15

Propensity and upsell scoring

Ranking 8,000 telecom customers by their likelihood to respond to upsell campaigns, enabling marketing teams to focus on the most promising prospects.

Random Forest Calibration Propensity scoring Decile analysis Telecom Marketing
2,155%
Targeted campaign ROI

Interactive dashboard

Four-page Streamlit application for propensity analysis and customer scoring

Overview
  • Dataset summary and response rate baseline
  • Feature distributions by response class
  • Model comparison metrics table
Segmentation
  • Customer segments by propensity decile
  • Revenue per tenure and usage intensity
  • Campaign targeting recommendations
Model and lift chart
  • Calibration curves for all three models
  • Cumulative lift chart vs. random baseline
  • Feature importance ranking
Scorer
  • Score individual customers in real time
  • Decile assignment with response probability
  • Campaign ROI: targeted vs. mass comparison
$ pip install -r requirements.txt && streamlit run app.py

Key results

Random Forest with isotonic calibration selected as the best model across three candidates

0.775
AUC-ROC
37.7%
Top decile response rate
63.2%
Top 3 deciles capture
2,155%
Targeted campaign ROI

Methodology

Built on 8,000 synthetic customer records with demographics, usage patterns, service subscriptions, and campaign response history at a 12% baseline response rate. Engineered features include revenue per tenure, usage intensity composite, service count, upsell headroom, and interaction terms. Three classifiers were trained with class balancing, then calibrated via isotonic regression to produce reliable probability estimates. Decile analysis ranks customers for targeted outreach.

Feature engineering
Revenue ratios, interactions, headroom
Model training
LR, RF, XGBoost with balancing
Calibration
Isotonic regression via CalibratedClassifierCV
Decile scoring
Campaign ROI, targeted vs. mass

Links