AI & prediction
How a contact-rate prediction engine actually learns
No industry benchmarks, no borrowed averages. A useful model is trained on your outcomes and retrained as your list changes.
Vapia Research · Prediction team · April 30, 2026 · 8 min read
Prediction in outbound is a supervised problem with a generous supply of labels: every dial produces an outcome. The engineering challenge isn't the model, it's making the labels trustworthy and the features honest.
Features that carry signal
- Day of week and hour, in consumer-local time
- Area code and regional behaviour clusters
- Historical behaviour of similar contacts
- Recent ANI reputation state
Labels you can trust
This is where normalized dispositions pay for themselves. A model trained on inconsistent outcome labels learns your tagging habits rather than consumer behaviour.
Retraining cadence
List mix changes weekly on most floors. A model retrained monthly is describing a population that no longer exists. Weekly retraining on rolling windows keeps the recommendations aligned with what you're actually dialing.
A prediction that can't state its expected lift and its confidence isn't a recommendation, it's a guess with better typography.