United States
Financial Sector Assessment Program-Stress Testing-Technical Notes

This Technical Note discusses key findings of stress testing on the United States under the Financial Sector Assessment Program. Several stress tests were used to quantify the potential impacts of risks and vulnerabilities in banking and non-banking sectors. The stress tests run by the authorities and by companies under the Dodd-Frank Act (DFA) suggest that most large bank holding companies (BHCs) are resilient to shocks similar to the last crisis. For BHCs, the IMF staff’s solvency stress tests over the initial stressed period are largely in line with the DFA stress testing results, and suggest that the system is generally robust, although some BHCs would fall below the hurdle rate in the stressed environment.


This Technical Note discusses key findings of stress testing on the United States under the Financial Sector Assessment Program. Several stress tests were used to quantify the potential impacts of risks and vulnerabilities in banking and non-banking sectors. The stress tests run by the authorities and by companies under the Dodd-Frank Act (DFA) suggest that most large bank holding companies (BHCs) are resilient to shocks similar to the last crisis. For BHCs, the IMF staff’s solvency stress tests over the initial stressed period are largely in line with the DFA stress testing results, and suggest that the system is generally robust, although some BHCs would fall below the hurdle rate in the stressed environment.


13. This note provides the methodology and results of stress tests of the financial sector carried out in the 2015 FSAP assessment for the United States. To obtain a more comprehensive assessment than possible with any single approach, the U.S. FSAP stress tests combined three groups of complementary approaches. The first group consisted of the IMF’s top-down stress tests for BHCs, insurance companies, and mutual funds. The second group included the Fed’s top-down (supervisory DFAST) tests for BHCs and bottom-up stress tests run by the companies (company run DFAST). The third group included a broad range of IMF’s top-down calculations using market-price data (Table 2 and Appendix I). The findings of the stress tests were used to provide quantitative support for the FSAP’s stability risk assessment by estimating the impact from the realization of key tail risks and to facilitate policy discussions on risk mitigation strategies and crisis preparedness.

Table 2.

Stress Testing: Overview of the Exercises Done by the IMF

article image
Note: for details, see the Stress Test Matrix (Appendix Table 2). Table 2 focuses on IMF-run stress tests and does not include the supervisory and companies-run stress tests that informed this exercise.

IMF Staff’s Solvency Stress Tests for Bank Holding Companies

14. This section explains the top-down solvency stress tests of the IMF FSAP team. The section covers: (i) overview of scope, (ii) the state of the banking sector; (iii) the macroeconomic scenarios; (iv) the capital definitions and standards that were used for calculating and reporting results; (v) the stress test methodology and the use of models to map the macroeconomic scenarios into credit losses, income projections, balance sheet items and risk weighted assets; (vi) the behavioral assumptions governing capital actions in the stress test scenarios, (v) the network analysis performed and (vii) the results of the stress test.

A. Scope of the Test

15. The top-down test followed the balance sheet-based approach. This assesses solvency of individual BHCs under the baseline and stress scenarios through changes in net income and risk-weighted assets. A range of sensitivity analysis was performed to supplement the two scenarios. This approach was comparable to the company run DFAST (bottom-up) and supervisory DFAST (top-down), despite some important differences (Table 12). It can be seen as complementary to the DFAST exercise.

Table 3.

Variables Used in the IMF Stress Test

article image
Source: IMF Staff.
Table 4.

Capital Standards


article image

GSIB surcharge factor (the factor was multiplied by GSIB surcharge).

Applied to intangible assets and DTAs.

For advanced approached BHCs only.

Source: FRB, IMF Staff.
Table 5.

Capital Standards for Advanced Approach BHCs and Other BHCs

article image
Source: FRB, IMF Staff.
Table 6.

Balance Sheet Mapping

article image
Source: FRB, IMF Staff.
Table 7.

Loan Portfolio Mapping

article image
Source: FRB, IMF Staff.
Table 8.

Projection Exercise

article image
Source: IMF Staff.
Table 9.

Models of Net Charge-Offs

article image
Source: IMF Staff.
Table 10.

Dividend Distribution Schedule\1

article image

GSIB surcharge was also taken into account (not shown in the table).

Source: IMF Staff.
Table 11.

Simulation Results of Credit and Funding Shock with Risk Transfers

United States: Simulation results of credit and funding shocks with risk transfers. Q32014

article image
Source: FRB, IMF Staff calculations.
Table 12.

Main Differences Between IMF’s Top-Down and FRB’s Top-Down Approach

article image
Source: IMF Staff

16. The stress test used publicly available, consolidated data. These cover individual BHCs from regulatory reports (FR Y-9C) provided by SNL Financial that take into account structural breaks resulting from bank mergers and acquisitions.

17. The coverage of the IMF top-down test was the same as in the DFAST, which increased the comparability of results. Results of the test were calculated by individual institution. The stress test covered 31 largest BHCs (with total consolidated assets of $50 billion and more), which account for about 85 percent of the BHC assets and 70 percent of total banking sector assets, defined as total assets of BHCs, savings and loans holding companies (SLHC) and commercial and savings banks that are not part of any BHC or SLHC (Figure 1). The network analysis was based on six largest BHCs, accounting for 52 percent of total BHC assets.

Figure 1.
Figure 1.

Structure of the U.S. Banking Sector

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: Federal Reserve Board, IMF Staff calculations.

18. While the test’s coverage in terms of total banking sector assets is comparable with other FSAPs, some large depository institutions were not included. BHCs with assets of $10 billion and more but less than $50 billion (which represent around 5 percent of the BHC assets) were not included. Most banking organizations (commercial banks and saving institutions) with assets of $50 billion and more that are regulated by OCC and the FDIC and are subject to bottom-up tests (company-run DFAST), were implicitly included in the top-down stress test, as they are subsidiaries of BHCs included in the top-down stress test. The exceptions are two banks (one regulated by the FDIC and the other by OCC) that are part of BHCs that will be subject to supervisory DFAST in the future and two banks that are part of large SLHCs (regulated by OCC)2. These institutions were stress tested as part of a robustness check (Figure 1). Moreover, one depository institution—the largest credit union—has assets exceeding $50 billion but was not included in the stress test, because it was not subject to risk-based capital requirements used by other federal banking regulatory agencies as of 2014Q3.3 Large SLHCs were not included in any stress test as they were not subject to capital requirements as of 2014Q3.4 SLHCs with assets of $50 billion and more represent 10 percent of total holding companies’ assets with assets of $50 billion and more. They will be required to perform DFA company-run stress tests in the future.

19. The cut-off date for the data was September 30, 2014. Minimum capital requirements used as hurdle rates were consistent with the revised capital regulatory standards that reflect Basel III capital standards including both the capital conservation buffer and a GSIB capital surcharge (calculated using the BCBS framework), on a phased-in basis, as minimums. The hurdle rates in the IMF stress test were more stringent than in the DFAST’s, which were consistent with the Basel III transition schedule and did not include capital conservation buffer or a GSIB capital surcharge.

B. Bank Holding Companies: An Overview

20. This section provides an overview of BHCs included in the stress tests (Appendix Figure 1). It analyzes the structure of balance sheets and income statements as well as off-balance sheet items. Moreover, it provides some detailed information on GSIBs.

21. Assets of 31 largest BHCs rose by 11 percent since the last FSAP and 18 percent since the crisis. Total assets of the largest BHCs represent 80 percent of nominal GDP. The increase in assets was primarily driven by increases in cash, federal funds bought and reverse repos and available-for-sales securities portfolio. Much of this growth reflects impacts of Quantitative Easing policies on bank balance sheets—impacts that are likely to be at least partially reversed with the upcoming unwind. Cash now accounts for 12 percent of total asset compared to 3 percent before the crisis. The increase in federal funds and repo mostly reflect a large increase in 2009. Securities holding have expanded by 60 percent since 2008 due to increases in holdings of available for sale (AFS) securities—as holdings of mortgage-backed securities (MBSs), Treasury securities and foreign debt securities increased—which increases BHCs exposure to interest rate risk.5 Half of the AFS portfolio pertains to MBSs, followed by Treasury securities (16 percent) and foreign debt securities (14 percent). As of 2014Q3 trading assets are lower by about 10 percent comparing to 2008 which is partly due to the implementation of the Volcker rule which severely restricts proprietary trading.

22. Total net loans are the largest asset category accounting for 40 percent of total assets, slightly lower than in 2008 or before the crisis. While total loans have increased by 10 percent the structure of loans has changed since the crisis. Real estate loans, the largest loan type, account for 28 percent of total loans, down from 36 percent in 2008. This was mainly due to lower revolving, open end real estate loans extended under lines of credit and close-end junior lien real estate loans as underwriting standards for those loans tightened considerably mostly by requiring more documentation and by imposing debt service ratio (OCC, 2014). On the other hand, loans to financial institutions increased their share to about 10 percent driven by increases in loans to non-depository financial institutions and loans for purchasing securities. Rapid loan growth since the beginning of 2013 driven by business loans calls for continued vigilance given evidence of weakening underwriting, especially in the leveraged loans market.6

23. Deposit growth, which accounted for the bulk of funding growth, supported the growth of assets. Deposits are 40 percent larger than in 2008 and account for 53 percent of total liabilities and 132 percent of loans. The deposit-to-loan ratio is 21 percentage points higher than in 2010 and 30 percentage points higher than before the crisis partly due to record corporate cash holdings. About 60 percent of deposits are money market deposits. Stable deposits7 account for almost 90 percent of total deposits and less stable deposits have significantly decreased since the crisis due to lower large, short-term domestic and foreign time deposits.

24. Deposit growth, along with deleveraging, has reduced BHCs’ reliance on wholesale funding. Non-deposit liabilities such as repos, trading liabilities and other wholesale funding are 10 percent lower than in 2009. Wholesale funding (defined as repos, trading liabilities, subordinated notes and brokered deposits) account for 30 percent of total liabilities. The bulk of wholesale funding pertains to other borrowed money (50 percent) and repos (25 percent). The maturity of other borrowed money has been extended since 2008 and most of other borrowed money (70 percent) in 2014 is related to unsecured liabilities and liabilities with maturity of 1 year and more.

25. Total equity has increased by 70 percent driven by retained earnings and surpluses, which have doubled since the crisis, largely in response to the higher regulatory requirements. CET 1 capital ratio has doubled since the crisis to 12 percent at the end of 2014Q3. The leverage ratio (defined as CET1 over total assets) has more than doubled to 8 percent since the end of 2008.

26. Off balance sheet activity has fallen since 2009 mostly due to lower holdings of derivatives, notwithstanding an increase in unused commitments. While the derivatives (credit equivalent) have fallen by 30 percent, unused commitments have increased by 5 percent since the crisis and still represent the largest off balance sheet item. The largest share of unused commitments pertains to consumer credit card lines (40 percent) and commercial and industrial loans (25 percent). The structure of unused commitments has changed since the crisis. Although unused credit card lines have fallen, unused commitments on commercial and industrial loans and loans to financial institutions have increased by 30 percent. Securities lending is the second largest off balance sheet item with a share of around 20 percent of total off balance sheet activity. Interest rate contracts are the largest component of derivatives portfolio (82 percent) followed by foreign exchange contracts (14 percent). Credit derivatives have been cut in half since 2009 and represent a small proportion of derivatives activities where most contracts are related to purchased or sold investment grade credit default swaps. Swaps and forward contracts dominate the derivative contracts. Almost the whole derivative portfolio is held for trading. In most of OTC derivatives transactions cash is the main collateral and major counterparties are banks and securities firms and non-financial corporate firms.

27. While BHCs have made material improvements in nonperforming loans, underwriting standards have continued to loosen since 2011 (Appendix Figure 2). Economic recovery has been conducive to further strengthening of the BHCs’ balance sheets. Delinquent and non-performing loans have continued to fall since their peak in 2009. Delinquency rates and NPLs have been cut in half since end 2009 and now stand at 3.5 and 2.5 percent respectively. Most of bad loans consist of residential mortgage loans. Non-real estate mortgage delinquent loans are at the levels before the crisis. Net charge-off rates are considerably lower than in 2009 but still higher than before the crisis due to higher charge-offs for consumer loans. However, regulatory surveys from the OCC suggest looser underwriting in commercial real estate, commercial and industrial loans, and auto loans with some banks having significant exposure to subprime auto loans. Moreover, LTVs for CRE loans are approaching their pre-crisis levels suggesting continuing monitoring is needed. The largest BHCs seem resilient to the recent oil price drop since their direct loan exposure to energy-related companies is only in the range of 1.2 to 5 percent their total loans.

28. While the BHCs have posted all time high profits in 2014Q3, there is a large dispersion of profitability indicators across BHCs. Net income has increased substantially since the last FSAP driven by lower provisions which have come down to pre-crisis levels and higher non-interest income. Net interest margins continue to compress as a result of protracted low interest rates, banks’ increased holding of liquid assets because of regulatory requirements as well as heightened loan competition. Non-interest expenses are 35 percent higher in 2014Q3 than in 2008, partly due to litigation-related charges and cyber security protection. Return on equity (ROE) is about 50 percent smaller than before the crisis mainly due to higher capitalization of BHCs. Return on assets are 20 bps lower than before the crisis. While profitability of all BHCs is relatively high there are large differences across BHCs, which, for some large BHCs, is also due to litigation charges stemming from BHCs’ business practices leading up to the crisis.8 Many BHCs are seeking to enhance their ROE by looking for new business, principally through new loan growth or reconfiguration of business models which increases the risk of relaxation of underwriting standards.

29. There are large differences in business models across BHCs. BHCs can be differentiated based on the largest asset category. Most of the BHCs are focused on lending as the main business activity. The second type of BHCs is more involved in capital market activities. The largest asset item of the third type of BHCs pertains to AFS securities. In general, BHCs with high proportion of loans have lower leverage, lower off balance sheet activities and are less involved in wholesale activities such as reverse repo and trading. They use deposits as a major source of funding and are more profitable than other types of BHCs. On the other hand, BHCs that have large trading activities are less involved in lending but more involved in reverse repo transactions. They are funded more on the wholesale market and have higher leverage. They also have higher NPLs, which might imply that they can’t compete with BHCs whose lending represents their core business or that they are less constrained by regulatory capital ratios and are searching for yield by targeting riskier loans. The third type of BHCs, of which the largest three BHCs have large operations as custodian banks, is less involved in lending, trading and repo transactions. They have higher leverage and off balance sheet activities but at the same time large deposit base.

30. Total assets of GSIBs have reached about $10 trillion and represent 75 percent of total assets of all BHCs included in the stress test (Appendix Figure 3). When derivatives positions and securities financing transactions are added their total exposure is about 35 percent higher than their total assets. Around 25 percent of their assets are related to foreign exposures and on average, they derive about ¼ of their total net revenue from foreign business. The largest component of foreign loans pertain to commercial and industrial loans. Securitization, mostly of mortgage loans, represents 11 percent of GSIBs assets but with large differences across BHCs. Securities (trading and AFS) account for 30 percent of their total assets. The structure of investment and AFS securities portfolios is very similar- MBSs (45 percent) and Treasury securities (around 20 percent) represent almost 2/3 of portfolios.

31. GSIBs are interconnected with the rest of the financial system. Intra-financial system assets represent 22 percent of GSIBs’ assets while intra-financial system liabilities represent 16 percent of their total liabilities with notable differences across GSIBs. The largest component of inta-financial system assets is the fair value and potential future exposure of OTC derivatives (52 percent), followed by deposits (20 percent). Most of the intra-financial liabilities pertain to deposits (47 percent; most of deposits were due to non-bank financial institutions) and OTC derivatives (35 percent). Almost half of OTC derivatives are cleared through a central counterparty.

C. Macroeconomic Scenarios

32. The solvency stress tests examined two macroeconomic scenarios: a baseline and a stress scenario over a five year horizon (Box 1, Appendix Figure 4). These scenarios were developed by the FRB in consultations with the OCC and the FDIC (over July and August 2014).9 The scenarios consisted of the future paths of 28 economic and financial variables (six measures of economic activity and prices, four measures of developments in equity and property prices, six measures of interest rates and variables for the euro area, the United Kingdom, developing Asia, and Japan).

33. The baseline scenario and the stress scenario over the initial three years reflected the supervisory baseline scenario and the severely adverse scenario under the Dodd-Frank Act Stress Tests (DFAST), respectively. The baseline scenario was very similar to the IMF’s latest WEO projections for the first three years of the horizon. The stress scenario reflected the severely adverse scenario under the DFAST10 for the first three years of the forecast horizon (up to 2017Q4). The scenarios were characterized by 10 variables from the DFAST and three additional variables not included in the DFAST (Table 3).

