Axioma In-Practice Series
Toward Dynamic Stress Tests
This note outlines an extension to Transitive Stress Tests. The main idea is to allow the stress test to “select” the proper period or regime for volatility/correlation estimations (using some common techniques from machine learning), i.e., a dynamic stress test. Typically, risk managers subjectively select periods of elevated correlation and volatilities, which are then used as inputs for transitive stress tests. Instead, we describe how the stress scenario itself determines the period when the shift was most likely to occur. Dynamic stress tests can be used alongside the traditional approach in which specific periods are explicitly specified.
Author: Iulian Cotoi and Robert Stamicar