Systemic risk and financial regulation
Huang, Qiubin
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date: 2019
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Huang, Q. (2019). Systemic risk and financial regulation. University of Groningen, SOM research school.
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
4
4
Impact of the Dodd-Frank Act on
Systemic Risk
∗
Abstract
This chapter introduces a two-step strategy — the synthetic con-trol method combined with the difference-in-differences method — to evaluate the treatment effect of the Dodd-Frank Act (DFA). We find no evidence from our counterfactual analysis in support of the DFA reducing systemic risk in the US banking system. Our results suggest that endogenous risk persistence is the main driver for the decrease in systemic risk in the post-DFA period. Additional analy-ses regarding the predictive power of the synthetic control method, endogeneity concerns and anticipation effects support the robustness of our findings.
∗
This chapter is based on Huang (2018). Part of this research was done during my internship at De Nederlandsche Bank. The views expressed do not necessarily reflect the views of De Nederlandsche Bank or the Eurosystem. I would like to thank Mark Mink, Jon Frost, Robert Vermeulen, Mastrogiacomo Mauro, Yue Li and Haizhen Yang for their productive discussions, and the participants at the De Nederlandsche Bank Lunch Seminar and the IFABS Oxford 2017 Conference for their valuable comments.
4
4.1
Introduction
The 2008 Global Financial Crisis (GFC) has led to a wave of financial regu-latory reform. In the US, the Dodd-Frank Wall Street Reform and Consumer Protection Act (known as Dodd-Frank Act) was passed on July 21, 2010 to bring a sweeping overhaul to the financial regulatory system. The Dodd-Frank Act (DFA) aims to promote US financial stability and to protect US taxpayers and consumers. To this end, Title I of the DFA establishes the Financial Sta-bility Oversight Council (FSOC) which has to identify and deal with the risks to US financial stability. Following the immediate establishment of the FSOC, various provisions of the DFA have been gradually put into practice.1 As of December 31, 2015, 267 out of 390 (68.46%) rules were finalized.2 However, the effectiveness of the new regulations remains unclear.
Recently, former US Federal Reserve’s Chairperson Janet Yellen expressed her appreciation for the regulatory reforms instructed by the DFA. In a speech at the Federal Reserve’s annual economic policy symposium in August 2017, she claimed that reforms in the past years have boosted the resilience of the financial system, promoted market discipline and reduced the problem of too-big-to-fail.3 Several studies seem to support her opinions (e.g., see Goodhart et al., 2013; Balasubramnian and Cyree, 2014; Duarte and Eisenbach, 2015; and Hirtle et al., 2016). In contrast, the Trump Administration and some researchers doubted the effectiveness of the DFA and called for improvement of the current regulatory system (e.g., see Allen, 2010; Fama and Litterman, 2012; and Acharya and Richardson, 2012). On April 21, 2017, President Donald Trump signed two presidential memoranda to direct the Secretary of the Treasury to review the FSOC and the Orderly Liquidation Authority (Title II of the DFA). He argued that the DFA may encourage excessive risk-taking by financial companies and fail to promote market discipline and reduce systemic risk.4 On June 8, 2017, the Financial CHOICE Act was passed in the 115th Congress which repealed
1 See www.treasury.gov/initiatives/fsoc/Pages/dodd-frank.aspx for an
implementa-tion roadmap of the DFA provided by the FSOC.
2 Source: Dodd-Frank Progress Report, available at https://www.davispolk.com/Dodd
-Frank-Rulemaking-Progress-Report/.
3See www.federalreserve.gov/newsevents/speech/yellen20170825a.htm for more
de-tails about Yellen’s speech.
4For a detailed description on the reasons and contents to review the FSOC
and the Orderly Liquidation Authority, see www.whitehouse.gov/the-press-office/2017/ 04/21/presidential-memorandum-secretary-treasury and www.whitehouse.gov/the-press -office/2017/04/21/presidential-memorandum-secretary-treasury-0, respectively.
4
several provisions of the DFA.5
Clearly, the regulatory reform is still ongoing although it has been seven years since the passage of the DFA and a decade since the onset of the crisis. However, knowledge on the effects of the regulatory reform remains limited, as Janet Yellen also pointed out in her speech. Against this backdrop, a systematic assessment on the effects of the DFA seems urgently needed and valuable. In this chapter, we aim to empirically explore the fundamental question: Has the DFA contributed to reduce systemic risk in the US banking system? Such an assessment can provide empirical evidence for the effectiveness of the DFA in achieving its goals and may reconcile the recent debate on the effects of the DFA and the necessity of further financial regulatory reform.
We study the evolution of systemic risk in the US banking system from 1998 to 2015 based on two market-based measures of systemic risk, ∆CoVaR of Adrian and Brunnermeier (2016) and marginal expected shortfall (MES) of Acharya et al. (2017). We find that systemic risk in the US banking system remained relatively low in the pre- and post-GFC periods, but experienced a steep increase in the second half of 2008 and a quick decrease in the first half of 2009. The question of interest is whether the relatively low level of systemic risk in the post-GFC period can be attributed to the passage and implementation of the DFA. We address this question with a counterfactual analysis. Relying on the synthetic control method introduced by Abadie and Gardeazabal (2003), we construct a synthetic control group (SCG) for the treatment group of large US banks and then evaluate the significance of the DFA’s treatment effects in a difference-in-differences (DID) framework. We find no evidence in support of the DFA reducing systemic risk in the US banking system.
Assessing the effects of the DFA is notoriously difficult for three reasons. First, the DFA contains various regulations which have been coming into force haphazardly and may affect each bank differently. Therefore, a clean DID anal-ysis among US banks seems impossible. Second, as these regulations gradually came into force, it is important to take into account the dynamic treatment ef-fects of the DFA. Third, it is difficult to disentangle the DFA’s efef-fects from those of other confounders (e.g., the GFC and the European sovereign debt crisis). We take up these challenges in the following ways.
First, we identify a treatment group of 24 large US bank holding
compa-5For a summary of the provisions of the DFA which are repealed by the Financial CHOICE
4
nies (BHCs) with $50 billion or more in total consolidated assets in 2010 as a representation for the US banking system. Unlike recent papers studying the effects of new financial regulations, there is no an explicit rule in our context to separate the treated banks from the non-treated banks. The papers surveyed in Table 4.1 study policies that aim for or only matter for a fraction of banks so that the researchers can use those unaffected banks in the same country as a control group. This way is not allowed in our context because: 1) small US BHCs are also affected by the DFA; and 2) systemic risk is mainly generated by large BHCs rather than small BHCs. Therefore, we consider large EU BHCs as a candidate control group.
The next step is to apply an appropriate method to estimate a counter-factual for systemic risk in the US banking system in the absence of the DFA and evaluate the change of systemic risk relative to the counterfactual. To this end, we apply the synthetic control method to assign weights to EU BHCs so that the weighted combination of selected EU BHCs has similar trends in sys-temic risk as that of the treatment group in the pre-DFA period. This weighted combination serves as a synthetic control group (SCG) for the treatment group. After this, we follow the papers surveyed in Table 4.1 to evaluate the treatment effect of the DFA in a DID framework. The use of the synthetic control method ensures that different systemic risk indicators of the SCG simultaneously and closely track those of the treatment group in the pre-DFA period, while the DID framework allows us to explicitly control for potential confounders during our sample period.
