According to various studies, the market for mobile apps with in-app purchases is growing steadily every year. Around the world, publishers and independent developers are releasing many apps in different categories, competing with each other.
As a result, the advertising market is in an all-out battle to scale their products and attract new users.
One of the most important elements in this regard is the ability to maximize the scalability of ad traffic from multiple channels and a well-thought-out return on ad investment strategy with a one-year or longer horizon.
Simply put, developers know they need to attract as many new users to their applications as possible for effective growth, but they need to guarantee users a return on investment.
Different methods of LTV Prediction forecasting
In order to understand at what price you can attract users to an application while staying in the payback zone, various forecasting techniques are widely used today. Their goal is to predict as accurately as possible the future revenue of the target group of users (subscribers).
So let's take a look at some LTV Prediction forecasting methods for your app growth strategy.
LTV Predictions based on historical (actual) data
The easiest way to predict LTV is if you have enough historical data in your application: from six months to several years for more accurate predictions.
In this case, you virtually know how different cohorts of subscribers have behaved in the past and can predict that new cohorts of users will behave similarly.
In this approach, the predicted LTV is calculated based on historical cohort conversion data for 1,2,...Nth subscription renewals. It is recommended that cohorts not be generic to the entire application, but that user groups be refined by geography, device model, gender/age (if available), and other attributes that may affect their subscription revenue.
In addition, to make the LTV forecast more realistic and not exaggerated, this approach often uses a correction (pessimization coefficient) that is chosen on a case-by-case basis.
The formula for this calculation method can be simplified as follows:
eLTV = Average subscription price * Estimated average renewals count * K
Subscription Price is the average subscription price of a user in the cohort;
Estimated Average Renewals Count is the estimated average number of subscription renewals in the cohort, and K is a correction factor (e.g., 0.9).
Advantages:
- Ease of implementation
- May be sufficient to obtain a baseline prediction (with properly selected cohorts and correction factor)
Disadvantages:
- Predictions are based on historical data from past cohorts
- Rough method of forecasting - correction factor is chosen by eye
- Possible serious errors in forecasting annual and 2-year renewals
Machine learning-based LTV predictions
Machine learning and intelligent models can significantly improve the accuracy of your forecasts and LTV metric estimates.
There are several fairly well-researched models and AI-based forecasting approaches on the market that have proven effective.
Historical data is also required for accurate forecasting, but the amount of historical data can be significantly reduced depending on the model used (relevant for new applications being launched).
Black Box model
An important advantage of machine learning is that we can work not only on subscription renewal events but also on other user actions and feed them into the model to get the most accurate prediction.
The model uses fuzzy rules to generate a prediction and acts as a kind of "black box", a system with a completely closed implementation.
The advantage of this approach is that we only need to prepare a dataset based on various behavioral data (analytical events) and the model itself will make a prediction.
The disadvantage is the difficulty in selecting the optimal dataset and its data for training the model, and potentially unpredictable results at the output, while it is not clear how exactly the model built this or that prediction.
Probabilistic approach
This approach is based on the theory of probability and is described in detail in open sources, and there are many scientific publications about it. At Apphud, we use this approach to build forecasts.
An important advantage of forecasting with probabilistic models is that we do not build a forecast based on the assumption that new subscribers will behave the same way as old users as in forecasting based on historical data. Instead, the model trained on historical data tries to make predictions using probability ratios about the behavior of exactly those new cohorts of users assuming it will be different from what we have observed in the past.
So let's summarize the effectiveness of making LTV predictions based on machine learning.
Advantages:
- High (and highest) accuracy with a properly tuned model
- Ability to train on a wide range of possible user events
- Great flexibility in customizing and configuring different counting techniques
Disadvantages:
- High implementation complexity
- Not many well-described, proven models available in the public domain
- Requires dedicated server infrastructure
- Requires specialized machine learning expertise
With Apphud, you get a robust and accurate solution for predictions of your app's revenue based on machine learning, without having to dive into all the complexities and intricacies of implementation!
LTV Predictions with respect to data granularity
Another important consideration when calculating predictions is the level of granularity at which predictions are made.
Data granularityThe simplest are aggregated cohort forecasts at the application level, the most granular are LTV forecasts at the user or specific subscription level.
