Using Predictions
Understand the time windows, probabilities, and tiers for purchase and churn predictions.
Context API predictions estimate how likely a user is to purchase or churn. The response includes a probability and tier for each model under predictions. This guide explains each model's predicted event and time window, and how to interpret probabilities and tiers.
Prediction targets and time windows
The purchase and churn models predict different events over different time windows, both measured from app installation.
App installation0
- 1
- 2
- 3
Days30
predictions.d3_purchase
Predicts the likelihood of the standard order-completed event, airbridge.ecommerce.order.completed, occurring within three days after app installation.
predictions.d3_churn
Predicts the likelihood of no activity events being observed after day three through day 30 after app installation.
The timeline after day 3 is compressed.
The models predict different events, so both tiers can be high.
Probabilities and tiers
Each model returns a probability and a tier.
| Result field | How to read it |
|---|---|
probability | Represents the model's predicted probability as a number between 0 and 1. A value of 0.723 means 72.3%. |
tier | Groups results using the distribution of predicted probabilities for that model's training data. The value is low, medium, or high. |
A purchase tier of high means a high predicted likelihood of purchase. A churn tier of high means a high predicted likelihood of churn. Each model has its own tier criteria, so the same probability can fall into different tiers in different models. There is no fixed probability threshold that applies to all models.
The response rounds probabilities to three decimal places. Each model determines tiers according to its own criteria. If your rule uses a tier, use the returned tier. Recalculating it from the rounded probability can produce a different tier from the one the model assigned.
Predictions and personalization effects
A predicted probability estimates how likely an event is to occur. It does not measure an increase in purchase probability caused by a message or additional results caused by personalization. To measure how much personalization improves results, run an experiment with defined comparison conditions.
Unavailable predictions
Treat unavailable predictions as a separate state from probability 0 or tier low. If an API error prevents you from receiving predictions, follow Error Handling and apply the default behavior you defined.
See Using Context to design your rules, or Context API for the full response and field definitions.