Specifically, we set a dummy to capture the effect of the GFC. Our results suggest that the GFC results in the increase of the systemic risk in both groups. We also set a dummy to capture the effect of the European sovereign debt crisis and a dummy to capture the effect of the financial regulatory change within the EU. We find no significant impacts of the European sovereign debt crisis and the financial regulatory change within the EU. To better take into account other ef-fects, we also control for several fundamental variables, such as economic growth, inflation, interest rate, government bond yield and financial development. We do not find consistent significant results of these variables. In addition, we in-clude the one-period-lagged systemic risk indicator as an explanatory variable to account for the endogenous risk persistence following L´opez-Espinosa et al. (2012). We find that systemic risk in the banking system is mainly driven by endogenous risk persistence. This finding is supported by Fahlenbrach et al. (2012) who show that there is strong persistence in banks’ risk culture. Next
4
to the above variables, we also control for group fixed effects to account for the intrinsic differences between the two groups and time fixed effects to account for the unobserved differences over time (cf. Berger et al., 2019). Taking into ac-count all of the above variables’ potential impacts, we do our best to disentangle the DFA’s effects from other effects of potential confounders.
However, our analyses fail to find a significant treatment effect of the DFA in reducing systemic risk in the US banking system during the 2011–2015 period. As the DFA’s regulations have been gradually coming into force, we examine whether the DFA’s effects become more apparent later. We address this by interacting the DFA dummy with each year dummy for the post-DFA period in our DID regressions following Juelsrud and Wold (2018). We do no find significant results to support the DFA’s dynamic treatment effects. Overall, our study appears to support the criticism on the ineffectiveness of the DFA (e.g., see Acharya and Richardson, 2012).
To check the robustness of our findings, we perform additional analyses to address several concerns. Firstly, we examine the out-of-sample predictive power of the SCM. Only when the out-of-sample predictive power is strong, the SCG we construct can provide counterfactual results. We use part of the pre-DFA data to reconstruct a SCG and the other part as a comparison for predictive power analysis. Our results show that the out-of-sample predictive power of the SCG is as strong as its in-sample predictive power.
Secondly, we reflect on whether our research design involves endogeneity concerns and explain why our research design is not subject to endogeneity problems. Furthermore, one might be concerned that banks may respond in advance to the passage of the DFA. To examine potential anticipation effects, we reconstruct two SCGs using information available up to different months before the passage of the DFA, but we find no significant anticipation effects. This suggests that banks did not respond to the DFA in advance regarding their systemic risk contributions.
4 T able 4.1. Recen t studies on effects of new financial regulations Study Regulation V ariable of in terest T reatmen t and con trol groups dynamic treatmen t effects Blac k and Hazelw o o d (2013 ) T roubled Assets Relief Program (T ARP) risk-taking T ARP recipien t and non-recipien t banks No Duc hin and Sosyura (2014 ) T ARP risk-taking T ARP recipien t and non-recipien t banks No Berger and Roman (2015 ) T ARP comp e titiv e adv an-tage T ARP recipien t and non-recipien t banks No Berger et al. (2019 ) T ARP systemic risk T ARP recipien t and non-recipien t banks No Juelsrud and W old (2018 ) Basel II I capital requiremen ts capital adjustmen t strategy lo w-and high-capitalized banks in Nor-w a y Y es Gropp et al. (2019 ) 2011 EBA capital exercise capital adjustmen t strategy banks w ere and w ere not selected in to the exerc ise No Iv onc h yk (2019 ) Do dd–F rank Act m unicipal b ond prices b on ds with and without m unicipal ad-visors No Kleymeno v a and Zhang (2019 ) Do dd–F rank Act v olu n tary di sclosures large BHCs and other financial institu-tions No
4
Lastly, we examine whether our results are dominated by the six largest US BHCs. We find that systemic risk contributions of the six largest BHCs and other BHCs in the treatment group do not diverge around or after the passage of the DFA. In fact, they are highly correlated in both the pre- and post-DFA periods. Therefore, the results suggest that the DFA does not have a significantly different impact on systemic risk contributions of the six largest BHCs and the other BHCs. In short, the above tests strongly support the robustness of our main findings.
The rest of the chapter proceeds as follows. Section 4.2 reviews the DFA and its implementation progress. In Section 4.2, we also survey related research and highlight our contributions to the literature. Section 4.3 illustrates our research design and Section 4.4 provides the counterfactual analysis. Section 4.5 conducts additional analyses to support our counterfactual analysis. Section 4.6 concludes.
4.2
Review of the DFA and related literature
The DFA was signed into federal law by President Barack Obama on July 21, 2010. As a response to the 2008 financial crisis, the DFA aims to promote US financial stability by a comprehensive reform of the financial regulatory system. In this section, we briefly review the main provisions of the DFA that aim to promote financial stability and examine its implementation progress. In addition, we survey recent research closely related to the DFA and highlight our contributions to the literature.
4.2.1 The DFA and its implementation
The DFA consists of 16 titles, of which Title I (Financial Stability Act of 2010) establishes the FSOC and charges it with the important tasks of identifying and mitigating risks to US financial stability. The FSOC aims to create a com-prehensive and collective financial regulatory framework. Prior to the GFC, the US financial regulatory framework focused narrowly on individual institu-tions and markets. There was no single regulator responsible for monitoring and addressing overall risks to financial stability. This drawback resulted in some supervisory gaps and regulatory inconsistencies which encouraged regula-tory arbitrage. To address these problems, the DFA authorizes the FSOC to 1).
4
facilitate regulatory coordination; 2). facilitate information sharing and collec-tion; 3). designate non-bank financial companies for consolidated supervision; 4). designate systemic financial market utilities and systemic payment, clear-ing, or settlement activities; 5). recommend stricter standards and 6). break up firms that pose a “grave threat” to financial stability.
Title II prescribes procedures for orderly liquidation of financial entities. Title III transfers powers of the Office of Thrift Supervision to the Comptroller of the Currency, the Federal Deposit Insurance Corporation, and the Board of Governors of the Federal Reserve to enhance financial institutions’ safety and soundness. Title IV amends the Investment Advisers Act of 1940 to repeal its exemption and applies registration requirements to a private fund investment adviser. Title V makes changes to regulation for insurance companies. Title VI improves the regulation of bank and savings association holding companies and depository institutions. Title VII makes regulations on swap markets to increase Wall Street’s transparency and accountability. Title VIII, for the first time, defines the terms “systemically important” and “systemic importance”, and directs the FSOC to designate and supervise financial market utilities or payment, clearing, or settlement activities which are, or are likely to become, systemically important. Titles IX to XIV are more about protections for in-vestors, consumers and taxpayers while the last two titles are miscellaneous provisions.
Although these titles focus on different elements of the US banking system, they share the same spirit of promoting US financial stability by enhancing su-pervision and regulation. Especially, the DFA requires much more stringent regulation on large, interconnected bank holding companies (LIBHCs) as they probably pose the greatest risks to US financial stability (FED, 2012). Relevant provisions are written in Section 165 of the DFA which requires the Federal Re-serve Board to impose enhanced supervision and prudential standards for BHCs with total consolidated assets of $50 billion or more to prevent or mitigate risks to US financial stability that could arise from these LIBHCs’ material financial distress or failure, or ongoing activities. According to Section 165, the Board needs to establish five required prudential standards, including risk-based capi-tal requirements and leverage limits, liquidity requirements, overall risk manage-ment requiremanage-ments, resolution plans and credit exposure report requiremanage-ments, and concentration limits. The Federal Reserve Board is also authorized to con-sider additional standards, such as a contingent capital requirements, enhanced public disclosures, and short-term debt limits.
