Credit Risk Spillovers among Financial Institutions around the Global Credit Crisis: Firm-Level Evidence

Transcription

1 Credit Risk Spillovers among Financial Institutions around the Global Credit Crisis: Firm-Level Evidence Jian Yang University of Colorado Denver Yinggang Zhou Chinese University of Hong Kong 1

2 Motivation Understanding the nature of the systemic risk is key to understanding the occurrence and propagation of financial crises One of the most important regulatory issues is to identify systemically important financial institutions (SIFIs). Primary criteria by IMF, BIS and FSB Size, Substitutability, and Interconnectedness This study concentrates on interconnectedness and credit risk transfer in financial network 2

3 Contributions It is perhaps the first to provide a data-determined identification of structure of credit risk transmission. It sheds light on identification of SIFIs from the perspective of interconnectedness. First, institutions to be identified as the SIFIs may be those prime senders of credit risk information (e.g., Lehman Brothers). Second, the institutions which are the exchange center of credit risk information could also be systemically important (e.g., Goldman Sachs, Bear Sterns). Third, the institutions which are prime receivers of credit risk information probably are not systemically important (e.g., JP Morgan). 3

4 Contributions (continued) We propose an innovative empirical framework of combining cluster analysis, the principal component analysis, the directed acyclic graph (DAG) (Bayesian Network) and structural vector autoregression (VAR) analysis to facilitate an in-depth search for credit risk transmission network. The empirical framework corresponds well to the theoretical work on credit risk network (e.g., Jarrow and Yu, 2001; Allen, Babus and Carletti, 2009; Cossin and Schellhorn, 2007). We also extend the data-determined structural VAR analysis of Swanson and Granger (1997) to a setting of a high-dimensional system. 4

5 Contributions (continued) We explore the credit default swap (CDS) data of international financial institutions. The CDS spread is considered as superior measure of credit risk. Relatively few empirical studies using firm-level CDS data and even fewer using international firm-level CDS data. Eichengreen, Mody, Nedljkovic, and Sarno (2009)use CDS spreads to study whether the contagion or economic fundamental linkages leads the subprime crisis to go global. 5

6 Data Following Eichengreen et al.(2009), we select 43 largest financial institutions in the US, the UK, Germany, Switzerland, France, Italy, Netherlands, Spain, and Portugal. From Bloomberg, the raw data are the end-of-day of mid-quotes for 5-year CDS spreads. We compute rolling-average, two-day changes of CDS spreads for two purposes: to smooth out sharp daily movements and irregular trading (Eichengreen et al., 2009) to control for the fact that CDS markets for financial institutions from different countries may not operate during the same trading hours (Forbes and Rigobon, 2002). 6

7 Data (continued) The sample runs from January 2007 to the early September of 2008 before the Lehman s failure. This period has seen the unfolding of the crisis until it infected the entire U.S. and global financial system (Brunnermeier, 2009). It provides an ideal time frame for us to investigate which SIFIs play an important role during the crisis. 7

8 Table 1: Summary statistics of CDS Spreads (Jan.1, 2007~Sep.9,2008) Country Name Mean Std. Dev Minimum Maximum US Fannie Mac US Freddic Mac US Lehman Brothers US Bear Sterns US Goldman Sachs US Merrill Lynch US Morgan Stanley US Bank of America US Wachovia US Citigroup US JP Morgan US Met Life US Safeco US American express US AIG US Chubb US Hartford UK Abbey UK Barclays UK HBOS UK HSBC

9 Table 1: Summary statistics of CDS Spreads (Jan.1, 2007~Sep.9,2008) (Cont.) UK Lloyds TSB UK RBS UK Standard Chartered France AXA France BNP Paribas France Credit Agricole France Societe Generale Germany Allianz Germany Commerzbank Germany HVB Germany Deutsche Bank Germany Dresdner Bank Germany Rueckversicherung Italy Banco Monte Dei Paschi Italy UniCredit SpA Netherlands ABN AMRO Netherlands ING Netherlands Rabobank Switzerland Credit Suisse Switzerland UBS Spain Banco Santader Portugal Banco Commercial Portugues

