Integration

Guideline for MAW Product

1. Introduction

This document provides a standardized framework to evaluate the effectiveness of Smile’s MAW product during backtesting and live POC phases, ensuring alignment between:

  • Blacklist identification logic
  • Risk segmentation performance
  • Business-specific approval strategies

2. Blacklist Definition & Data Handling

2.1 Blacklist Decision Rule

  • True Blacklist:
    • L5 level signals in SMS and voice represent confirmed high-risk / fraud / delinquency users

2.2 Seat Data Treatment

  • Seat data must be excluded from blacklist count calculations

Reason:

  • Seat data includes mixed operational activities:
    • Marketing outreach
    • Customer service calls
    • Payment reminders
    • Collection activities
  • These introduce label noise and distort blacklist precision

However:

  • Seat data should still be used as behavioral signals in risk analysis (not labeling)

3. Scenario-Based Risk Evaluation Framework

3.1 New Customer Acquisition (Data Coverage ≥ 80%)

3.1.1 Indicators for Increasing Approval (Pass Rate ↑)

Profiles considered lower risk:

  • Presence of:
    • Multiple L1 / L3 signals in SMS or voice
    • No L4 / L5 signals
  • Seat data signals:
    • L5 exists with valid call duration
    • Requires validation via:
      • Connection frequency
      • Call duration consistency

Interpretation:

  • Likely normal financial behavior / engaged users

3.1.2 Indicators for Reducing Approval (Pass Rate ↓)

Profiles considered higher risk:

  • Abnormal recent activity (last ~7 days):
    • Multiple outbound attempts
    • No successful connections (0 duration)
  • Risk signals in telco data:
    • Multiple L4 hits
    • Even small number of L5 hits
  • Bullish / aggressive borrowing behavior:
    • High distinct CID count
    • Example threshold:
      • more than 10 unique CIDs

Interpretation:

  • Indicates over-leveraging, potential fraud, or credit stress

3.2 Existing Customers (Credit Limit Increase / Repeat Loans)

3.2.1 High-Quality Customer Signals

  • Seat data:
    • L5 with normal call duration (healthy engagement)
  • Borrowing behavior:
    • Controlled CID exposure:
      • Suggested range: 3–5 CID hits
  • Risk outcome indicators:
    • No signs of default risk
    • CID behavior not in extreme ranges

Interpretation:

  • Eligible for credit line increase / retention strategies

3.2.2 Fraud / Uncertain Risk Signals

  • Very low CID activity:
    • 0–1 CID hits

Interpretation:

  • Insufficient behavioral data
  • Must be cross-validated with other data sources (e.g., device, KYC, bureau)

4. Key Evaluation Parameters

4.1 Time Window Configuration

Must be aligned with product type:

  • New Customers:
    • Short-term loans:
      • Last 6–12 months
    • Installment products:
      • Full historical dataset preferred
  • Existing Customers:
    • Based on:
      • Loan tenure
      • Repayment cycle
      • Credit review frequency

4.2 Dynamic Threshold Setting

Thresholds must not be static and should be calibrated per client:

Examples:

  • Number of dial attempts
  • Connection rate
  • Distinct CID count
  • L4/L5 hit frequency

Adjustment drivers:

  • Target segment risk profile
  • Approval rate targets
  • Local market behavior (PH / ID / LATAM / Africa differences)

5. Backtesting Methodology

5.1 Sample Segmentation

Split test population into:

  • Approved vs Rejected (client decision baseline)
  • Good vs Bad (actual repayment outcome)
  • MAW risk tiers (L1–L5 distribution)

5.2 Core Evaluation Metrics

  • Hit Rate (Blacklist Detection):
    • % of bad users correctly identified via L5
  • KS / AUC (if combined into score)
  • Approval Rate Impact:
    • Pass rate change when applying MAW rules
  • Bad Rate Improvement:
    • Compare:
      • With MAW rules
      • Without MAW rules

5.3 Rule Simulation

Simulate scenarios:

  • Reject:
    • Any L5 (SMS/voice)
    • High CID count (> threshold)
    • Multiple L4
  • Approve:
    • Only L1–L3
    • Stable seat engagement

Then compare:

  • Approval rate vs Bad rate tradeoff

6. POC / Live Testing Recommendations

  • Start with:
    • Partial traffic rollout (10%–30%)
  • Monitor:
    • Early delinquency (D7, D14)
    • Approval rate shift
  • Gradually adjust:
    • CID thresholds
    • L4/L5 sensitivity
    • Time windows

7. Summary

  • MAW is behavioral telco-derived data, not a standalone credit score
  • Best performance achieved when combined with:
    • Device intelligence
    • Credit bureau
    • Alternative data (e.g., Smile's Footprint Score)
  • Seat data:
    • Valuable for behavior
    • Not suitable for labeling

Updated about 3 hours ago