Methodology

The balance sheet-based network analysis follows Espinoza-Vega and Sole (2010) framework to simulate the default of one bank at a time and to track how the contagion spreads to the rest of the network. This approach not only considers contagion through direct bilateral connections, but also indirectly through third parties by accounting for potential “cascade effects” after the initial round of distress in the network. Cascade effects continue until there is no more subsequent failures in the network. The model looks at both credit and funding channels of the contagion from a given bank. To analyze the effects of a credit shock, the exercise simulates the individual default of each bank (with probability of default=1), for a given loss-given-default parameter (λ), where the counterparties’ capitals absorb the losses on impact. Then, bank i is said to experience acute distress if its capital buffer is insufficient to fully cover its losses due to bank h defaulting:

• that is if k_i – λx_hi < 0, where x_hi stands for bank i loans to bank h and ki stands for i’s capital buffer.

As for the funding shock, in this stylized exercise, it is assumed that banks are unable to replace all the funding previously granted by the defaulting bank, which, in turn, triggers a fire sale of assets. In this setup, bank i is able to replace only a fraction (1-ρ) of the lost funding from bank h, and its assets trade at a discount, so that bank i is forced to sell assets worth (1+δ) ρx_ih in book value terms, where x_ih stands for bank i borrowing from bank h. The funding shortfall induced loss, δρx_ih is absorbed by bank i’s capital. Then, bank i is said to experience acute distress if its capital buffer is insufficient to fully cover the funding shortfall induced loss due to bank h defaulting:

• that is if k_i – δρx_ih < 0.

When the credit and funding shocks are combined, the condition causing acute distress can be formulated as:

• k_i – (λx_hi + δρx_ih)< 0.

In the subsequent rounds, if there are new banks experiencing acute distress, the losses need to be accumulated over all the rounds in order to test the inequalities above.

In terms of results, this exercise generates four main outputs for each bank (bank i):

• a) Induced failures: the number of subsequent system-wide failures if bank i fails first;

• b) Vulnerability level: total number of independent simulations (one per each bank’s failure) under which bank i falls into distress as a consequence;

• c) Index of contagion: averages the percentage of loss of other banks due to the failure of bank i:

• Contagion index of bank i: ${\text{CI}}_{\text{i}}=100*\raisebox{1ex}{${\mathrm{\Sigma }}_{\text{j}\ne \text{i}}{\text{L}}_{\text{ji}}$}\!\left/ \!\raisebox{-1ex}{${\mathrm{\Sigma }}_{\text{j}\ne \text{i}}{\text{K}}_{\text{j}}$}\right.$

where K_j is bank j’s capital and L_ji is the loss to bank j due to the default of bank i.

• d) Index of vulnerability: averages the percentage of loss of bank i due to the failure of all other banks:

• Vulnerability index for bank i: ${\text{VI}}_{\text{i}}=100*\raisebox{1ex}{${\mathrm{\Sigma }}_{\text{j}\ne \text{i}}{\text{L}}_{\text{i}\text{j}}$}\!\left/ \!\raisebox{-1ex}{$(\text{n}-\text{1})*{\text{K}}_{\text{i}}$}\right.$

where K_i is bank i’s capital, L_ij is the loss to bank i due to the default of bank j, and n is the number of banks in the network.

The main adverse scenario in this exercise assumes λ=0.6, ρ=0.5, δ=1, which means that loss-given-default is 60 percent, the fraction of funding shortfall is equal to 50 percent with a 50 percent discount rate on the assets that a bank may be forced to sell. This can be described as a moderately severe shock scenario to stress-test the banking system from the perspective of contagion risk. Since the credit risk mitigation measures amount to 15 percent of gross exposures, and all remaining exposures can be assumed to be at risk in the case of a default, 60 percent loss given default assumption is suitable for a moderately severe scenario. Given the challenges in calibrating the other parameters based on actual data, assumptions similar to those considered in previous FSAPs were used to simulate an adverse scenario and a wide range of sensitivity checks were conducted.

Bank-Level Data

The data on interbank exposures and Tier 1 capital can be obtained from COREP templates. The two main supervisory data sources are:

• COREP Large Exposures template shows the breakdown of each bank’s assets by counterparty. A large exposure is defined as an exposure that is 10 percent or more of a bank’s eligible capital base vis-à-vis a single borrower or a group of connected clients. For qualifying exposures vis-à-vis a group of connected clients, all exposures vis-à-vis each client in the group must be reported regardless of the 10 percent threshold. For the network analysis, a comprehensive dataset was built by combining the data reported by each bank in the sample. Due to the dataset size as well as the imperfect nature of the reported metadata, the biggest task involved reconciling all the counterparty level data into a standard form where the counterparties as reported by different banks could be matched and further filtering can be performed. Large exposures data was used to fill blocks A and B on Appendix II Figure 1.