10 Empirical Methodology The empirical methodology used in this study is the combination of cluster analysis, the principal component analysis, the DAG (Bayesian Network) and structural (VAR) models Classify large worldwide financial institutions into several clusters Extract the major driving force behind the changes of CDS spreads in each cluster into principal components Apply DAG technique to find contemporaneous causalities among inter-cluster and intra-cluster financial institutions Run structural VAR analysis based on DAG to find economic significance of contemporaneous, short and longer-run dynamics 10

11 Empirical Methodology Cluster Analysis and PCA Cluster analysis: attempt to determine whether or not a data set contains distinct groups or clusters of variables (or observations). First, hierarchical clustering methods use a distance matrix as their starting point. Second, there are a variety of ways to measure how different the two clusters are. This depends on the distance between cluster pairs. Third, an index that can be used for choosing the number of clusters is the cubic clustering criterion (CCC). PCA is an orthogonal linear transformation for dimensionality reduction. Given a set of data, the first component (the eigenvector with the largest eigenvalue) corresponds to a line that passes through the mean and minimizes sum squared error with these points. 11

12 Empirical Methodology VAR Models and Innovation Accounting Let denote a vector of stationary changes in CDS spreads, which can be modeled in a vector autoregressive model (VAR): (1) (2) t = 1,2,,T. (3) (4) k 1 t i t i + + et ( t = i= 1 X t = A i ε t i i= 0 n t, n = Alε t+ n l l= 0 n ' Cov(ξ t, n ) = Al ΣAl l= 0 X = Γ X µ ξ 1,..., T ) where Σ is the variance-covariance matrix of the error term, as in equation (1). The remaining basic problem is how to orthogonalize the VAR residuals.

13 Empirical Methodology Directed Acyclic Graphs Analysis The DAG technique represents the recent advance in causality analysis. There are five possible edge relationships: No edge (X Y) indicates (conditional) independence between two variables. Undirected edge (X Y) signifies a covariance that is given no causal interpretation. Directed edge (Y X) suggests that a variation in Y, with all other variables held constant, produces a (linear) variation in X that is not mediated by any other variable in the system. Directed edge (X Y) has an analogous interpretation as (Y X). Bidirectional edges (X Y) denote the bidirectional causal interpretation between the X and Y. 13

14 Empirical Methodology Directed Acyclic Graphs Analysis Contrast with the well-known Granger causality Time series v.s. contemporaneous relation Bivariate v.s. multivariate relation The DAG (PC) algorithm In the elimination stage, the algorithm removes edges from the undirected graph, based on unconditional correlations between pairs of variables: Edges are removed if they connect variables that have zero correlation. The remaining edges are then checked for whether the first-order partial correlation, etc. Once the elimination stage is completed, the algorithm proceeds to the orientation stage. The notion of sepset is then used to assign the direction of contemporaneous causal flow between variables remaining connected after we check for all possible conditional correlations. 14

15 Empirical Results Table 2: Cluster Analysis This table reports 4 clusters of 43 largest financial institutions based on a hierarchical cluster analysis. The average linkage method is used to measure the distance between cluster pairs and the cubic cluster criterion (CCC) is used to choose the number of clusters. Cluster 1 (US GSEs) Cluster 2 (US Banks) Cluster 3 (US Insurances) Cluster 4 (EU Financial Instituitions) Fannie Mac Lehman Brothers Met Life Abbey Freddic Mac Bear Sterns Safeco Barclays Goldman Sachs American express HBOS Merrill Lynch AIG HSBC Morgan Stanley Chubb Lloyds TSB Bank of America Hartford RBS Wachovia Standard Chartered Citigroup JP Morgan AXA BNP Paribas Credit Agricole Societe Generale Allianz Commerzbank HVB Deutsche Bank Dresdner Bank Hannover Rueckversicherung Banco Monte Dei Paschi UniCredit SpA ABN AMRO ING Rabobank Credit Suisse UBS Banco Santader Banco Commercial Portugues 15

