn prediction in-house 


Absolutely. It's not just about driving installs. It's about what we do with those installs after. So we are very proud of the models that we've created at Adikteev. But we know that this is a problem that a lot of gaming studios and other marketing companies have tried to address. So of course, we want you to use our model, but we also want to share some actionable tips from our data science folks.

If you want to try creating a model in-house, we've actually run head-to-heads of our model versus in-app models created by studios. And we've won out simply because this is really our area of focus— predicting churn as a retargeting company. 

The first tip is do not over-engineer things. The way our team approached it is we used an existing model from outside of the adtech sphere and then applied rules to it and adjusted it. There's no need to really invent something wholly new. Use a baseline that exists and make the proper adjustments. 

We also think it's important to measure accuracy at the segment level versus the user level. It's really hard to know what one individual user is going to do, but when you zoom out and look at it at a segment basis, that's when your predictions become more solid. Of course, you’d want to validate your predictions: always validate them at the segment level and not the user level. 

Lastly, in our experience, we've found it much better to focus on predicting churn instead of predicting LTV. It's really hard to know exactly how much a user is going to pay and for how long they're going to do so. But we do have enough data points to say if a user will leave or not. And when we're talking about very valuable payers, we can use this churn model as a proxy for what LTV will look like. So again, just some quick tips there.


No, those are great tips. I think when you're really looking at these predictions and working with a partner such as Adikteev, you can really utilize the data set that you have available, the historical dataset you see within Adjust, in really looking at your baseline.

I think a really good strategy in general is to really look at the retention rate by channel; looking at the churn rate by channel. This is something that we should always be keeping updated on a moving window basis. As you look at these predictions and implement some of these strategies, you can really see how that compares to the baseline that your data providers are highlighting in some of the data segments.

How to score users for churn


Absolutely. The data is a key part of this. We are relying on MMPs like Adjust to build our model. We are pulling in all of the organic postbacks that Adjust is sending for a given partner, so they'll send us all that session data, etc. And then we're reusing that to feed the individual model for the individual app that we are working with. 

To take a step back, what is the output of this churn prediction product that we're talking about? Basically, we take a look at all of your users and we come up with a score for each of those users: a number assigned per user from 0 to 1. Zero means this user is staying in the app. They will not churn. They're great and they're going to keep paying. Don't worry about retargeting them. A score of 0.8, 0.9, or 1 means a big red warning: that user is at risk of churning. Let's get a push notification in front of them. Let's serve them an ad. Let's keep them engaged, particularly if they are a valuable user. 

We use those scores to basically define audience buckets for retargeting. We have a bucket of users in that 0.1, that 0.2 to 0.3, etc., and that 0.8, 0.9 and 1 score audience are the ones that we would run campaigns on for retargeting.

Churn prediction and its impact on retargeting


Now, I am not going to walk you through all of these charts displayed here. Don't worry, I saw some worried faces in the crowd! Basically, what I want to talk about here is how churn prediction impacts the retargeting strategies that you use today. 

So many folks in this room are probably working with a retargeting strategy that is rule-based. Like we’ve mentioned earlier, I want to retarget users who have lapsed for seven days. After seven days of not opening the app, I will get an ad in front of them. What we've found in all of our studies is that inactivity window-based rules like this are actually a pretty good proxy. We found, and you can see some data on that in the first chart here, that it works.

Users who are seven days lapsed are likely to have already churned, but you're leaving a real missed opportunity on the table by waiting for those seven days. As seen in this graph, there's this big pool of users that you're missing out on who have this high, high likelihood to churn, this 0.9, this 1, who are excluded from your audience lists.

We’ve talked about the problem of churn, predicting churn, and getting an ad in front of those users. Your next question may be: does it actually work? Can you prevent churn from getting an ad in front of the user? And the answer is yes. We've done many incrementality studies to say what is the difference between the users who are in these various buckets, who do not see an ad versus those that do. We saw a positive lift across all audiences, but the very strongest lift on the users that have the highest likelihood to churn. Those are the ones that it's very important to get an ad in front of. 

Testing the accuracy of churn prediction


Accuracy, accuracy, accuracy. Indeed, I think this is the year of predictive analytics. You're hearing that probably at many of the talks today. For every time you hear the word 'prediction,' I encourage you to challenge whoever says that to you. Ask if they can also validate those predictions. If we tell you that these users are going to churn, you can come back to us a month later and ask, “Did those users actually churn?”

We are really strict with what we're working on and we need more than 85% accuracy on all of the models that we're using, and very proud to say that the models are achieving that. We can say with above 85% certainty whether a user will churn or not. Below 0.8, don't touch it. Don't go after that segment.


Yeah. Companies like Adjust can also, as a third party, help provide this data to understand if those predictions are actually accurate. Are you seeing those churn? It’s very easy to visualize them within your MMP dashboards so you can really see the improvements pretty easily in real time.


More features of Adikteev's churn prediction


Absolutely. We talked about using this for retargeting with a DSP. You can also use these audience lists across other channels that you're working with. We create this audience list for you of users who are very likely to churn. You can use that for Meta retargeting, TikTok retargeting, Twitter… across the board, depending on what your use case is.


I think that's a really good feature there. And you know, there's a lot of talk about preventing churn, which is great. That's just moving forward. But what about all those people that have already churned? You know, I spent millions and millions of dollars and lost tons and tons of users. I think you can also utilize audiences, especially from a company like Adjust just to create segments of audience or users that have already churned and try to get those people reengaged as a parallel strategy with preventing churn moving forward.


Absolutely. Absolutely. It's kind of a full holistic approach here. We're not saying that you should only use this kind of churn prediction audience. You really want to have a holistic retargeting strategy, which is where partners like Adjust and Adikteev work together to solve all of your churn problems. 

In summary, what we're suggesting here is segment your users by churn risk and predict your future revenue. We have a very strong accuracy. It works at scale. Churn prediction amplifies a rule-based strategy that you already have in place, and it can also strengthen your retargeting on social media. We see this as the evolution of retargeting, going from rule based to this very cool predictive technology.