• COREP Concentration of Funding by Counterparty (C 67.00) template reports the top ten largest counterparties either as a single creditor or a group of connected clients from which funding obtained exceeds a threshold of 1 percent of total liabilities. In order to expand the scope beyond EA banks, large exposures reporting on the asset side was complemented with largest funding sources on the liabilities side of banks’ balance sheets. Completing the funding dataset was relatively less complex task as the data reported on the top ten largest counterparties by its nature is limited to a small number of counterparties and metadata reporting is of higher quality. Data from this template was used to fill block C in Appendix II Figure 1. Because of the limited nature of the data on the funding sources, analysis that relies on block C can only provide a partial picture and the resulting conclusions can only underestimate the associated spillovers.

Country-Level Data

The country-level analysis comprises 208 jurisdictions. These are then divided into key regions including Euro Area (EA), non-EA European Union (nonEAEU), non-EU Europe (nonEUEUR), EM Asia, Latam, Other Advanced Economies (AE), and Rest of the world (ROW). The focus is on the first six clusters which capture the majority of the exposures and have reliable aggregate data. The index for each cluster is taken as an average of the constituent countries. The analysis is based on 2008Q4, 2013Q1 and 2017Q1.

Cross-border banking exposure claims data are based on BIS locational banking statistics on a residence basis. For those countries, which do not report the exposure claims data directly, a partial map is created using the cross-border banking exposure liabilities. Core capital data are taken from IMF’s FSI Statistics and central bank authorities. Some countries where country level country data is not available are assumed to have a high capital level so that they never fail under the simulations. Most of these countries are relatively smaller in size, and most likely do not impact the final outcome of the analysis.

Implementation

Contagion analysis relied on Espinosa-Vega and Sole methodology for the analysis of supervisory data on banks’ large exposures and funding sources provided in a secure room at the ECB.

• For the intra-EA analysis, the initial data collection focused on the main 25-bank sample, where a 25 by 25 matrix was constructed (block A on Appendix II Figure 1) amounting to a total of €125 billion.

• For the international contagion analysis, the 25-bank sample was expanded to incorporate significant counterparties within EA and outside. The large exposure data was complemented with the 10 largest counterparties who provide funding. The scope of the network is contained to the counterparties classified as credit institutions. Furthermore, in order to have a consistent sample, all the individual counterparty level data was aggregated to the level bank holding groups to the extent possible but excluding exposures to nonbanking clients within each group. The exposures vis-à-vis clients amounting to less than €100 million were filtered out. After the aggregation and filtering, the final network dataset comprised: (i) the 25 EA reporting banks; (ii) 28 additional banks in EA (not part of the main sample); (iii) 101 banks outside EA. The global network including intra-EA exposures total to about €524 billion.

• The data collection focused on gross exposures after deducting exemptions but before credit risk mitigation measures given limited information and resources to analyze underlying collateral for each counterparty. While this is a limiting factor, overall credit risk mitigation measures (substitution effect and funded credit protection) amount to about 15 percent of gross exposures between banks and the exercise tests a wide range of loss-given-default ratios as a sensitivity check.

• Two different assumptions were considered with respect to capital buffer in this exercise: one that allows banks to deplete all their CET1 surplus (CET1 in excess of the minimum 4.5 percent) before an acute distress occurs, and a narrower buffer, where a decline in CET1 corresponding to 5 percentage points of risk-weighted assets would cause an acute distress in a bank. For all non-reporting banks, only the Tier 1 measure was available publicly. The average ratios of these two buffers in relation to full CET1 (0.67 and 0.33, respectively) for reporting banks were applied to Tier 1 capital of non-reporting banks to approximate their buffers based on proportionality.

• Regional categories are used to decompose the spillover indices on a geographical basis. Quantifying the contribution of each group to these indices facilitates a broad comparison between the results from intra-EA analysis and those from the global network analysis. Increasing the number of banks in the network generally causes a downward bias on the indices due to averaging. Hence, the index values alone do not serve as suitable measures for comparing networks of different sizes, namely the intra-EA network versus the global network. However, the contribution of EA25 group to the overall index values can serve as a basis to compare the contributions from other groups in the global network. The regional categories are formed as follows: EA25: member states of the euro area where the 25 reporting banks are located; Other EA: member states of the euro area where the other, non-reporting banks are located; EU (extra-EA): Denmark, Poland, Sweden, United Kingdom; Europe (extra-EU): Norway, Russia, Switzerland, Turkey; Advanced Economy: Australia, Canada, Japan, Korea, Singapore, United States; and Other: Algeria, Azerbaijan, Brazil, Chile, China, Egypt, Hong Kong, India, Mexico, Morocco, Qatar, Thailand, Tunisia, United Arab Emirates, Vietnam.

Euro Area: Large Exposures Dataset

Citation: IMF Staff Country Reports 2018, 231; 10.5089/9781484369586.002.A002

Sources: ECB, and IMF staff calculations.

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Euro Area: Large Exposures Dataset

Citation: IMF Staff Country Reports 2018, 231; 10.5089/9781484369586.002.A002

Sources: ECB, and IMF staff calculations.

• Download Figure
• Download figure as PowerPoint slide

Euro Area: Large Exposures Dataset

Citation: IMF Staff Country Reports 2018, 231; 10.5089/9781484369586.002.A002

Sources: ECB, and IMF staff calculations.

• Download Figure
• Download figure as PowerPoint slide