Robust Knockoffs for Controlling False Discoveries With an Application to Bond Recovery Rates

TL;DR

This paper introduces a robust knockoff framework for variable selection in highly correlated data, particularly in finance, improving false discovery control and identifying new relevant factors while maintaining predictive accuracy.

Contribution

We develop a robustified knockoff method that handles high dependence and time correlation, with a subsampling strategy and false discovery control over multiple levels.

Findings

Identified new relevant factors for bond recovery rates.

Achieved similar predictive power to state-of-the-art models.

Enhanced false discovery control in correlated data environments.

Abstract

We address challenges in variable selection with highly correlated data that are frequently present in finance, economics, but also in complex natural systems as e.g. weather. We develop a robustified version of the knockoff framework, which addresses challenges with high dependence among possibly many influencing factors and strong time correlation. In particular, the repeated subsampling strategy tackles the variability of the knockoffs and the dependency of factors. Simultaneously, we also control the proportion of false discoveries over a grid of all possible values, which mitigates variability of selected factors from ad-hoc choices of a specific false discovery level. In the application for corporate bond recovery rates, we identify new important groups of relevant factors on top of the known standard drivers. But we also show that out-of-sample, the resulting sparse model has similar predictive power to state-of-the-art machine learning models that use the entire set of predictors.