← Back to all Predictive Models
PREDICTIVEThis model is designed to predict the probability that a customer will churn.
Retention Campaigns
Deploy win-back strategies before churn occurs
Target Group Marketing
Enable strategic customer targeting to:
- Allocate promotional budgets optimally based on risk levels
- Take proactive measures to combat churn in real-time
- Customize retention strategies by churn probability tiers
- Population Lifecycle Stages: Customers in all "live" lifecycle stages (New, Active, Reactivated)
- Training: Runs every iteration (unless data drift/changes require retraining)
- Inference: Runs daily in the daily process
The following outputs are stored in the customer profiles (as internal fields):
- Churn Probability Score: Probability of this customer to migrate into Churn within the next 2 iterations
-
Percentile rank in Churn Probability:
Percentile rank among all live customers by Churn probability.
On a scale of 1 (lowest) to 100 (highest) -
Permille rank in Churn Probability:
Permille rank among all live customers by Churn probability.
On a scale of 1 (lowest) to 1000 (highest), providing higher-resolution segmentation than the percentile ranking -
Percentile rank in LCS Churn Probability:
Percentile rank among same lifecycle stage customers by Churn probability.
On a scale of 1 (lowest) to 100 (highest) -
Permille rank in LCS Churn Probability:
Permille rank among same lifecycle stage customers by Churn probability.
On a scale of 1 (lowest) to 1000 (highest), providing higher-resolution segmentation than the percentile ranking
- Training Window: default = 20 iterations
- Prediction Horizon: default = 2 iterations
- Catboost classifier which is the state-of-the-art model for tabular data
- Individual models per lifecycle stage for optimal accuracy
- Hyperparameters tuning in the first model training for optimal model parameters
Data Cleaning & Imputation: Handles missing values, removes constant features, and groups rare categorical values
Feature Engineering & Encoding: Converts categorical variables to numerical format using target-based encoding and applies scaling transformations
Feature Selection & Optimization: Removes low-variance and highly correlated features, selecting the most predictive variables
- Historical customer profiles (customer activity data, demographic data, campaign history, etc.)
- Minimum 3 iterations of historical data required for training (1 iteration for feature generation and 2 iterations for target definition, given the 2-iteration prediction horizon)
- Minimum number of 500 customers per lifecycle stage required for training
- Minimum number of 50 customers per lifecycle stage that has been churned required for training