34. For the first three years of the horizon (from 2014Q4 to 2017Q4), the IMF staff adhered to the supervisory scenarios. For the additional years (from 2018Q1 to 2019Q4), the paths for a selected subset of key indicators (GDP growth, unemployment, short- and long-term interest rates) were extended based on the latest WEO projections for the baseline. For the stress scenario, the paths for the key indicators were extended so as to converge to the baseline by the end of the horizon. The other variables from the DFAST (house prices, commercial real estate prices, VIX, Dow Jones stock price index, BBB corporate yield and mortgage rates) were extended using simple OLS regression models and projections for key indicators as exogenous variables (Table 3). Business and consumer interest rates11 and federal funds rate were added to the set of variables from the DFAST and were projected from 2014Q4 using regression models and projections of variables from the DFAST.

35. The baseline scenario was very similar to the average projections of economic forecasters.12 It reflected a sustained, moderate expansion of U.S. economic activity converging to a growth rate of about 2 percent and unemployment rate reaching 4.4 percent by the end of 2019. Gradual normalization in federal funds rate and Treasury yields starts in second quarter of 2015. Interest rates on mortgage loans, consumer loans and business loans follow broadly the dynamics of short-term rates. All assets prices (equity, house and commercial property) rise steadily accompanying the modest expansion of economic activity.

36. The stress scenario was similar in severity to the 2007–09 recession. The stress scenario was based on the severely adverse scenario of the DFAST, which was deemed appropriately stressful from the FSAP viewpoint. Nonetheless, the scenario horizon was expanded to 5 years, bringing it closer to recent FSAPs. Following the approach adopted in the DFAST, the trajectories and co-movements of key variables were informed by post-war U.S. recessions, with scenario severity calibrated to be similar to the 2007–09 recession. The unemployment rate was used as the primary basis for specifying the scenario13 and the other variables were set using a combination of economic models, typical paths of these variables in past recessions, and informed judgment.14 The severely adverse scenario in the 2015 DFAST (the shock was applied from 2014Q4) was characterized by a 4 percentage point rise in the unemployment rate over a two-year period. It was assumed that: real GDP would be on average 6.6 percentage points lower than the baseline in 2015 (Figure 2);15 equity prices would fall by 60 percent in the first year; house prices would decline by 25 percent over the first two years; corporate spreads would rise significantly in 2015, reflecting a deterioration of U.S. corporate credit quality; mortgage rates would increase by 80 basis points; and market volatility would rise to levels the same as the peaks reached in the 2007–09 recession. Short-term interest rates would remain at zero by end of 2017, reinforcing the negative effects from protracted period of low interest rates, after which normalization would start. Long-term Treasury yields would first drop to 1 percent in 2014Q4 and then edged up slowly over the remainder of the stress testing horizon. The scenario also included a rise in oil prices to about $110 per barrel possibly reflecting a materialization of geopolitical risks. After 2017, most of the variables were assumed to converge to the levels in the baseline scenario. The stress scenario was complemented with sensitivity analyses to estimate the marginal impact of individual risks not captured by the scenario (the interest rate spike in particular). The supervisory DFAST calculations of net income losses also incorporated projected losses generated by operational risk events such as fraud, computer system, or other operating disruptions.

Figure 2.
Figure 2.

GDP Growth in the Baseline and Stress Scenario

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: FRB, IMF Staff calculations.

D. Capital Standards

Capital definitions

37. The capital definition applied in the stress test corresponded to Basel III capital standards. This was applied to all BHCs, recognizing that only advanced approaches BHCs were subject to Basel III capital rules in 2014, and non-advanced approached BHCs became subject to the rules from January 1, 2015.

38. Hurdle rates included the CET1 minimum requirement, the capital conservation buffer, and the GSIB surcharge. The solvency stress test assessed the level of BHCs common equity Tier 1 ratios of both advanced approaches and non-advanced approaches BHCs against the regulatory threshold consistent with the Basel III transition schedule but also accounting for capital conservation buffer and a G-SIBs capital surcharge (calculated using the BCBS framework), as minimums (Table 4). The phase-in for the Basel III framework that began during 2014 and the revised capital framework that introduced a new standardized approach to RWAs starting in 2015 were also considered (Table 5). A common equity surcharge associated with G-SIB status was also taken into account. It ranged from 1.0 to 2.5 percent, following Financial Stability Board (FSB) buckets corresponding to required level of additional loss absorbency, and it was phased-in between January 1, 2016 and end of 2018.16,17

39. CET1 capital for the base period was estimated for non-advanced approaches BHCs. Non-advanced approaches BHCs became subject to Basel III capital rules from 2015Q1 and did not report CET1 capital (on Schedule HC-R, Part I.B. of FR Y-9C) as of September 2014. Therefore, SNL’s estimate of CET1 (after deductions and adjustments) were used. SNL calculates the CET1 as: Tier 1 Capital - Non-qualifying Perpetual Preferred Stock - Preferred Stock & Surplus - Qualifying: Non Controlling Interests—Qualifying: Restricted Core Capital Elements—Qualifying: Mandatory Convertible Securities.18

40. Deductions from CET1 were needed to calculate phase in of deductions from CET1. Most deductions pertained to goodwill, intangible assets, and deferred tax assets (DTAs).19 To calculate phase in of deductions from CET1 the following strategy was implemented:

  • The deductions were reported by the advanced approaches BHCs only (in Schedule HC-R). For non-advanced approaches BHCs intangible assets deducted from CET1 were approximated by intangible assets other than goodwill and mortgage servicing assets (MSA) (from Schedule HC-M) adjusted for deferred tax liabilities (DTLs) associated with intangible assets.20 A deduction related to goodwill, net of deferred tax liabilities, were reported by all BHCs. A deduction related to DTAs for non-advanced approaches BHCs was approximated by DTAs deducted from Tier 1 capital.

  • The nominal value of all deductions was assumed to stay constant over the stress testing horizon (as in the supervisory DFAST). Each deduction had to be considered separately since there was no transition provision for goodwill while intangible assets and DTAs followed the Basel III transition provisions.21 Deductions were calculated by multiplying intangible assets and DTAs by the transition provision factor. Only the difference between the deduction in period t+1 and t was subtracted from CET1 capital in period t since CET1 in period t was already defined as CET1 after adjustments and deductions.22 No assumptions were made about banks’ behavioral responses to phase-ins.

41. The treatment of accumulated other comprehensive income (AOCI) reflected the Basel III transition arrangements. Consistent with Basel III transition arrangements, only 20 percent of AOCI was incorporated into CET1 capital in 2014 and additional 20 percent in every year after 201423 for advanced approaches BHCs. Consistent with the supervisory DFAST, it was assumed that non-advanced approaches BHCs would opt out of including AOCI. In comparison to Fed’s stress test that held the components of AOCI other than unrealized gains (losses) on AFS securities constant over the planning period, in the IMF top-down test the aggregate AOCI was modeled as the structure of AOCI is not publicly available information. As in the case of deductions, only the difference between AOCI in period t+1 and t was added to CET1 capital in period t since CET1 in period t was already defined as CET1 after AOCI.

Risk-weighted assets

42. Total risk weighted assets (RWAs) were projected for each BHC that participated in the stress test. The two components of RWAs (credit RWAs for total assets and off-balance sheet items and market RWAs24) were challenging to model separately with publicly available data,25 and making simplified assumptions about each component of total RWAs could yield misleading results.26 Nonetheless, the dynamics of total RWAs followed closely the dynamics of total assets which was projected in the exercise. Therefore, the year-on-year growth rate of total RWAs was modeled in a panel regression model with fixed effects as a function of year-on-year growth rate of total assets. Interest rates were added as an exogenous variable to reflect the assumption that the credit portfolio’s underlying risk features does not remain constant27 throughout the horizon thereby making the projection of RWAs risk sensitive. While BHCs can qualify for using the advanced approach credit risk RWAs from January 1, 2016,28 it was assumed that the relationship between RWAs and total assets found before 2016 would hold also after 2016.

43. Operational risk RWAs were included in calculation of total RWAs for advanced approaches BHCs that exited the parallel run, given the requirements of the Collins amendment.29 Operational risk capital charge for non-advanced approaches banking organizations and advanced-approaches BHCs that have not exited the parallel run was not applied. Since there was no meaningful way to project RWAs for operational risk, it was assumed that the share of operational RWAs in total assets stays the same over the stress testing horizon. RWAs for advanced approaches BHCs that exited the parallel run were projected without operational RWAs. Projection of operational RWAs was then added to projected RWAs to calculate projected total RWAs.

44. The increase in risk weighted assets due to the implementation of standardized approach was applied to projected total RWAs.30 Credit RWAs under the standardized approach were not possible to calculate as publicly available data were not granular enough to apply the new weights to calculate credit RWAs.31 The increase in credit risk RWAs, due to introduction of standardized approach, was applied in 2015 onwards based on the calculated average increase in RWAs reported for the 2014 DFA stress testing exercise.32 The average increase of RWAs due to implementation of standardized approach was 9 percent which is in line with BCBS Basel III monitoring exercise estimates of RWA changes due to Basel III rules, as per Table A.13 in BCBS (2014).33

E. Models and Behavioral Assumptions

45. Quarterly data from 1991 to 2014Q3 from FR Y-9C report and a set of panel regression models were used to forecast each BHCs’ main components of balance sheets and income statements (Figure 3).34 Projections of balance sheets (Step 1, Table 8) over the stress testing horizon were used for the purposes of projecting total RWAs and income statement items (Step 2). Projections of RWAs and net income, with assumptions on dividend distribution, Basel III deductions and AOCI determined capital requirements over the stress testing horizon (Step 3). In comparison to the DFAST, asset disposals and acquisitions over time were not considered.

Figure 3.
Figure 3.

IMF Stress Testing Framework: Bank Solvency

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF Staff.

46. The models used were intended to capture how the balance sheet, RWAs, and net income of each BHC are affected by the macroeconomic and financial conditions (that served as independent variables) described in the scenarios. In those cases where the panel modeling approach was not appropriate, due to highly volatile individual bank data or insignificant relationships with macroeconomic and financial variables, modeling the particular variable at the aggregate level was tried. Projections of aggregate variables were then distributed to each BHC based on their market share or 2014 DFAST results.

Balance sheet growth projections

47. The growth rate of total assets was assumed to be equal to the growth rate of the largest asset category, accounting for smaller volatility of the growth rate of total assets.35 In most cases, this meant that assets grew in line with total loans. In several cases, that meant that assets grew in line with trading assets. The projection of BHCs’ total assets and loans was used for projecting income statement items and RWAs.36

48. Modeling individual BHC’s total loans proved to be more straightforward than modeling each loan portfolio item (residential real estate, CRE, business loans, consumer loans, loans to foreign governments, loans to financial institutions and other loans, Table 6 and 7).37 A panel, fixed-effects model of a year-on-year growth rate of net total loans was estimated and projected at the bank by bank level. It was assumed that lower economic activity, increases in interest rates and higher market turbulence (as a proxy for risk aversion) would lower the demand for loans. Therefore, independent variables included: year-on-year growth rate of real GDP, year-on-year changes in interest rates, and the VIX. The model also included lagged dependent variable to account for persistence in the growth rate of loans.38 The growth rate of loans was used in projection of net interest income and total assets and deposits. Based on a strong historical relationship between year-on-year growth rates of loans and deposits (correlation equal to 0.85) it was assumed that the growth rate of deposits is equal to the growth rate of loans over the stress testing horizon. In the sensitivity analysis different growth rates of loans and total assets (in line with Fed’s top down approach) in the stress scenario were analyzed.

49. Total assets of Goldman Sachs and Morgan Stanley were primarily determined by the dynamics of their trading assets, which were projected separately.39 Trading assets were projected in a panel regression with fixed effects where year-on-year growth rate of trading assets of both companies were determined by VIX, real GDP growth rate, and the federal funds rate. The first two regressors were expected to control for factors that affect trading activity whereas the interest rates was included as a regressor to control for the effects on asset prices. The growth rates of total assets for the two companies over the stress horizon were assumed to be equal to the growth rate of projected trading assets.

Loan losses and net income projections: methodology

50. The projections of revenues, expenses, and loan losses were based on the IMF’s projections of the balance sheet for each BHC over the planning horizon. Most components of pre-provision net revenue (including components of net interest income, noninterest income, and noninterest expenses, Table 8) were modeled using data on historical revenues and operating and other non-credit-related expenses reported on the FR Y-9C report in a simple panel regression model framework (Tables 23 and 24).40 Projections of all independent variables were taken from the scenarios.

51. Provisions for loan losses.

  • In the first approach, which was used to calculate the results in the benchmark case, aggregate provisions were modeled as the ratio of total provisions over aggregate net loans as a function of real GDP growth, differenced unemployment rate, credit spreads, growth rates of house prices, growth rate of VIX41 taking into account that effect on provision in periods of stress are more pronounced.42 Projected aggregate provisions were distributed among BHCs using individual bank’s share of provisions in total provision in the 2014 DFAST exercise. This approach was used to benchmark projections of provision using other approaches as the model of aggregate provisions managed to capture the spike in provisions during 2008/2009 crisis.

  • In the second approach, used in the sensitivity analysis, it was assumed—as in the last FSAP’s stress testing exercise—that provisions are equal to net charge offs. Models of net charge off rates by loan types43 that capture the historical behavior of net charge-offs over corresponding type of loans relative to changes in macroeconomic and financial market variables were considered (Table 9). The predicted net charge-off rates were multiplied by loan balances. The growth rate of each loan type was assumed to be equal to the projected growth rate of total loans. While there was a close relationship between provisions and net charge off rates during normal times, the disadvantage of this approach was that provisions increase more quickly than realized net charge-offs during stress times.44 The IMF team’s analysis showed that while the net charge offs dynamics compares well to the dynamics of provisions, they lag 2 to 3 quarters.45 In the sensitivity analysis, the effect of projecting total net charge-offs and total provisions (instead of by loan types) was also explored.

52. Net interest income. Net interest income was projected using fixed effects panel regression and the annual difference of net interest income as the dependent variables and the annual difference of a product of total net loans and loan interest rates and the annual difference of a product of total interest bearing deposits and deposit rate46 as explanatory variables. By including loans and deposits as an independent variable the macroeconomic environment as well as bank specific characteristics were taken into account. A projection of loans was taken from projection exercise of BHC’s balance sheets and interest bearing deposits were assumed to grow at year-on-year growth rate of loans. Deposit interest rates were assumed to be equal to the Federal funds rate. For the purposes of estimation, the loan interest rate was defined as a weighted average of mortgage, business, consumption lending rate and federal funds rate that approximated the inter-banking interest rate adjusted for each BHC’s loan portfolio structure.47

53. Non-interest income excluding trading income. This item was projected using a panel regression model with non-interest income (excluding trading income) over total assets as the dependent variable and the growth rate of VIX, unemployment rate and lending rates as independent variables. It was expected that during period of market turbulence (higher growth rates of VIX) trading from fees and commissions goes down as brokerage, underwriting, securitization fall. The same would be true if unemployment is high implying economic activity is low. Lending rates were included in the regression to control for substitution effect—it was expected that higher lending rates would make BHCs shift from non-interest income activities to interest income activities.

54. Trading income. Trading income was modeled as aggregate trading income that includes gains on AFS and hold to maturity (HTM) securities over total aggregate assets in a regression with the following independent variables: year-on-year growth rate of VIX, the interaction term between year-on-year growth rate of VIX and a dummy variable that took the value of 1 when the growth rate of VIX was positive to account for any non-linearity between market volatility and trading losses in times of stress, the change in credit spread48 and the change in term spread. Projected trading income was distributed among BHCs based on their 2014 share of trading income and gains on AFS and HTM securities in total trading income that includes total gains on AFS and HTM securities.

55. Non-interest expenses. This item was modeled as a year-on-year growth rate in a panel regression with year-on-year growth rate of total assets as the only independent variable and fixed effects. The assumption was that non-interest expenses depend on the size of the business which is ultimately related to the size of the balance sheet.

56. Taxes. Taxes were set at 28 percent—the pooled average level of the tax rate over the last 25 years.

57. Extraordinary items and minority interest. It was assumed that these items are equal to zero as, in general, this item did not contribute much to the net income.

58. AOCI. Aggregate total AOCI49 was modeled as the ratio of AOCI to aggregate assets in a regression with yearly change in BBB yields, 10 year Treasury bond, real GDP growth and the interaction term between real GDP growth and a dummy that takes value of 1 when GDP growth rate was negative to account for potential “non-linear” effects on unrealized losses in downturns. Projected total AOCI was distributed among BHCs using proportions of individual BHCs’ AOCI losses in total AOCI losses in the 2014 DFAST exercise.

