I study two related questions: when do machine-learning systems fail the people they are meant to serve, and what does rigorous causal identification reveal about structural discrimination in financial markets?

My published and submitted work spans healthcare ML fairness, racial disparities in mortgage lending, market microstructure, and cross-domain deployment reliability — united by a commitment to external validity, honest negative results, and methods that hold up under scrutiny.

Shift type, not shift magnitude, determines machine-learning failure modes under deployment shift.

One pre-registered audit protocol applied to four dissimilar real-world domains (clinical, mental-health NLP, mortgage lending over 42M records, network intrusion). The finding: the type of distribution shift — not its magnitude — predicts which axis of trustworthiness (discrimination, calibration, or subgroup reliability) breaks at deployment, and it is diagnosable in advance with cheap, label-free probes.

Comprehensive Evaluation of Machine Learning for Type 2 Diabetes Risk Prediction: Large-Scale External Validation and Fairness Analysis _{Rajveer Singh Pall, Sameer Yadav, Siddharth Bhalerao, Sourabh Sahu, Ritu Ahluwalia, Bhaskar Awadhiya · IEEE Conference}

XGBoost trained on NHANES 2015–2020 (n = 15,685), externally validated on BRFSS 2020–2022 (n = 1,285,783). Internal AUC 0.794 degraded to 0.717 under distribution shift; a 13.5 AUC-point gap between young (0.742) and elderly (0.607) adults means the highest-risk population receives the weakest algorithmic performance.

XGBoost SHAP DeLong CI Algorithmic Fairness External Validation TRIPOD-AI

Persistent Racial Disparities in U.S. Mortgage Approval: Evidence from 42 Million Applications, 2020–2024 · _{Journal of Housing Economics}

42.3M applications, 5,500 lenders. Black applicants face a raw 14.95 pp approval gap; after DFL reweighting on all observable financials, 98.6% remains unexplained. Within-lender fixed effects attribute 74.6% of the gap to within-institution decisions, rising 66.8% (2020) → 78.3% (2024). RDD at the 80% LTV/PMI boundary and a DiD around the 2022 Fed tightening isolate two specific channels.

Regression Discontinuity Difference-in-Differences DFL Decomposition HMDA Fixed Effects

The Transaction Cost Trap: Why ML Stock Prediction Fails Economically Under Realistic Frictions · _{Quantitative Finance and Economics}

A regime-filtered CatBoost/RF/DNN ensemble reaches 73.3% conditional directional accuracy in bear regimes yet returns −42.49% annually (Sharpe −2.83) after 5 bps costs, versus buy-and-hold's +34.77%. A closed-form breakeven shows profitability needs 88% accuracy — an explicit case against publication bias in financial ML.

CatBoost DNN Walk-forward Validation Market Efficiency Ensemble Methods

TrustShift: Shift Type, Not Shift Magnitude, Determines ML Failure Modes · _{Applied Intelligence} — see Flagship above.

Causal Inference     Causal Forests · Double ML · RDD · DiD · DFL Decomposition · Manski Bounds
Fairness             Subgroup AUC · DeLong CI · ECE · Equalised Odds · Disparate Impact
ML                   XGBoost · LightGBM · CatBoost · PyTorch · scikit-learn · Random Forest
NLP                  FinBERT · HuggingFace Transformers · RAG · SHAP
Federated / Privacy  Flower (flwr) · FedAvg · FedProx · FedNova · Opacus (DP)
Data at Scale        HMDA 42M · BRFSS 1.28M · NHANES · 17,773 stock-days · 14,584 transcripts
Languages            Python · R · SQL

🐍  Contribution Graph