Subscription Addons Prepaid

Problem & Business Impact

The prepaid segment often represents 50-70% of telecom operator customer base, but generates significantly lower ARPU than postpaid customers (typical ratio 1:3). The monetization opportunity via addons (data packs, international minutes, roaming options) is under-exploited, with average adoption rates of only 15-20%. Most prepaid customers recharge only the minimum necessary, missing value offers that could improve their experience while increasing revenue.

The challenge is to identify the right time, the right addon, and the right customer to maximize conversion. An untargeted approach suffers from marketing fatigue and low conversion rates (<2%). Our ML system analyzes consumption patterns, credit depletion moments, and usage behaviors to propose relevant addons at the optimal time.

Measured impact:

• 42% increase in addon adoption rate
• 18% increase in average prepaid ARPU (€8.50 → €10.03)
• Generation of €3.2M in additional annual revenue
• 650% ROI on ML-targeted campaigns

Data & Key Features

Main data sources

• Recharge history: Amounts, frequency, channel (USSD, app, store)
• Detailed consumption: Data usage, national/international voice, SMS
• Critical events: Credit depletions, service interruptions, notifications
• Digital engagement: Mobile app usage, self-service, offer consultation
• Demographic profile: Age, location, socio-economic segment

Engineered features

• Recharge frequency and average amount (90 days)
• Credit depletion pattern (predictability, regularity)
• Data vs voice vs SMS consumption ratio
• Distance to typical addon user profile (clustering)
• Affinity score by addon type (data, voice, international, roaming)
• Optimal proposal timing (post-recharge, pre-depletion)
• Estimated price sensitivity (offer response history)
• Potential for postpaid upgrade

Models & Methods

Personalized recommendation approach

1. Random Forest (Propensity scoring): Subscription probability by addon

   • Average AUC-ROC: 0.76 (varies by addon: 0.72-0.81)
   • Top 15% precision: 38%
   • Key features: Consumption pattern (28%), recharge frequency (22%), app engagement (18%)
   • Specialized model by addon type (data, voice, international)

2. Collaborative Filtering: Similarity-based recommendations

   • "Customers like you also subscribed to..."
   • Detection of sequential adoption patterns (addon A → addon B)
   • Enables bundle opportunity identification

3. Association Rules (Apriori/FP-Growth): Purchase pattern discovery

   • Identification of addon combinations frequently subscribed together
   • Bundle and package optimization
   • Confidence >60% for active recommendations

Prepaid profile segmentation

• Data-hungry: High data consumption, frequent depletions (32%)
• International: Regular international calls (18%)
• Basic: Minimal usage, low and irregular recharges (35%)
• Occasional: Variable usage, promotion-sensitive (15%)

Real-time Integration

Contextual recommendation pipeline

Trigger event (recharge, depletion) → Real-time scoring (``<100ms``)
    ↓
Optimal addon selection → Offer personalization
    ↓
In-app/USSD/SMS display → One-click subscription
    ↓
Conversion tracking → Model feedback

Optimal contact moments

• Immediate post-recharge: Addon proposal during recharge flow
• Pre-depletion: Proactive alert 24h before predicted depletion + addon
• Expiry end: Existing addon renewal with upsell
• Special events: International travel, weekends, local events

Message personalization

• Highlight of profile-specific benefit (e.g., "Your usual 10 GB for €5")
• Comparison with current consumption ("Save €3 vs your separate recharges")
• Urgency and scarcity ("Offer valid today only")
• Social proof ("Chosen by 15,000 customers like you")

KPIs & Performance Targets

| Metric | Target | Current | |--------|--------|---------| | Addon adoption rate | >22% | 15.2% → 21.6% | | Average prepaid ARPU | >€10 | €8.50 → €10.03 | | Campaign conversion rate | >8% | 3.2% → 9.4% | | Monthly addon revenue | >€250K | €267K | | Model precision (top 15%) | >35% | 38% | | Addon renewal rate | >55% | 58% |

Deployment & Monitoring

Recommendation architecture

1. Real-time feature store: Consumption feature aggregation (streaming)
2. Scoring API: ML endpoint for on-demand scoring (<100ms latency)
3. Rules engine: Final addon selection + message personalization
4. Channel integration: USSD menu, mobile app, SMS, stores (POS terminal)
5. Attribution tracking: Conversion and revenue measurement by channel and segment

Performance monitoring

• Real-time dashboard: Proposals, conversions, revenue by addon
• Cannibalization analysis (addons vs traditional recharges)
• Addon renewal rate tracking (recurrence)
• Continuous A/B testing of messages and timings
• Weekly retraining with new conversions

Offer optimization

• Price elasticity analysis by segment
• New addon testing (validation before mass launch)
• Bundle optimization based on association rules
• Promotion calibration (optimal discount to maximize revenue)

FAQ & Prerequisites

Q: How to avoid marketing fatigue on the prepaid segment? A: Frequency limitation (max 2 proposals/week), high relevance scoring (top 15-20% only), strict personalization (addon aligned with actual usage). The model integrates a "message fatigue" score.

Q: What's the difference from generic promotions? A: Classic promotions target broadly with low conversion (2-3%). ML specifically targets customers with high propensity (30-40% in top decile) and personalizes the addon to actual need.

Q: How to handle low-income customers? A: Specific segmentation with reduced-price addons but high volume. The goal is to increase engagement and loyalty, even with lower margins. These customers can also be targets for eventual postpaid migration.

Q: Can we propose multiple addons simultaneously? A: Yes, but generally we recommend presenting max 1-2 addons to avoid choice paradox. The model ranks addons by propensity and presents the top 1-2.

Q: What are the technical prerequisites? A: Data warehouse with consumption and recharge history (6+ months), real-time integration with recharge systems (USSD, app), performant ML scoring API (<100ms), A/B testing capability on digital channels.