4
Subsequent to the passage of the DFA, the FSOC, the Board and other regu-latory agencies have issued various regulations, policies and guidelines mandated by the DFA and taken necessary actions to carry out the DFA. For example, the FSOC made its final rule on July 27, 2011 to designate financial market utilities as systemically important. The Board issued a consolidated supervision frame-work for large financial institutions on December 17, 2012. According to the Dodd-Frank Progress Report, there are in total 390 rulemaking requirements in the DFA, of which 267 (68.46%) have been met as of the end of 2015. In addition, 271 rulemaking requirement deadlines have passed as of the end of 2015, but only 204 (75.3%) have been met.
Although the implementation progress of the DFA seems a bit slow, the DFA has brought many changes to the development of the US banking system. One of the most apparent changes is the increased capital levels at BHCs due to the enhanced capital requirements. The common equity tier 1 ratio of the 31 LIBHCs started to decrease from 7.07% in the fourth quarter of 2005 to 4.82% in the fourth quarter of 2008, but steadily climbed up to 11.82% in the fourth quarter of 2015. Capital levels at LIBHCs were about two third higher than their pre-crisis levels. BHCs other than LIBHCs also experienced an increase in their capital levels to 12.1% as of the fourth quarter of 2015, about one fifth higher than that before the GFC (see FSOC, 2016). The higher capital levels are expected to enable banks to better absorb losses and reduce bank insolvency, thereby enhancing financial stability in the US banking system. Berger and Bouwman (2013) show that higher equity capital increased medium and large US banks’ survival rates during banking crises. However, banks with higher capital levels may be also associated with higher risk-taking and distress probability. Duchin and Sosyura (2014) find that US banks receiving capital infusions during the 2008-2009 period took greater risk in their retail and corporate lending and investment practices. These banks appeared safer according to regulatory capital ratios, but showed an increase in volatility and default risk. The benefit of banks’ higher capital levels may just be offset by their greater risk-taking. In the next subsection, we provide a detailed survey of recent research related to the DFA and highlight our contributions to the literature.
4.2.2 Related literature and our contributions
The passage of the DFA has led to research aiming to examine its effectiveness. A few studies show that the DFA, to some extent, has taken effect in
promot-4
ing market discipline and reducing risk in the financial system. For instance, Balasubramnian and Cyree (2014) find that the discount for size on yield spreads of subordinated debt was reduced by 47% and the too-big-to-fail discount was reduced by 94% after the adoption of the DFA, concluding that the DFA has improved market discipline. Akhigbe et al. (2016) find that the largest financial institutions experienced the greatest reduction in individual risk and market risk following the passage of the DFA, reflecting the effectiveness of stricter reg-ulations on large banks. Acharya et al. (2018) find that banks subject to DFA Stress Tests reduced credit supply to relatively risky borrowers to decrease their credit risk, uncovering a positive impact of the DFA on financial stability.
In contrast, some studies find limited or even negative effects of the DFA. For instance, Andriosopoulos et al. (2017) report that financial institutions’ market risk significantly increased following the passage of the DFA. Acharya et al. (2016) find that credit spreads were sensitive to risk for most financial institutions, but not for the largest financial institutions due to their systemic importance. They argue that the DFA has not changed investors’ expectations of implicit government support for large financial institutions. Poghosyan et al. (2016) document similar results by exploring the relationship between size and support ratings of US banks. They even find that the link became stronger after the passage of the DFA, implying the ineffectiveness of the DFA to end too-big-to-fail. Gao et al. (2018) report that larger financial institutions experienced more negative abnormal stock returns and more positive abnormal bond returns during the legislative process of the DFA, but this phenomenon disappeared around the passage of the DFA. The authors argue that this is because market participants worried about the stringent regulations of the DFA at the beginning, but they started to doubt the effectiveness of the DFA when its final version came out. Dimitrov et al. (2015) demonstrate that the DFA had a negative impact on credit rating agencies, causing them to issue lower ratings, give more false warnings and issue downgrades that are less informative following the passage of the DFA. Bouwman et al. (2018a) and Bindal et al. (2017) explore the impact of the DFA’s regulatory size threshold and uncover its indirect impact on asset growth strategies of banks with assets below but close to the size threshold.
Our research differs from these studies in two ways. First, previous studies focus on stock returns or individual and market risk measures, but the effective-ness of the DFA in reducing systemic risk has not been empirically examined. We contribute to the literature by directly examining the effectiveness of the DFA in reducing systemic risk. Second, most previous studies focus on the
leg-4
islative process of the DFA while we examine the period since the passage of the DFA. Therefore, our paper assesses the effectiveness of the DFA’s imple-mentation rather than market participants’ expectations during the legislative process. Our study is more meaningful given that many provisions of the DFA have been gradually implemented during the 2010-2015 period and the Trump Administration has started to reflect on the DFA and has taken some actions to revise it.
Our research makes an innovation regarding the methodology. We show how to use the SCM integrated with the DID method. Although the SCM and the DID method have been widely used in various fields, such as health policy, politics and economics, they are rarely used together to address their respective weaknesses. The integration of the SCM and the DID method can address the common trends assumption under the DID method and the fact that the SCM does not account for post-treatment confounders. Recent literature on financial policy evaluations mainly relies on the DID method (e.g., Beck et al., 2010; Duchin and Sosyura, 2014; Berger and Roman, 2015; Dinger and Vallascas, 2016; Berger and Roman, 2017; Bouwman et al., 2018a; and Acharya et al., 2018), with one exception: Acharya et al. (2018) use the difference-in-differences-in-differences method. To the best of our knowledge, our research is the first to integrate the SCM with the DID method to conduct policy evaluation.6 In this respect, our research contributes to the above literature by proposing a new methodology for policy evaluation.
4.3
Research design
This section elaborates our research design for evaluating the impact of the DFA on systemic risk in the US banking system. We exploit the passage of the DFA as a quasi-experiment and identify the treatment group and the candidate comparison group in Section 4.3.1. Section 4.3.2 introduces two market-based systemic risk measures and Section 4.3.3 describes our techniques to evaluate the DFA’s effectiveness.
6We would like to point out that another possible method is the regression discontinuity
design as there is a clear size threshold (i.e., $50 billion) for banks to subject to stricter regulation. However, the regression discontinuity design requires that there is a large sample around the threshold. This requirement cannot be met in our sample. See Imbens and Lemieux (2008) and Lee and Lemieuxa (2010) for details about the regression discontinuity design.
4
Table 4.2. US BHCs in the treatment group
This table presents US BHCs with consolidated assets greater than $50 billion in 2010. We require the BHCs to have continuous stock price information since 1997 for estimating their systemic risk contributions and for comparative analyses. Name, ticker and asset information are obtained from Datastream. The last column indicates bank assets ($ billion) in 2010.
