Predict tells you how likely your users are to convert before they do. Grades are given to users based on their previous behavior. All you need to do is select an action and Predict will tell you how likely it is that they will complete it.
Additionally, Predict goes a step beyond just a grade and will tell you just how likely it is your users will convert. For each letter grade, the lift at any given time indicates how much more (or less) likely a customer with that grade will convert vs. the average customer (ie. a “A” customer with a 40% chance of conversion will have a lift of 4 if the average customer has a 10% chance of conversion).
By hovering over the you can see how much more likely than the average user these users are to convert.
As we keep updating the Predict model the lift we give you will change since there will be new data.
Building a Predict Model
Building a Predict model takes just a few steps. Select your conversion event goal and then decide if you want to see users who will do this event for the first time or users who will convert but have also done it before. Once the model is built, Predict will send you an email of the results.
By default, the "for the first time ever" box will not be checked. This means that Predict will check for any people who will do this action at any time. If you do check the box, the model will try to predict users who will do the action for the first time.
Imagine you have a music streaming app that allows users to also purchase songs, you probably want to know who is going to buy a song who has not purchased one before. Or who isn’t likely to buy one, so you can nudge users. Predict is perfect for improving leading metrics, like a user buying a song for the first time. In this example, you want to know which users are buying a song for the first time, so you will want to make sure you have checked "for the first time ever," as shown in the above image.
If we cannot accurately calculate the Predict model, Mixpanel will send an email explaining the error and why the model was rejected. Why would a Predict model not be accepted?
Predict Model Degradation
While rare, it is possible for successful predict models to degrade to the point of no longer being predictive. This will happen if customer behavior drastically deviates from the behavior used in the existing model or if data quality degrades.
If this happens, Mixpanel will automatically turn off the model, send you an email notification alerting you of the change, and turn off any existing notification campaign that use the Predict grades.