AppQuantum: How to build predictive analytics

Fedor Loktionov

1. Why predictive analytics is needed

Predictive analytics is in demand today in many areas. Including forecasting revenue from mobile games and apps. This is what will be discussed today. But first, let’s focus on when it might come in handy.

Let’s say you have created a mobile application that is monetized through a paid subscription:

• every new user gets free access to the service for the first 7 days;
• after launching the application, you start promoting it, launch an advertising campaign on Facebook;
• real data about purchases will appear only after 7 days, when the free period ends for the first users, and some of them will leave the bank card binding to the service.

It turns out a situation where money goes to traffic, people use the application, but you don’t know if the marketing expenses will pay off.

At this point, predictive analytics comes to the rescue. In some cases, on the second day she can demonstrate how effective the UA campaign is by showing how much you will earn in the current campaign.

Thanks to this, it is easier for the marketer to make a decision, for example, regarding the further promotion strategy.

Or another example. Suppose you have a free-to-play mobile game:

• it was released a year ago and it has a base of active users, as well as paying players;
• the project is growing, the amount of spending on marketing is growing rapidly;
• at the same time, the risks and costs that need to be reduced are growing.

And here it is necessary to turn to predictive analytics again. Without it, it will not be possible to optimize the attraction strategy, nor to know in advance what to do with the budget in order to lose as little money and time as possible.

But, of course, as a tool it is useful not only for marketers. In our AppQuantum, for example, project producers and media buyers usually use predicates:

• it is important for producers to track the effectiveness of filling for individual countries, groups of people and sources;
• it is important for buyers to determine the effectiveness of specific sources, and then be able to scale the campaign in time.

2. What should be remembered when referring to predictive analytics

From the first chapter, it became clear that predictive analytics can show future earnings from the current campaign. But there is an important nuance here.

When using predictive analytics to estimate future revenue, it is not the actions (that is, whether purchases were made by specific users) that are predicted, but the LTV of the average user received within a specific campaign.

Of course, ideally, any mobile marketer, developer or publisher would like to know exactly how each of its users will behave. Will it pay and if so, how much. However, it is impossible to predict user behavior with such accuracy.

3. Building a predictive model

Forecasts can be built for any applications and games. That’s just the complexity of the calculations strongly depends on the type of application.

For example, it is most difficult to build predicates for midcore and hardcore mobile games. It is much easier to predict revenue from hyper-casual titles.

In general, the following rule works: the higher the percentage of advertising monetization in the application, the higher the accuracy of the predicates.

Also, the accuracy is strongly influenced by the age of the project. The older the game is and the more users it has, the more data about it — and therefore the future forecast is more accurate.

But let’s take a closer look at how the predictive model is built.

I)

The very first stage of building a predicate is data collection. With the help of tracking in applications, we get the necessary information: from which sources, countries and platforms our users come. Data is also collected during the user’s lifetime interval (24 hours) inside the application.

II)

The second stage is validation and data processing, during which:

• invalid payments are deleted and the full revenue from the user’s in-game purchases is determined;
• the part of users with too large volume of payments is separated. The data of such people can spoil the quality of the prediction;
• the total revenue from ad views for each day of the user’s life is determined;
• cohorts are selected, according to which payment profiles will be built for a period of 15 days or more. Plus, determine the minimum size of the cohort, the number of paying users in it, the total number of ad views.

Before the second stage, do not forget to pre-configure the payment validation mechanism.

III)

At the third stage, a predictive model is calculated.

To do this, you need to multiply the revenue from the cohort by the coefficient corresponding to the platform/ country/ source/ traffic optimization and the day of the cohort’s life.

The coefficient is calculated independently based on trends. We are talking about graphs that display the ratio of the LTV accumulated by the nth day to the LTV of a certain fixed day (for example, the 90th).

Trends

The values corresponding to trends are commonly referred to as Payment Profile or payment profiles.

Profiles are broken down by platform, country, traffic source and fill optimization.

 

For example, let’s take a profile on an abstract country. Suppose the cohort earned $1,500 in the first month of life. From the payment profile for traffic from this country, let’s say we learned that, by day 30, users are monetized by 30% of the monetization for the entire lifetime

Therefore, for all the time this cohort will work $1500 * 100% / 30% = $5000.

And then we predict LTV and calculate ROI.

For reference: the initial construction of a payment profile requires at least 1000 installations, at least 75 paying users and the age of the cohort from 7 days of using the application. But of course, the more data, the better. The main thing to remember is that we need to build a profile for each of the breakdowns.

IV)

And last but not least is the visualization of the data obtained so that each producer and media buyer can evaluate traffic and adjust the purchase in real time.

For example, we use Tableau for visualization.

4.Predicates in User Acquisition

In User Acquisition, the work with predicates can be divided into two large blocks.

1. Evaluation of current traffic

The UA evaluates the predictive LTV by the purchased traffic, that is, how much money these users will potentially bring to the application.

We take data on any advertising campaign at the level of an individual user, grouped by platform /source/ campaign/country. Using business intelligence (BI) tools, we find detailed information about this campaign.

We focus on ROI. This metric makes it clear what kind of profit we get from attracted users, if there is one at all. That is, their LTV should be higher than the CPA (Cost per Acquisition, not to be confused with Cost per Action).

In any case, an analysis can be carried out and a conclusion can be drawn about what to do next. Perhaps some creative has gone very well and you can continue to work with him or vice versa. Maybe the wrong GEO or campaign optimization is selected.

2. Preparation for further traffic purchase

The global task is simple: to buy traffic at the installation price cheaper than LTV.

Here the algorithm is more complicated: using the filter settings in the predicate, we select the type of campaign optimization and GEO. Let’s say we are going to launch an advertising campaign with optimization for purchases in the USA.

According to the calculation of the predicate, we see that a campaign with such optimization in this country will give the LTV of the user $ 10. We take note of this information, put the appropriate KPI. That is, you need to redeem up to $ 10 so that the campaign does not go into negative territory. At the same time, users bring $10 only with such a certain number of purchases – respectively, and this information is embedded in the KPI.

The predicate works in two ways: it allows you to scrupulously evaluate the purchase of traffic both separately for each campaign and comprehensively. When users are purchased for $ 700 thousand, it makes no sense to analyze each campaign in parts. It is difficult to micromanage such a large amount, and turning to predicates will save resources. In a couple of clicks, you can evaluate the entire purchase at once.

***

As in any other area of analytics, predictive analytics has its own difficulties — data quality and sample size. It can be difficult to collect high-quality data and isolate the most relevant ones to build a predicate. But at the same time, predictive analytics provides easy scaling of traffic and gives an understanding of how effectively we buy it — right now, here and now.

With the help of predicates, we save money, time, nerves and reduce the chances of error. Therefore, if you are developing a mobile game and still do not use predicates, we advise you to think hard about getting started.