NO. Name Ticker Assets
1 American Express Company AXP 143.65 2 Bank of America Corporation BAC 2264.91 3 Bank of New York Mellon Corporation BK 247.26
4 BB&T Corporation BBT 156.23
5 BMO Financial Corp. BMO 474.64
6 Capital One Financial Corporation COF 197.50 7 Charles Schwab Corporation SCHW 92.40
8 Citigroup Inc. C 1913.90
9 Comerica Incorporated CMA 53.67
10 Fifth Third Bancorp FITB FITB 111.01 11 Goldman Sachs Group Inc. GS 905.09 12 Huntington Bancshares Incorporated HBAN 53.82 13 JPMorgan Chase & Co. JPM 2117.61
14 KeyCorp KEY 91.41
15 M&T Bank Corporation MTB 68.02
16 Morgan Stanley MS 807.70
17 Northern Trust Corporation NTRS 83.84 18 PNC Financial Services Group Inc. PNC 264.28 19 Regions Financial Corporation RF 132.35 20 State Street Corporation STT 158.72
21 SunTrust Banks Inc. STI 172.87
22 U.S. Bancorp USB 307.79
23 Wells Fargo & Company WFC 1258.13
24 Zions Bancorporation ZION 50.49
4.3.1 The treatment and control groups
The DFA aims to limit systemic risk contributions of large financial institutions which probably pose the greatest risks to the US banking system and the econ-omy (FED, 2012). Several steps have been taken to limit the incentives and abilities of these large financial institutions to take risk (see FSOC, 2013, 2015). Therefore, this research concentrates on large financial institutions to examine the effect of the DFA. For our research purpose, we require that the consolidated assets of a BHC in 2010 must at least be $50 billion to ensure that the
provi-4
sions of the Dodd-Frank Act apply.7 This is because many provisions of the DFA are issued to limit the incentives and abilities of financial institutions with consolidated assets equal to or great than $50 billion to take risk (FSOC, 2013). Besides, a BHC should have continuous stock price information since 1997 so that we have sufficient data to estimate systemic risk and perform comparative analyses. As a result, we have 24 BHCs in the treatment group (see Table 4.2), accounting for more than 70% of total assets of top-tier BHCs in the US (FSOC, 2015).
Since 2012, the Federal Reserve conducted annual stress tests for 31 BHCs as required by the DFA, of which 23 are included in our sample. The other eight BHCs are filtered out because there is not enough information on their stock prices. These BHCs are Ally Financial, Citizens Financial Group, Discover Financial Services and BBVA Compass Bancshares which were listed after 2007, and Deutsche Bank Trust Corporation, HSBC North America Holdings, MUFG Americas Holdings Corporation and Santander Holdings USA which have not been listed. We also include Charles Schwab Corporation in the treatment group, as it has more than $50 billion in total consolidated assets although it is not included in the Federal Reserve’s annual stress tests.
We need to form a control group of banks unaffected by the DFA but with similar trends in systemic risk as banks in the treatment group prior to the passage of the DFA. A natural way is to use US BHCs with consolidated assets less than $50 billion as a comparison. We do not adopt this approach because these BHCs are also subject to the DFA (see Peirce et al., 2014; Bindal et al., 2017; and Bouwman et al., 2018a). Instead, we consider large BHCs in the European Union as candidates, because they are hardly affected by the DFA and play a similar role in the European financial system as large BHCs do in the US banking system. Therefore, we apply the same rules for selecting US BHCs to European BHCs, which results in 42 BHCs from 15 European countries (see Table 4.3). These BHCs cover several global systemically important banks identified by the Financial Stability Board and the Basel Committee on Banking Supervision and cover most banks which are (potentially) global systemically important institutions identified by the European Banking Authority.
7 For example, the DFA requires the Federal Reserve to conduct an annual stress test of
BHCs with $50 billion or more in total consolidated assets and all non-bank financial companies designated by the FSOC for Federal Reserve supervision. The Federal Reserve Board adopted rules implementing this requirement in October 2012.
4
Table 4.3. EU BHCs in the candidate control group
This table presents EU BHCs to be considered as control banks. † indicates BHCs that are identified by the European Banking Authority as (potentially) global systemically important banks (G-SIBs) in 2015. ‡ indicates BHCs that are G-SIBs identified by the Financial Stability Board and the Basel Committee on Banking Supervision in 2015. Name, ticker and asset information are obtained from Datastream. The last column indicates bank assets ($ billion) in 2010.
NO. Name Ticker Country Assets
1 Erste Group Bank AG† ERS Austria 275.13
2 Dexia SA DEXB Belgium 755.05
3 KBC Groep NV† KBC Belgium 426.31
4 Danske Bank AS† DAB Denmark 576.49
5 BNP Paribas SA†, ‡ BNP France 2662.54
6 Societe Generale SA†, ‡ SGEX France 1508.98
7 Natixis SA KN France 608.63
8 Commerzbank AG† CBK Germany 1005.00
9 Deutsche Bank AG†, ‡ DBK Germany 2539.88
10 Alpha Bank AE PIST Greece 88.85
11 Eurobank Ergasias SA EFG Greece 116.04
12 National Bank of Greece SA ETE Greece 161.01
13 Piraeus Bank SA PEIR Greece 76.66
14 Bank of Ireland Group BKIR Ireland 222.68
15 Permanent TSB Group Holdings PLC IL0A Ireland 98.50
16 Allied Irish Banks plc ALBK Ireland 191.22
17 UniCredit SpA†, ‡ UCG Italy 1229.19
18 Intesa Sanpaolo SpA† ISP Italy 873.87
19 Banca Carige SpA CRG Italy 53.08
20 Banca Popolare di Milano Scarl PMI Italy 71.80 21 Banca Popolare dell’Emilia Romagna Sc BPE Italy 77.71 22 Mediobanca Banca di Credito Finanziario SpA MB Italy 108.56 23 Banca Monte dei Paschi di Siena SpA† BMPS Italy 321.51
24 ING Groep NV†, ‡ INGA Netherlands 1657.16
25 DNB ASA† DNB Norway 319.48
26 Banco BPI SA BPI Portugal 60.55
27 Banco Comercial Portugues SA BCP Portugal 132.96 28 Banco Bilbao Vizcaya Argentaria SA† BBVA Spain 732.53
29 Banco Santander SA†, ‡ SCH Spain 1606.98
30 Bankinter SA BKT Spain 72.37
31 Banco Popular Espanol SA POP Spain 172.96
32 Nordea Bank AB†, ‡ NDA Sweden 777.19
33 Svenska Handelsbanken AB† SVK Sweden 320.65
34 Skandinaviska Enskilda Banken AB† SEA Sweden 324.36
35 Swedbank AB† SWED Sweden 255.32
36 Credit Suisse Group AG‡ CSGN Switzerland 1094.84
4 Table 4.3 (continued)
38 Royal Bank of Scotland Group PLC †, ‡ RBS UK 2257.72
39 Barclays PLC†, ‡ BARC UK 2320.01
40 HSBC Holdings PLC †, ‡ HSBA UK 2459.13
41 Lloyds Banking Group PLC † LLOY UK 1540.42
42 Standard Chartered PLC†, ‡ STAN UK 515.60
4.3.2 Systemic risk measures
To capture systemic risk in the US banking system, we use two different mea-sures, ∆CoVaR of Adrian and Brunnermeier (2016) and MES of Acharya et al. (2017). ∆CoVaR is the change in the financial system’s value-at-risk con-ditional on an institution being in distress relative to its median state. MES is the expectation of an institution’s equity return when the financial system is in distress. By definition, ∆CoVaR and MES capture systemic risk from dif-ferent perspectives, but both measures highlight the tail-dependency between individual institutions and the financial system. They have been demonstrated to timely capture the increase of individual banks’ systemic risk contributions during the 2007-2009 financial crisis (see Adrian and Brunnermeier, 2016; and Acharya et al., 2017) and have been widely used to study systemic risk in various financial systems.8
In addition, both measures employ market data to estimate systemic risk and have not been monitored by central banks. Therefore, they are not subject to Goodhart’s law that “when a measure becomes a target, it ceases to be a good measure” (Strathern, 1997, p. 308). Below we introduce the ∆CoVaR and MES methods to estimate individual banks’ systemic risk contributions. After calculating each bank’s ∆CoVaR and MES, we take asset-weighted aver-age ∆CoVaR and asset-weighted averaver-age MES across banks as two aggregate measures to indicate the degree of systemic risk in the financial system. These aggregate measures can also serve as indicators of financial stability (FSOC, 2011).
