Intelligence Models

How the intelligence models work

An overview of LiftSignals' 12 predictive models — how they train, how accurate they are, and when to trust them.

What are intelligence models?

LiftSignals trains 12 machine learning models on your specific store's order and customer data. Unlike generic industry benchmarks, these models learn your store's unique patterns — your typical purchase cadence, seasonal rhythms, and customer behaviour.

The 12 models at a glance

ModelWhat it predictsUpdates
Churn DetectionProbability a customer will not return (0–100 score)Every 2 hours
CLV ForecastPredicted revenue from a customer over the next 12 monthsDaily
Win-back TimingThe optimal day to contact a lapsed customerEvery 2 hours
RFM SegmentationCustomer segment based on Recency, Frequency, Monetary valueDaily
Propensity to BuyLikelihood of purchasing in the next 30 daysDaily
Repurchase PredictionProbability of a customer buying again based on their historyDaily
Customer Health ScoreOverall engagement health (0–100)Daily
Cohort RetentionRetention curves by acquisition cohortWeekly
Channel AttributionWhich acquisition channels produce the best long-term customersWeekly
LTV MaximizerActions most likely to increase a customer's lifetime valueDaily
Spend OptimizerWhere to focus retention spend for maximum ROIWeekly
AI RecommendationsSynthesises all 11 models into prioritised daily actionsEvery 2 hours

How models train

On first connection, LiftSignals trains all models on your complete historical data. After that, models retrain automatically when:

  • A new sync completes and adds significant new data
  • LiftSignals detects a meaningful shift in your store's patterns
  • You manually trigger retraining from Model Studio

Model accuracy

Accuracy is displayed in Explore → [Model Name] for each model. The Churn Detection model typically achieves 87–94% precision across our customer base. Accuracy is higher with:

  • More order history (12+ months gives the best results)
  • More complete data (high email fill rate, complete order dates)
  • Consistent purchase patterns (models struggle with very erratic buying behaviour)
⚠️
New stores: Models need at least 500 customer records and 90 days of order history to produce reliable scores. With less data, scores are provisional — treat them as directional guidance, not hard predictions.

Model Studio

Advanced users can adjust model thresholds, run scenarios, and compare model versions in Model Studio. This is covered in detail in the Model Studio documentation.

Was this article helpful?