Credit & Lending

The Z-score model using MDA; 10,000+ citations and still the benchmark for bankruptcy prediction.

Introduced logistic regression to default prediction; established logit as industry workhorse.

Structural model treating equity as call option; foundation for KMV/Moody's EDF.

Explains equilibrium credit rationing from adverse selection; essential for lending market theory.

Dominant ML algorithm for credit scoring; consistently outperforms logistic regression.

Definitive benchmark comparing 41 classifiers across 8 datasets; establishes that ensemble methods outperform logistic regression.

First major paper showing transaction data + ML dramatically improves default prediction; estimates 6-25% cost savings.

Uses account-level data from 6 major US banks; finds substantial heterogeneity—no single model works for all.

Seminal Boston Fed study documenting lending discrimination; foundational for fair lending enforcement.

Latinx/African-American borrowers pay 7.9 bps more; FinTech reduces but doesn't eliminate discrimination.

ML models create distributional impacts favoring advantaged groups even without using race.

Credit scores are noisier for minority borrowers; quantifies how data disparities translate to lending disparities.

Proves impossibility of satisfying calibration and error rate parity simultaneously—THE foundational fairness theorem.

Independently derives impossibility result with explicit disparate impact focus; addressed ProPublica/COMPAS controversy.

Documented sticky credit card rates; introduced adverse selection explanations establishing the field.

First direct evidence of adverse selection using randomized solicitations.

Demand-and-cost curve framework for analyzing selection; widely applied to credit markets.

Intuitive graphical framework for understanding selection and welfare in credit/insurance.

Benchmark comparing survival methods for WHEN default occurs—critical for IFRS 9 and lifetime expected credit loss.

Landmark study finding increased default propensity independent of risk composition; suggests declining stigma.

Comprehensive LGD estimation review; go-to reference for Basel II/III recovery models.

Industry-standard LGD model using debt type, seniority, and macro factors.

JPMorgan Chase's 18-year study of 3,761 defaults establishing key LGD drivers.

Mobile behavioral data outperforms credit bureaus for thin-file borrowers; foundational fintech paper.

Alternative data identifies 'invisible primes' overlooked by traditional scores.

Mobile metadata predicts socioeconomic status; opened mobile-based credit scoring in developing countries.

Documents 45 million 'credit invisible' Americans; foundational regulatory framework.

Digital footprints (device, email domain, typing) match credit bureau accuracy; foundational fintech credit study.

The foundational imbalanced-learning method used in virtually every fraud detection system.

Authoritative taxonomy of fraud detection methods; still cited for conceptual foundations.

Definitive survey covering statistical, ML, and proximity-based anomaly methods.

Addresses realistic fraud detection challenges: class imbalance, concept drift, and delayed feedback.

Transaction aggregation features improve fraud detection by 40%; widely adopted in industry.

Survey on using graph structure to detect fraud rings and coordinated attacks.

Network propagation improves fraud detection by leveraging transaction graph structure.

Embedding GLMs into neural networks improves insurance pricing while maintaining interpretability.

Evolutionary trees optimize premium segmentation while respecting regulatory constraints.

Comprehensive comparison of fraud detection methods for insurance claims.

Telematics data (speed, braking) improves claims prediction; foundational usage-based insurance paper.

Neural networks improve reserve estimation over traditional chain-ladder methods.

Foundational empirical study of reputation dynamics and their impact on seller behavior.

Field experiment showing reputation inflation and limits of feedback systems.

Comprehensive survey of reputation system design and effectiveness in platforms.

Natural experiment measuring the causal impact of reputation visibility on market outcomes.

SHAP values unify feature attribution methods; now standard for credit model explanations.

Local interpretable explanations for any black-box model; widely used for adverse action notices.

Argues interpretable models match black-box accuracy for credit; influential regulatory perspective.

Generalized additive models with interactions; template for interpretable credit scoring.

Comprehensive survey on detecting and adapting to changing data distributions in credit.

Identifies sample selection bias in scorecard validation; essential for production monitoring.

Reviews and advances reject inference methods for handling missing data from declined applications.

Active learning reduces labeling costs for streaming fraud detection systems.