CoVaR can be defined as Value-at-Risk (VaR) of the financial system
con-8For example, L´opez-Espinosa et al. (2012) apply the ∆CoVaR measure and Weiß et al.
(2014) apply the MES measure to international banks while Huang et al. (2017) apply both measures to Chinese banks. Both measures have been widely used not only in academia but also among central banks (Benoit et al., 2017). We refer to Bisias et al. (2012) and Benoit et al. (2017) for a detailed discussion of these systemic risk measures.
4
ditional on a bank being in distress:
P [(rs≤ CoV aRqs|i)|ri= V aRqi] = q, (4.1) where rsand ri indicate returns of the financial system and bank i, respectively; and q is the quantile to be selected. Adrian and Brunnermeier (2016) propose to estimate a bank’s systemic risk contribution to the system as:
∆CoV aRqs|i= CoV aRqs|r
i=V aRqi
− CoV aRq
s|ri=V aR50%i
. (4.2)
In this chapter, we estimate CoVaR using quantile regressions with the quantile q set to 5% and use the same set of state variables adopted by Adrian and Brunnermeier (2016). These state variables include the change in the three-month yield, the change in the slope of the yield curve, a short-term TED spread, the change in the credit spread, the market return, the real estate sector return and equity volatility. In our analysis, we use the same set of state variables sam-pled from the US market as common state variables for US banks and European banks, which seems reasonable due to the strong degree of globalization in the financial industry and the predominance of the US economy (L´opez-Espinosa et al., 2012). We estimate CoVaR at a daily frequency and take the negative value of ∆CoVaR to translate it into an increasing measure of systemic risk. Further, we take the equally-weighted average of daily ∆CoVaR in a month as ∆CoVaR in that month for our research purpose. We then take the asset-weighted monthly ∆CoVaR across US BHCs in our sample as the first measure of systemic risk in the financial system.
Contrary to the CoVaR approach that examines the system’s distress con-ditional on individual institutions’ distress, Acharya et al. (2017) propose to estimate systemic risk from individual institutions’ distress conditional on the system’s distress. The definition of the banking system’s expected shortfall (ES) is as follows:
ESq = −E[rs|rs≤ V aRq]. (4.3) We decompose the banking system’s return into the sum of each bank’s return (ri) and rewrite Equation (4.3) as:
ESq = − X
i
yiE[ri|rs≤ V aRq] (4.4) where yi indicates the weight of bank i in the banking system and rs=
P iyiri. Acharya et al. (2017) define a bank’s marginal expected shortfall (MES) as:
M ESi= ∂ESq
∂yi
4
We calculate a bank’s average equity return during the worst 5% market outcomes at daily frequency as an estimator for MES. Acharya et al. (2017) and Weiß et al. (2014) also adopt this approach as it is easy to apply and allows to avoid model misspecification. We calculate MES in a 12-month rolling window in order to obtain a monthly observation for MES for each bank. We then take the asset-weighted MES across banks as our second measure of systemic risk in the financial system. For US banks, the CRSP value-weighted index is used as a proxy for the US stock market. For European banks, the MSCI Europe Index is taken as a proxy for the European stock market. Data on stock returns are obtained from Datastream.
4.3.3 Methodology for evaluating the DFA’s effect
In the field of policy evaluation, the DID method has been widely used since the early 1990s while the SCM is the most important innovation in the literature on policy evaluation in the last 15 years (Athey and Imbens, 2017). In practice, the use of the DID method usually faces the challenge of common trends that the outcomes of the treated and control groups should have followed parallel trends in the absence of the treatment. This assumption is not directly testable and difficult to meet in many cases (Xu, 2017). The SCM, which has first been used by Abadie and Gardeazabal (2003) and later formalized by Abadie et al. (2010, 2015), gets rid of the parallel trends assumption by directly constructing a synthetic group that can assemble the covariates and outcomes of the treatment group in the pre-treatment period. In the post-treatment period, the gap of the outcomes between the treatment and control groups is considered as the treatment effect. To evaluate the significance of the treatment effect, Abadie et al. (2010, 2015) propose to compare this effect with the placebo effects obtained by randomly assigning the treatment to control units. This procedure, however, does not take into account potential confounders in the post-treatment period. In this chapter, we propose a simple two-step strategy to perform policy evaluation. That is to use the SCM to construct a synthetic control group whose pre-treatment trend in outcomes is similar to that of the treatment group. Based on the treatment and synthetic control groups, we then apply the regression-based DID method to estimate the treatment effect, controlling for potential confounders in the post-treatment period.9 By integrating the DID method
9We are more concerned with confounders in the post-DFA period in view of the European
4
with the SCM, this two-step strategy can address the weaknesses mentioned above when using these two methods separately. In addition, the use of the SCM can help to reduce discretion in the choice of control units and guard against extrapolation usually applied in the DID method (Abadie et al., 2010, 2015). Below we introduce the techniques in more detail.
The synthetic control method
Suppose that there are N + 1 units and only the first one is exposed to the treatment of interest. We observe outcomes of interest for T periods where the treatment starts in T0+ 1 (T0 < T ). In the context of this research, unit, treatment and outcome of interest are bank, the DFA and systemic risk, respec-tively. Below we keep using the terms “unit”, “treatment” and “outcome” for simplifying the exposition.
Denoting the outcome vector as Yi = (Yi1, · · · YiT0, · · · , YiT)
0, the observed outcome for unit i at time t is
Yit= YitU + αitDit, (4.6) where YitU is the observed outcome for unit i at time t in the absence of the treatment; Dit is a dummy variable that takes value one if unit i is subject to the treatment at time t, and value zero otherwise; and αitdenotes the treatment effect. As we assume that only the first unit is exposed to the treatment, we aim to estimate α1t for t > T0:
α1t= Y1t− Y1tN. (4.7)
To obtain α1t, we only need to estimate Y1tN because Y1t is observed. The strategy is to construct a weighted combination of untreated units to reassemble the characteristics of the treated unit before the treatment and then use the outcome of the weighted combination in the post-treatment era as an estimate of the counterfactual outcome (i.e. Y1tN) for the treated unit. To this end, Abadie et al. (2010) assume that YjtN is given by a factor model:
YjtN = δt+ θtZj+ λtµj+ εjt (4.8) where δt is a time fixed effect, indicating an unknown common factor with constant factor loading across banks; Zj is a vector of observed time-invariant
4
factors with a time-varying coefficient vector θt; λt is a vector of unobserved common factors with a factor loadings vector µj; and εjt are unobserved tran-sitory shocks with zero mean. Consider a N × 1 vector of weights W = (w2, · · · , wi, · · · , wN +1)0 where wi is the contribution of unit i to the synthetic control group and PN +1
i=2 wi= 1. According to Equation (4.9), the outcome for the synthetic control group can be written as:
N +1 X i=2 wiYit= δt+ θt N +1 X i=2 wiZi+ λt N +1 X i=2 wiµi+ N +1 X i=2 wiεit (4.9)
Suppose that there is W∗ = (w2∗, · · · , w∗i, · · · , wN +1∗ )0 such that for 1 ≤ t ≤ T0, N +1 X i=2 wi∗Yit= Y1t, (4.10) and N +1 X i=2 w∗iZi= Z1. (4.11)
Then the counterfactual outcome of the treatment unit is estimated as: ˆ Y1tN = N +1 X i=2 wi∗Yit (4.12)
and α1t for t > T0 is estimated as: ˆ
α1t= Y1t− ˆY1tN. (4.13) Abadie et al. (2010) demonstrate that ˆα1t is an approximately unbiased estima-tor of α1tas long as the number of pre-treatment periods is large relative to the scale of idiosyncratic shocks and the outcome is a linear function of potential confounders.