59. The following income statement items were not considered: (i) losses related to operational risk events, mortgage repurchases, or OREO (ii) HFS/FVO loan losses as the data on these items were not publicly available and (iii) deferred tax assets (DTAs).

Capital action assumptions

60. A dividend distribution rule was defined where dividend distribution depends on the CET1 ratio. It was also assumed that BHCs do not issue new shares or make repurchases during the stress test horizon similar to the supervisory DFAST.50 In comparison to DFAST assumptions on dividend payments where common stock dividend payments continue at the same level as in 2014, the following rule for determining dividend payments was assumed:

  • Dividend payouts were payable out of the current year’s profit using the Basel III capital conservation rule taking into account transition provisions and GSIB surcharge (Table 10). Dividends were assumed to be paid out of current period net income after taxes by BHCs that were in compliance with the capital requirement equal to the hurdle rate. A maximum allowed dividend payout was assumed to be equal to the dividend payout ratio (dividends over net income after taxes) in 2014. If a bank fell below the hurdle rate before dividend distribution, it was considered capital constrained and followed a schedule of dividend payouts per Table 10. If a bank fell below the hurdle rate because of dividend distribution, it was assumed that the bank’s dividend payout was limited to a level that ensures the hurdle rate is not breached. This rule applied only if a bank earned a positive net income. If net income was negative it was assumed that there was no dividend payout. If a bank was above the threshold it paid a maximum allowed proportion of dividend.

  • In the sensitivity tests, an assumption that dividends remained fixed (in nominal terms) at their 2014 value was explored.

F. Sensitivity Analyses

61. To account for some of the differences between the IMF top-down stress test and the supervisory DFAST a range of sensitivity analyses were performed. These included: (i) assuming that loans and balance sheet grow at the similar rate as in the supervisory DFAST; (ii) assuming constant dividend distribution in the stress test like in the supervisory DFAST; (iii) assuming hurdle rates from DFAST and assuming that total RWAs are a sum of credit risk RWAs and market risk RWAs (and not including operational RWAs) and (iv) assuming all the elements of DFAST in the same scenario (constant dividends, fixed loan supply and total RWAs defined as a sum of credit and market RWAs only). Moreover, additional sensitivity analyses were performed: (i) including an oil price shock in the stress scenario; (ii) calculating the impact of a large interest rate shock in first year on the stress scenario; (iii) using different measures to calculate provisions in the stress scenario and (iv) extending the scope of the stress testing exercise to include all banking organizations with consolidated assets of $50 billion or more that are subject to Basel III capital requirements.

G. Results

62. The solvency stress test suggests that no BHCs would fall below the hurdle rate during the first year of severe economic distress (2015). This was due to high BHCs’ capital position in the base year (11.7 percent) relative to the CET1 regulatory threshold (4.5 percent) and large profits in the base period. The system-wide CET 1 ratio fell by 2½ percentage points in 2015 (Figure 4) relative to the base year or 2.2 percentage points relative to the baseline scenario in 2015.

Figure 4.
Figure 4.

CET1 Ratio Under the Baseline and Stress Scenario

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Note: The whisker boxes for each year display the distribution of CET1 ratios, projected using the IMF methodology, for 31 BHCs in terms of distribution’s moments: the bottom and top of the whisker box are the first and third quartiles, the band inside the box is the second quartile (median), the diamonds represent an un-weighted average CET1 ratio for 31 BHCs, the lower and the upper whisker represent the minimum and the maximum CET1 ratio respectively. The system-wide CET1 ratios are defined as projected total CET1 capital over total RWAs (sum over all BHCs).Source: IMF Staff calculations.

63. The results in the year of downturn were mainly driven by the increases in provisions and trading losses (Appendix Figure 5).

  • Compared to the base period (2014Q3), the system wide CET1 remained intact in the first quarter of the stress testing horizon (2014Q4).51 Net income fell sharply from its annualized, cumulative all-time high level in 2014Q3 of about $120 billion to a $30 billion net loss due to the impact of the initial negative shock in 2014Q4 on credit losses that doubled in comparison to 2014Q3 levels. On the other hand, lower RWAs due to slowdown in lending and total assets affected by the initial negative shock cushioned the negative effect of net losses on the CET1 ratio in 2014Q4. The modest negative impact of dividend distribution was not felt due to the capital conservation rule kicking in from 2014Q4 onwards.

  • Provisions for credit losses and trading losses play the major role in dynamics of CET1 ratios in 2015. The lower CET1 ratio in 2015 was mainly a result of higher provisions, which were four times higher than at the end of 2014 (subtracting -2.3 percentage points from CET1 ratio which is 8 times higher than in the base period), trading losses (-0.6 percentage points), negative AOCI and higher deductions from CET1 (each contributing with -0.3 percentage points). Net interest income fell by 5 percent compared to the end of 2014 due to lower loan demand and despite higher spreads. Noninterest income also fell by more than 10 percent. The impact of the implementation of the standardized approach to calculate RWAs for credit risk (which raised total RWAs by about 9 percent in 2015) was cushioned by lower loans and total assets and RWAs stayed stable in comparison to 2014.

64. Two BHCs would fall below the hurdle rate in the first year of the recovery (2016). The system-wide CET1 would fall by additional 0.3 percentage points in 2016. One BHC fell below the hurdle rate mainly due to relatively low profitability in the baseline.

65. Eleven additional BHCs would fall below the hurdle rate by the end of the stress testing horizon due to “too rapid” expansion of their balance sheet. Notwithstanding the favorable economic environment, system-wide CET1 would fall by additional 0.6 percentage point in 2017–2018 after recovering in the last year of the stress testing horizon (by 0.1 percentage points). Despite a recovery in BHCs’ capital position in 2019 a number of BHCs would fall below the hurdle rates. The BHCs that fell below the hurdle rate are the ones that on average had higher projected and historical (as reflected in the fixed effects in the loan equation) credit and total assets growth. Although this reflects the unrealistic assumption that BHCs would increase balance sheets even if this would take them below the regulatory threshold, this still suggests that these BHCs may need to raise additional capital if they are to be in a strong position to support a recovery in the face of an adverse scenario.

66. Changes in CET1 ratios in the recovery period were mainly driven by higher RWAs due to the expansion of BHCs’ balance sheets.

  • The first year of the recovery (2016) was still characterized by net losses (after taxes) which together with higher RWAs and deductions from CET1 resulted in lower CET1 ratios. Provisions fell comparing to 2015 but stayed at relatively high level. Provisions and higher non-interest expense due to expansion of balance sheets primarily contributed to net losses. The growth rate of loans, total assets and RWAs picked up due to higher economic activity and lower market volatility that also contributed to higher non-interest income, including trading income.

  • Higher RWAs weigh on CET1 ratios in the last three years of the stress testing horizon. While the cushioning impact of net interest income and non-interest income got stronger—reflecting the favorable economic environment—capital ratios deteriorated by increasing RWAs subtracting 80bps from CET1 in each of the last three years and increases in non-interest expenses due to the expansion of BHCs’ balance sheets (by 10 percent annually). While provisions and deductions from CET1 still played a role, their impact became very small by the end of the stress testing horizon.

67. Recapitalization needs are manageable. Recapitalization needed to bring all BHCs to the hurdle thresholds peaks in 2019 at 180 percent of their annualized 2014Q1–Q3 net income—which corresponds to 1 percent of 2019 nominal GDP.

68. CET1 ratios were projected to fall in the baseline; this is driven by increases in the RWAs despite positive net incomes. The system-wide CET1 ratio fell by 90bps in 2019 comparing to the base period. The negative impact of higher RWAs and expansion in total assets due to favorable economic environment was larger than the positive impact of all time high profits reported by the BHCs in the period 2015–2017. The impact of higher RWAs throughout the stress testing horizon was reinforced by the implementation of the standardized approach in 2015. Moreover, the impact of net income got weaker from 2015 due to lower contribution of net interest income which came down as a result of tighter spreads and lagged effects of rising policy rates on lending rates. Dividend distribution played a role too, subtracting on average 40bps from CET1. Even under the baseline scenario 3 BHCs fell below the hurdle rate in 2018 and seven more in 2019. These are the same BHCs that failed the test under the stress scenario mainly due to rapid expansion of their balance sheets in the second part of the stress testing horizon.

69. Sensitivity analyses were performed with respect to loan and total assets dynamics, dividend distribution rule, oil price shock, interest rate shock and different models of provisions (Figure 4). The results show that:

  • Loan dynamics: By using an assumption that loans and total assets do not fall during the period of downturn52 CET1 ratios fall more in 2015 and less in subsequent periods. System-wide CET1 ratio fell by additional 80bps due to higher RWAs when compared to the stress scenario but increased on average by 40bps a year in the last three years due to lower increases in RWAs resulting from lower growth rate in total assets then in the stress scenario.

  • Dividend distribution: Holding dividends constant at their 2014Q3 level (as in the DFAST) would subtract additional 30bps from the system wide CET1 ratio each year.

  • DFAST hurdle rate, no operational RWAs: Despite lower hurdle rates and higher initial capital ratios, the same two BHCs would fall below the thresholds. Additional 3 BHCs would fall below the threshold in the recovery period.

  • Constant loans and dividends, no operational RWAs: The system wide CET1 fell by 380 bps in 2015. Under this scenario two additional BHCs would fall below the regulatory minimum of 4.5 percent mainly due to dividend distribution which, in the benchmark, are cut to zero due to negative net income in 2016.

  • Oil price shock: Including oil prices as one of the determinants of provisions and applying the oil price shock where oil prices fell by 25 percent 2014Q4 and additional 60 percent by the end of 2015 would increase total provisions by around 3 percent.

  • Interest rate shock: Applying the interest rate shock to the 3-month Treasury yields of 450bps in 2015 would not change the results significantly since higher short term rates would have both positive and negative effects on income statements and balance sheets. On one hand, higher short-term rates would reduce loans growth rate and subsequently total assets and RWAs. On the other hand, higher short-term interest rates would result in higher credit losses (by 11 percent in 2015) and losses on AOCI (almost twice as higher in 2014Q4–2015Q4 than in the stress scenario).53 While these losses would be higher than in the stress scenario they would not be large enough to make a material impact on the results since the other variables have more pronounced effects on credit and trading losses than interest rates. However, the effects of higher interest rates on GDP growth were not analyzed.

  • Models of provisions: The only model that can capture the spike in provisions in 2008/2009 is the model of aggregate provisions which projected provision in the period 2014Q4–2016Q4 at the level of $390bn. A panel model of total provisions and a panel model of net charge offs resulted in the same projection of total provision—$280bn over the same period. Comparing to the benchmark stress scenario, this would correspond to system-wide CET1 ratio increase of about 80bps in 2015 and 15bps in 2016. While panel models project lower provision in the first two years the projections are more persistent and results in higher total provision over the whole stress testing horizon than the model of aggregate provisions.

  • Structure of loan losses: Modeling net charge-offs by loan type allows comparison of losses by loan type. An increase in credit losses is mainly driven by an increase in default rates and the size of exposures on the household sector. Around 35 percent of losses in 2015–2016 come from consumer loans and 25 percent comes from residential real estate exposures. While the share of commercial real estate and business loans is similar to 2008–2009, loans to financial institutions (and other loans) would account for much higher proportion of losses given their size in 2014 which is much higher than before the crisis.

  • Expanded coverage of the stress test: Expanding the coverage of the stress test to include all the deposit taking financial institutions with asset size of $50 billion and more that reported Basel III capital ratios as of 2014Q3 requires including additional two BHCs54 and two large savings banks.55 The same methodology was used to assess the capital adequacy of the four financial institutions as in the main stress test.56 The results show that one financial institution would fall below the threshold in 2016 mainly due to higher provisions and relatively low capitalization in the base period. Total recapitalization needs would increase by 5 percent due to the recapitalization of this financial institution.

H. Network Analysis for Large BHCs

70. To assess potential spillovers among the six largest U.S. G-SIBs, FRB staff implemented an updated version of the network stress-test methodology developed in Espinosa and Solé (2011). The network stress-tests were conducted by the staffs of the Federal Reserve Board (FRB) and the IMF in order to assess contagion risks among six U.S. BHCs designated as globally systemic.57 This methodology consists of simulating credit and funding shocks within a network of institutions and then tracking the contagion effects in terms of capital losses and path of bank failures. In addition, the methodology also allows for the assessment of the systemic impact arising from existing off-balance sheet financial linkages (e.g., credit default swaps).

71. To preserve data confidentiality, Fund staff provided the FRB the software necessary to implement the network stress-tests but had no access to the actual data. In turn, the staff of the FRB ran several simulations and robustness checks for a range of model parameters. The output of these simulations was reported back to the Fund.

The data

72. The FRB maintains a dataset that was used to execute the Espinosa and Solé algorithm for six systemic BHCs in the United States. The confidential dataset contained information on these BHCs’ capital levels, credit and funding exposures, as well as credit default swaps (CDS) contracts. The six institutions under consideration hold capital of around 9 to 12 percent of their total assets. Further details of the data are as follows:

  • Credit exposure data: For each of the six BHCs the FRB has an estimate of the credit loss that would be borne if one of the other five institutions went into default. This credit exposure estimate incorporates both direct credit losses—i.e., the losses that result from the default on a loan—and indirect credit losses—i.e., which would include losses that result from replacing defaulted derivative positions and losses on owned securities that have been issued by the defaulting BHC.

  • Funding exposure data: For each of the BHC, the FRB collects data on the amount of funding received from the other five counterparties. This amount includes secured and unsecured funding, as well as repo transactions of all maturities. Thus, the entire amount of a BHC’s funding exposure represents an estimate of the total amount of borrowings that would have to be replaced by that BHC if the counterpart entered default.

  • Risk transfer data: The FRB also has data on the amount of credit risk that has been transferred between the six BHCs. In particular, the data measure the amount of single-name notional CDS exposure that each bank has with respect to the five other BHCs in the sample (i.e., the notional amount of CDS protection that bank i has sold to bank j on reference entity bank h). Hence, note that the data do not include any CDS index trades nor does it contain data on more complex exposures such as CDS options.

Simulation and results

73. The network stress-test exercise comprised four different sets of simulations designed to capture key dynamics at play during the 2007–08 financial crisis. The first set of simulations examines the domino effects triggered if each of the six BHCs defaulted (one at a time) on their respective credit commitments. The second set of simulations assesses the effects of a credit-plus-funding event, where the default of an institution also leads to a liquidity squeeze for those institutions funded by the defaulting institution. In this case, the credit shock is compounded by a funding shock and the associated fire sale losses.58 The third and fourth sets of simulations build on the previous two by incorporating the credit default exposures of each BHC.59 The simulations were conducted with quarterly data for each quarter between 2013Q1 and 2014Q3.

74. The results indicate that the six BHCs hold enough capital to sustain a range of credit and funding shocks to individual counterparties within the network. As reported by the FRB, most simulations did not trigger contagion chains among the six institutions under consideration. This result likely emerges from the fact that direct exposures within the six-BHC network are not large enough (relative to the initial capital of each institution) to lead to second-round spillovers. Nonetheless, the positive results are also suggestive of the need to expand the data on exposures included in the network (e.g., exposure of the six BHCs to money market funds), as well as consider richer market dynamics in the simulations (e.g., downward spirals in the value of certain financial assets).

75. The simulation results show that the six BHCs hold enough capital to sustain shocks to a single counterparty within the network. In the four simulations considered, all BHCs appear to have enough capital to sustain the credit and funding losses individually impinged by the other BHCs in the network. For example, Table 11 shows results for the credit-plus-funding shock with risk transfers: in all instances the capital losses born by each BHC are not large enough to trigger a second round of contagion. Nonetheless, the losses could be substantial for some BHCs (e.g., BHC 4 could suffer losses of up to 2.5 percent of its initial capital), and could in turn lead to further funding difficulties for that institution if market concerns arise given the relatively large loss of capital.

76. The results also illustrate the importance of monitoring (and stress-testing) off-balance sheet exposures. Risk transfers such as credit default swaps can alter dramatically the risk profile of financial institutions. The simulations conducted allow an assessment of the potential impact that CDS exposures could have on each BHC’s capital. Figure 5 shows, for example, that the losses to BHC 5 from the failure of each of its counterparties (one at a time) would be dramatically different depending on whether CDS are taken into account. The chart suggests that BHC 5 is actually hedging its exposures to BHCs 1, 2, and 3 via CDS, but that it is actually increasing its exposure to BHCs 4 and 6 via this market. This type of analysis could be useful to regulators to monitor the interaction between on- and off-balance sheet exposures.