The more granular the forecast we have, the more flexibility we have in using cohort analysis and calculating predicted LTV for different metrics such as ARPU, ARPPU, and ARPAS.
Factors affecting LTV Predictions accuracy
When using user revenue forecasting, many factors can affect the accuracy of eLTV. Let's take a look at the most important ones.
User country
Users from different geographies may have different subscription prices and different behavior in terms of app usage and in-app purchases. This factor needs to be taken into account when making a prediction.
Traffic smoothness
A typical pattern in modern applications is to attract users through different traffic channels. At the same time, the quality of the traffic itself can vary significantly. Therefore, it is important to take this into account when forecasting and to double-check the forecasts made for different traffic channels.
Seasonality
One of the most common mistakes in forecasting is not taking seasonality into account. If you see in the historical data that the revenue of different cohorts has a pronounced seasonal trend (seasons, holidays, etc.), this should also be taken into account when making forecasts.
New app features
If significant changes have been released to the app in terms of functionality (redesign of core features or addition of major new features), this can also have a significant impact on retention. Therefore, a prediction based on early users who have not seen the changes will be irrelevant.
New app interface
This is almost the same as the point above, but here we are talking about significant changes in the interface: redesign of the app or new layout of the main sections. If such updates have a significant impact on the use of the product, this should also be taken into account in the forecast.
LTV Predictions with Apphud
Apphud offers comprehensive LTV metric forecasting capabilities to provide valuable insight into future revenue for different customer cohorts. Forecasts are available for 30, 90, 180, and 365 days.
Implementation Features
We use the algorithms underlying the shifted-beta-geometric (sBG) model to generate forecasts in Apphud. Our predictive model uses a sophisticated machine learning algorithm to estimate the future value of each subscription.
Predictions are based on a variety of data, including
- Purchase history
- Free trial data
- User demographics
The model is especially effective for weekly and monthly subscriptions.
A more detailed mathematical justification and description of the algorithms can be found in this blog post.
The process of calculating subscription forecasts includes the following steps:
- Model. This step is used to calculate projected rebill rates (subscription renewals) for different segments based on transaction history.
- Prediction. This step applies the predicted rebill rates to each subscription based on the segment and the current renewal iteration.
Immature Cohorts and Early Predictions
For different prediction situations, we may have different amounts of historical data available for training. For example, if we segment users by geography, there may be too little traffic for some non-major countries for prediction.
Therefore, some user segment limitations in modeling make it impossible to estimate model parameters:
- The segment has a too-short funnel
- The segment has an insufficient number of subscriptions
- The segment has inconsistent data.
This means that we need to cascade segment calculations at different levels and then apply them hierarchically when making predictions.
Our cascade of segments is represented in the model as follows:
duration_intervalduration_interval, app_idduration_interval, app_id, is_trialduration_interval, app_id, is_trial, country
As a result, we maintain the values of the predicted return rates throughout the cascade of segments.
The most useful predictive metrics for optimizing ad campaigns and revenue
The most basic predictive metrics that can be useful:
- Revenue
- LTV
The most common metric used to evaluate the ROI of paid traffic is cumulative LTV.
Cumulative LTV chart with predictions in ApphudThis is what the LTV forecast looks like in Apphud. With this data, you can:
- Segment cohorts using dozens of different segments
- Apply filters to sort data
- Select a metric to calculate LTV: ARPU, ARPPU or ARPAS
Another useful option is to calculate complete or incomplete cohorts. This option affects the display of cohorts that have lived N days within the selected cohort period.
When we calculate the cumulative LTV, since all cohorts in the selected period must live a minimum of N days to be shown on the graph, this makes the projections more accurate.
By default, if this option is unchecked, we count all cohorts including incomplete data for the selected period - this speeds up the process of displaying data for new cohorts, but the predictions may change over time.
More details on the implementation of LTV forecasts in Apphud can be found in our documentation.
Conclusion
In this article, we have outlined why LTV predictions are critical for app growth strategies today.
We've also looked at the different approaches to forecasting and the nuances to consider when choosing one way or another to forecast revenue metrics.
At Apphud, we have considered many of your LTV forecasting needs and present a solution that covers all your needs and provides you with accurate forecasts to effectively scale your ad channels. Sign up for free and contact us to try Apphud LTV predictions!