To look for the optimal vector of weights, Abadie and Gardeazabal (2003) suggest to minimize the discrepancy between predictors of the outcomes for the treatment group and the synthetic group before the treatment as follow:
f (W ) = min{p(X1− X0W )0V (X1− X0W )} (4.14) s.t. : wi ≥ 0(i = 2, 3, · · · , N + 1) and PN +1i=2 wi= 1,
4
where X1 is a (K × 1) vector of pre-treatment values of K covariates and out-comes of the treated unit, X0 is a (K × N ) matrix with values of the same variables as those in X1 for the N candidate control units, and V is a (K × K) positive definite and diagonal matrix which assigns weights according to the relative importance of the variables. We restrict the weights to be non-negative and sum to one to prevent extrapolation. Without this restriction, X1 would be perfectly fitted as long as the rank of X0 is equal to K, no matter how far is X1 from X0 (Abadie and Gardeazabal, 2003). V can be decided upon based on subjective knowledge of the importance of the variables. In our analysis, we find that adding covariates (such as asset size and profitability of banks) does not improve the predictive power compared to only using the outcome variable. Therefore, we only minimize the discrepancy between systemic risk indicators of the US banking system and the synthetic control group before the treatment. In Section 4.5.1, we examine the out-of-sample predictive power of the synthetic control group constructed in this way and find that it is as good as the in-sample predictive power.
In addition, we have two indicators of systemic risk (∆CoVaR and MES). We can construct a synthetic group that has a similar trend in ∆CoVaR as the treatment group and construct another synthetic group that has a similar trend in MES with the treatment group following Equation (4.14). Alternatively, we can construct a synthetic group that simultaneously has similar trends in ∆CoVaR and in MES as that of the treatment unit by changing Equation (4.14) to Equation (4.15):
f (W ) = minnq(X∆CoV aR
1 − X0∆CoV aRW )0(X1∆CoV aR− X0∆CoV aRW ) + q (XM ES 1 − X0M ESW )0(X1M ES− X0M ESW ) o (4.15) Theoretically, the latter way is stricter as it requires the synthetic group simul-taneously matches the treatment group in terms of ∆CoVaR and MES, while the former way only considers one dimension of systemic risk.10 Therefore, we adopt the latter approach in this chapter.
Notice that the above illustration of the synthetic control method is for the case that there is only one treated unit. For the case that there are multiple treated banks, we first aggregate the outcomes of treated banks and then
per-10Empirically, we find that the results obtained from these two approaches are very similar.
4
form the above procedure following the suggestion of Abadie et al. (2010).11 In other words, we take asset-weighted average ∆CoVaR and asset-weighted aver-age MES across US banks as two indicators of systemic risk in the US banking system. We then use the SCM to assign weights to candidate EU banks to con-struct a synthetic control group which can generate similar trends in ∆CoVaR and MES as that of the US banking system. In the next subsection, we ex-plain how to use the DID method to estimate the treatment effect based on the treatment and synthetic control groups.
The difference-in-differences method
After constructing the synthetic control group, we apply the DID method to estimate the treatment effect of the DFA and evaluate its significance after controling for potential confounders. We adopt the following baseline model with fixed effects:
Yit= α + βGiDF A + ρYit−1+ Xit−1γ + δi+ ηt+ εit, (4.16) where Yitindicates the systemic risk indicator (∆CoVaR or MES) of the financial system; Gi is a group dummy with value of 1 for the US banking system and 0 for the synthetic control group; DFA is a treatment dummy with value of 1 in the period from July 2010 to December 2015 and 0 before July 2010; and the coefficient β captures the treatment effect of the DFA. In Equation (4.16), δi is the group-specific fixed effects that capture the difference between two groups even when there was no DFA. ηt indicates the year-specific fixed effects to capture the difference among years even when there was no DFA. In the following analyses, we use cluster-robust standard errors to account for within-cluster correlation and heteroscedasticity that the fixed-effects estimator does not take into account (Cameron and Miller, 2015).
Billio et al. (2012) argue that systemic risk is an endogenous consequence when banks lend to similar industries, or due to inter-bank loans, derivatives, and other transactions. These stylized facts let banks become too-interconnected-to-fail and encourage the buildup of systemic risk. Therefore, we include Yit−1 in the right-hand side of Equation (4.16) to account for endogenous risk persis-tence following L´opez-Espinosa et al. (2012). We also control for a set of state variables (indicated by Xit−1) that may be related to the change of systemic risk
11One benefit to examine the DFA’s effect at the system-level is that there is less estimation
4
to reduce the impact on model estimation of omitted variables. These variables are economic growth, inflation, the 3-month interbank interest rate, the gov-ernment benchmark bond yield and financial development.12 Economic growth and inflation capture the macroeconomic fundamentals while interest rate re-flects the liquidity in the money market, government benchmark bond yield reflects a country’s sovereign risk and financial importance captures the devel-opment of the group of BHCs relative to the economy. Overall, these variables jointly capture economic and financial developments.
Table 4.4. Summary of control variables
Variable Symbol Description Data source Economic
growth
GDPGR Quarterly real GDP is converted to monthly real GDP by the cubic spline interpola-tion. GDPGR is calculated as the change of monthly real GDP (year-on-year, YoY)
Datastream, and the authors’ calculation Inflation INFLA INFLA is calculated as the change of
monthly CPI (YoY)
Datastream, and the authors’ calculation Interest
rate
IBOR The 3-month interbank offered rate Datastream Government
bond yield
YIELD Yield on the 10-year government benchmark bond
Datastream Financial
develop-ment
FINDEV For US, it is the sum of 24 US banks’ assets relative to US GDP; for EU, it is the sum of 42 EU banks’ assets relative to EU GDP
Datastream, and the authors’ calculation Global
financial crisis
GFC Taking value of 1 from August 2007 to June 2009 and 0 otherwise
The authors’ calculation Sovereign
debt crisis
SDC Taking value of 1 from November 2009 to December 2013 for the synthetic control group and 0 otherwise
The authors’ calculation CRD IV CRDIV Taking value of 1 from July 2013 to
Decem-ber 2015 or the synthetic control group and 0 otherwise
The authors’ calculation
In addition to these fundamental variables, we also consider three dummies: the GFC, the European sovereign debt crisis and the Capital Requirements
12
There is little research on macroeconomic determinants of systemic risk at the system-level. We find some studies about bank-specific determinants of systemic risk at the bank level, but there is no consensus on its determinants. For example, Laeven et al. (2016) find that systemic risk is related to bank size and capital while L´opez-Espinosa et al. (2012) find that short-term wholesale funding is a key determinant of systemic risk. In contrast, Weiß et al. (2014) find that bank size, leverage, non-interest income or the quality of the bank’s credit portfolio are not persistent determinants of systemic risk, but the regulatory regime dominates the change of systemic risk. As we focus on systemic risk at the system-level, we only control for macroeconomic factors in our regressions rather than bank-specific factors.
4
Directives (CRD) IV in the EU. The first dummy is used to account for the shock of the GFC. The latter two dummies are used to account for the changes of systemic risk in the EU due to the shocks from the sovereign debt crisis and the regulation change in the post-crisis period. Table 4.4 summarizes these variables.
