Beyond Aggregate Calibration: Decomposing Income-Conditional Recall Disparities in Automated Credit Default Prediction

Technical Summary: Decomposing Income-Conditional Recall Disparities in Automated Credit Default Prediction

Problem Statement
The paper addresses a critical intersection between data-centric AI curation and algorithmic fairness in consumer credit underwriting. While data pipelines frequently utilize model confidence scores to filter "noisy" or mislabeled training instances, and while fairness literature acknowledges that removing protected attributes rarely prevents the reconstruction of bias via correlated features, the specific interaction between confidence-based label filtering and socioeconomic disparities remains unexplored. The central problem investigated is whether an algorithm's tendency to treat prediction failures (false negatives) as data noise skews systematically along income lines, and whether standard "blind" fairness interventions (removing sensitive attributes) are sufficient to eliminate these disparities.

Methodology
The study utilizes a large-scale, publicly accessible dataset from LendingClub (N = 1,344,936 records, 2007–2018), focusing on loans with terminal outcomes ("Fully Paid" or "Charged Off"). Borrowers are categorized post-hoc into three income brackets: Low (<$50k), Middle ($50k–$100k), and High (>$100k).

The research employs a sequential feature-blinding methodology to isolate the drivers of disparity across three model specifications:

1. Full Model: Includes all features, explicitly including self-reported annual income.
2. Income-Blind Model: Excludes annual income.
3. Double-Blind Model: Excludes both annual income and the origination interest rate (to remove upstream institutional bias encoded by human underwriters).

The author evaluates performance using Equal Opportunity (Hardt et al., 2016), focusing specifically on the True Positive Rate (TPR/Recall) among actual defaulters, rather than aggregate accuracy. To explain the mechanisms behind residual disparities, the study uses out-of-sample signed SHAP (SHapley Additive exPlanations) values extracted strictly from a held-out test partition to prevent in-sample memorization. The findings are validated across two model families: Gradient-Boosted Trees (XGBoost) and Regularized Logistic Regression.

Key Contributions

1. Empirical Audit of Label Cleaning: The paper demonstrates that standard confidence-threshold filtering introduces a statistically significant, income-conditional skew. High-income defaulters are disproportionately flagged as "label noise" (2.40% discard rate) compared to low-income defaulters (0.38% discard rate), a disparity confirmed by a chi-square test ($\chi^2 = 1192.26, p < 0.0001$).
2. Fairness Metric Reframing: By applying the Equal Opportunity criterion, the study reveals a substantial recall deficit for high-income borrowers that aggregate calibration metrics obscure. There is a raw 16.86 percentage point gap in recall between high-income (55.98%) and low-income (72.84%) defaulters.
3. Mechanism Decomposition: The author traces the disparity to three distinct mechanisms:
   • Direct Reliance: Direct use of self-reported income accounts for roughly half the initial gap.
   • Inherited Institutional Bias: The origination interest rate acts as a proxy for human underwriter judgment. High-income defaulters received lower interest rates (median 14.49% vs. 15.31% for low-income), leading the model to incorrectly classify them as lower risk when income is removed.
   • Residual Structural Proxies: Even after blinding both income and interest rates, a statistically significant gap persists (2.56% on held-out test data). SHAP analysis identifies loan amount and home ownership status as the primary drivers, acting as structural proxies that reconstruct the omitted income signal.
4. Methodological Verification: The results are robust across model architectures (XGBoost and Logistic Regression) and validation schemes (5-fold cross-validation vs. strict held-out test), confirming the disparity is embedded in the data structure rather than an artifact of a specific algorithm.

Results

• Confidence Filtering: High-income defaulters are removed as "noise" at a rate approximately 6 times higher than low-income defaulters.
• Recall Disparity: The initial TPR gap between high and low-income groups is 16.86 percentage points.
• Blinding Effects:
  • Removing income reduces the gap to 7.49 percentage points.
  • Removing both income and interest rates reduces the gap further to 3.55 percentage points (cross-validation) and 2.56 percentage points (held-out test).
• Residual Drivers: The remaining 2.56% gap is driven by loan amount and home ownership. Middle-income borrowers, who borrow larger absolute amounts than low-income borrowers, show slightly higher aggregate recall than low-income borrowers in the double-blind model, but this is a composition artifact; when analyzed by loan-amount deciles, high-income borrowers consistently exhibit the lowest recall.

Significance and Claims
The paper claims that simply blinding an algorithm to sensitive attributes (like income) is insufficient to ensure fairness in regulated financial contexts. The study highlights that institutional pricing decisions (interest rates) and downstream behavioral proxies (loan size, asset ownership) collectively reconstruct the signals of omitted variables.

The author concludes that fairness audits for data-centric AI workflows in financial institutions must extend beyond the removal of explicit protected attributes. They must account for upstream institutional variables that encode prior human judgment and downstream transaction geometries that systematically reconstruct omitted signals. The findings suggest that without addressing these structural proxies, automated credit systems will continue to produce income-conditional disparities in both data cleaning heuristics and minority-class recall.