Figure 5.
Figure 5.

Credit and Funding Shock

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Note: Credit and funding shock – impact of CDS exposures on capital losses of BHC 5 (in percent of initial capital) Source: FRB, IMF Staff calculations.

77. The results appear robust to further simulations with stressed values of the model’s parameters. The FRB conducted robustness tests by assigning extreme values to the model’s parameters that measure the severity of an institution’s funding squeeze vis-à-vis the defaulting BHCs, the loss of asset value due to the fire sales associated with the funding squeeze, and by increasing the capital level below which an institution is considered under distress.60

78. Additional calculations by the FSAP team identified a combination of severe factors under which contagion would take place. A situation where funding markets are severely impaired and assets trade at heavy discounts could lead to a chain reaction of BHCs going into distress. For this severe scenario, the FSAP team assumed that only 65 percent of the short-term funding provided by a defaulting institution is rolled over by other market participants and that asset fire sales take place at 25 percent of book value. Given the team’s lack of direct access to confidential supervisory information, it was assumed that all BHCs have initial capital levels equivalent to 8 percent of risk-weighted assets and that if a bank loses more than 5 percent of its initial capital in one round, it suffers “distress”, which triggers the next round of contagion. Under these simplifying, though admittedly stark, assumptions, it is possible to trace which institutions would be more vulnerable through the contagion chain. For example, as shown in Figure 6, one such path would be triggered by distress in BHC 1 and would lead to three successive rounds of contagion affecting BHC 4 first, then BHCs 5 and 6 in a second round, and finally BHC 3. Note that BHC 2 would not go into distress in this scenario. These calculations are for illustration only, and more data on, for example, initial capital levels and exposure of the BHCs to specific financial instruments (e.g., reliance on short-term funding) to construct more accurate scenarios.

Figure 6.
Figure 6.

Contagion Path Triggered by BHC 1 Distress

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff.

79. While somewhat reassuring, the results also point to the need to expand the FRB’s dataset and consider richer market dynamics in the simulations. As explained, the FRB’s dataset only includes data on direct exposures among the six systemic BHCs. Thus, the stress-tests are unable to assess the potential impact of contagion feedbacks arising from other segments of the financial sector (e.g., if the six BHCs are exposed to a common funding source prompt to runs, such as money market funds; or a common credit exposure, such as CDS contracts referenced to a beleaguered sovereign). Similarly, new simulations that comprise downward spirals in the value of certain financial assets (e.g., asset-backed securities) could be designed and added to the network stress-test. Fund and FRB staffs have held conversations in this regard.

Discussion of Supervisory and Company-Run Solvency Stress Tests

80. This section summarizes the solvency stress tests conducted for the 2015 DFA stress testing exercise and discusses the differences vis-à-vis the FSAP team’s analysis. The stress testing frameworks of both company-run DFAST and supervisory-run DFAST for the 31 BHCs are publicly available information. Details of the company-run solvency stress test for BHCs can be found in Board of the Governors of the Federal Reserve System, 2013, “Comprehensive Capital Analysis and Review 2014 Summary Instructions and Guidance”, November 2013 and Board of the Governors of the Federal Reserve System, 2013, “Capital Planning at Large Bank Holding Companies: Supervisory Expectations and Range of Current Practices”, August 2013, whereas the details of the supervisory DFAST can be found in Board of the Governors of the Federal Reserve System, 2015, “Dodd-Frank Act Stress Test 2015: Supervisory Stress Test Methodology and Results”, March 2015.

A. Supervisory Stress Tests

81. There are important differences between the IMF approach and the authorities’ approach, with potentially significant impacts on results (Table 12). The hurdle rates in the IMF stress test were more stringent: there are differences in the calculation of risk-weighted assets, and capital ratios, such as the inclusion of operational RWAs in the IMF stress tests, that on average resulted in lower ratios for advanced approaches BHCs. The level of the granularity of the data and subsequently the methodology used by different approaches was different—the authorities used much more granular data. The dividend distribution rule used by the IMF was a function of the level of capital which was a less conservative assumption than the one used by the authorities. Loans and total assets were projected in the IMF stress test whereas they were forecasted conditional on the assumption that credit supply is maintained at long-run historical levels in the DFAST. Some losses were not covered by the IMF stress test due to lack of historical data. Finally, while the authorities used a 9-quarter stress horizon and 3-year forecast scenarios, the IMF test focused on stress testing horizon that spanned over 5 years. However, IMF forecasts over the first three years were consistent with the forecasts of the FRB under the baseline and stress scenarios.

82. An in-depth analysis of the results of the DFAST scenario is limited by constraints on publicly available information. Detailed information on capital, RWAs and income statement items were not publicly available. Only the minimum and the ending capital ratios over the stress testing horizon of nine quarter were published. RWAs were published as of the end of the stress testing horizon. Income statement items were published as a cumulative value of pre-provision revenue, other revenue, provisions, realized losses/gains on securities (AFS/HTM), trading and counterparty losses, other losses/gains, net income before taxes and other comprehensive income. AOCI included in capital was published as of the end of the stress testing horizon. Loan losses (cumulative over the nine quarter period) and loan losses structure was published by type of loan. All the results were presented both in the aggregate for the 31 BHCs and for individual BHCs for severely adverse and adverse scenario.

83. With this caveat, the results of the supervisory DFAST suggest that, in the aggregate, BHCs are resilient to shocks for the severely adverse scenario. All BHCs stay above the regulatory minima. Even if the IMF hurdle rates were used, no BHC failed the stress test. Over the nine quarter of the stress testing horizon the system-wide CET1 would fall to 7.6 percent (its minimum) and to 7.8 by the end of 2016. If compared to the IMF estimate of the system wide CET1 in 2014Q3 (of 12.4 percent) the CET1 ratio at its minimum would fall by 480 basis points, compared to 2014Q3. This is both due to increase in RWAs, which mainly reflects the implementation of standardized approach in 2015 and the assumption on the loan supply, as well as net income losses that were projected to be -$222 billion.

84. The losses were mainly driven by loan losses ($340 billion) and trading and counterparty credit losses from a global market shock ($103 billion). Projected losses on mortgage 61and consumer loans62 represent 56 percent of projected loan losses driven by higher unemployment rate and lower house prices. The largest losses pertained to credit card losses ($83 billion). The nine-quarter cumulative loss rate of 6.1 percent, with significant differences across BHCs, is high by historical standards and more severe than any recession since the 1930ties. Trading losses at the six BHCs and counterparty losses at the eight BHCs ranged between $1bn and $24 billion across the eight BHCs.

85. Low level of pre-provision revenue mainly reflected low projected net interest income and non-interest income. This is consistent with low interest rates and flattening of the yield curve in the first part of the stress testing horizon and falling asset prices, rising equity market volatility and falling economic activity.

86. Under the adverse scenario BHCs would report moderate declines in capital ratios. The adverse scenario simulates a mild recession but with a sharp increase in short term rates that affect BHCs’ funding costs. The projected capital ratios are smaller than those under the severely adverse scenario. The main difference is higher pre-provision revenue driven by higher net interest income due to higher interest rates. In the publicly available results, the authorities did not indicate any impact of higher interest rates on loan delinquency. However, AOCI is three times larger than in the severely adverse scenario due to higher interest rates.

87. In aggregate, the CET1 projections in the DFAST and IMF top-down approach are similar (Figure 7). The minimum CET1 ratio in the benchmark IMF stress test is higher than in the DFAST, mainly reflecting less conservative loan dynamics and lower dividend distributions for firms constrained by the hurdle rate. However, when DFAST assumptions on loan supply, dividend distribution and operational RWAs were introduced in the IMF stress test the CET1 ratio in the benchmark case came down close to the DFAST CET1 ratio.

Figure 7.
Figure 7.
Figure 7.

DFA vs. IMF Stress Test Results

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF Staff calculations.

88. However, while the aggregate capital ratio may be similar, RWAs in DFAST were higher than in IMF stress tests. Total RWAs in DFAST were 6 percent higher than in the IMF benchmarks case (again reflecting the difference in credit dynamics that outweigh the addition of the operational RWAs to the IMF model) but also 4 percent higher than in the IMF stress test with the DFAST assumptions. If the DFAST estimates of the 2015 standardized RWAs increase for individual BHCs were applied in the IMF stress test the total RWAs would be very similar.

89. Correspondingly, the aggregated capital level in the DFAST was higher than in IMF’s stress test. Although it was not possible to decompose the underlying factors driving this result based on publicly available information, this difference could be due to a number of reasons. First net income losses could have been smaller in DFAST. However, this was not the case, and in fact net income losses in the two stress tests were very similar due to the fact that projected provisions were almost the same and that the sum of pre-provision revenue and trading, counterparty and other losses was very similar to pre-provision revenue in the IMF stress test. Therefore, one or a combination of the following factors could have led to a higher estimate of capital in the DFAST: (i) taxes; (ii) deductions from CET1; (iii) dividends; (iv) extra-ordinary items; (v) change in valuation of allowances. Deductions and dividends were presumably the same in both stress tests in the case where IMF stress test took DFAST assumptions. However, since the details of these other factors were not part of the published results, an accounting of which of these led to the higher aggregate capital estimate could not be ascertained.

90. Moreover, bank-specific capital ratios differed significantly between the two exercises. About one third of BHCs in the DFAST have either higher or lower CET1 ratio by 2.5 percentage points. For a few large BHCs IMF estimates of CET1 ratios were higher due to high trading and counterparty default losses in the DFAST for these BHCs. For the rest of the BHCs, IMF estimates of net income losses were mainly higher, including for the two BHCs that would fall below the capital hurdle rates. This may be due to the modeling approach of the IMF stress testing framework, which was different from the DFAST due to granularity of the data.

B. Company-run DFAST for 31 BHCs

91. Three bottom-up stress tests were reviewed by the agencies: one by the FRB63 one by the OCC and one by the FDIC. All the tests relied on banking companies’ internal consolidated data to assess solvency of individual companies under different macroeconomic scenarios through changes in net income and risk-weighted assets. The cut-off date of both tests of the data was September 2014.64

92. The DFA company run stress test covered BHCs with consolidated assets of $10 billion or more, which account for about 90 percent of total BHC assets. OCC’s stress tests covered national banks and federal savings association with total consolidated assets over $10 billion. FDIC’s stress tests covered FDIC-insured state banks that are not members of the Federal Reserve System and FDIC-insured state-charted savings associations with total consolidated assets of more than $10 billion. Instructions to companies, together with scenarios, were issued on October 23, 2014.

93. This note focuses on the results of the company run DFAST for 31 BHCs. Company-run stress testing results were reported to the primary supervisor on January 6, 2015 for companies with assets size of more than $50 billion and on March 31, 2015 for companies with assets size of more than $10 billion but less than $50 billion. While the results were not published by companies’ supervisors, banking organization with assets of $50 billion or more, including 31 BHCs included in the supervisory DFAST, disclosed a summary of the results of the bottom-up stress test, under the severely adverse scenario in March 2015. Other companies will be required to publish the bottom-up stress testing results in the period from June 15 to June 30. To ensure comparability with the IMF top-down stress test and the supervisory DFAST, and given the appropriate coverage of the top-down tests and the timing of the FSAP (the second mission took place during the last week of February and the week of March), only the results of the DFA bottom up stress test for largest BHCs (with total assets of $50 billion and more) are presented in this note.

94. The capital definition applied in the stress tests corresponded to that required by local regulation, i.e., Basel III65 (subject to phase-in) for advanced approaches BHCs and non-advanced approaches BHCs (from January 1, 2015) and Basel I for non-advanced approached BHC for the first quarter of the stress test horizon (last quarter of 2014).66 In order to assess the potential impact of negative shocks on the capital requirement metrics over the stress horizon, companies were required to assume consistency with the Basel III transition schedule. The stress tests incorporated the transition arrangements and minimum capital requirements from the revised regulatory capital framework implementing the Basel III capital reforms from January 2014 and Basel I capital standards for non-advanced banking organization applied for 2014 only. Starting in 2015, the revised capital framework introduced a new standardized approach for risk weighting assets, which replaced the calculation of risk weights using the general risk-based capital approach.

95. The results of the company run stress test were disclosed by BHCs in the supervisory DFAST format. As in the supervisory DFAST, detailed information on capital, RWAs and income statement items were not publicly available. While capital ratios and income statement items were published in the same format as in the supervisory DFAST for all BHCs, six BHCs did not publish their RWAs which precluded the analysis of the results in the aggregate for the 31 BHCs.

96. On average, the individual results of the company run test were more optimistic than the results of the supervisory DFAST or the IMF stress test. The tests suggest that BHCs are resilient to shocks for the severely adverse scenario. All BHCs stay above the regulatory minima, even if the IMF hurdle rates were used. Over the nine quarter of the stress testing horizon the un-weighted system-wide average CET1 would fall from 12.9 percent to 9.4 percent (its minimum). The average CET1 ratio in the company run stress test is 90bps higher than in the supervisory DFAST and 175bps higher than in the IMF stress test. Higher CET1 ratios are both due to lower increase in RWAs, for the BHCs that reported their projection of RWAs, as well as lower net income losses that were projected to be -$190 billion (compared to -$222 billion in the supervisory DFAST and -$224 billion in the IMF stress test).

97. The differences in the net income losses were mainly driven by loan losses ($250 billion). While trading and counterparty losses ($108 billion) were almost the same as in the supervisory DFAST, projected loan losses were much lower by the BHCs than by the FRB or the IMF. Moreover, the structure of loan losses was different in the company run stress test than in the supervisory DFAST or the IMF stress test. For example, losses on consumer loans represent 45 percent of projected loan losses in the company run stress test, compared to 35 percent in the supervisory DFAST or the IMF stress test (Figure 7).

98. While the correlation between BHC-specific capital ratios for the company run and the supervisory stress test is high,67 some BHCs reported significantly higher CET1 ratios. Comparing to the supervisory DFAST results, in the company run DFAST five BHCs have higher CET1 ratio by 2.5 percentage points. This is due to lower losses and RWAs projected by the BHCs. Many BHCs argue that the differences in projected net income statement items and RWAs may be due different modeling approaches employed by the BHCs.

IMF Staff’s Liquidity Risk Analysis for BHCS

99. A liquidity risk analysis was done by the IMF team in order to assess the resilience of the banking sector with respect to sudden, sizable withdrawals of funding. The analysis was done as of 2014Q3 on a bank-by-bank basis and included the same BHCs as in the solvency stress test.

100. Due to data constraints, the LCR or the NSFR was not possible to calculate.68 The liquidity metric calculated by the IMF team was defined and calculated based on publicly available data reported in FR Y-9C report (Schedules HC, HC-B, HC-D, HC-E, HC-L and HC-M).

A. Liquidity Metric

101. The liquidity metric measured whether BHCs have adequate levels of liquid assets that can be converted into cash to meet their liquidity needs. The liquidity metric was defined as the ratio between the stocks of liquid assets to the total cash outflow. While the items that were included in the numerator and denominator of the metric were informed by the LCR definitions used in BCBS (2013) (Table 13, 14) there was no attempt to replicate the LCR calculation based on publicly available data.69 Haircuts were taken from the LCR and run-off rates were calibrated based on the 2008/2009 episode.70 Two stress horizons were assumed over which the withdrawal of funding took place: 1 quarter and 3 quarters. In the sensitivity analysis the run-off rates were calibrated based on the LCR.

Table 13.

Liquid Assets

article image
Source: IMF Staff
Table 14.

Outflow Items

article image
Source: IMF Staff

B. Results

102. The results of the analysis give some comfort that the system is able to meet liquidity requirements, but there are pockets of vulnerability. The analysis suggests that most, but not all, BHCs have enough liquid assets to meet a liquidity shock similar to 2008/2009 event (Figure 8).71 Several BHCs would face liquidity pressures due to deposit outflows in the short run and large unused commitments over a longer stress horizon. If faced with a much larger shock, as characterized by the LCR run-off rates, liquid assets for many BHCs would not be sufficient to meet liquidity needs due to large withdrawal of wholesale funding. Wholesale funding plays an important role in this case since the run-off rate on wholesale funding is much larger than what happened in 2008/2009.

Figure 8.
Figure 8.

Liquidity Metric, Historical Run-Off Rates, and Approximation of LCR Run-Off Rates

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Note: The whisker boxes display the distribution of the liquidity metric, calculated using the IMF methodology, for 31 BHCs in terms of distribution’s moments: the bottom and top of the whisker box are the first and third quartiles, the band inside the box is the second quartile (median), the diamonds represent un-weighted average of liquidity metrics for 31 BHCs, the lower and the upper whisker represent the minimum and the maximum liquidity metric, respectively.Source: IMF Staff calculations.

