Technical Summary: Application of Alternative Data in Credit Assessment for the Unbanked Population

1. Problem Statement

The study addresses the challenge of "credit invisibility," a condition affecting individuals who lack traditional credit histories (e.g., credit card usage, formal loan records). In conventional credit reporting systems, these individuals are often misclassified as high-risk or automatically denied credit due to data gaps, forcing them into subprime loans or excluding them from financial services entirely. The core problem is that traditional scoring models conflate a lack of data with a lack of repayment capacity, thereby limiting financial inclusion and preventing institutions from identifying potential high-quality clients.

2. Methodology

Data and Experimental Design

The research utilizes real-world subprime auto loan application data from 2021 to 2023. The dataset comprises:

• Training Sample: 50,000 observations.
• Test Sample: 20,000 observations.
• Target Variable: Default status over a 12-month observation period.

The study constructs a comparative framework involving two primary groups:

1. Baseline Models: Trained exclusively on traditional credit variables (credit card limits, historical balances, outstanding accounts, income, debt-to-income ratio, account tenure, delinquency flags, and recent inquiries).
2. Fusion Models: Trained on traditional variables augmented with Alternative Data.

Feature Engineering

Alternative data features are derived from 24-month observation windows of non-credit payment behaviors, specifically:

• Rent Payment History: Successive lease counts, settlement rates (e.g., last 90 days), and arrears duration.
• Utility and Telecom Bills: Bill settlement rates, payment delay days, and payment amount volatility.
• Stability Metrics: Mobile number stability over 365 days and the rate of change in payment amounts.

Crucially, the study treats missing records not as errors to be imputed with mean/median values or deleted, but as informative missing indicators. This preserves the distinction between applicants with limited formal credit history and those with unstable repayment behaviors.

Model Architecture

The study employs a three-layer framework:

1. Baseline Layer: A regularized logistic regression model using traditional variables to establish a linear risk baseline.
2. Fusion Layer: Integration of alternative data features into three algorithms:
   • Logistic Regression
   • Random Forest
   • XGBoost
3. Decision Layer: A joint discrimination mechanism that simultaneously optimizes for risk control (default probability), approval coverage, and group fairness.

Algorithms and Optimization:

• Logistic Regression: Used for interpretability and linear baseline comparison.
• XGBoost: Used for non-linear feature interaction and ensemble learning. The objective function minimizes loss between true default labels and predicted probabilities while controlling model complexity via leaf node weights and regularization parameters.
• Fairness Constraints: The decision layer incorporates constraints for Demographic Parity (DP) and Equal Opportunity (EO) to prevent the model from increasing approval rates at the cost of demographic equity.

Evaluation Metrics

Performance is assessed using:

• Predictive Power: AUC, KS statistic, and F1 score.
• Operational Metrics: Loan approval rates and 12-month default rates.
• Fairness Metrics: Demographic parity differences and equal opportunity differences across racial and income groups.
• Explainability: SHAP (SHapley Additive exPlanations) values for global and local feature importance, and LIME for local consistency checks.

3. Key Results

Predictive Performance and Approval Rates

The introduction of alternative data significantly improved model performance across all algorithms compared to the traditional-data-only baselines:

• AUC Improvements:
  • Logistic Regression: Increased from 0.682 to 0.731.
  • Random Forest: Increased to 0.758.
  • XGBoost: Increased to 0.781.
• Approval Rates: The XGBoost model with alternative data increased the approval rate from 41.6% (traditional baseline) to 63.7%. This represents an increase of 22.1 percentage points, approving an additional 4,420 applicants in the test set.
• Risk Control: Despite the significant increase in approvals, the 12-month default rate rose only marginally from 6.8% to 7.2% (a 0.4 percentage point increase), which was found to be statistically insignificant. This indicates that the additional approvals were not concentrated in the high-risk tail.

Feature Importance

Feature importance tests identified the following alternative data variables as critical determinants of repayment capacity for credit-invisible borrowers:

• Timely rent payments.
• The number of timely bill payments.
• Long-term consistency in phone numbers.

Fairness Analysis

Fairness tests conducted across racial and income groups showed no evidence of worsened equity:

• Racial Groups: The difference in demographic equality changed by 0.005 (p-value: 0.312), and the equal opportunity difference changed by 0.004 (p-value: 0.347).
• Income Groups: The demographic equity gap changed by 0.005 (p-value: 0.286), and the equal opportunity difference changed by 0.003 (p-value: 0.401).
  The results suggest that the expansion of approval coverage did not come at the expense of increased discrimination against minority or low-income groups.

Ablation Study on Missing Data

The study found that treating missing traditional credit records as informative signals (rather than imputing them or deleting samples) was essential. Imputation compressed heterogeneous missing patterns, weakening the boundary between information scarcity and genuine risk, while sample deletion reduced the representativeness of the credit-invisible population.

4. Key Contributions

1. Empirical Validation of Alternative Data: The study demonstrates that alternative data (rent, utilities, telecom) can effectively supplement traditional credit information, converting unobserved payment continuity into measurable repayment capacity evidence.
2. Risk-Equity Joint Optimization: The proposed framework successfully balances the expansion of financial inclusion (higher approval rates) with risk management (minimal default rate increase) and fairness constraints, avoiding the "risk spillover" often associated with loosening approval standards.
3. Methodological Rigor: By comparing alternative-data models against a strong traditional-data XGBoost baseline, the study isolates the incremental value of alternative data from the benefits of using more complex algorithms alone.
4. Handling of Missing Data: The paper advocates for a specific data processing strategy where missingness is treated as a structured signal, which is shown to be more appropriate for credit-invisible populations than standard imputation techniques.

5. Significance and Claims

The paper claims that its approach offers a broader scope of financial service coverage compared to conventional credit-risk assessment practices. It posits that alternative data does not merely lower approval standards but provides critical behavioral information that traditional credit scores miss.

The authors assert that this methodology:

• Enhances financial inclusion by enabling credit-invisible individuals to access formal credit rather than relying on higher-cost or opaque borrowing channels.
• Refines risk assessment by identifying stable payment patterns that correlate with repayment capacity.
• Maintains demographic equity, ensuring that the benefits of alternative data are distributed fairly across racial and income groups.

The study concludes that while the findings are specific to the subprime auto loan context, they provide a robust empirical basis for using alternative data in inclusive credit assessment. The authors note limitations regarding the generalization of these results to other loan types and suggest future work could involve longer-term data, multi-scenario integration, and privacy-preserving computation techniques.