Predictions and Predictive Audiences
This domain tests mParticle's machine-learning scoring: how a prediction is defined, what data it needs, and how a score becomes a predictive audience.
How a prediction is defined
You select an outcome event that represents the behavior to predict, a training window of historical data, and a prediction window describing how far ahead you are scoring. mParticle trains a model and emits a score per user, typically on a 0-100 scale, stored on the profile. Models retrain on a schedule.
Rules
- The outcome must be instrumented and consistently named; the model learns exactly what you send.
- Training needs enough examples of both outcomes (did and did not). A short or one-sided window produces a weak model.
- The prediction window defines the question: will this user perform the outcome in the next N days?
- A predictive audience is a normal audience whose criteria include the prediction score, for example score at least 70.
- New users with no history are cold-start cases; their score is unreliable.
Worked example
A subscription app wants fewer cancellations. It instruments a Subscription Cancelled event, trains on 180 days of history with a 30-day prediction window, and scores active users. A predictive audience "High churn risk" uses score at least 70 and feeds a win-back journey. Because the outcome is cancellation, a high score means risk, not loyalty.
Traps
- Predicting an event that is not in the data plan.
- Expecting scores immediately after creating the model; training takes time.
- Treating the score as a calibrated probability without checking.
- Building the audience before the model has produced scores.