Median Regression for Heavy-Tailed Noise

Reported in interviews at Citadel

Problem statement

You are fitting a regression model, and the errors are independent and identically distributed but have very heavy tails -- think Cauchy-level noise where even the mean may not exist.

Ordinary least squares breaks down in this setting. Why does it fail, and what should you use instead?

Explain the alternative approach, its loss function, why it is robust, and what practical trade-offs it introduces compared to OLS.

Show hint

What summary of the conditional distribution of $y$ does each candidate loss function actually estimate?

Relevant roles: Data Scientist, Applied Scientist, ML Engineer and Forward Deployed Engineer.

Related free problems