Essays on Stress Testing

Author(s):
Chayaluck Garun Chompucot, PhD
Keywords:
Stress Testing, Supervisory Models, Machine Learning

Abstract :

Stress Testing: A Plea for Model Simplicity and Explanatory Variable Extension. The Federal Reserve supervisory stress test model endeavors to gauge the extent of capital required to withstand unexpected crises. However, the complex multi-equation modelling approach similar to the approaches normally used by the Federal Reserve produces the worst prediction inaccuracies as compared to the simpler approaches over the 2008 financial crisis period. This paper introduces the approaches as alternatives to the ones similar to those of the Federal Reserve. In the proposed approaches, the complex net income models akin to the ones in the supervisory models are simplified. The paper also introduces the use of financial explanatory variables to reflect banking and financial systems besides macroeconomic variables given by the Federal Reserve. Results verify that the proposed financial variables visibly improve model fit for in-sample and out-of-sample model specifications. The conclusion is that the supervisory models should be simplified and explanatory variables extended for greater model accuracy and a more robust insurance of banking capitalization in stressed periods.

Stress Testing: Machine Learning to the Rescue for Scenario Severity Measurement.  As scenario severity is the keystone of stress testing, the design of a supervisory scenario framework is a challenge for the Federal Reserve. In the context of domestic supervisory stress test scenarios, there is a controversy within the Federal Reserve as such analyses confirm that supervisory scenarios are less conservative than scenarios designed by industry practitioners. Yet, in the context of international supervisory stress test scenarios, no scenario severity measurement has been done. This paper attempts to do so by observing the magnitude of the impact of the international supervisory stress test scenarios on the U.S. real GDP growth and unemployment rate. With forward-stepwise regression and feedforward backpropagation neural network in batch mode techniques, results from the model estimation and projection suggest the less extremity in the severely adverse scenarios for the U.S. real GDP growth and unemployment rate.  Thus, the Federal Reserve needs to revisit the scenario design framework and the severity measurement process for supervisory stress test scenarios.

 

Publication date of the thesis
04-03-2022

Thesis committee

Supervisor: Abraham Lioui, EDHEC Business School 

External reviewer: Dimitris Korobilis, University of Glasgow  

Other committee members: Emmanuel Jurczenko and Enrique Schroth, EDHEC Business School