16 Empirical Results Table 3: Principal Component Analysis Cluster 1 (US GSEs) Cluster 2 (US Banks) Cluster 3 (US Insurances) Cluster 4 (EU Financial Institutions) 1st factor 90.04% 62.56% 66.13% 70.82% First 2 factors 100% 72.32% 80.23% 80.17% First 3 factors 79.24% 89.44% 82.50% First 4 factors 85.27% 94.14% 84.52% First 5 factors 89.54% 97.49% 86.49% 16

17 Empirical Results Figure 3 Contemporaneous causal flow patterns among US banks as well as other clusters Bank of America US GSEs US Insurance Citigroup Goldman Sachs Lehman Brothers Morgan Stanley JPMorgan Wachovia Bear Sterns Merrill Lynch 17

18 Empirical Results Table 6: Forecast error variance decomposition results (percentage) for US banks and first principle components of other groups Days US GSEs US Insurances EU financial Institutions Lehman Brothers Bear Sterns Goldman Sachs Merrill Lynch Morgan Stanley Wachovia Citigroup JP Morgan Bank of America Variance of US GSEs explained by the shocks of the 12 CDS spreads Variance of US insurances explained by the shocks of the 12 CDS spreads Variance of EU financial institutions explained by the shocks of the 12 CDS spreads Variance of Lehman Brothers explained by the shocks of the 12 CDS spreads Variance of Bear Sterns explained by the shocks of the 12 CDS spreads

19 Empirical Results: Table 6 (Cont.) Days US GSEs US Insurances EU financial Institutions Lehman Brothers Bear Sterns Goldman Sachs Merrill Lynch Morgan Stanley Wachovia Citigroup JP Morgan Bank of America Variance of Goldman Sachs explained by the shocks of the 12 CDS spreads Variance of Merrill Lynch explained by the shocks of the 12 CDS spreads Variance of Morgan Stanley explained by the shocks of the 12 CDS spreads Variance of Wachovia explained by the shocks of the 12 CDS spreads Variance of Citigroup explained by the shocks of the 12 CDS spreads

20 Empirical Results Table 6 (Cont.) Days US GSEs US Insurances EU financial Institutions Lehman Brothers Bear Sterns Goldman Sachs Merrill Lynch Morgan Stanley Wachovia Citigroup JP Morgan Bank of America Variance of J.P. Morgan explained by the shocks of the 12 CDS spreads Variance of Bank of America explained by the shocks of the 12 CDS spreads Note: The variance decomposition is based on the directed graph on innovations given in Figure 3. 20

21 Empirical Results Figure 4 Contemporaneous causal flow patterns among US insurance companies as well as other clusters EU financial Institutions US GSEs American Express Chubb AIG US Banks Safeco Hartford Met Life 21

22 Empirical Results Table 7: Forecast error variance decomposition results (percentage) for US insurance companies and first principle components of other groups Days US GSEs US Banks EU financial Institutions American Express AIG Chubb Met Life Hartford Safeco Variance of US GSEs explained by shocks to the nine CDS spreads Variance of US Banks explained by shocks to the nine CDS spreads Variance of EU financial institutions explained by shocks to the nine CDS spreads Variance of American Express explained by shocks to the nine CDS spreads Variance of AIG explained by shocks to the nine CDS spreads

23 Empirical Results Table 7 (Cont.) Days US GSEs US Banks EU financial Institutions American Express Variance of Chubb explained by shocks to the nine CDS spreads AIG Chubb Met Life Hartford Safeco Variance of Met Life explained by shocks to the nine CDS spreads Variance of Hartford explained by shocks to the nine CDS spreads Variance of Safeco explained by shocks to the nine CDS spreads Note: The variance decomposition is based on the directed graph on innovations given in Figure 4. 23

24 Empirical Results Figure 5 Contemporaneous casual flow patterns among European Societe financial institutions as well as other clusters Credit Hannover ING HSBC Generale Agricole Rueckversicherung US GSEs US Insurances Banco Santader BNP Paribas Robobank Lloyds TSB ABN AMRO US banks Deutsche HVB Barclays Abbey AXA Bank Standard Chartered Banco Monte Dei Paschi RBS HBOs Dresdner UBS Bank UniCredit SpA Commerzbank Allianz Banco Commercial Portugues Credit Suisse 24

