Calculating LTV post-IDFA: Solving the need for predictive analytics
When evaluating the effectiveness of a campaign, the lifetime value of a user in relation to the cost of acquiring that user (LTV/CAC), is a key metric in mobile marketing measurement. Working with Apple’s SKAdNetwork (SKAN) post-IDFA has made predictive analytics for metrics like LTV on iOS more challenging, as anonymized data based on activity from the first 24 hours is the only feedback received from campaigns. This is also why it’s essential that marketers set up their conversion value schemas to maximize insights. The entire mobile advertising ecosystem is becoming more privacy-centric and Adjust is embracing these changes, remaining agile, and developing next-generation solutions that ensure our clients’ continued success.
While some marketers are working to overcome this challenge by using methods like coefficients (which we explain below), this is complicated and hard to get right, which is why Adjust recommends leveraging predictive modeling to launch and measure successful campaigns on iOS. Our model feeds large amounts of (SDK) data into machine learning algorithms to sift through the noise to better understand the layers of data that exist and determine what their correlations to one another are. From here we can predict long-term outcomes early on and provide conversion value models and flows that suited to your app vertical and key KPIs.
In this blog, we outline the difficulties iOS 14.5+ has brought to campaign measurement, the pain points marketers are experiencing, and how Adjust is approaching predictive LTV (pLTV) to address and solve these challenges.
iOS 14.5+ and the challenges brought to predicting LTV and evaluating campaign success
Estimating and finding LTV needs to be done strategically and early in a campaign’s lifecycle to ensure that all marketing decisions are optimized to drive top-line revenue. For example, if users brought in by campaign A spent big initially but later churned, while users from campaign B were slow to start but made purchases of an ultimately higher value, examining day 1 activity would be poor in predicting behavior throughout a 30 day cycle. This is why marketers need a good way of predicting LTV.
Working on iOS 14.5+ within the SKAN framework has increased the complexity of predictive analytics. Previously (and now only with AppTrackingTransparency (ATT) opted-in users), we were able to tie iOS campaigns to IDFAs and device-level data, sent via our SDK. From there, we could see the actions taken and revenue generated at user-level, apply predictive modeling to associate that user with a cohort of users, and ultimately estimate LTV.
With SKAN, we receive anonymized user data based only on activity from within the first 24 hours (and then delayed up to 24 hours), as feedback from marketing campaigns on iOS. This complicates the process of predicting user LTV, because now:
- We only receive SKAN postbacks, using a schema that must be defined by the mobile app, and which cannot be tied to a specific device.
- We can’t measure revenue or proxy metrics directly, and must work with SKAN values 0-63.
- We don’t receive the information in real-time.
Before we can start making predictions for iOS campaigns, it’s essential that marketers and developers first set up their conversion value schemas. Then, when running a new marketing campaign, SKAN data must be decoded before being used in predictive models.
Adjust’s solution to predictive modeling post-IDFA
Some marketers have attempted to circumvent issues around predictive modeling by using the coefficient (D0 revenue/DX revenue) per user from historical data. This number is then multiplied by the real D0 revenue to get a predicted DX LTV. The problem with this approach, and similar approaches, is that coefficients can be wildly inaccurate, making predictions just as unreliable.
The solution we have at Adjust uses machine learning to analyze layers of trends that help to predict a user’s future behavior. This way, a user’s historical data, and patterns learned from other similar users—such as those who have provided consent via the ATT opt-in—can help predict the value of that user on, say, day 30, from data supplied on day 1. By working with large data sets (collected by our SDK) fed into machine learning algorithms, we can extrapolate and correlate to paint a picture of long-term outcomes for non opted-in users.
Adjust’s predictive models are custom built for each specific app, meaning that they learn and are trained from the real (SDK) data of each specific app. By pairing predictive modeling with cohort analysis and aggregated SKAN data, marketers are able to extract meaningful insights and make informed decisions. They’re also empowered to understand the future value of a campaign early into its runtime (bypassing the SKAN waiting period), and surface data relationships that might have otherwise been missed.
iOS 14.5+ has made predictive analytics significantly more complex, but not impossible. By leveraging the data that’s available in the most effective ways, marketers and advertisers can gain the insights needed to increase efficiency, increase returns, and improve campaign performance on iOS with the same data-driven mindset as before. What’s crucial is that you have the right tools to enable that performance.