103. While the liquidity metric offers some insights into the liquidity risks, the results come with many caveats. While ideally liquid assets should include unencumbered liquid assets only, the liquidity metric here includes both types of assets which will inflate liquid assets.72 Also, the liquidity metric does not include pledged assets to BHCs which deflates the liquid assets measure. Moreover, inflows were not considered as part of liquidity analysis which would make the liquidity metric smaller than what it would otherwise be. Outflows on derivatives contract were not considered. Due to unavailability of granular data, the actual run-off rates included in the sensitivity analysis had far smaller variation than those in the LCR. Finally, current regulatory report lack important elements to run an accurate liquidity stress test such as liability tenor information, inflows resulting from maturing transactions and relevant contractual terms embedded in derivatives contracts.

IMF Staff’s Solvency Tests for Insurance

104. This section reports on the top-down solvency stress tests for insurance companies performed by the IMF FSAP team. In addition, it also includes a summary of a top-down stress test performed by the National Association of Insurance Commissioners (NAIC) for this FSAP.

A. Scope of the Test

105. The scope of the exercise was broad in terms of risk categories included and methodological approaches. The stress test included a sample of 44 insurance groups, of which 22 are predominantly active in the life insurance business, 15 in property & casualty (P&C) business, five in health insurance, and two in credit and mortgage insurance. All of the groups in the sample are publicly listed. The stress test included the three groups designated both as global systemically important insurers by the Financial Stability Board (FSB) and as systemically important by the Financial Stability Oversight Council (FSOC). Overall, the sample represents 40 percent of the domestic insurance sector in terms of gross written premiums. The separate top-down stress test performed by the NAIC, which also informed the IMF’s sensitivity analyses, covered all U.S. life and P&C insurers filing with the NAIC.

106. The stress test used publicly available, consolidated data of insurance groups from regulatory returns provided by SNL Financial and Bloomberg. Further sector-wide data was provided by the NAIC to enhance the granularity of the exercise. The separate stress test performed by the NAIC used legal entity data filed with the NAIC, including both publically available information as well NAIC risk-based capital data filed exclusively with the NAIC and state insurance regulators. The risk-based capital data was only made available to the IMF in aggregate format and not on an individual company basis.

B. Scenario

107. The insurance top down stress test was built on the DFA stress test specifications. Given the nature of insurance business and its balance sheet structure, the main focus of the stress test was on investment assets and, therefore, the market risk parameters of the DFAST. The market risk stresses (Table 15) included shocks to bond holdings (sovereigns, municipals, and corporates), securitizations, equity, property and other investments (such as hedge funds and private equity). In addition, like in the severely adverse scenario of the DFAST, lower swap rates were assumed, in a range of minus 21 to minus 143 basi–s points for maturities of one year and 30 years, respectively. Broadly speaking, the market risk parameters reflect a severe market distress similar to the situation observed at the height of the financial crisis in 2008–09. All stresses were assumed to occur instantaneously.

Table 15.

Market Risk Parameters

article image
Source: IMF Staff assumptions based on DFAST.

108. The lack of detailed data on the investments of insurance undertakings made it necessary for the FSAP team to simplify some of the stresses provided in the DFAST, or to re-arrange some risk factors. As an example, the equity shock was not calculated on a per-country basis, but an overall weighted shock was generated based on the geographical breakdown of the equity exposures of the U.S. insurance sector. In a similar vein, the shock for non-U.S. sovereign bonds was derived, split into the rating categories defined by the NAIC.73 For the property shock, it was assumed that the price developments in the commercial real estate sector as defined in the DFAST for the twelve quarters from 2014-Q4 until 2017-Q4 would occur within just one period.

C. Valuation and Capital Standard

109. An important point of context for the stress tests is that statutory accounting of U.S. insurers is based on U.S. GAAP, which in some instances differs from a “fully market-consistent” approach to the valuation of assets and liabilities. Under statutory accounting, the liabilities of P&C insurers are generally not discounted which adds a layer of conservatism. Also under statutory accounting, life insurance liabilities are discounted with a rate that is set at the time when the policy is sold to the policyholder or with a discount rate based on the expected return of assets associated with the insurance liabilities. The NAIC calculated that the discount rate on all life insurance policies averaged approximately 4 percent in 2013, which is above current market rates.74 Under statutory accounting, amortized cost is the predominant accounting regime for fixed income assets,75 which means that neither unrealized gains nor losses are recognized. For the stress test, this results in a significant difference in the impact of a shock to the risk-free interest rate. In a truly economic balance sheet with a fully market-consistent valuation of both assets and liabilities lower interest, as specified in the scenario, mean that the liabilities of a life insurer increase more than its assets, given a structural mismatch of assets and liabilities that is very common in that type of business. While the duration mismatch is usually smaller for non-life insurers, the same mechanics apply. State insurance regulation requires that companies perform an asset adequacy analysis at least annually to measure the structural mismatch of assets and liabilities under a range of different interest rate scenarios.

110. Under statutory accounting, also the impairment rules for life insurers differ from a fully market-consistent regime. Investment assets are impaired only when the fair value loss is deemed to be other than temporary. Once impaired, a bond cannot be written back up to its original fair value after recovery.

111. In the absence of a group capital requirement for insurers, the hurdle rate for the IMF’s top-down stress test was set, generously, as the complete extinction of shareholder equity. This means that the capital deemed to cover unexpected losses and serving as “the first line of defense” would no longer be in place. Groups with negative shareholder equity after stress clearly failed the stress test. However, this perspective holds true only in a full fair-value regime. With negative shareholder equity, assets are smaller than liabilities. While such a non-coverage of liabilities might potentially cause policyholders to surrender their life insurance policies (a “run” situation), such a behavior is partly disincentivized by surrender penalties, loss of insurability, and the federal income tax treatment of life insurance. Further, a company’s insurance book can be run off in a relatively ordered way, because of the powers the state insurance regulator has in a troubled company context, thereby widely limiting contagion effects.

D. An Overview of Insurance Companies Soundness76

112. Balance sheet and asset quality. The aggregated balance sheet of the 44 insurance groups in the stress testing sample increased from 2009 to 2013 by 25 percent to $4.5 trillion. The largest group within this sample consists of life insurers, with aggregated assets of $3.1 trillion. Non-life insurers and health insurers complete the sample with aggregated assets of $1.1 trillion and $0.3 trillion, respectively. The share of the non-life sector fell, owing to the continuing restructuring of AIG, which divested parts of its business in 2010 and 2011. In the aggregate, the ratio of shareholder equity to balance sheet assets was 8.3 percent for life insurers in 2014, and was 22.9 percent and 33.1 for non-life insurers and health insurers, respectively. Non-life insurers hold higher amounts of shares and also shorter durations in their bond portfolio than life insurers. Health insurers have in comparison the most conservative asset allocation, with relatively high allocations in sovereign bonds.

113. In recent years, there was a tendency among life insurers to invest in longer maturities. In 2014, the median duration reached 7.6 years for the life insurers in the sample (an increase from 7.2 years in 2009). In comparison, median asset durations for non-life companies and health insurers were 5.4 years and 5.2 years, respectively. Also, the median share of non-investment grade bonds in the bond portfolio of life insurers reached 5.4 percent, compared to 4.1 percent for non-life insurers and 4.4 percent for health insurers. The share of non-investment grade bonds among life insurers has been decreasing since 2009, while it has been increasing for non-life, albeit from a lower base.

114. Income statement. In terms of revenues (premiums, capital gains and investment income being the main components), health insurers form the largest group in the stress test sample. Their share reaches 42 percent, mainly coming from recurring premiums. Life and non-life insurers account for 32 percent and 26 percent, respectively. Between 2009 and 2014, revenues of life insurers increased by 36 percent. Non-life insurers increased their revenues by 11 percent and health insurers by 46 percent.

115. Nearly all insurers in the stress test sample reported a positive net income in 2014. Returns on equity were highest in the health sector with a median of 14.0 percent. Non-life insurers reported 11.7 percent and life insurers 8.2 percent. Compared to 2013, the return on equity improved substantially for non-life and health insurers, while declining for life insurers.

116. Life insurers have benefitted from positive capital market developments since 2009 with both stock markets and bond markets improving, but have struggled in the face of lower interest rates. Investment yields have been declining over the last years as higher-coupon bonds expire and are replaced by lower-yielding new issues. For P&C insurers, super storm “Sandy” in 2011 was the latest major catastrophe event, resulting in a median combined ratio77 above 100 percent. Since then, the combined ratio has stabilized below 100 percent, with the median in 2014 being 94 percent. Over the last five years, changes of the combined ratio have predominantly been driven by the loss ratio, while operating expenses have been stable on average, around 32 percent.

117. While the implementation of the Affordable Care Act resulted in rising premium income of health insurers, uncertainties remain about the future profitability of policies sold under this program. The IMF staff observed that higher premiums and positive profit margins have benefitted health insurers after the first full year of business under the Affordable Care Act. Already since the enactment of the new system in March 2010, stock prices have significantly outperformed the S&P 500. Nevertheless, risks continue to exist for the health insurance industry, notably administrative risks with regard to regulated prices and legal challenges to subsidies provided to policyholders. Also the medium-term behavior of policyholders is unknown, especially lapse rates could be higher than in other health insurance lines. Finally some uncertainties about future claims exist as the Affordable Care Act has led to a shift in the average risk profile of the policyholder by attracting higher-risk cohorts of customers.

E. Modeling Assumptions

118. Assuming a fully market-consistent valuation impact of the shock, the value of investment assets and ultimately shareholder equity declines substantially.78 Credit spread increases are multiplied with the duration of the respective asset class, derived from the maturity buckets provided in statutory reporting, also resulting in a lower value of investment assets.

119. Separate accounts have not been included in the stress test as investment losses are generally passed on to policyholders. Such accounts are offered by 17 companies in the sample (with a median share of total reserves of 53 percent amongst those companies offering separate accounts). No breakdown of investments or detail on the guarantees provided in these accounts is available in the consolidated public filings, so it was assumed that the asset allocation matches the group-wide asset allocation and no economic loss remains with the insurance company. The market risk shocks were accordingly applied only to the investment assets held in the general account. This simplifying approach might underestimate the effect of the stress scenario in some cases, especially when an insurance company has issued a guarantee for the separate accounts (or parts of them).

120. In an economic balance sheet approach, the shock to the risk-free rate needs to be applied to both assets and liabilities. While the duration on the asset side can be approximated based on maturity buckets, no detailed information on the duration of liabilities is available. It was therefore assumed, based on suggestions by market participants, that the duration of liabilities exceeds the duration assets by two years in the case of life insurers, and by one year in the case of non-life and health insurers.

121. The stress test does not take into account any mitigating effect from hedging. Insurance companies usually apply a sophisticated hedging strategy with regard to their interest rates (mainly via swaps and swaptions), and also hedge against declines in the stock market via options and futures. As these hedging activities can vary substantially among companies, it is difficult to estimate the mitigating effect in times of stress. In any case, it is very likely that the stress test gives a maximum impact.

122. For the sensitivity tests, the IMF team built on various approaches developed by the NAIC. This was with regard to modeling the effect of (a) major natural catastrophes, (b) a pandemic, and (c) a prolonged period of low interest rates. The results of these analyses were not added to the outcome of the main stress scenario, although it is possible to assume that the stress scenario occurs at the same time as a major natural catastrophe or a pandemic.

123. The IMF team has specified three types of natural catastrophes, for which the NAIC has provided approximate results. These were based on re-assessing historic events of a similar type and by cross-checking this against newly introduced amendments to RBC filings of companies. The catastrophe events included a hurricane in Florida, similar to but worse than hurricane Andrew in 1992, causing an industry-wide loss of at least $40 billion. As a second event, an earthquake in California, similar to but worse than the Northridge earthquake in 1994 was specified; the earthquake should have a magnitude of at least 7.2 and the insured loss should be greater than $35 bn. Finally, a series of three major tornados should be assumed to occur in the Midwest of the United States, each causing an insured loss of at least $4 billion. Each of these events, which have been modeled to occur independently from each other, would reduce the capital position of exposed P&C insurers.

124. Similarly to catastrophic events which mainly affect the P&C sector, a shock to mortality rates in the form of a pandemic was modeled in the life insurance sector. This mortality shock is calibrated as a pandemic with 1.5 additional deaths per 1,000 which is considered a 1-in-200 year event (Swiss Re 2007). While the increased mortality rate is well below numbers reported for the influenza pandemic in 1918–19, when more than 5 per 1,000 people ceased in the United States, the situation is not directly comparable to today’s standards of healthcare and governments’ responsiveness. The results for both the catastrophe shock and the pandemic shock do not include potential macroeconomic implications of such an event which in turn could have a further negative effect on capital markets, as recent events like e.g. the SARS outbreak in Hong Kong in 2002–03 have shown.

125. With regard to the prolonged period of low interest rates (“low for long”), the NAIC has provided an analysis that compares the net investment yield of life insurers against the average credited rate. The horizon for this analysis (Figure 9) was from 2014 until 2018. The investment yield was assumed to decline linearly based on the trend observed since 2006. 79 The average credited rate would also decline, as new business would be issued with lower contractual guaranteed interest rates. For example, currently issued annuities contain an interest rate guarantee of 1 percent. The spread between investment yield and credited rate, multiplied with the projected reserves gives the impact on the profitability of the insurer.

Figure 9.
Figure 9.

Medium-term Projections in a Low Interest Rate Environment, 2006–18

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: NAIC

F. Results

126. The stress test shows a significant impact on the U.S. insurance sector, especially in the life business (Figure 10). The results for the life insurance companies show a wide dispersion, but 11 out of 22 groups would report negative shareholder equity after stress if a fully market-consistent accounting regime was in place. The effect is larger than what has been observed historically, e.g., in the recent global financial crisis life insurers had the ability to hold their investment assets until maturity without the need to sell them at depressed price—Box 2 provides further analysis on how stresses unfold under statutory accounting. The other insurance segments are much less affected, given their lower exposure to investment risks. No nonlife or health group would be in distress, suggesting that the respective sectors are in a more robust shape. Overall, the losses in shareholder equity amount to $267 billion, or 45 percent of pre-stress equity. Life insurers contribute to this amount by $187 billion, while $72 billion come from P&C insurers, $6 billion from health insurers, and $2 billion from credit insurers. Out of the total loss, $104 billion is attributed to the distressed companies.

Figure 10.
Figure 10.

Top-Down Insurance Stress Test

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: SNL Financial and IMF Staff calculations

127. The main contribution to the overall loss in shareholder equity, calculated in a fully market-consistent way, comes from the credit spread increase in the corporate bond portfolio. For the full sample, this corporate bond shock (also applied to hybrids) accounts for 62 percent of the overall loss. Further notable contributions come from the shock to sovereign bonds and GSEs, the shock to the securitization portfolio, and the shock to share prices (each with 7 percent). Given their larger holdings in shares, P&C insurers are relatively more affected by the equity market shock, while health insurers would record larger losses stemming from their sovereign bond portfolio. For credit insurers, the full impact of higher credit spreads and higher default rates would likely be higher than the numbers suggest as also their liabilities would be affected.

128. Among the distressed companies, smaller and medium-sized institutions are in a majority. The aggregated balance sheet assets of the 11 companies amount to around $490 billion (11 percent of the full sample). Their pre-stress shareholder equity declines from $55 billion to a negative $48 billion under stress.

129. A large catastrophic event or a pandemic, seen in isolation, are likely manageable for both the life and the non-life sector (Table 16). The most expensive event would be the Florida hurricane which, modeled as a 1-in-250 year event could cause insured losses of around $80 billion. An earthquake in California with the same expected occurrence frequency could result in a loss of $34 billion, since the vast majority of California earthquake exposure for residential properties is through the California Earthquake Authority. A series of three severe tornados in the Midwest of the United States shows only rather contained effects, substantially below the claims expected in the other two scenarios. A pandemic with 1.5 additional deaths per 1,000 which is considered a 1-in-200 year event (Swiss Re, 2007), could cost the U.S. life insurance industry between $20 and $25 billion. The hurricane and earthquake are estimated net of reinsurance. The pandemic is estimated on a gross basis, not taking into account the mitigating effect of reinsurance and alternative risk transfer, so that only parts of these amounts would ultimately be borne by U.S. insurance sector. However, the results for both the catastrophe shock and the pandemic shock do not include potential macroeconomic implications of such an event which in turn could have a further negative effect on capital markets.

Table 16.

