Vik Shirvaikar

PhD Student in the StatML CDT
Department of Statistics, University of Oxford

Hello! I'm a fourth-year PhD student in statistics and machine learning. My research focuses on Bayesian prediction, model uncertainty, and causality (supervised by Chris Holmes). I grew up in Dallas, Texas and completed my bachelor's degrees in mathematics and economics at the University of Texas at Austin. Before coming to Oxford, I worked for two years as a forensic data analyst, partnering with U.S. government agencies to investigate financial crimes. My hobbies include football (both British and American), science fiction, baking, and theater.

I'm looking for full-time opportunities in applied data science and machine learning. Contact me at vik.shirvaikar (at) gmail.com or on LinkedIn!

Resume

Research Interests

  • Bayesian Theory and Methods
  • Predictive Analytics
  • Causal Inference
  • Model Uncertainty

Selected Projects

A general framework for probabilistic model uncertainty

With Stephen Walker and Chris Holmes. Submitted. [arXiv] [Github]

We propose a novel approach to model uncertainty, framing it as a missing data problem that can be solved via predictive resampling. This eliminates the need for subjective prior elicitation by leveraging one-step-ahead predictive densities and consistent model selection criteria. We demonstrate applications to hypothesis testing, density estimation, and variable selection.

Targeting relative risk heterogeneity with causal forests

With Xi Lin and Chris Holmes. Submitted. [arXiv] [Github]

We develop a novel methodology to estimate heterogeneous treatment effects (HTE) by modifying causal forests to target relative risk. Through simulations on realistic clinical trial data, we show that our approach can improve statistical power and enhance detection of clinically relevant subgroups.

A critical review of causal reasoning benchmarks for large language models

With Linying Yang, Oscar Clivio, and Fabian Falck. AAAI (2024), workshop paper. [arXiv]

We conduct a comprehensive review of benchmarks for evaluating LLM capabilities in causal inference. We identify limitations in existing benchmarks and propose criteria for designing robust assessments that incorporate interventional and counterfactual reasoning.

Rethinking recidivism through a causal lens

With Choudur K. Lakshminarayan. Submitted. [arXiv] [Github]

We propose a novel reframing of recidivism prediction that allows us to assess the effect of incarceration on recidivism using modern causal methods. We find evidence that longer prison sentences increase the likelihood of re-offending, offering insights for criminal justice policy.

Contact Me

vik.shirvaikar (at) gmail.com / LinkedIn