25 Robustness Checks We conduct many robustness checks on the main results, including: re-estimate all the models by including second principal components of each cluster conduct a small scale simulation to assess the effectiveness of the PC algorithm in DAG analysis in this study consider an alternative sample period before the fall of Bear Sterns. Random errors, if substantial, should be biased toward no significant relationship between credit risk transfer and potential economic factors. Diamond and Rajan (2009) and Acharya and Viswanathan (2011) emphasize the role of short-term debt in affecting asset market liquidity under the crisis scenario. 25

26 An important index from perspective of interconnectedness: Table 8 Country Name I-index US Fannie Mae 1 US Freddie Mac 1 US Lehman Brothers 3 Country Name I-index France AXA 1 France BNP Paribas 3 France Credit Agricole 1 France Societe Generale 1 Germany Allianz 1 Germany Commerzbank 2 Germany HVB 2 Germany Deutsche Bank 0 Germany Dresdner Bank 3 Germany Hannover Rueckversicherung 1 Italy Banco Monte Dei Paschi 1 Italy UniCredit SpA 1 Netherlands ABN AMRO 0 Netherlands ING 0 Netherlands Rabobank 0 Switzerland Credit Suisse 1 Switzerland UBS 3 Spain Banco Santader 1 Portugal Banco Commercial Portugues 1 US Bearn Sterns 2 US Goldman Sachs 2 US Merrill Lynch 1 US Morgan Stanley 3 US Bank of America 2 US Wachovia 0 US Citigroup 0 US JP Morgan 0 US Met Life 2 US Safeco 3 US American express 1 US AIG 3 US Chubb 3 US Hartford 0 UK Abbey 1 UK Barclays 2 UK HBOS 1 UK HSBC 1 UK Lloyds TSB 1 UK RBS 2 UK Standard Chartered 1

27 Why are some financial institutions systemically important? Regress the index of systemic importance on size and leverage ratios Table estimate nobs Adj-R 2 estimate nobs Adj-R 2 estimate nobs Adj-R 2 Log market value (million USD) (-1.11) % (-1.25) % (-0.25) % Total debt/ Total asset 0.01 (1.58) % 0.01 (0.92) % 0.01 (0.77) % Short-term debt/ Total asset 2.00** (2.29) % 1.87* (1.89) % 1.09 (1.02) % Total debt/ common equity 0.05** (2.43) % 0.04* (1.78) % 0.03* (1.91) % Short-term debt/ common equity 0.08*** (3.32) % 0.06*** (2.61) % 0.04* (1.80) %

28 Why are some financial institutions systemically important? (continued) Regress the index of systemic importance on various corporate governance indictors Table 9 (continued) estimate nobs Adj-R 2 estimate nobs Adj-R 2 estimate nobs Adj-R 2 G-index 0.11 (1.10) % E-index (0.00) % 0.10 (0.54) % Board size (-1.23) % % independent director (-0.85) 17-5%

29 Why are some financial institutions systemically important? (continued) Regress the index of systemic importance on liquidity risk and write-down (counterparty risk) measures Table 9 (continued) estimate nobs Adj-R 2 estimate nobs Adj-R 2 estimate nobs Adj-R 2 Average Ask-bid 0.28 (1.09) % 0.11 (1.21) % Average normalized Ask-bid (-1.31) % (-0.64) % Write-down (billion USD) 0.02 (0.82) % (-0.46) % % of write down / total asset 0.23 (0.71) % 0.08 (0.77) % % of write down /market value 0.01 (0.68) % (1.05) %

30 Conclusions First, institutions to be identified as the SIFIs may be those prime senders of credit risk information, such as Lehman Brothers, Morgan Stanley, Sefeco, Chubb, and possibly AIG in the US and BNP Paribs, Dresdner bank, and UBS in the Europe Second, the institutions which are the exchange center of credit risk information could also be systemically important. Goldman Sachs, Bear Sterns, Bank of America, and Metlife in the US and Barclays, RBS, Commerzbank, and HVB in the Europe belong to this category. Leverage ratios and particularly the short-term debt ratio are significant determinants of identified different roles of financial institutions in credit risk transfer, while no such evidence is found for corporate governance indexes, size, liquidity and asset write-downs.