Impact of Natural Catastrophes and Pandemics

article image
Source: NAIC

130. A scenario of prolonged low interest rates poses a slow burning risk which could become a solvency risk for life insurers in a few years. For the period from 2006 to 2013, the industry-wide spread between the net portfolio yield and the guaranteed credited rate declined by 57 basis points. Assuming that interest rates remain at their current levels, the spread could, linearly projected, continue to decline further as the lower rates influence the average guaranteed credited rate at a significantly slower speed than the portfolio yield. The modeling is based on several restrictive behavioral assumptions, especially with regard to the asset allocation, but clearly indicates that the risk of low interest rates requires intense monitoring as it influences the business model of life insurance substantially. While at the moment, the risk only reduces profitability of the companies, potential negative spreads between investment yields and guaranteed rates could result in wide-spread losses ultimately resulting in a weaker capital position of the sector.

NAIC Top-Down Stress Test

In parallel to the top down exercise performed by the IMF team, the NAIC has also run a top down stress test based on end-2013 statutory data for the whole U.S. insurance sector. Also the NAIC stress test shows a substantial impact, though less pronounced as no shock is applied to investment grade corporate bonds and sovereign bonds are expected to increase in value (combining the effect of potentially higher credit spreads and offsetting lower risk-free interest rates).

In addition to immediate losses in the general accounts, also guaranteed components of separate accounts were included. In those cases where an insurance company guarantees certain benefits of separate account products, a shock is modeled by applying a factor of 10 percent to the company’s maximum possible guaranteed amount that would be diverted from the general account to the separate account.

The exercise further included two natural catastrophes, a Florida hurricane and a California earthquake, each with a 1-in-250 year probability and both occurring at the same time.

article image

Based on statutory accounting, the U.S. insurance industry as a whole shows some robustness although the capital impact is significant. While there are some companies which would fall below the regulatory thresholds, the average decline in total adjusted capital amounts to $166 billion in the life sector (-35 percent) and $267 billion in the P&C sector (-33 percent). Even with stressed capital, both sectors would, on average reach RBC coverage ratios of 659 percent and 447 percent, respectively.

The main contribution to the overall changes comes from the securitization portfolio and other investment assets in the life sector, while for P&C insurers, the equity shock and the catastrophic events have the largest impact.


Contribution to loss in total adjusted capital

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: NAIC

131. The results show the life insurance sector in a very challenging position. Not only could market turbulences cause huge losses in terms of fair value accounting (less so in the current statutory accounting as long as the market stress is not prolonged or substantial), but vulnerabilities persist. The low-yield environment could, if continued, erode the profitability of the sector and ultimately also deteriorate the capital position. Net investment yields would need to remain above 4 percent to ensure profitability. The sector has already reacted and is actively changing its product mix, offering more policies where investment risks are partially or fully passed on to the policyholder.

132. The low interest rate environment is challenging especially for life insurers and caused many companies to change their investment behavior. As insurers search for yield, durations in the bond portfolio have increased since 2009 (Table 17). The share of bonds below investment grade has declined for the life companies in the stress test sample (from 7.5 to 5.4 percent). Over the same period, the share went up significantly for non-life and health insurers, albeit from a low base. Also within the investment grade bond portfolio there was a clear tendency to take on more credit risk by expanding the relative share of BBB-rated assets. Some movement into alternative investments (hedge funds, private equity) has been observed, but this is still of a smaller dimension and rather restricted to larger insurance companies who are expected to have an adequate risk management in place.

Table 17.

Changes in Duration and Credit Quality of Bond Portfolio

article image
Source: SNL Financial and IMF staff calculationsNotes: based on insurance stress test sample

133. While life insurance companies are exposed to the risk of prolonged low interest rates, also a sharp upward shock to interest rates poses a material risk. The market value of the bond portfolio would decline, even more so with the longer durations insurers are holding now. In statutory accounting, however, life insurers would be able to carry many of these assets at amortized cost, unless their decrease in value was determined to be other than temporary. Rising interest rates would likely lead to an increase in policy surrenders when policyholders switch into higher-yielding assets (within and outside the insurance sector), leaving companies with a liquidity drain and potentially some losses for those products where no surrender penalty applies. While the share of companies that fund themselves on the capital market is rather small in the insurance sector, higher interest rates drive up funding costs for those who issue bonds on a regular basis.

134. P&C insurers showed greater robustness in the stress tests. After some years with only few large catastrophes, the capital positions have improved to weather times of stress. However, competition and price pressure have increased with some products, and investment yields have decreased, requiring companies to improve their cost structure and also to remain prudent in their underwriting.

135. Insurance companies are able to mitigate some of the effects of the stress. A range of life insurance products includes profit sharing features between the insurance company and the policyholder which allow the insurer to (partially) pass on investment losses by reducing discretionary benefits. In a risk-based solvency regime, it is also possible for the companies to de-risk their investments in order to reduce their capital requirements, resulting in higher solvency ratios; similarly ceding risks to a reinsurance company could be considered. Finally, dividend policies, both upstream from subsidiaries to the top (holding company) level and from the top level to shareholders can be actively managed, especially with larger and diversified groups. As the range of management actions is very broad, no general modeling result can be provided based on publicly available data.

Additional IMF Stress Test Based on Statutory Accounting

To evaluate the effect of differences in accounting methods, the IMF team complemented its mark-to-market stress test with an additional statutory accounting exercise, based on the same sub-samples of life and P&C insurers1 and end-2014 data. The difference in the results of both approaches reveals the impact the valuation regime can have on an insurance company’s capital position.


Reduction in shareholder equity

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

The following modeling steps have been taken:

  • Mark-to-market impairment, based on the DFAST severely adverse scenario, for holdings in equity, corporate bonds below investment grade, and other investment assets;

  • Default losses in the corporate bond and securitizations portfolio (default rates are based on observations from 2008-092, with an assumed 50 percent loss given default rate);

  • A catastrophic event (for P&C) and a pandemic causing higher mortality rates (for life) which result in net cash outflows of $25 billion in each of the two sectors (outflows are assumed to be distributed among companies according to their respective market share);

  • The need to liquidate investment assets to match the cash outflows, realizing losses at distressed market levels; it is assumed that companies would sell U.S. treasury bonds, municipals and GSE issues first.

The aggregated reduction in statutory capital for life and P&C companies under this modeling approach amount to $111 billion, which represents 22 percent of the pre-stress statutory capital. The median loss in the life sample is 26 percent, while for P&C insurer it is 23 percent. One life insurer would be in distress as its statutory capital turns negative, for one other life insurer the reduction amounts to more than 75 percent.

Most of the overall impact can be attributed to the impairment of investments which accounts for 69 percent of the change in statutory capital. The catastrophic (for P&C) or pandemic (for life) event contributes 18 percent, the defaults in the investment grade corporate bond and securitizations portfolio 13 percent and the realized losses in the forced sell-off only contribute marginally (1 percent). The differences between the life and the P&C sector are not very pronounced.

The results of this exercise are broadly in line with those of the NAIC stress test, which was also based on statutory accounting. Some divergence however exists due to a different scenario design; as an example, the NAIC stress test does not include the impact of a pandemic event in the overall results of life insurers.

Notes:/1 Health and credit insurers have not been included in this exercise./2 Default rates are calculated as averages of default rates provided by Fitch, Moody’s and Standard&Poor’s for the period from mid-2008 to mid-2009 via the Central Repository (CEREP) set up by the European Securities and Markets Authority (ESMA).

IMF Staff’s Liquidity Analysis of Mutual Funds

136. Open-ended mutual funds’ investments are exposed to redemption risk.

  • Such funds may be more susceptible to runs when their investments are directed into markets that are less liquid. Their liabilities are liquid due to a regulatory obligation to meet investor redemption demand in-cash within 7 days, yet they may possess neither the balance-sheet liquidity capacity nor access to robust-to-severe stress back-up lines of liquidity. This could leave funds with no recourse but to sell assets in the open market, even at a steep discount.

  • The rapid expansion of such funds into fixed-income markets has raised concerns about the potential for investor runs on funds to exacerbate asset market stress. The markets for corporate bonds, emerging market debt, bank loans, and municipal bonds are less liquid than equities, U.S. Treasury, and GSE securities. Moreover, there has been a notable decrease in trading liquidity in corporate bonds since the crisis.

137. IMF staff has performed an analysis of liquidity in U.S. mutual funds, and important caveats apply. This was the first time this type of analysis was performed in an FSAP context, and while publicly available data are voluminous they are incomplete in important respects. Moreover, the analytical basis for analyzing the liquidity position of mutual funds is still nascent. The analysis could therefore be only exploratory in nature, but still provides interesting results that help point to possible vulnerabilities and areas where further analysis may be warranted.

138. The analysis was geared to measuring whether markets would be able to absorb severe redemption pressures wherein these funds are forced to liquidate positions. Ideally, an asset pricing model would be deployed to examine the (marginal) impact of redemption spikes on asset prices or (bid-ask) spread measures of market liquidity. Absent such a model, the approach taken in this exercise was to assess whether a standard metric of available trading liquidity, dealer inventory in specific assets markets80, is sufficient to absorb redemption demand in a tail risk scenario. Dealer inventory was selected as a proxy for trading liquidity because it was believed to provide a useful indication of market appetite for a variety of asset classes.

139. Specifically, a top-down liquidity risk analysis was performed. Close to 9,000 mutual funds representing around 80 percent of the industry were analyzed. The funds were divided into 69 styles81 capturing their investment objectives. For each market under consideration (mortgage, corporate, municipal, government bonds), the universe of funds captured by this approach includes both index funds and hybrid funds. Opting for this style-based approach means greater coverage of funds investing in each market of interest, but it also entails a greater aggregation in the markets analyzed. For example, it is possible to stress market liquidity in corporate bonds markets but not separately for investment grade and high yield. The choice of such an approach is justified by the fact that it is important to capture a full universe of funds. The calculations were based on granular data on mutual funds from the Center for Research in Security Prices (CRSP). The cut-off date for the analysis was the third quarter of 2014.

140. The IMF staff compared assets sold by mutual funds hit by a shock to data on dealers’ inventory. If dealers’ inventory would be smaller than assets sold by mutual funds this would indicate potential liquidity pressure on mutual funds that invest in the assets sold in that particular market. This might also give rise to fire-sale risks on that particular market and might imply that investors in the funds exposed to those markets have to take a haircut on their investment.

141. The shock was defined as a one time, tail event redemption shock. The distribution of net flow rates, including both net inflows and net outflows, by fund style was analyzed. Net flow rates were defined on the monthly basis as a simple average of net flow rates of all mutual funds of the same style over the period 1998-2014Q3. The first percentile of net flow rate distribution was taken as the stress redemption shock. Averaging redemptions across component market segments (such as investment grade and high yield) was a simplifying assumption that may affect the results of the analysis.

142. Once a mutual fund is hit by a redemption shock it would have to sell its assets to meet redemptions. The following two sets of assumptions on redemption induced assets sold were made:

  • Approach 1 (“pro rata”): Pro-rata selling of assets was assumed i.e. assets were sold to meet the redemptions by making sure that the structure of assets is intact. This assumption is a natural one to adopt for the case of index funds which would be expected to sell assets to meet redemption demand in a way that seeks to keep portfolio weights unchanged to continue minimizing tracking error relative to their benchmark.

  • Approach 2 (“waterfall”): Mutual funds were assumed to rank order assets held by their liquidity characteristics, as captured by the LCR haircut hierarchy, selling assets to meet redemptions in descending order of liquidity. Specifically, the assets were assumed to sold in the following order (Figure 11):82 cash is the first asset to be used to meet redemptions, government securities are second, MBSs are third (assuming most are GSE-backed MBSs), then 20 percent of equity and corporate bonds, then municipal bonds (whole portfolio) and then an additional 40 percent of equity and corporate bonds.

  • Under both approaches, realized assets sales due to the tail event shocks were added up across all funds included in the exercise and for each asset market and then compared to dealer inventory, which is was used as a proxy for assessing market makers’ (dealers) willingness and/or ability to make markets in, and as an indicator of general market demand for, a given asset class.

Figure 11.
Figure 11.

Mutual Fund Liquidity Analysis: The Waterfall Approach

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF Staff

143. The results of the analysis, under both approaches, suggest that municipal bonds and corporate bonds markets may face significant stress when faced with tail event redemption shocks. The analysis illustrates the danger that funds that invest in corporate and municipal bonds might sell these assets at a fire-sale discount to meet redemptions. Under the same tail event shock, municipal bonds that might be sold to meet the redemptions could be three (in the case of the assumed ordering of assets sold) to four times (in the case of pro-rata asset selling) larger than the dealers’ inventory of municipal bonds (Figure 12). Similarly, the analysis shows that the volume of corporate bonds sold under severe stress by mutual funds could be up to seven times larger than what dealers currently hold in inventory.

Figure 12.
Figure 12.

Results of the Liquidity Risk Analysis

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Note: “Sold assets- pro-rata” represent asset sold by mutual funds hit by a tail event redemption shock that have to sell their assets pro-rata i.e., by making sure that the structure of assets is intact (approach 1).“Sold assets- assumed ordering” represent asset sold by mutual funds hit by a tail event redemption shock that have to sell their assets in descending order of liquidity (approach 2)Source: IMF Staff calculations.

144. A number of caveats precluded a more detailed analysis of the liquidity risks in the mutual fund industry. First, not all mutual funds were included in the analysis, and a larger sample would make the assets sold under stress larger, underscoring the vulnerabilities identified. Second, the data used in the analysis may not have been granular enough (both on the mutual funds’ asset structure and on the structure of dealers’ inventories) to fully capture the liquidity dynamics of a potential asset fire-sale. For example, both domestic and foreign corporate bonds were bundled in the same category of corporate bonds and compared to dealers’ inventories of corporate bonds. This prevented the analysis from discriminating among potential differences in the liquidity profiles of the two bonds types.83

145. To address the caveats and more closely examine the potential illiquidity of mutual fund assets, IMF staff performed a separate security-level analysis of mutual fund portfolio holdings. This was done by combining portfolio security information from CRSP with security issue(r) information from the CUSIP Global Service’s CUSIP Master File and the Mergent Fixed Income Securities Database, and then aggregating security holdings into homogenous asset classes, based upon their perceived liquidity characteristics. Unfortunately, the issuer information from the latter two sources was found to contain a significant domestic equity reporting bias or to only cover a small fraction of the securities present in the CRSP holdings database.84 Accordingly, the aggregated asset classes were deemed unrepresentative of the broader mutual fund holdings universe, and liquidity assessments using these asset classes as inputs were not pursued further.

146. In the absence of data constraints, the “ideal” liquidity analysis would utilize detailed supervisory data on fund holdings and explicitly take into account individual fund characteristics and market-level liquidity information. This includes information such as investment mandates and trading volumes, when estimating redemption risk and the potential for asset fire sales. For each individual fund, such an analysis would: (i) determine tailored liquidity waterfalls based upon investment mandates, (ii) apply extreme redemption shocks to assess the immediate demand for cash, (iii) use the liquidity waterfall to estimate which, and how much of, each security should be sold to meet redemptions, and (iv) use cumulative information on security bid-ask spreads and trading volumes to estimate what the sales price of each security sold would be. The liquidity analysis would then aggregate the value of the securities sold across funds into granular asset classes, and compare these values to what would have been obtainable under ordinary market conditions in order to determine which particular asset classes may be most prone to fire-sale dynamics.

147. The authorities are encouraged to step up work to assess susceptibility of markets to extreme mutual fund redemptions. The authorities should further clarify the guidance on liquidity risk analysis performed by the industry. It is important that the authorities mobilize the resources necessary to regularly conduct liquidity risk analyses as part of their overall approach to mutual fund industry oversight.

IMF Staff’s Market-Price Based Stress Tests

148. This section presents the results of IMF staff’s market-price based analysis and stress tests. This exercise is intended to complement the analysis above by taking into account the information about risk that is embodied in market prices, which allows consideration of correlations between institutions and higher frequency and more timely assessments. There are important caveats that must be acknowledged, however. The findings are necessarily sensitive to: methodological issues or choices – such as simplifying assumptions; the use of imperfect proxies; the selection of sample periods; heterogeneity across model, data, and variable definitions; and potential endogeneity.85 Accordingly, in many cases these findings should be interpreted as informative approximations, as opposed to precise estimates, which primarily highlight the co-behavior of risk factors rather than causal relationships between them. Despite these caveats, the market-price based analysis and stress tests are useful tools for examining the financial stability landscape. For this reason, they have been used extensively in previous work by IMF staff as well as many central banks and others analyzing systemic risk.

