Multi-Source Credit Scoring Framework

1. Executive Summary

This document defines a credit scoring framework that addresses two distinct customer segments (Individuals and SMEs) using a confidence-weighted, multi-source approach 1. The framework is designed for the Iraqi market where traditional credit bureau data is sparse or nonexistent 14.

Core Innovation: Two-Dimensional Confidence

Every data source in this framework is evaluated on two dimensions:

1. Expected Performance Benchmark - Typical predictive power observed for this data type in academic literature and industry deployments (measured in Gini coefficient / AUC), treated as a prior expectation not a hard ceiling
2. Data Availability Confidence - How much data we have for this specific customer, relative to the minimum threshold needed for reliability

Design Decisions Made

2. Segment-Specific Scoring Framework

2.1 Individual Scoring

Individuals are scored primarily on behavioral signals that indicate willingness and capacity to repay.

Data Sources for Individuals

Individual Scoring Formula

Fairness Note on Network Score: The Ahl Score (family/network financial health) must be implemented using behavioral metadata only (e.g., contact call/payment patterns), never identity attributes (tribe, sect, region). Requires explicit fairness testing before production to ensure no proxy discrimination. May be removed if regulatory or fairness concerns arise.

2.2 SME/Company Scoring

SMEs require fundamentally different data because risk drivers differ from individuals.

Data Sources for SMEs

SME Scoring Formula

2.3 Segment Comparison

3. Data Source Tiers & Accuracy Benchmarks

Each data source has an expected performance benchmark—the typical Gini coefficient range observed in academic literature and industry deployments. These are prior expectations, not hard ceilings; actual performance varies based on label definition, product terms, segment mix, feature engineering, and macroeconomic conditions.

Tier 1: Highest Predictive Power (Gini 0.35-0.50)

Tier 2: Medium Predictive Power (Gini 0.25-0.40)

Tier 3: Supplementary/Thin-File (Gini 0.15-0.25)

Ensemble Lift Effect

Combining sources breaks individual ceilings:

4. Confidence Scoring & Data Requirements

4.1 Confidence Calculation

For each data source, confidence is calculated based on data availability relative to minimum thresholds:

Note: Minimum_Viable (see Section 4.2) is a gating threshold; below it, exclude the source entirely (or set Data_Confidence ≈ 0) even if the formula yields a non-zero value.

4.2 Minimum Data Thresholds by Source

4.3 Ground Truth Requirements

For model training and validation:

Notes:

• Default definition: X DPD (e.g., 60+ for short-tenor microloans; 90+ commonly used in bank portfolios), to be finalized during pilot based on product terms and regulatory guidance.
• Sampling: If observed portfolio default rate is below 5%, use case-control sampling (oversample defaults) for training, then recalibrate PDs to the true base rate before deployment (AUC/Gini unaffected; calibration is).

5. Cold Start Strategy

5.1 Baseline: Psychometric + ID Verification

When a customer has zero history, the system uses:

1. Identity Verification (Gatekeeper)

   • INID (Unified National Card) biometric match
   • Liveness detection (anti-spoofing)
   • SIM-binding verification

2. Psychometric Assessment (Scoring)

   • 15-30 minute gamified quiz
   • Measures: Conscientiousness, Locus of Control, Impulsivity, Fluid Intelligence
   • Expected Gini: 0.25-0.35 standalone
   • Accuracy: ~70% classification (AdviceRobo benchmark)

5.2 Cold Start Score Interpretation

5.3 Progression Pathway (Ladder Model)

Based on Tala/Branch industry standard:

*Assumption Warning: Default rates shown are illustrative benchmarks from comparable emerging markets (Kenya, Philippines). Actual rates in Iraq will vary significantly based on underwriting policy, economic conditions, and customer selection. These figures should be recalibrated after 6-12 months of portfolio performance data.