4.4
Main results
This section first shows the results for systemic risk in the US banking system and then applies the SCM to construct the synthetic control group. Next, we apply the DID method to evaluate the treatment effect of the DFA in terms of reducing systemic risk in the US banking system.
4.4.1 Systemic risk in the US banking system
We apply the ∆CoVaR and MES measures to 24 US BHCs (see Table 4.2) to calculate their systemic risk contributions and take the asset-weighted average ∆CoVaR and asset-weighted average MES as two indicators of systemic risk in the US banking system. Both ∆CoVaR and MES suggest that systemic risk in the US banking system slowly increased between January 1998 and November 2002 and tended to decrease between December 2002 and May 2006. However, systemic risk began to rise since June 2006 and increased to an extremely high level at the end of 2008 following the outbreak of the GFC. After 2008, systemic risk decreased quickly and returned to pre-crisis levels (see Figure 4.1).
The abnormal movements of systemic risk during the 2007-2009 period coincided with the evolution of the subprime mortgage crisis. Borrowers found it more difficult to refinance their loans when US interest rates began to rise and home prices steeply declined after peaking in mid-2006. The subprime mortgage crisis became apparent by July 2007 when the investment bank Bear Stearns announced that two of its hedge funds had imploded. The crisis hit a critical point in September 2008 with many financial institutions being in distress, such as the sudden bankruptcy of Lehman Brothers, the buyout of Merrill Lynch and the seizure of Washington Mutual by federal regulators (see Hellwig, 2009; and Duchin et al., 2010). Figure 4.1 shows that systemic risk in the US banking system slowly increased between May 2006 and June 2007 and quickly went up between July 2007 and November 2008, corresponding to the evolution of
4 0% 10% 20% 30% 40%
Jan-98 Feb-00 Mar-02 Apr-04 May-06 Jun-08 Jul-10 Aug-12 Sep-14
US ∆CoVaR US MES
US ∆CoVaR is asset-weighted average ∆CoVaR while US MES is asset-weighted average MES across 24 US BHCs shown in Table 4.2. The vertical line indicates the passage of the DFA in July 2010.
Figure 4.1. Systemic risk in the US banking system
the financial crisis. With the passage of the Troubled Asset Relief Program on October 3, 2008 to alleviate the crisis, systemic risk in the US banking system started to decrease after November 2008, as indicated in Figure 4.1.13
Figure 4.1 also shows that ∆CoVaR decreased in the year after the en-actment of the DFA in July 2010 and then increased again to a local peak in September 2011. After September 2011, ∆CoVaR tended to decrease until July 2014 and then started to increase. MES has an analogous movement in the post-DFA period. It seems that systemic risk has decreased to pre-crisis levels before the enactment of the DFA, but did not remain at a constant level in the post-DFA period. A critical question is how much of the decrease in systemic risk can be attributed to (the expectation on) the enactment of the DFA.
4.4.2 Constructing the synthetic control group
To quantify how much of the decrease in systemic risk in the post-DFA period can be attributed to the enactment of the DFA, we resort to the methodology
13For details of a series of financial rescue plans in response to the crisis, see
CNN-Money.com’s bailout tracker at: http://money.cnn.com/news/storysupplement/economy/ bailouttracker/index.html.
4
outlined in Section 4.3.3. We first apply the SCM to determine the weights for candidate control banks shown in Table 4.3, using information up to June 2010. The SCM selects six EU banks with weights varying from 0.0264 to 0.2805 and assigns zero weights to other banks (see Table 4.5). The weighted combination of these six banks is going to serve as the comparison group for the US banking system. We calculate the weighted average ∆CoVaR and the weighted average MES across these six EU banks as two indicators of systemic risk in this synthetic control group (SCG).
Table 4.5. Weights assigned to candidate control banks Ticker Weight Ticker Weight Ticker Weight Ticker Weight Ticker Weight ERS 0.0000 PIST 0.0000 ISP 0.2504 BBVA 0.0000 UBSG 0.0000 DEXB 0.0000 EFG 0.0000 CRG 0.0000 SCH 0.0000 RBS 0.0000 KBC 0.1764 ETE 0.0000 PMI 0.0000 BKT 0.0000 BARC 0.0000 DAB 0.0000 PEIR 0.0000 BPE 0.0000 POP 0.0000 HSBA 0.0000 BNP 0.0000 BKIR 0.0000 MB 0.0000 NDA 0.0000 LLOY 0.0000 SGEX 0.0000 IL0A 0.0000 INGA 0.2026 SVK 0.0000 STAN 0.0637 KN 0.0264 ALBK 0.0000 DNB 0.0000 SEA 0.2805
CBK 0.0000 UCG 0.0000 BPI 0.0000 SWED 0.0000
DBK 0.0000 BMPS 0.0000 BCP 0.0000 CSGN 0.0000 Total 1.0000 0% 10% 20% 30% 40%
Jan-98 Mar-02 May-06 Jul-10 Sep-14 US ΔCoVaR SCG ΔCoVaR 0% 4% 8% 12% 16%
Jan-98 Mar-02 May-06 Jul-10 Sep-14 US MES SCG MES
US ∆CoVaR (US MES) is asset-weighted average ∆CoVaR (MES) across 24 US BHCs shown in Table 4.2. SCG ∆CoVaR (SCG MES) is weighted average ∆CoVaR (MES) across EU BHCs using weights shown in Table 4.5. The vertical line indicates the passage of the DFA in July 2010.
Figure 4.2. Systemic risk in the US banking system and the SCG Figure 4.2 shows ∆CoVaR (MES) of the US banking system and of the
4
SCG. SCG ∆CoVaR (SCG MES) closely tracks US ∆CoVaR (US MES) in the pre-DFA period. The Pearson correlation coefficient between SCG ∆CoVaR (SCG MES) and US ∆CoVaR (US MES) in the pre-DFA period is 0.98 (0.94), suggesting a strong positive relationship between SCG ∆CoVaR (SCG MES) and US ∆CoVaR (US MES). This result implies nearly common trends in SCG ∆CoVaR (SCG MES) and US ∆CoVaR (SCG MES), allowing us to perform the DID analysis for the effect of the DFA in reducing systemic risk in the US banking system with less concerns about the common trends assumption required by the DID model. Visually, we cannot see much difference between SCG ∆CoVaR (SCG MES) and US ∆CoVaR (US MES) in the post-DFA period from Figure 4.2, but a systematic evaluation controlling for potential post-DFA shocks is essential. We show the evaluation based on the DID method in Section 4.4.3 while illustrating the benefits to use the SCG as a comparison in the following paragraphs.
Except using the SCM to construct the control group, one might suggest to use asset-weighted average ∆CoVaR (MES) across all or part of the candidate EU banks. The latter is straightforward, but does not guarantee that the group constructed in this way has similar trends as that of the treatment group. Take ∆CoVaR as an example. We find that the correlation between US ∆CoVaR and ∆CoVaR of EU banks varies from 0.62 to 0.97 in the pre-DFA period. If we take asset-weighted average ∆CoVaR of the six EU banks that have the lowest correlations with the US ∆CoVaR (hereafter, Low6 ∆CoVaR, for short) as a comparison, and the six EU banks that have the highest correlations with the US ∆CoVaR (hereafter, High6 ∆CoVaR, for short) as another comparison, their pre-DFA correlations with the US ∆CoVaR are 0.87 and 0.98, respectively. Figure 4.3 displays US ∆CoVaR and SCG ∆CoVaR with Low6 ∆CoVaR in the pre-DFA period at the upper panel and with High6 ∆CoVaR at the bottom panel. Obviously, Low6 ∆CoVaR tracks US ∆CoVaR less closely than SCG ∆CoVaR, especially in the period between May 2003 and June 2010. High6 ∆CoVaR tracks US ∆CoVaR almost as closely as SCG ∆CoVaR.