149. The section is divided into two complementary parts. First, a broad survey of available systemic risk measures is presented and their historical evolution is examined. Second, a stress testing analysis of the market-implied interaction between default risk and the macroeconomic environment is conducted using the Contingent Claims Analysis (CCA) framework. Both parts use high-frequency, forward-looking market consensus information to “cross-check” the findings of IMF staff’s other stress test methodologies. They also extend stress-test coverage to sectors which are not traditionally subject to microprudential oversight, such as non-bank non-insurance financial institutions, and help to compensate for a lack of access to supervisory data.

A. Systemic Risk Dashboard

150. To limit model risk, this technical note adopts a multi-model approach to identifying systemic risk in the United States. There exist numerous definitions of systemic risk, and the methodologies employed to measure such risk are diverse as well. Accordingly, this technical note adopts a multi-faceted approach, called a “systemic risk dashboard”, which employs a range of measures in an effort to identify the numerous dimensions in which a threat to financial stability may arise. The dashboard helps to inform the risk-based assessment process and to guard against another important type of risk known as “model risk”, which is the excessive reliance on a single modeling framework.86

Key Dashboard Findings for the United States

  • Market-based measures point to a reduction in the systemic risks of banks. Systemic RISK87 (“SRISK”)—a well-known measure of market-implied capital shortfall for a given bank or given banking system—suggests that systemic risk posed by banks has declined towards to its pre-crisis average. SRISK capital shortfalls peaked at approximately one trillion dollars in early 2009 but have now fallen to 300 billion dollars. This level is roughly commensurate with average pre-crisis shortfall estimates and constitutes approximately 2 percent of GDP (Figure 13). One caveat is that SRISK exclusively measures systemic risk in the banking sector. Risks posed by nonbanks are covered in subsequent analysis.

  • Standard early warning indicators of banking distress are also reassuring. Both financial cycles88 and credit-to-GDP gap89 measures (Figure 13)—widely-used early warning indicators of impending domestic financial crises—signal that the United States banking system is relatively healthy from a cyclical perspective. Downward inflections in financial cycles and positive credit-to-GDP gaps often coincide with periods of financial distress.

  • Equity price indicators suggest that valuations may be stretched relative to fundamentals. Percentage deviations of observed equity prices from theoretical prices based on Asset Pricing Theory (left) and the Equity Composite Z-score90 (right) (Figure 14) indicate the degree to which equity prices are misaligned with economic fundamentals. Figure 14 shows that these indicators are either positive or have been trending into positive territory, which suggests that equity price levels may be approaching unsustainably high levels.

  • Housing price indicators are at normal levels, but need to be closely monitored. Turning to measures of credit and housing price sustainability (Figure 16), credit-based measures (top) are still well below their pre-crisis levels, indicating that excessive credit growth does not pose an immediate financial stability risk to the United States. However, growth in housing prices (bottom) has fully regained its pre-crisis momentum and needs to be carefully monitored.

  • The potential threat posed to the financial system from credit risk migration has declined significantly since 2007-09, but remains a concern. An analysis of U.S.- and foreign entity-based credit risk networks (Figure 17) suggests that there has been material decline in the susceptibility of institutions to credit risk migration, as denoted by the decrease of dense, red connections in the current period relative to 2007–09. Nonetheless, the role played by GSIBS in the transfer of credit risk appears to remain important.

Figure 13.
Figure 13.

SRISK Market Implied Capital Shortfalls

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: NYU Stern Volatility Lab, as of end 2014Q
Figure 14.
Figure 14.

Early Warning Indicators

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations.
Figure 15.
Figure 15.

Equity Price Misalignment Measures

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations.
Figure 16.
Figure 16.

Fundamentals in Housing and Credit

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations.
Figure 17.
Figure 17.

Credit Risk Networks

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations.Note: The credit risk networks shown were created by applying bivariate Granger-Causality tests to time-series of 198 institutions’ first-differenced default probabilities. These tests were applied 21 times at differing specifications and the color the connecting lines reflects the robustness of the identified Granger-casual connections. Red connections were found in 95 percent or more of trials, dark gray connections were found in 70 percent or more of trials, and light gray connections were found in at least 50 % of trials. Thick black connections represent membership in the network’s minimum spanning tree. Node size represents the GSIB vs non-GSIB (large vs. small) network dimension and node color (blue vs. gold) reflects the domestic vs. foreign dimension. The 21 testing specifications are combinations of three different test significance levels (1%, 5%, and 10%) and seven different lag orders (orders 1 through 7).
  • Individual financial system sectors are exposed to each other, and to the rest of the world, through a common set of financial instruments (Figure 18).

    • Asset managers, insurers, pensions, and households are sizable net claimants on the corporate equities and corporate bonds issued by other sectors, such non-financial corporations (“corporate”) and nonbank financial institutions (NBFIs).

    • Asset managers, insurers, pensions, and households, in addition to banks, also serve as notable net claimants on the government through their holdings of treasury, agency, and municipal debt securities.

    • The largest net claims on fund shares, the primary instrument used by asset managers to raise capital, are attributable to the households and pensions, although banks, insurers, and non-financial corporate also possess net claims on this sector as well.

    • Money market instruments act an important conduit between households, asset managers and banks.

    • Banks, corporates, households, and GSEs have net exposures to all other sectors, and to abroad, via an assortment of deposits, loans, and mortgages.

    • Government and pensions are significantly exposed to households and corporates via pension entitlements.

    • On a net basis, the rest of the world is exposed to the U.S. financial system primarily through net claims on money market instruments, corporate equities, and other direct investment. Foreign net claims on the United States take the form of shares in investment funds, government securities, corporate bonds, loans, mortgages, and miscellaneous deposits.

    • Publicly available locational data on OTC credit-default-swaps (CDS) shows that both the rest of the world and the aggregate U.S. financial system have sizable outstanding exposures to these instruments. Unfortunately, additional detail on the nature of these credit risk exposures, or other OTC derivatives exposures, cannot be obtained without access to confidential trade information.

  • Comprehensive measures of systemic risk suggest that nonbanks contribute more to systemic risk than banks. An important analytical tool in this regard is the Systemic Risk Indicators (SyRIN) framework, which uses distress-based metrics91 to assess the level of systemic risk present in a given set of financial sectors. SyRIN has been used in earlier studies, for example in the October 2014 Global Financial Stability Report (IMF, 2014b).92 An examination of the U.S. financial system using this framework finds that although systemic risk in the United States appears to be declining towards pre-crisis level, confirming the findings from narrower measures such as SRISK, areas of concern remain. The framework highlights that in terms of asset size, banks account for less than 30 percent of the U.S. financial system, with more than 70 percent accounted for by nonbanks, and these asset shares were broadly unchanged between 2010Q4 and 2014Q4 (Figure 19, top chart). But the SyRIN framework also shows that there are parts of the nonbank sector that contribute to systemic risk more than one could expect based on their size. The marginal contribution to systemic risk (MCSR) of the U.S. high yield and insurance sectors is disproportionately large relative to their share of financial system assets. At the same time, for example, the pension sector’s ratio of MCSR to size is approximately 0.6. This means that the pension sector contribution to systemic risk is lower relative to its size, hence this sector appears to diversify risk from the financial system (Figure 19, bottom chart). A longer-term analysis of the ratio of MCSR relative to size suggests that these ratios for High Yield Bond Funds, IG Bond Funds and Hedge Funds have shown a significant increasing trend.

Figure 18.
Figure 18.

Financial Account Net Exposures

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: FRB; BIS; IMF staff calculationsNote: Yellow nodes denote sectors and blue nodes denote instruments. Node size represents the total value of balance sheet assets held by each sector or cumulatively outstanding for each instrument as of 2014Q3. The directed lines between each node, and the color of these lines, represent the net exposure of a given sector to a given instrument. Black lines denote the net claims of a sector, and red lines denote net liabilities of a given sector, in the form of a given instrument. Line thickness signifies the magnitude of net claims (red lines) and net liabilities (black lines). The thicker (thinner) a line, the larger (smaller) the size of a given net exposure. Net exposures were calculated for each sector-instrument combination by subtracting (netting) liabilities from asset claims.All lines are directed, but arrowheads may not always appear due to formatting limitations. Black lines without visual arrowheads are always directed from a sector to an instrument. Similarly, arrow-less red lines are always directed from an instrument to a sector. The dashed line effects do not have a semantic significance and are used only as a visual aid. The instrument “CDS” represents the aggregate, over-the-counter (OTC) credit default swap (CDS) exposure attributable to different domiciles as calculated by official BIS statistics. The box titled “U.S. CDS Exposure” represents the aggregate OTC CDS exposure of U.S. counterparties that cannot to be attributed to specific domestic sectors. Detailed sector and instrument definitions are presented in Appendix Table 7.
Figure 19.
Figure 19.

Systemic Risk Indicators Framework, 2010Q4 and 2014Q4

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations based on methodology described in Segoviano and Goodhart (2009) and expanded in the October 2014 Global Financial Stability Report (IMF 2014b); and Segoviano and others (2015).Note: The top chart shows sectors’ shares of financial system total assets. The bottom chart shows each sector’s contribution to systemic risk relative to its asset size. Ratios exceeding 1.0 indicate that the systemic importance of a given sector is greater than what would be suggested by asset size alone. In the SyRIN framework, “distress” is defined as an event whereby the index used to represent returns of a given sector falls to a level that should only be observed 1 percent of the time based on history. The SyRIN definition of “distress” is independent of the Contingent Claims Analysis (CCA) definition of “default.” HY = high yield, IG=investment grade, MMMFs = money market mutual funds.

151. In sum, the findings of the Systemic Risk Dashboard suggest that although threats to financial stability have diminished since the peak of the crisis, they remain worthy of continued close monitoring, especially with regard to nonbanks. Financial stability indicators in the United States have broadly improved, but areas of concern do exist, especially outside the banking system. Equity prices levels, housing price growth, and the vulnerability of banks to distress originating in the high yield and insurance sectors should be assessed more closely. The metrics presented thus far focus on historical developments in the financial system and serve as a baseline assessment. Against this background, the next section uses a hybrid market/balance-sheet framework known as Contingent Claims Analysis (CCA) to explore how financial risks may evolve under the IMF/DFAST stress scenario.

Systemic Risk Indicators (SyRIN) Framework: A Primer

Contagion through interconnectedness across financial institutions and sectors plays an important role in the realization of systemic risk. The recent crisis underlined that proper estimation of contagion risks among financial institutions and sectors in a financial system is essential for effective financial stability assessment. The realization of simultaneous large losses in various entities would affect a financial stability, and thus represents a major concern for regulators. Thus, the analysis of systemic risk should aim at understanding these contagion risks (due to direct and indirect linkages across financial institutions) and their changes across the economic cycle.

The SyRIN framework used in the market-price based systemic risk assessment draws upon recent financial stability literature and a wealth of previous analytical work to estimate systemic risk. Earlier versions of this approach were used in the 2010 FSAP as well as the October 2014 Global Financial Stability Report (IMF, 2014b). The SyRIN employs distress-based metrics to assess the level of systemic risk present in a given set of financial sectors. The SyRIN framework is independent of the CCA analysis, and their findings, definitions, and approaches to measuring systemic risk should not be conflated.

The SyRIN framework conceptualizes the financial system as a portfolio of entities, which spans banks and non-banks alike, and incorporates the effect of interconnectedness into its risk estimates. Most empirical literature on systemic risk measurement has tended to focus on a single sector, typically the banking sector, without taking into account the rest of the financial system. Research concerned with analyzing risks associated with nonbank financial institutions has only recently started to emerge. By providing a comprehensive treatment of both bank and non-bank financial sectors, SyRIN is able to analyze system-wide risks in a manner consistent with empirical facts (Espinoza and Segoviano 2014). For the purposes of this Technical Note, the SyRIN portfolio includes the largest domestic banks, insurance companies, pensions, mutual fund and hedge fund sectors.

Contagion through interconnectedness in the financial system can happen due to direct and indirect linkages across sectors. Direct linkages occur through the direct exposure channel via inter-sector exposures (interbank deposits, lending, syndicated loans), derivative transactions and exposures counterparty bankruptcy. Indirect linkages can be due to exposures to common risk factors which surface in periods of economic and financial distress, either through the asset liquidation channel or through liquidity-induced fire-sales. Other associated factors include general increases in investor risk aversion during stress periods, the effect of “herding” on investor portfolio reallocation decisions, and actions taken by short-sellers to drive down stock prices for financial firms which follow complex or opaque business models or rely on embedded leverage.

SyRIN estimates are drawn from a multivariate joint probability density which accounts for financial system interconnectedness. This density models the distribution of asset price returns for a financial system portfolio that explicitly accounts for the dependence structure (defined in terms of direct and indirect linkages) of its portfolio components. This approach allows SyRIN estimates to capture the joint effect of changes in asset values for sectors throughout a financial system. To infer this density, SyRIN uses a robust, non-parametric approach that incorporates endogenous changes in sector co-dependence. When economic conditions deteriorate, the density’s “distress” dependence structure increases, thereby increasing its probabilistic estimates that sectors will fall into distress.

SyRIN uses a nonparametric multivariate density to measure financial stability from several complementary perspectives. Measurement takes the form of estimation (via simulation) of distressed losses to the financial system, and is the basis for the calculation of a given sector’s MCSR. The MCSR captures the effect of portfolio sectors’ interconnectedness (co-dependence) and relative size, and can be used in conjunction with complimentary joint and conditional probability estimates, which are also drawn from SyRIN’s multivariate density, to round out its assessment of systemic risk.

B. Contingent Claims Analysis: Stress Testing for Systemic Risk

152. Contingent Claims Analysis (CCA)93 uses equity prices and accounting information to measure the credit risk of institutions with publicly traded equity. The CCA framework is useful because it provides forward-looking default probabilities which take into account both leverage levels and market participants’ views on credit-quality. Another benefit is that it provides a standardized benchmark of credit risk, known as default probabilities, that facilitate cross-sector and cross-country comparisons. However, CCA does suffer from some theoretical shortcomings. Namely, it can only be applied to entities with either publicly-traded equity or very liquid CDS spreads, and it cannot capture liquidity or (financing) roll-over risk. CCA risk measures are calculated at the firm-level and complement the high-level survey of risk measures presented in the systemic risk dashboard with micro-based information on financial risks.

153. The central idea behind CCA is that an institution’s risk of default is driven by the level and uncertainty in its asset values relative to the promised payments on its debt obligations. Assets of a financial institution or corporation are uncertain and change due to factors such as profit flows and risk exposures. Default risk over a given horizon period is driven by uncertain changes in future asset values relative to promised payments on debt– where these payments are often referred to as the “default barrier”. As first introduced by Merton (1973), the key financial insight used to quantify this asset/debt inter-relationship is that equity values can be modeled as an implicit call option on assets, with an exercise price equal to a default barrier, and that risky debt can be modeled as the default-free value of debt minus an implicit put option (i.e. an expected loss due to default). Accounting information on an institution’s debts and market information on the price of its outstanding equity are used to calibrate CCA balance sheet risk indicators and to estimate forward-looking probabilities of default.

154. The CCA and SyRIN are two different frameworks that are independent of and complementary to each other; their findings, definitions, and approaches to measuring systemic risk should not be conflated. Importantly, CCA uses “default probabilities” as its measure of credit/default risk, whereas SyRIN uses “distress probabilities” in its financial risk estimates. CCA default probabilities are derived from a structural model (Merton, 1973) and signify the chance that the value of a given financial institution’s (market-adjusted) assets will fall below a specific liability threshold (its default barrier). SyRIN distress probabilities are defined as the chance that the value of a given financial sector’s market equity will fall below the 1st percentile value of its historical equity price distribution. Also, the sectors analyzed in the SyRIN and CCA approaches are not necessarily identical. For example, only SyRIN’s analysis includes sectors such as pensions and high yield bonds funds and only CCA’s analysis includes foreign banking and insurance sectors. (See relevant sections of the market-price based stress tests for additional details.)

CCA Stress Test Data

155. The historical default probability estimates used in the CCA stress tests were acquired from CreditEdge.94 CreditEdge follows several broad steps in its production of this key stress testing input.

  1. For each institution, daily equity values, equity volatilities, and default barriers are calculated.

  2. The inputs from step (i) are then used to simultaneously solve two structural CCA equations and estimate market-implied asset values and market-implied asset volatility.

  3. “Distance-to-default” indicators are computed using the inputs from the first two steps and are then mapped to empirically observed one-year default probabilities using Moody’s extensive historical default database.

