Statistical Evaluation of Algorithmic Decision Systems
Algorithmic fairness · LLM evaluation · Computational social science · Advised by
Prof. Subhabrata Majumdar
This work studies large language models as decision aids in hiring pipelines. Using controlled synthetic-resume experiments, the project examines whether LLMs exhibit differential shortlisting behaviour across region, gender, institutional prestige, and caste-associated name signals in the Indian hiring context. The broader goal is to understand when apparently merit-based automated evaluations become sensitive to social identity signals, prompt framing, or incomplete candidate evidence.
Semi-parametric Extreme Quantile Estimation
Extreme-value theory · Non-parametric statistics · Machine learning · Advised by
Prof. Soudeep Deb
This work focuses on estimating very high conditional quantiles and tail risks when ordinary data are sparse in the extreme region. The methodological emphasis is on combining non-parametric statistical ideas, extreme-value theory, and machine-learning-based function approximation to study rare-event regimes where standard prediction methods can fail or provide weak uncertainty guarantees.
Rare-event and Systemic-risk Modelling under Partial Information
RST problem · Large deviations · Systemic risk · Partial information · Advised by
Prof. Anand Deo
This line of work is concerned with rare systemic-risk events in networked financial systems, especially when the full agent-asset allocation structure is only partially observed. The problem connects large-deviation ideas, partial-information modelling, and stress-testing logic, with the aim of characterizing how extreme losses can emerge from incomplete knowledge of the underlying network.
Importance Sampling for Tail-risk and Rare-event Estimation
Rare-event simulation · Variance reduction · Coherent risk measures
This project studies importance sampling methods for estimating rare-event probabilities and tail-sensitive risk functionals such as VaR, CVaR, expected shortfall, and distortion-based risk measures. The broader aim is to design simulation procedures that improve estimation accuracy in the tail, where naive Monte Carlo methods often require prohibitively large sample sizes.