Unlock Triggers:

• On-time repayment of current loan
• Early repayment accelerates progression
• Adding new data sources (bank link, employer verification)
• Platform engagement (using wallet features, bill pay)

5.4 Bootstrapping via Simulation

Since no historical default data exists, initial model weights will be:

1. Literature-based: Use published Gini coefficients as starting weights
2. Regional benchmarks: Adapt models validated in similar markets (Kenya, Philippines, Jordan)
3. Conservative bias: Start with tighter approval thresholds, loosen as data accumulates
4. Learning portfolio: Small initial loans to generate outcome data within 6-12 months

6. Dynamic Weight Adjustment

6.1 Weight Rebalancing Logic

Weights are not static—they adjust based on:

1. Data availability per customer: More data → higher weight for that source
2. Cross-source validation: Conflicting sources → both get reduced weight
3. Outcome feedback: Sources that predict well get increased weight over time

6.2 Customer Journey Model

Progressive Unlock + Milestone-Based (Industry Standard)

Re-scoring Triggers:

• Each loan application
• Each repayment (on-time, early, or late)
• New data source added
• Quarterly periodic review (for active customers)

6.3 SME Owner Credit Decay Schedule

For SME scoring, owner personal credit weight decreases as business matures:

Transition Logic: Weight shifts when business demonstrates:

• 12+ months of transaction history
• Positive trade payment record
• Stable or growing revenue trend
• No owner-level delinquencies

7. Conflict Resolution Approaches

When data sources provide conflicting signals (e.g., psychometric says "trustworthy" but transactions show erratic spending), three approaches exist:

Approach A: Hybrid Fusion (Best Performing)

How it works:

1. Combine related sources early (e.g., all telecom signals → single telecom score)
2. Keep dissimilar sources separate (telecom vs psychometric vs transactions)
3. Meta-learner resolves conflicts using learned weights

Pros:

• Best empirical performance in research
• Handles non-linear relationships
• Adapts to data characteristics

Cons:

• Complex to implement
• Requires sufficient training data
• Less interpretable

Approach B: Information Value (IV) Weighted Averaging

How it works:

1. Calculate IV for each data source during model training
2. Higher IV = higher weight in final score
3. Conflicts resolved by mathematical averaging

Typical IV-based weights:

Pros:

• Transparent and explainable
• Easy to implement
• Regulatory-friendly

Cons:

• Assumes linear relationships
• May not capture complex interactions

Approach C: Implicit ML Resolution

How it works:

1. Feed all features to ensemble model (XGBoost, LightGBM)
2. Model learns optimal feature interactions automatically
3. SHAP values provide post-hoc explanation

Pros:

• Often best predictive accuracy
• Handles feature interactions
• Discovers unexpected patterns

Cons:

• "Black box" concerns
• Requires explainability layer (SHAP/LIME)
• Risk of overfitting

Recommendation

Start with Approach B (IV-weighted) for interpretability and regulatory acceptance. Transition to Approach A or C as data volume and technical capability mature.

8. Model Architecture Decision

8.1 Chosen Architecture: Shared Core + Segment Overlays

Based on McKinsey 9 and FICO 7 best practices, the recommended architecture is:

8.2 Why This Architecture

8.3 Cascading Data Logic

When data is missing, the system cascades:

9. Open Questions & Future Research

9.1 Ground Truth Validation

Challenge: Without historical default data, initial model weights are assumptions.

Proposed Approach:

1. Launch with literature-based weights
2. Deploy conservative approval thresholds (reject borderline cases)
3. Track actual defaults over 6-12 months
4. Recalibrate weights based on observed outcomes
5. Expand approval gradually as confidence increases

9.2 Concept Drift Monitoring

Credit risk changes over time (economic conditions, fraud evolution). Required:

• Monthly model performance monitoring (PSI, KS stability)
• Trigger-based retraining when drift detected
• COVID-style stress testing for economic shocks
• Fraud pattern updates as new schemes emerge

9.3 Iraq-Specific Calibration Needs

9.4 Data Partnership Priorities

To improve model accuracy, pursue data partnerships in order:

