HBB Cancellation (Home Broadband)

Problem & Business Impact

Home Broadband (HBB) subscription cancellations represent a critical challenge for telecom operators, with particularly high customer acquisition cost (CAC) in this segment (€150-300) and typical contract duration of 12-24 months. HBB churn is often linked to service quality issues (speed, stability), aggressive competition, or life changes (relocation, bundle consolidation).

The average annual churn rate in the HBB sector ranges between 15-20%, with peaks during contract end periods. Each cancellation results not only in direct revenue loss (average ARPU €35-45/month), but also potential loss of additional services (TV, fixed telephony) and cross-sell opportunities. Our prediction system identifies at-risk customers 30-60 days before actual cancellation, maximizing retention action effectiveness.

Measured impact:

• 22% reduction in HBB churn rate
• Preservation of €6.8M in annual revenue
• 34% increase in retention campaign success rate
• 480% ROI on proactive interventions (network upgrade, offers)

Data & Key Features

Main data sources

• Network quality: Actual speed, latency, failure rate, connection stability
• Technical support: Open tickets, complaints, resolution time, satisfaction
• Service usage: Data volume consumed, usage patterns, TV/VOD engagement
• Contract data: Tenure, offer type, contract end, upgrade history
• Competitive environment: Competitive offers in area, fiber/5G coverage

Engineered features

• Composite network quality score (speed × stability × availability)
• Promised vs actual speed gap (over 30 days)
• Technical incident frequency and severity
• Engagement trend (consumption evolution over 90 days)
• Distance to contract commitment end
• Price/performance ratio vs local market
• Customer satisfaction score (NPS, CSAT on support)
• Additional services usage rate

Models & Methods

Multi-model approach for risk segmentation

1. Gradient Boosting (LightGBM): Global churn risk prediction

   • AUC-ROC: 0.84
   • Top 15% precision: 62%
   • Key features: Network quality (31%), support tickets (22%), contract end (18%)
   • Segmentation: Technical risk vs competitive risk vs contract end

2. Deep Learning (LSTM): Temporal analysis of degradation patterns

   • Detection of pre-churn behavioral sequences
   • Identification of degradation trajectories (rapid vs progressive)
   • Time window accuracy: ±8 days

3. Survival Analysis (Cox Model): Time-to-cancellation estimation

   • Enables intervention prioritization by urgency
   • Calculation of survival probability at 30/60/90 days
   • Integration of temporal covariates (quality evolution)

Churn cause segmentation

• Service quality: Recurring technical issues (38%)
• Price competitiveness: More attractive competitive offers (28%)
• Contract end: End of contract period without renewal (24%)
• Relocation: Change of residence (10%)

Real-time Integration

Continuous monitoring pipeline

Network data (real-time) → Hourly aggregation → Feature store
    ↓
Daily scoring (batch) → Anomaly detection (streaming)
    ↓
Segmentation by cause → Prioritization by urgency
    ↓
Action routing (CRM/NOC) → Personalized intervention
    ↓
Feedback loop → Monthly retraining

Differentiated activation by segment

• High technical risk: Proactive NOC intervention, equipment upgrade
• Price sensitivity: Targeted commercial offer, plan upgrade
• Contract end: Anticipated renewal contact, loyalty bonus
• Relocation: Transfer facilitation, multi-site offer

Proactive technical interventions

• Automatic line optimization (DLM, vectoring)
• Network profile change (priority, QoS)
• Defective CPE equipment replacement
• Migration to superior technology (ADSL→VDSL→Fiber)

KPIs & Performance Targets

| Metric | Target | Current | |--------|--------|---------| | Annual HBB churn rate | <14% | 18.2% → 14.1% | | Top 15% precision (PPV) | >60% | 62% | | Save rate (retention) | >30% | 36% | | ARPU preserved | €6M/year | €6.8M | | Average detection lead time | <40 days | 35 days | | At-risk customer NPS | >+20 pts | +24 pts |

Deployment & Monitoring

Production pipeline

1. Data collection: Multi-source aggregation (network, CRM, support) - real-time + batch
2. Feature engineering: Feature calculation on sliding windows (7/30/90 days)
3. Daily scoring: Risk + timing + cause prediction for 100% HBB base
4. Segmentation & prioritization: Classification by urgency and intervention type
5. Intelligent routing: NOC (technical), Commercial (offers), Support (customer contact)
6. Outcome tracking: Actual cancellation monitoring, offer conversion, NPS improvement

Network quality monitoring

• Real-time dashboard of quality indicators per customer
• Automatic alerts on >20% degradation vs baseline
• Correlation network incidents ↔ churn score increase
• Network investment prioritization by area (anti-churn ROI)

A/B Testing

• Test group: At-risk customers with ML-guided intervention (40%)
• Control group: At-risk customers with standard process (20%)
• Holdout: Observation without intervention (10%)
• Metrics: Churn rate, NPS, ARPU, intervention cost

FAQ & Prerequisites

Q: How to differentiate a dissatisfied customer from a relocating one? A: Cross-analysis: technical issues + support tickets = dissatisfaction. No incidents + offer search in new area = probable relocation. The model integrates geographic and search data (web/app).

Q: What is the cost of a proactive intervention? A: Variable by type: commercial upgrade (€10-20/month discount for 6 months) = €60-120, technical intervention (CPE, optimization) = €40-80. Average new HBB customer acquisition cost is €200-300, so ROI is favorable.

Q: Can we act on detected network issues? A: Yes, many automated or semi-automated actions: DLM adjustment, profile change, traffic re-routing. For complex cases, escalation to NOC with ML context (high priority).

Q: How to handle contract commitment ends? A: Proactive contact 60 days before deadline, personalized renewal offers, loyalty bonus. Renewal rate increases 40% with anticipated contact vs reactive contact.

Q: What are the data prerequisites? A: Real-time network data (SNMP, syslog), support ticket history, contract data, usage metrics (DPI for aggregated traffic). Network data quality is critical for model accuracy.