Post-to-Pre Migration

Problem & Business Impact

The migration of postpaid customers to prepaid represents a significant loss of revenue and long-term customer value. These customers generate on average 3 to 4 times more revenue than prepaid customers and have a significantly higher lifetime value (LTV). Early warning signals of this migration include progressive consumption decline, recurring payment issues, and changes in usage behavior.

Without proactive intervention, approximately 8-12% of postpaid customers migrate annually to prepaid, leading to substantial base value erosion. Our early detection system identifies at-risk customers 30 to 90 days before actual migration, providing an optimal intervention window to propose adapted plans, resolve friction points, and preserve the contractual relationship.

Measured impact:

• 35% reduction in post-to-pre migration rate
• Preservation of €4.2M in annual recurring revenue
• 28% increase in retention offer success rate
• 520% ROI on targeted retention campaigns

Data & Key Features

Main data sources

• Consumption data: Voice, data, SMS traffic evolution over 3-6 months
• Billing history: Billed amounts, payment delays, incidents
• Customer support: Contact frequency, complaint types, satisfaction
• Demographic profile: Customer segment, tenure, income, location
• Product behavior: Service usage, option activation, plan evolution

Engineered features

• Consumption trend (linear regression over 90 days)
• Actual consumption vs subscribed plan ratio
• Payment stability score (12-month history)
• Complaint rate weighted by severity
• Distance to typical prepaid profile (clustering)
• Monthly bill volatility
• Product engagement score (feature activation, mobile app)
• Comparison with peers in same segment

Models & Methods

Three-tier cascade approach

1. XGBoost Classifier (Main model): Migration risk prediction

   • AUC-ROC: 0.87
   • Top 10% precision: 68%
   • Key features: Consumption trend (-24%), plan ratio (19%), payment history (18%)
   • Alert threshold: Score >0.65 for proactive intervention

2. Random Forest (Timing model): Migration window estimation

   • Time horizon prediction (30/60/90 days)
   • RMSE: 12 days
   • Enables prioritization of urgent interventions

3. Logistic Regression (Segmentation model): Classification by migration reason

   • Segmentation: Economic (52%), Dissatisfaction (28%), Usage change (20%)
   • Enables personalized retention offers

At-risk profile segmentation

• Budget reduction: Income decrease, cost optimization (45%)
• Chronic under-utilization: Consumption <50% plan for 3+ months (32%)
• Service issues: Recurring technical complaints (15%)
• Life transition: Professional/personal situation change (8%)

Real-time Integration

Daily scoring pipeline

Customer data (D-1) → Feature engineering → ML scoring
    ↓
Segmentation by reason → Prioritization (score × timing)
    ↓
CRM/Campaign Manager → Personalized intervention
    ↓
Results tracking → Model retraining (weekly)

Multi-channel activation

• Score 0.65-0.75 (Moderate risk): Personalized email + optimized plan offer
• Score 0.75-0.85 (High risk): Proactive advisor call + bill audit
• Score >0.85 (Critical risk): Manager intervention + premium offer

Offer personalization

• Intermediate plans between postpaid and prepaid (hybrid)
• Plan adjustment to actual consumption
• Targeted temporary discounts (3-6 months)
• Proactive technical issue resolution

KPIs & Performance Targets

| Metric | Target | Current | |--------|--------|---------| | Post-to-pre migration rate | <6% annual | 7.8% → 5.1% | | Top 10% precision (PPV) | >65% | 68% | | Offer conversion rate | >25% | 31% | | Revenue preserved | €4M/year | €4.2M | | Average detection lead time | <45 days | 38 days | | Model AUC-ROC | >0.85 | 0.87 |

Deployment & Monitoring

Production pipeline

1. Data extraction: Daily feature aggregation (last 3 months)
2. Batch scoring: Score calculation for 100% of postpaid base
3. Segmentation: Classification by reason and urgency
4. CRM routing: Export to campaign tools (Salesforce, Adobe Campaign)
5. Tracking: Action and outcome monitoring (conversion, actual migration)

Quality and drift monitoring

• Daily score distribution monitoring
• Alerts on >15% deviations vs baseline
• Continuous A/B testing of intervention strategies
• Weekly retraining with updated data
• Monthly performance validation (hold-out test set)

A/B Testing

• Test group: At-risk customers with ML-guided intervention (30%)
• Control group: At-risk customers with standard process (20%)
• Holdout: No intervention for calibration (10%)
• Metrics: Migration rate, revenue preserved, customer satisfaction

FAQ & Prerequisites

Q: What is the optimal intervention window? A: Between 30 and 60 days before predicted migration. Earlier, the customer hasn't made their decision yet. Later, options are limited and intervention cost increases.

Q: How to avoid costly false positives? A: Use of a high score threshold (>0.65), validation by the timing model, and intervention personalization by segment. Offers are calibrated to remain profitable even with 30% false positives.

Q: Can we identify customers who have already decided to migrate? A: Yes, via the timing model. Customers with score >0.85 and horizon <30 days require immediate premium intervention (manager, exceptional offers).

Q: What are the data prerequisites? A: Consumption history (6 months minimum), billing and payment data (12 months), customer support interactions, demographic and contractual profile.

Q: How to measure ROI? A: (Revenue preserved - Retention offer cost - Operational cost) / Total cost. Our average ROI is 520%, with an intervention cost of €15-25 per customer and average preserved revenue of €180/year.