Churn Management – Hayyak+

Problem & Business Impact

The Hayyak+ prepaid plan represents a high-value segment for telecom operators. Failure to detect customers at risk of non-renewal results in significant loss of recurring revenue.

Measured impact:

• +8% overall retention on Hayyak+ segment
• 25% reduction in intervention time
• 3:1 ROI on targeted retention campaigns

Data & Key Features

Data sources

• Recharge history: amount, frequency, recency
• Voice/data usage: volume, trends, temporal patterns
• Account balance: current level, evolution
• Network behavior: location, roaming, technical data
• Demographics: tenure, segment, region

Engineered features

• Recharge trend (declining, stable, growing)
• Usage/recharge ratio (satisfaction indicator)
• Regularity score (predictable behavior)
• Credit depletion proximity
• Multi-channel engagement

Models & Methods

Multi-model approach

1. XGBoost: Main model for churn prediction

   • Accuracy: 87%
   • Recall: 82%
   • AUC: 0.91

2. LightGBM: Complementary model for speed

   • Inference time: <50ms
   • Accuracy: 85%

3. Random Forest: Ensemble for robustness

   • Used for cross-validation
   • Feature importance identification

Prediction windows

• 0-7 days: Emergency intervention
• 7-14 days: Personalized offer
• 14-21 days: Preventive communication

Real-time Integration

Decisioning architecture

Kafka Stream → Online Feature Store → Model API → CRM/Campaign
     ↓
Event triggers (critical thresholds)
     ↓
Automated campaign orchestration

Activation modes

• Pull API: On-demand scoring via CRM
• Push events: Automatic alerts on thresholds
• Batch: Daily score refresh

KPIs & Performance Targets

| Metric | Target | Current | |--------|--------|---------| | Overall retention | +8% | +8.2% | | Model accuracy | >85% | 87% | | Inference time | <100ms | <50ms | | Campaign coverage | >70% | 76% | | Uplift vs. control | >15% | 18% |

Deployment & Monitoring

MLOps Pipeline

1. Training: Quarterly retraining or triggered by drift >5%
2. Validation: Systematic A/B testing (70/30 split)
3. Monitoring:
   • Drift detection (data + predictions)
   • Performance metrics dashboard
   • Automatic alerts

Drift management

• Continuous monitoring of feature distributions
• Alerts if Jensen-Shannon divergence >0.15
• Automatic retraining if drift confirmed

FAQ & Prerequisites

Q: What minimum data is required? A: 6 months of recharge and usage history, basic demographics.

Q: Time to production? A: 3-4 months (Foundation phase) including data pipeline and CRM integration.

Q: Multi-vendor compatible? A: Yes, tested with Ericsson, Huawei, ZTE, Nokia.