Prediction

Beta

# Introduction

Prediction shows how some action events can “predict” the probability which users will move from a source cohort to a target cohort. This statistics-heavy correlation analysis feature masterfully identifies key user actions in the app.

# Requirements

• The Prediction feature is only available to Kubit Enterprise customers.
• Extensive usage of Cohorts to understand user behavior.

# Create a Prediction

• Select a Source Cohort and a Target Cohort • Optionally:
• Specify filter conditions
• Select Action Events and Time Window (within X days) to be considered
• Choose Date Range # Execute

Default view shows All Events’ prediction. For each event, you can see for each N days after the action was taken, what’s the “Correlation” value which predicts users moving from Source Cohort to Target Cohort.

The frequency (the number of times) an action was taken matters a lot in prediction. Thus only the frequency of highest predictive value is displayed in each cell where the “n in (>n)” indicates that frequency. ## Detailed View

Details of each cell: a correlation table with detailed statistics and the different predictive value for each frequency of the Action Event.  ### Predictive Level

• High: Correlation Value ≥ 0.4
• Moderate: 0.4 > Correlation Value ≥ 0.3
• Slight: 0.3 > Correlation Value ≥ 0.2
• Not Predictive: -0.2 ≤ Correlation Value < 0.2
• Slight Negative: -0.3 ≤ Correlation Value < -0.2
• Moderate Negative: -0.4 ≤ Correlation Value < -0.3
• High Negative: Correlation Value < -0.4

### Correlation Value

Pearson correlation is the geometric mean of how predictive X is of Y and how predictive Y is of X. In this case, X is performing at least the threshold number of events and Y is being retained.

### Correlation Table

This table displays the following raw user counts, starting clockwise from the upper left: True Positive, False Positive, True Negative, False Negative.

### Detailed Statistics

The detailed statistics are a series of ratios that can be helpful in interpreting the correlation score. They can all be generated directly from the user counts in the correlation table.

#### Correlation

Pearson correlation is the geometric mean of how predictive X is of Y and how predictive Y is of X. In this case, X is performing at least the threshold number of events and Y is being retained.

#### Positive Predictive Value

Proportion of users performing at least the threshold number of events who were also retained.

#### Negative Predictive Value

Proportion of users performing less than the threshold number of events who were not retained.

#### Sensitivity

Proportion of retained users who performed at least the threshold number of events.

#### Specificity

Proportion of not retained users who performed less than the threshold number of events.

#### Proportion Above Threshold

Proportion of users performing at least the threshold number of events.

## Statistics Filter

You can toggle to see specific statistics on the detailed view.  