1. Critical: Telecom APIs (Zain, Asiacell, Korek)
2. Critical: National ID verification (INID integration)
3. High: Mobile wallet data (Zain Cash, AsiaHawala)
4. High: Qi Card salary data
5. Medium: Utility payment history
6. Medium: Employer verification networks

9.5 SME Entity Resolution Challenge

SME identity in Iraq is fragmented (inconsistent registration, multiple owner identities, cash-based revenue, loan stacking risk). Required capabilities: Entity resolution linking Business ↔ Owners ↔ Devices ↔ Wallets ↔ Bank Accounts, cross-application duplicate detection, beneficial ownership verification, and velocity rules to prevent stacking. Technical specification to be defined in architecture phase.

10. Citations

Academic Sources

1. Djeundje et al. (2021). "Enhancing Credit Scoring with Alternative Data." Expert Systems with Applications. ScienceDirect

2. Hlongwane et al. (2024). "Enhancing credit scoring accuracy with a comprehensive evaluation of alternative data." PLOS ONE. PMC

3. Weng et al. (2024). "Class imbalance Bayesian model averaging for consumer loan default prediction." Research in International Business and Finance. ScienceDirect

4. Feng et al. (2019). "Dynamic weighted ensemble classification for credit scoring using Markov Chain." Applied Intelligence. Springer

5. MDPI (2023). "Character Counts: Psychometric-Based Credit Scoring for Underbanked Consumers." MDPI

6. World Bank (2022). "Evening the Credit Score: Impact of Psychometric Credit Scoring on Women-Owned Firms." World Bank

Industry Sources

7. FICO Blog. "How to Build Credit Risk Models Using AI and Machine Learning." FICO

8. Experian (2025). "Blended Credit Scores: A Smarter Approach to Small Business Lending." Experian

9. McKinsey & Company (2021). "Designing Next-Generation Credit-Decisioning Models." McKinsey

10. CGAP (2019). "Credit Scoring Technical Guide." CGAP

11. LenddoEFL. "Scoring Methodology." LenddoEFL

12. Cenfri/i2i (2017). "Advancing Financial Inclusion Case Study: Branch." Cenfri

13. BIS Papers No. 148. "Digital Innovation for SMEs." BIS

14. AFI (2025). "Alternative Data for Credit Scoring." AFI

15. Nav. "FICO SBSS Score Explained." Nav

Residual Uncertainties

• Exact IV weights for Iraq: Literature values are from other markets; Iraq-specific calibration needed
• Ahl Score effectiveness: Family/tribal scoring is theoretical; requires validation
• Regulatory acceptance: CBI stance on psychometric and alternative data scoring unknown
• Telecom partnership terms: Data access and pricing not yet negotiated

Appendix A: Understanding the Gini Coefficient

A.1 Historical Origin

Original Economics Meaning:

• Gini = 0: Perfect equality (everyone has same income)
• Gini = 1: Perfect inequality (one person has everything)

A.2 What Gini Measures in Credit Scoring

The Gini coefficient measures discriminatory power — how well a score separates defaults from non-defaults.

Gini Interpretation Scale

A.3 Relationship to AUC

The Gini coefficient is directly related to AUC (Area Under the ROC Curve):

A.4 Combining Multiple Data Sources (Ensemble Gini)

Critical Insight: Ensemble Gini is NOT additive.

Why it's not additive: Combined predictive power depends on correlation between sources.

Simplified Approximation (for uncorrelated sources):

A.5 Practical Ensemble Calculation

In practice, you don't calculate combined Gini mathematically — you measure it empirically:

A.6 Typical Ensemble Lifts (Empirical)

Based on industry research, here are typical Gini improvements when combining sources:

A.7 Key Takeaways

The Golden Rule for Adding Data Sources:

Example: Bank transactions are so valuable because they have:

• High Gini (0.40-0.50)
• Low correlation with psychometric/telecom
• Capture unique financial behavior signals