156. Default probabilities can also be mapped to credit-risk boundaries, such as credit spreads or ratings. Empirical research suggests that an approximate investment-grade “safe zone” for financial institutions corresponds to an expected one-year default probability of 0.5 percent or less.95

157. The CCA stress tests covered 210 institutions from a mixture of domestic and foreign domiciles and eight different sectors (Table 18). The eight sectors consisted of: domestic banks (“banks”); life, health, and property/casualty insurers (“insurers”); investment management companies, REITS, and private equity firms (“asset managers”); large, publicly-traded nonfinancial firms (corporates); other nonbank financial institutions (NBFIs), government sponsored housing enterprises (“GSEs”);96 all non-U.S. domiciled G-SIBs (“foreign banks”), and all G-SIIs as well as other large foreign insurers (“foreign insurers”). In addition, an overall U.S. financial system sector was created using data pooled from among all the domestic U.S. sectors, excluding GSEs.

Table 18.

CCA Stress Test Sample Data

article image
Source: IMF staff calculations

158. The CCA stress tests utilized a connectivity variable and a number of credit risk and macroeconomic variables (Table 19). For comparability purposes, the macro variables used in the stress testing scenarios were identical to those used in the top-down balance-sheet based stress tests run for banks. Daily historical data was used in estimations and covered a ten-year period spanning end 2004Q3 to end 2014Q3. Additional information is listed below:

  • Daily default probability data was acquired from CreditEdge and daily, monthly, and quarterly macroeconomic data was from obtained from Haver Analytics and Bloomberg.

  • Inter-connectivity data was calculated by IMF staff, using the process outlined in the subsequent section on the stress testing methodology, on a rolling monthly basis over the period of 2004–2014.

  • All quarterly and monthly variables, including connectivity, were temporally disaggregated to a daily frequency using the Chow-Lin Max-Log methodology. This includes all data contained in the baseline and stress scenarios as well.

  • The connectivity measure used for projections was extended forward in time using the following two assumptions: (i) the connectivity trend which was observed in 2008–2010 would repeat itself, then (ii) remain at its post-crisis historical average once this value has reached.

  • All default probabilities used in the CCA stress test models are median one-year, expected default probabilities. For each sector, daily median default probability time series were computed from observations pooled at the sector level. Daily U.S financial system median default probabilities were generated from observations pooled from all domestic sectors, excluding GSEs.

Table 19.

Variables Used in CCA Stress Tests

article image
Source: IMF staff.

CCA Stress Test Methodology

159. To estimate the relationship between connectivity, macroeconomic factors, and median default probabilities, the CCA stress tests utilized a class of models known as General Adaptive Models of Location, Shape, and Scale (GAMLSS). GAMLSS, described in detail in Stasinopoulos and Rigby (2007), is an extremely flexible model class which allows one to: (i) utilize a wide variety of distributions to characterize the response variable and (ii) explicitly model the first four moments of these distributions as functions of exogenous conditions. As a result, the framework is well suited to address the presence of tail-risks, nonlinearities, and deviations from the normality assumption. The default probability data used in the CCA exercise exhibited all of these latter characteristics, thereby motivating the choice of the GAMLSS model class for stress testing purposes.

160. The GAMLSS modeling process follows an iterative approach consisting of eight steps.

  1. Fit approximately 100 different distributions to the response variable (i.e., median 1-year default probabilities).

  2. Compare the quality of each fit using the Generalized Akaike Information Criteria (GAIC) and select the distribution with the best score.

  3. Using the distribution chosen in step (i), run information criteria based selection procedures97 to identify the independent variables with the most linear explanatory power.

  4. For variables not selected in step (iii), transform them using additive terms (e.g., orthogonal polynomials, penalized basis splines, etc.) and repeat the prior step’s selection procedure.

  5. Once a final set of independent variables has been determined, experiment with different additive terms to enhance model fit.

  6. Compare all experimental models generated in step (v) using the GAIC and select the model with the best score.

  7. Repeat steps (i) through (vi) for each distributional parameter (i.e., the mean, variance, skewness, and kurtosis).

  8. viii. Use diagnostics to assess whether the model residuals are supportive of the assumed response distribution. If not, return to step (i) and repeat the process using the next best GAIC-identified distribution.

161. To avoid over-fitting and to test the out-of-sample predictive power of stress test projections, stiff penalties98 were imposed on model complexity when computing GAIC-based model selection scores and a back-testing regime was used throughout the modeling process. The back-testing regime consisted of (i) using a quantile-based sampling algorithm99 to partition the overall dataset into separate validation, testing, and training datasets, and (ii) using the validation dataset for calculating initial distributional fits; using the training dataset to estimate the regression coefficients and for variable selection; and using the testing dataset to gauge out-of-sample model performance.100

162. The GAMLSS model for the overall U.S. financial system suggests that macroeconomic, sector credit risk, and interconnectivity factors influence credit risk levels in the United States in significant, and often non-linear, ways. The model finds that median 1-year default probabilities for the U.S. financial system are expected to fall linearly with rises in long-term Treasury yields, housing and commercial real-estate prices and with appreciation of the U.S. dollar against the Euro. Default probabilities are expected to increase linearly with BBB-rated corporate bond yields and bank, insurer, asset manager, other non-bank financial institution, foreign bank, and foreign insurer median default probabilities. The magnitude of these co-movements differs by for each variable, but changes in BBB corporate bond yields, long-term Treasury yields, and asset manager default probabilities have some of the greatest effects on financial system credit risk. The model also finds many non-linear relationships. U.S. credit risk slightly rises with euro-area consumer inflation, and falls with declines in real GDP growth and the VIX and Dow Jones Industrial indices. Overall credit risk in the U.S. also rises with increases in corporate credit risk, but only in non-linear statistically significant manner. A simplified summary of the final specification for the overall U.S. financial system model is presented in Appendix Table 5. Appendix Figure 8 shows the estimated relationships for the model’s four non-parametric additive terms.

163. A single connectivity measure was calculated and used as an explanatory variable in each of the estimated GAMLSS models. This connectivity measure was calculated using institutional-level default probability time-series data from the following five domestic sectors: banks, insurers, asset managers, other non-bank financial institutions, and nonfinancial corporates. This measure, formally known as a “global clustering coefficient”, was derived using the following three-step process:

  • Perform Spearman Rank Correlation Tests to identify institutional default probabilities that are correlated at the .0001 percent significance level.

  • Construct an adjacency matrix from the test results, and use this matrix to derive a “correlation network.”

  • Calculate the network’s global clustering coefficient score.

The above three-step process was repeatedly applied to one-month rolling windows spanning the period of 2004Q3 to 2014Q3. Monthly values were then temporally disaggregated using the Chow-Lin Max-Log methodology to generate a daily connectivity time series. Figure 20 shows the corresponding quarterly series, which was calculated using both domestic and foreign sectors. The figure underscores the increase in connectivity over time. Statistical tests suggest that a structural break occurred in 2008. After the break, connectivity has a higher mean and lower variance, suggesting that the financial system is more “consistently connected” post-2008.

Figure 20.
Figure 20.

Financial System Connectivity

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations based on data from Moody’s CreditEdge.Note: The clustering coefficient measures network connectivity on a scale from 0 to 1 and its values reflect the probability that a set of points within a given network will share a direct relationship. The clustering coefficient scores for the U.S. financial system were calculated by first applying correlation tests to one-year, rolling windows of firm-level measures of default risk and assessing the clustering present in these test results. The correlation tests are based on two-sided, non-parametric Spearman rank correlation tests, performed at the 0.001 percent significance level. The score at each point in time represents the connectivity present between firms with highly correlated default risk over the prior year-to-date. The firm-level measures of default risk used in the underlying calculations were monthly default probabilities for approximately 210 domestic and foreign entities. The red line is a structural break, as identified by a pruned exact linear time (PELT) mean-variance test performed at a 1 percent significance level. Connectivity has a higher mean and lower variance after the break which suggests that the financial system is more “consistently connected” post-2008.

CCA Stress Test Results

164. The CCA stress tests estimated the relationship between connectivity, the macroeconomic environment, and median default probabilities using separate GAMLSS models for each sector and for the overall U.S. financial system. These models were used to assess the impact of macroeconomic changes and fluctuations in connectivity under the “baseline” and “stress” scenarios and to individually project default probabilities for the overall U.S. financial system and for five domestic and two foreign sectors. The CCA stress test scenarios are identical to those used in the balance sheet stress tests for banks (i.e., identical to the IMF/DFAST scenarios). In order to isolate the impact of potential spillovers, the CCA stress tests controlled for a number of different factors. The use of quantile-based default probabilities served as a control for idiosyncratic risk at the firm level. The presence of macro variables with high explanatory power in the estimated regressions was used to control for macroeconomic risks and the inclusion of a statistically significant connectivity measure was used as a control for other relational changes between sectors. 101 Given these controls, the relationships estimated between sector default probabilities are believed to effectively capture the effect of intersectoral spillovers.

165. Under the stress scenario, expected one-year median default probabilities for the U.S. financial system are projected to increase from 0.20 percent to 0.65 percent, about two-thirds the level seen at the height of the 2008–09 financial crisis. Figure 21 shows the historical evolution of overall system median default probabilities along with forecasts under the baseline and stress scenarios. On average, default probabilities for all sectors (Appendix Figure 9), are expected to increase to about two thirds of their 2008–09 levels, with the exception of corporates and asset managers which are expected to experience smaller increases. Using a threshold of 0.5 percent or lower as a low credit risk boundary,102 under the stress scenario banks, insurers, NBFIs, and foreign banks exit this “safe zone” during the period of peak stress which occurs in 2015–16. Corporates, asset managers, and foreign insurers never breach this boundary though, even under severely adverse macroeconomic conditions. Only projections for the overall U.S. financial system model explicitly take into account changes in the estimated default probabilities of other sectors. Individual sector projections (Appendix Figure 9) were generated exclusively using macroeconomic and connectivity factors.

Figure 21.
Figure 21.

U.S. Financial System Median Default Probability Forecast

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations.Note: Blue lines indicate 25th and 75th percentile values of the distribution of historical estimates of U.S. institution 1-year ahead default probabilities. The dashed black line denotes the median value of the distribution of historical estimates of U.S. institution 1-year ahead default probabilities. The solid red and black lines denote median 1-year ahead default probabilities projected by the CCA stress tests under the stress and baseline scenarios, respectively. To better show projection details, the y-axis has been truncated at 1 percent. The line denoting the 75th percentile reached a maximum value of 2.5 percent in 2008-09. Only projections for the overall U.S. financial system model explicitly take into account changes in the estimated default probabilities of other sectors. Individual sector projections (Appendix Figure 9) were generated exclusively using macroeconomic and connectivity factors.

166. The analysis shows the importance of spillovers across sectors. Spillovers from the United States to the rest of the world were found to be large, but spillbacks from rest of the world to the United States appear relatively limited. The estimated effect of credit risk shocks (Figure 22) varies from sector to sector, but several observations are worth noting.

  • Shocks to the credit profiles of U.S. asset managers, NBFIs, insurers, and foreign banks increased U.S. bank credit risk the most.103

  • Increases in overall U.S. credit risk and credit risk in foreign insurers most negatively impacted the foreign banking sector, whereas the foreign insurance sector is most negatively affected by adverse changes in the credit profile of U.S. asset managers and insurers.

  • The effect of a rise in default risk in foreign banks and foreign insurers on U.S. default risk appears small. Excluding the impact of macro factors, non-U.S. domiciled banks and insurers were found to collectively account for only 3 percent of median U.S. default probabilities. Under the 2015–16 peak period of the stress scenario, this contribution is projected to fall by an additional one-third, to 2 percent of the total attributable to non-macro factor components.

Figure 22.
Figure 22.

Spillover Map

Citation: IMF Staff Country Reports 2015, 173; 10.5089/9781513591506.002.A001

Source: IMF staff calculations.Note: The figure shows the marginal contribution (in percent) of a one-time shock on a recipient sector’s default probabilities given a severe 50 basis point (0.5 percent) shock to an originating sector’s default probabilities. The results were calculated using the GAMLSS inter-sector relationships estimated as of end 2014Q3. Sector end of period default probabilities were separately increased by 50 basis points and shocks were measured as the deviation in a recipient sector’s probability of default from its unshocked historical baseline. Dashed red lines denote cross-border effects of credit risk shocks. For this exercise, spillover effects were captured by incorporating historical estimates of median sector default probabilities as independent variables into the final specification of each of the sector projection models used to generate the results shown in Appendix Figure 8.

Recommendations for Improvement

167. While the authorities’ solvency stress test for BHCs are well advanced, and state-of-the-art in many respects, there is scope for enhancement. It would be useful to try to link liquidity, solvency and network analysis in a systemic risk stress testing framework. For example, Bank of Canada’s Macro Financial Risk Assessment Framework captures the various sources of risk (solvency, liquidity and spillover effects) within a single stress testing framework (Bank of Canada, 2014). Moreover, reexamining some of the solvency stress test assumptions to make them consistent with historical evidence would be useful. For example, there may be merit in reexamining the stress test assumptions on loan and asset growth as well as dividend distribution.

168. Increasing the coverage of the tests would be helpful. In particular, the FRB should include the largest SLHCs in the supervisory stress tests once they start performing company-run stress tests (from 2017).

169. Establishing a regular liquidity stress testing framework for banks will be an important further step. The announced Comprehensive Liquidity Analysis and Review (CLAR), that is expected to be launched by the end of 2015, is a step is the right direction. This will complement the solvency testing under the DFA.

170. Another area for improvements relates to modeling network contagion. The network contagion exercise here illustrates the need to expand the FRB’s data on interbank exposures to include a richer set of dynamics and a broader range of counterparties.

171. In insurance, the focus should be on developing and performing insurance stress tests on a consolidated, group-level basis. This is especially important for groups that are (i) designated as systemically important; (ii) engaged in material group-internal risk transfer, e.g., via captives; or (iii) exposed to non-linear market risks through the sale of products which include guarantees or optionalities, e.g., variable annuities. It would also be useful to improve public disclosure by requiring insurance companies to disclose market risk sensitivities in a more harmonized manner.

172. Regular system-focused liquidity risk analysis for the mutual fund industry should be done on a regular basis. At present, a considerable range of bottom-up analyses is performed by the industry. According to the authorities, rule-making is forthcoming to standardize stress tests by mutual funds with consolidated assets of $10 billion or more. The authorities are encouraged to further clarify the guidance to the industry on liquidity risk analysis, and to start conducting regular top-down analysis to provide a more holistic picture of the industry’s contribution to systemic risk.

173. The authorities are encouraged to conduct more intensive monitoring of systemic financial sector risks, including the use of market-based solvency and shortfall measures. Market-price based stress tests employ forward-looking, higher-frequency, market consensus information that, when used appropriately, can add value to traditional stress tests in a variety of ways. While the market-price based analysis has its limitations, it can be a useful “cross-check” to corroborate the findings of other stress test methodologies. They can also be readily extended to assess the safety and soundness of sectors which are not traditionally subject to bank-like supervisory oversight.


  • Acharya, Viral, Robert Engle, and Matthew Richardson, Capital Shortfall: A New Approach to Ranking and Regulating Systemic Risks, AEA, January 7, 2012—SRISK Model, NYU Lab (New York: New York University).

    • Search Google Scholar
    • Export Citation
  • Adrian, Tobias, Michael Fleming, Jonathan Goldberg, Morgan Lewis, Fabio Natalucci, and Jason Wu, 2013: “Dealer balance sheet capacity and market liquidity during the 2013 selloff in fixed-income markets”, Liberty Street Economics, 16 October.

    • Search Google Scholar
    • Export Citation
  • Bank of Canada, 2014, “Stress Testing the Canadian Banking System: A System-Wide Approach”, Financial System Review, June (Bank of Canada)

    • Search Google Scholar
    • Export Citation
  • Basel Committee on Banking Supervision (BCBS), 2010a, “Guidance for national authorities operating the countercyclical capital bufferhttp://www.bis.org/publ/bcbs187.pdf (BCBS: Basel).

    • Search Google Scholar
    • Export Citation
  • Basel Committee on Banking Supervision (BCBS), 2010b, “Basel III: International Framework for Liquidity Risk Measurement, Standards and Monitoring,December (BCBS: Basel).

    • Search Google Scholar
    • Export Citation
  • Basel Committee on Banking Supervision (BCBS), 2013, “Basel III: The Liquidity Coverage Ratio and liquidity risk monitoring tools,January (BCBS: Basel).

    • Search Google Scholar
    • Export Citation
  • Basel Committee on Banking Supervision (BCBS), 2014, “Basel III Monitoring Report,September (BCBS: Basel).

  • Blancher, Nicolas, Srobona