Predictive analytics

Predictive analytics of myTracker is advanced analytics tools that expand your knowledge about app audience, enriching facts with «data from the future».

Forecast of various metrics, scorings, event correlation, assessment of user quality, etc. — all of that is predictive analytics tools. One of the important is the prediction of financial metrics.

With prediction, you can take a more informed decision on app promotion and not wait for actual data to be accumulated. For example, you can remove ineffective channels in time and rebalance the budget, resulting in reduction of expenses and improving of user acquisition.

How it works

Predictive analytics is based on a large amount of data collected by myTracker. Forecast accuracy is ensured by several groups of predictive models continually trained and updated.

There are the following stages of myTracker predictive analytics:

  1. Data gathering. Embedded SDK collectsdata on your app. The forecast is more accurate when the data amount is large.
  2. Models training. All users grouped by shared characteristics are divided into cohorts. Models analyze historical data on cohorts and calculates coefficients for future forecasts. The training period depends on the model group.
  3. Model using. When a model has trained, myTracker starts to build the forecast. Everyday incoming traffic is divide into tiny cohorts, for which myTracker builds a forecast. In the first eight days after the app installation, every 24 hours forecasts are refined and rewritten with new data. On the ninth day results are fixed for each cohort it is the final eighth prediction. So you can see the forecast the day after users installed your app, and get more accurate data with each new day.
  4. The first eight days are indicative to make a strategic decision like buying traffic. So we update prediction in the first eight days and give the final forecast on the 9th day.

  5. Prediction assess. It is important to compare predictions with facts. You can do this when the amount of tracked data exceeded the forecast horizon (for example, on the 23rd of August, you can compare «LTV180d» and «Prediction LTV180» for cohort installed the app on the 22nd February). Results will convince you to keep using predictions in the work. If facts have a seriously deviated from prediction, contact our support team  — we will carefully study the instance and improve prediction models.

Predictive models

myTracker forecasts are based on several groups of predictive models:

  • First — models that show not bad prediction just the day after users installed your app. This models predict on little data. The first group gives more accurate forecast with some app history.
  • Second group — models that show good prediction only on a large amount of current data.
  • Third group — models that show accurate forecast on a large amount of historical data.

Any prediction is a combination of three models group. This mix-model gives prediction the next day after app installation, can work on little data and give an accurate forecast on 30 days of history.