Learning to track user behavior and leverage this data to tweak or improve a strategy will help businesses gain a competitive edge in the long run. Access to large amounts of historical data gives companies the advantage of being able to predict user actions.
With billions of dollars poured into the development of powerful predictive models, methods that were once only accessible to a handful of industry behemoths are increasingly embraced by other players. Some of them are offered for free in MyTracker.
In this article, we will walk you through the benefits of MyTracker's predictive analytics model.
Applications of predictive models are remarkably diverse:
Mobile apps are no exception to this. With an ability to make predictions of revenue, churn rate and other audiences behaviour metrics, apps offer a helpful way to propel your project forward in the most cost-effective way.
A customer’s LTV (Lifetime Value) is where the power of predictive insights comes into play. This key metric indicates the total revenue a business will generate from each particular customer over the length of their relationship.
It can give you an idea of what channels will not work for your ad campaigns as users will generate less revenue than you spend to attract them.
Being able to predict a customer’s LTV is something many dream of. But, as Tech in Asia points out in its overview of MyTracker:
Traditional methods used to calculate LTV can be extremely complicated and are often unable to account for any sudden shifts in customer behaviour. Additionally, attempts to predict LTV in-house can be challenging, due to the amount of resources required to do so. This is where MyTracker wants to make a difference.
Accurate LTV prediction can be challenging for a number of reasons:
1. The behavior of customers can vary substantially. One person may pay a large lump sum, while another will make smaller payments over a longer period of time. To get an accurate forecast for different types of users, one needs to have them segmented.
2. Chaotic and unpredictable, user behavior cannot be easily defined through a single function.
3. There are many factors at play here, such as the app itself, the means of monetization, the target audience, etc.
There are two ways to predict LTV:
1. Based on historical data. The historical datasets for different groups (cohorts) of previous users can be fed into a machine-learning model to train it to make predictions for a new cohort. The more similar the two cohorts, the more accurate the prediction.
2. By tracking the behavior of a cohort. The system tracks a group of users for the first couple of days and then tries to approximate the LTV using the function that is the most appropriate for the given cohort.
Unfortunately, each model has its drawbacks.
1. The historical data-based approach does not take into account the factors that may impact the revenue from a new cohort. These can include promos and ad campaigns that may change user behaviour dramatically. As this method draws on historical data, the system will not be able to make accurate predictions for users from new geographies or ad sources (i.e. a completely new cohort of users you have never seen before).
2. The tracking-based model is also unable to properly address a number of factors such as ad campaigns, promos, features and fading interest from customers. Without material historical data, human behavior remains unpredictable for algorithms in many respects.
This is why MyTracker relies on a much more sophisticated technology.
First, MyTracker classifies users into different segments (cohorts), with users from each cohort sharing a number of common traits. The model identifies the high-level factors that impact user behavior, breaking them down by:
We identify cross sections that provide the most effective categorization of users by LTV. Users are grouped into hundreds of cohorts that are taken into account for prediction purposes.
MyTracker then uses a number of linear, coefficient-based and gradient tree boosting models along with the historical data and tracking-based models with machine-learning features described above.
The choice of model depends on an array of factors such as:
A model may be replaced from time to time by another one that describes the behavior of your users more accurately.
Algorithms can make predictions for various types of revenue (in-app payments, subscriptions, ads) and select the best model to analyse each cohort in-app. Over the day, several different models are used to make a prediction of your LTV.
Models are selected automatically to fit the audience of a specific app. The more data is accumulated, the better the algorithm is trained and the more accurate the prediction will be. To achieve even higher accuracy, LTV predictions are updated during the first eight days of the app installation.
Powered by this technology, MyTracker helped video-editing app Efectum grow its audience significantly through general advertising sources like Google Ads and TikTok. This enabled Efectum to more than double its return on marketing spend by focusing on the channels with customers who have a higher LTV (based on the predictions).
One of the key weaknesses of predictive analytics is the poor ability to factor in qualitative shifts or fluctuations after the bifurcation points. As all these algorithms are based on quantitative probability methods, they face difficulties when it comes to predicting user behaviour after sudden and significant changes.
Say you announced a big discount or promo, resulting in more customers buying your products. Is this only a temporary effect? If yes, when will it be reversed? Will it lead to a higher or lower LTV of an average user?
Predicting such effects is a challenging task for an algorithm, which may produce a less accurate LTV prediction after a successful promo. Another factor likely to impact the analysis is a change in ad monetization or overall app development strategy. The system detects unexpected activity and makes relevant adjustments to models by flattening the peaks. Yet, 100% accuracy cannot be guaranteed in this case.
One of the greatest strengths of MyTracker’s model is its ability to predict revenue generation from any group of users. Thanks to the model mix, we are less dependent on historical data for a similar cohort. Even if no historical data is available, the system makes calculations based on user behavior in similar apps.
This means that you can have an accurate estimate for any cross section (by age, sex, country or add campaign).
Different types of payments (in-app payments, subscriptions, ad monetization) may vary substantially. Subscription payments can drop, while ad revenues may increase. Their trends may often move independently of each other.
This is why MyTracker employs different approaches to tackle each type of payment. Some of them would need a regression tree model, while for others a simple approximation of the logarithm function works better.
MyTracker can predict Lifetime Value (total revenue from a user since app installation) with an accuracy of 70–80% on the 30th, 60th, 90th and 180th day of the app installation.
The more data we collect by app, revenue type or other metrics, the more accurate MyTracker's prediction will be. All methods that we use in MyTracker are regularly validated. On top of that, the prediction quality is controlled on an ongoing basis, with some models discarded automatically and others becoming more dominant.
LTV prediction is a valuable marketing tool to analyze apps with in-app purchases or subscriptions and assess revenues by advertising channel and campaign before you spend money on them.
Relying on billions of actions, MyTracker models are capable of making up-to-date predictions from the first days of use. As your app data accumulates and the models are trained to adapt to a specific audience, predictions become increasingly accurate.
Integrate MyTracker SDK into your app. More instructions on how to do this are available here.
Go to our Constructor and select the prediction period in the Metrics → LTV Prediction section.
More information about MyTracker’s approximation and linear LTV prediction models can be found here.
MyTracker does not limit itself to LTV predictions. We are now working to improve the tools to deliver more predictive insights. You will soon have access to predictions of churn rate (Churn), DAU and MAU. These will be most helpful in predicting the customer reaction to new features or changes in the product marketing strategy.