Get Back to Blog

Main Approaches to Forecasting Revenue for Mobile Applications in 2023

By Madiyar Khamzanov

Several approaches can be used to forecast in-app, advertising and subscription revenue. When examining them closely, three approaches stand out as the main ones:

  • Forecasting using coefficients (Bottoms Up ARPPU)
  • Forecast using Machine Learning models (Regression)
  • Forecast using a mixed approach - combination of Bottoms Up and ML forecast

Different monetizations have their own nuances, but if we consider the main approaches that big mobile companies use, they can be boiled down to the three mentioned above.

Method 1 - Bottoms up ARPPU

Let the chosen forecast horizon be 1 year (365 days). When forecasting, we use the assumption that the accumulated payment curve for each payment type (in-app, subscriptions, advertising) has a logarithmic shape, and we should extrapolate it to the forecasting period. Thus, we get a logarithmic curve or a trend. A Logarithmic curve form - the general trend of a logarithmic family of functions, one example of which is the dotted orange curve in the figure above.

The general formula for the curve is $y = log2(ax+b) + c$

The accumulated payment curve is the trend of payment accumulation for the selected monetization averaged per user.

Once we get this trend, we can calculate the coefficients of income growth on this trend curve from each day to 365 days (the ratio of the trend curve value in 365 days to the selected day), thus obtaining the coefficients we want.

These coefficients will serve as a multiplier to the already accumulated sum of payments of a particular cohort, and the resulting product of two numbers will produce the desired forecast.

Note that the extrapolation of the trend is not always a trivial task. It largely depends on the number of unique payers participating in the construction of the accumulated payment curve, as well as on the period of users lifetime and payment accumulation.

The best approach to extrapolate the accumulated payment curve for each country, thus making it more sensitive to the heterogeneity of users' payment behavior depending on the origin of the traffic.

Method 2 - Machine Learning

If in the first method we predicted revenue by cohort, now we predict revenue by each user. First, we must determine the forecast horizon - usually it’s 90, 180, 365, 730 days. Let’s choose 180 days. In addition, we should determine the period or the life time from which we will build the forecast for the 180th day. Let’s choose day 7 as an example. So we will make a forecast for day 180 based on the user’s behaviour during the first seven days.

Now we must collect a training sample, i.e. select the specific users who have a life time of more than 180 days. Then we create a vector of attributes describing the user's behavior during the first 7 days and select the user's income on the 180th day as the target event. The more such users we have, the better. Then, we teach the regression on these attributes and apply the model to new data.As a rule lightgbm or catboost are used as a regression algorithm.

Method 3 - Mix

This method is used when the number of installs is large enough, from 1 million per month, and the life of the project is not less than 3 months. In this case, you want to build a quality forecast in terms of the user for 360 days, but taking into account that the life of the project is not less than 90 days.

We build it this way: for each user based on payment history and user inputs in the first 3–7 days we build a unique forecast for the 30th, 60th, 90th day. Then, having received this forecast, we multiply it by the coefficient, which was calculated using Bottoms up ARPPU forecast.

In conclusion:

Bottoms up ARPPU:

  • Easier to implement;
  • Has satisfactory accuracy with the necessary training, about 90% from day 7 for a period of more than 6 months;
  • You can always find out why there was over- or under-forecasting as the coefficients are available.

Machine learning:

  • More difficult to implement and you need much more data;
  • Accuracy can be higher with proper training and adjustment to 95% from day 7 for a period of more than 6 months;
  • It is difficult to understand why there was over- or under-forecasting, as it is a black box.

Mad Curve is currently using Bottoms up ARPPU, ML is in testing.

Our platform already considers many nuances when working with different forecasting methods and monetization. Our models are already running on a per-country basis. There are options of model fixing or daily auto-learning for the last 1-6 months, depending on the stage of development of your project.

When working with revenue forecast, it is important to keep in mind that the revenue prediction is based on actual data. None of the predictive models can take into account the factors such as your future app content, monetization events, live ops, or support quality. Your team should know how to manage and operate a product in a long run to achieve your payback period.

We use methods that work in large corporations and are developed under the supervision of strong teams of professionals. Apply for beta today: we'll help you achieve your marketing and business goals!

More Blog Articles

What Is a Profit and Loss Statement and How Does It Aid in Business Planning?

Any creation of a business, studio, or startup begins with writing a business plan.

Read more

Revenue forecast: why do you need to predict the future of your business?

Generally marketing and product departments go a long way to arrive at the best forecasts.

Read more