Predictive analytics of MyTracker is advanced analytics tools
that will expand your knowledge about the site or app audience, enriching gathered information with "data from the future".
Forecasts of various metrics, scoring, event correlation, user quality assessment, the event-important prediction of financial metrics — these are all predictive analytics tools.
Prediction enables you to take a more informed decision on app promotion without having to wait for actual data to accumulate. For example, you can remove inefficient channels before they become a real problem, or rebalance your budget to reduce costs and improve user acquisition.
How it works
Predictive analytics is based on a large amount of data collected by MyTracker.
A combination of several groups of continually evolving and updated
predictive models works to give you the most accurate forecast possible.
Below are the common stages of predictive analysis in MyTracker:
Gathering the data. The embedded SDK collects data on your site or app. The more data it has,
the more accurate the forecast.
Learning. All users are divided into cohorts based on shared characteristics.
The models analyze historical data for each cohort and calculate ratios for future forecasts.
The learning time depends on the model group.
Using the models. Once a model’s learning process completes,
MyTracker starts making its forecast.
Every day all incoming traffic is divided into tiny cohorts,
with each to be given its own forecast in MyTracker.
During the first several days after users first visit
the site or install the app, forecasts are updated every 24 hours based on newly received data.
On the eighth or ninth day (depending on prediction metrics), the results are fixed for each cohort, and the final, seven or eighth forecast is made.
You can see the forecast the day after users first visit the site or install the app.
With each passing day, the accuracy of these forecasts will increase.
The first several days provide a solid enough foundation to make a strategic decision on things like buying traffic.
This is why we update our predictions during the first seven or eight days and give the final forecast on the eighth or ninth day (depending on prediction metrics).
Prediction assess.
Predictions should be compared with facts. You can do this when the amount of tracked data goes beyond the forecast horizon (for example, on August 23, you can compare "LTV 6m" and "LTV Prediction 6m" for the cohort that installed the app on February 22). You’ll see that the decisions made based on the forecasts paid off, so it’s safe to continue using them. If a forecast considerably deviates from the facts, please contact our support team. We will carefully review each case and improve the predictive models.
Predictive models
LTV Prediction
MyTracker LTV forecasts are based on a combination of predictive models
that can be distilled into three groups:
Models that make decent predictions as early as the day after users first visit the site or install the app.
Although they do not need much data, a forecast will be more accurate if there is some site or app history.
Models that make good predictions only based on a large amount of current data.
Models that give an accurate forecast using a large amount of historical data.
Any prediction is based on a combination of these three groups.
This mixed model makes a prediction as early as the day after users first visit the site or install the app,
and it can work with little data and provide an accurate
forecast based on a history of just 30 days.
SKAN LTV forecasts are based on a combination of the predictive model trained on SKAN postback data and the three groups of models for LTV Prediction listed above.
The prediction maintains user privacy, does not identify devices,
and provides data only at the level of partners and ad campaigns.
