Systemic Risk and Stress Testing

Systemic risk is recognised as a crucial risk component in financial markets which both regulators and market players are interested in quantifying.

The number of financial transactions happening every second is ever-increasing, occurring on a variety of marketplaces, performed by different market players, and related to more and more interdependent complex instruments.

Ensuring financial stability and monitoring systemic risk require making sense of multiple data structures and data sources, describing our societies and economies, not only our markets. Stochastic modeling remains a fundamental benchmark, but complex market behaviours require data-driven modeling hardly captured by standard theory. 

Network theory is a powerful framework for modeling dynamic multi-layered relationships between financial institutions and for training non-linear multivariate models -including machine learning ones- needed to capture the complex interdependencies of financial variables and to provide timely and informative indicators about the financial system. 

We propose a systematic algorithmic reverse-stress testing methodology to create “worst case” scenarios for regulatory stress tests by accounting for losses that arise from distressed portfolio liquidations. First, we derive the optimal bank response for any given shock. Then, we introduce an algorithm which systematically generates scenarios that exploit the key vulnerabilities in banks’ portfolio holdings and thus maximize contagion despite banks’ optimal response to the shock.

We apply our methodology to data of the 2016 European Banking Authority (EBA) stress test, and design worst case scenarios for the portfolio holdings of European banks at the time. Using spectral clustering techniques, we group 10’000 worst-case scenarios into twelve geographically concentrated families.

Our results show that even though there is a wide range of different scenarios within these twelve families, each cluster tends to affect the same banks. An “Anna Karenina” principle of stress testing emerges: Not all stressful scenarios are alike, but every stressful scenario stresses the same banks.

These findings suggest that the precise specification of a scenario is not of primal importance as long as the most vulnerable banks are targeted and sufficiently stressed. Finally, our methodology can be used to uncover the weakest links in the financial system and thereby focus supervisory attention on these, thus building a bridge between macro-prudential and micro-prudential stress tests.

This paper studies how firm failures and the resulting disruptions to supply chains can amplify negative shocks. We develop a non-competitive model where customized supplier-customer relations increase productivity, and the relationship-specific surplus generated between firms and their suppliers is divided via bargaining. Changes in productivity alter the distribution of surplus throughout the economy and determine which firms are at the margin of failure. A firm’s failure may spread to its suppliers and customers and to firms in other parts of the production network. We provide existence, uniqueness, and a series of comparative statics results, and show how the response of the equilibrium production network may propagate recessionary shocks.

We construct an empirical measure of expected network spillovers that arise through default cascades for the US financial system for the period 2002-2016. Compared to existing studies, we include a much larger cross-section of US financial firms that comprises all bank holding companies, all broker-dealers and all insurance companies, and consider their entire empirical balance sheet exposures instead of relying on simulations or on exposures arising just through one specific market (like the Fed Funds market) or one specific financial instrument (like credit default swaps). We find negligible expected spillovers from 2002 to 2007 and from 2013 to 2016. However, between 2008 and 2012, we find that default spillovers can amplify expected losses by up to 25%, a significantly higher estimate than previously found in the literature.