How the estimation model works to attract audience and increase revenue

Yandex Games, an example of the importance of models image attend guess

Crypta experts have developed a predictive model to analyze users from mobile apps on the infrastructure of AppMetrica, an analytics and marketing platform. Using the model, Yandex Games, for example, can attract more interested users who have spent more time on the app and contribute to revenue growth, as well as increased LTV.

Yandex Games is a category of browser games that can run on both mobile phones and computers. Yandex will build a game catalog on the homepage, as well as on the Yandex Browser and the Yandex app. Users on Yandex Games appreciate the ability to play without the Internet and synchronize their achievements and progress between devices.

Yandex Games compared its own ad optimization event to the model from Crypta

The app has optimized ads for the AppMetrica event, which is fired if the user spends more than 10 minutes in the game.

At the same time, Crypta experts are developing a predictive analytics model on top of the AppMetrica infrastructure, called predictive LTV. Predictive LTV is a machine learning model that predicts from in-app user behavior on the first day how much time they will spend on similar apps and how that will affect business growth. potential revenue in the future. The model evaluates all new users and sends events to AppMetrica for those in the top 5%, top 20%, or top 50%. Ads can then be optimized for these events.

To build a predictive LTV model, Crypta collected historical data on app installs and user activity on the first day after installation. Then calculate revenue from ad views by the appropriate user category for 28 days after installation. It also trains a model that estimates, based on first day activity, the revenue the app will receive from each user’s ad view over the next 28 days.

At the beginning of December 2022, Crypta experts asked Yandex Games to compare which method would be more effective.

Every day, Crypta experts track new installs and user activity in the app on day one. Based on this activity, they predict how much money certain users who see an ad will bring to the app over the next 28 days, then form the top 5%, 20%, and 50% of users among them. The user groups are most interested in the application and therefore best suited to the LTV of the model prediction and anonymous data submission to AppMetrica.

Can the ads be optimized to attract more engaged audiences ?

To see if ads could be optimized to attract more engagement, Crypta experts ran A/B tests to compare optimizations based on the “user played” event at least 10 minutes” and event-driven optimization based on the Crypta model for the top 20% of solvent users.

The campaigns are identical in terms of settings and budget – they differ only in the optimized event. The first ten days of the campaign were spent learning and collecting data, after which Crypta experts collected a week’s worth of installation data, on which they calculated statistics.

The predictive model attracted users to spend more time in the app and contributed to increased revenue

The results of A/B tests are evaluated based on how much money the app earns in the first few days of use by a certain group of users. Statistics show that ROAS and APRU were stable on day 7 and reached +19% and +25% respectively.

One conclusion was determined about how more accurate forecasting affects Yandex Games profits: Optimizing the predictive model helped attract an active gaming audience, while the cost was similar. action does not change.

“The predictive LTV model would be useful for any app that offers monetization. This could be a game, e-commerce, or subscription-based fitness app. Any app with results. AppMetrica connections can use predictive models,” said Vladislav Titov, head of the Crypta machine learning team.

In addition, ad views of users attracted using this model generated more revenue in the first week after installation than users attracted using this method. earlier in the same period. This trend continues for the next two weeks. App usage time increased by 10.5% on average, meaning the predictive model attracted more engaged and loyal audiences.

“Setting the optimization goal as predictability, we could take a running campaign and change the event used right now to an LTV event. We hope that optimizing the likelihood. predicted to bring more interactive audience, which in turn will increase revenue,” added Vladislav Titov.