To quantify and compare the lack of fit of SCG ∆CoVaR, High6 ∆CoVaR and Low6 ∆CoVaR for US ∆CoVaR, we calculate their root mean square pre-diction errors (RMSPE) following Abadie et al. (2010).14 The RMSPE of SCG ∆CoVaR for US ∆CoVaR, High6 ∆CoVaR for US ∆CoVaR and Low6 ∆CoVaR for US ∆CoVaR are 0.88%, 1.41%, and 2.69%, respectively, suggesting that SCG
14RMSPE is defined as follow: RM SP E =q1 T
Pt=T
t=1(Y1t− Y0t)2, where Y0tand Y1t
4 0% 10% 20% 30% 40%
Jan-98 Sep-99 May-01 Jan-03 Sep-04 May-06 Jan-08 Sep-09 US ΔCoVaR SCG ΔCoVaR Low6 ΔCoVaR 0% 10% 20% 30% 40%
Jan-98 Sep-99 May-01 Jan-03 Sep-04 May-06 Jan-08 Sep-09 US ΔCoVaR
SCG ΔCoVaR High6 ΔCoVaR
US ∆CoVaR and SCG ∆CoVaR are copied from Figure 4.2. Low6 and High 6 are asset weighted average ∆CoVaR across the six banks with the lowest and the highest correlations with US ∆CoVaR, respectively.
Figure 4.3. ∆CoVaR of US and other groups in the pre-DFA period
∆CoVaR has the best fit for US ∆CoVaR in the pre-DFA period.
In the above example, we show that by choosing banks whose ∆CoVaR are highly correlated with US ∆CoVaR, the asset weighted average ∆CoVaR across these banks can track US ∆CoVaR almost as good as SCG ∆CoVaR, but this has several limitations. First, it is arbitrary to determine the threshold of the correlation, or how many banks should be selected. We use six banks in the above example just for illustrative purposes. We also test the asset weighted average ∆CoVaR (MES) across all EU banks and find that it does not fit US ∆CoVaR (US MES) as good as SCG ∆CoVaR (SCG MES). Second, it is possible that we do not have a sample of banks whose ∆CoVaR (MES) are highly correlated with US ∆CoVaR (MES). In this case, the SCM would be
4
very useful because it does not require the candidate control banks’ ∆CoVaR (MES) to be highly correlated with US ∆CoVaR (US MES).
4.4.3 Evaluating the treatment effect of the DFA
We run the DID model (see Equation 4.16) for the US banking system and the SCG to quantify the effect of the DFA in reducing systemic risk in the US banking system.15 Table 4.6 shows the regression results when the dependent variables are ∆CoVaR and MES. We perform different model specifications and different regression techniques to examine the effect of the DFA. Columns I, II, V and VI show the estimation results for fixed effects regressions (i.e., using the ‘xtreg, fe’ command in Stata) while the other columns show estimation results for biased-corrected least square dummy variable (LSDVC) regressions (using the ‘xtlsdvc’ command of Bruno (2005) in Stata). The former method is widely used for panel data with small number of periods. However, our data contains two groups and 216 periods (months) which is a long panel. In addition, we include the lagged dependent variable as one of the explanatory variables, which makes the model to be a long dynamic panel. In this case, the LSDVC method appears to be a better method than within regression, difference or system generalized method of moments (see Kiviet, 1995; Judson and Owen, 1999; and Bruno, 2005).
Column I of Table 4.6 reports the results when the model only includes fundamental explanatory variables. We find that the coefficient of Gi ·DFA is 0.45, which means that the passage of the DFA caused an increase of 45 basic points of ∆CoVaR, but this effect is statistically insignificant. As to other explanatory variables, only the lagged dependent variable Yt−1 and the GDP growth rate (GDPGR) are statistically significant. The GDP growth rate has a coefficient of -0.41, being significant at the 10% level, which suggests that economic growth might be helpful for suppressing the buildup of systemic risk in the financial system. The coefficient of the lagged dependent variable is positive and significant at the 1% level, indicating the existence of endogenous risk persistence. This finding supports using dynamic panel models to examine the effect of the DFA.
15 We also examine the DFA’s dynamic treatment effects by interacting the DFA dummy
with each year dummy for the post-DFA period in our DID regressions. We do not find significant effects of the DFA in any year between 2011 and 2015. The results are not reported here for brevity.
4
Table 4.6. DFA’s effect on systemic risk in the US banking system
Yt−1 represents one period lag of corresponding dependent variables. Models I, II, V and VI
are estimated using fixed effects within regressions. Models III, IV, VII and VIII are estimated via the LSDVC dynamic regression technique (see Bruno, 2005). All models take group-specific and time-specific effects into account. For simplicity, we do not report the constant and the goodness of fit.
Panel A: ∆CoVaR Panel B: MES
I II III IV V VI VII VIII
Gi ·DFA 0.45 0.51 0.34 0.6 0.14 0.30 0.15 0.19 Yt−1 0.70*** 0.58** 0.84*** 0.73*** 0.93** 0.91** 0.96*** 0.93*** GDPGR -0.41* -0.30* -0.14 -0.04 -0.07 -0.05 -0.07** -0.04 INFLA -0.45 -0.74 0.06 0.06 -0.08 -0.14* -0.03 -0.03 YIELD -0.56 -0.69 -0.21 -0.15 -0.10 -0.13*** -0.07 -0.05 IBOR 0.47 0.53 0.2 0.17 0.08** 0.07* 0.11* 0.08 FINDEP 0.22 0.17 0.03 -0.12 0.05 0.00 0.02 -0.03 GFC 3.84** 3.55** 0.82** 0.88** SDC 0.75 0.64 0.50 0.23 CRDIV 0.03 0.39 0.16 0.13
Column II reports the results when we also control for the GFC, and the European sovereign debt crisis and regulatory change within the EU. The GFC dummy is used to account for the shock of GFC on US and EU banks. The European sovereign debt crisis dummy (SDC) and the regulatory change dummy (CRDIV) are used to account for the sovereign debt crisis and regulation change on banks in the synthetic control group. We find that the coefficient of Gi ·DFA increases from 0.45 to 0.51, but is still statistically insignificant. The other fundamental explanatory variables remain almost the same. The global financial crisis caused a significant increase of 384 basic points in ∆CoVaR, which is in line with our expectation. In contrast, the sovereign debt crisis and the regulatory change within the EU have no significant impact on EU ∆CoVaR, which reduces our concern that the SCG based on EU banks might be not comparable due to shocks from the sovereign debt crisis and the EU regulatory change in the post-crisis period.
Columns III and IV report the results when we use the LSDVC method to estimate the models. The results are almost the same, which suggests that the DFA did not have a significant impact on ∆CoVaR. Panel B shows the results when we use MES as the dependent variable. We find that the results in Panel B are similar to that in Panel A. Overall, our results suggest that the passage of the DFA has not contributed to reducing systemic risk in the US banking
4
system. This evidence adds to literature doubting the effectiveness of the DFA (Fama and Litterman, 2012; Acharya and Richardson, 2012; Acharya et al., 2016; Poghosyan et al., 2016; Andriosopoulos et al., 2017; and Gao et al., 2018). Our results also suggest that endogenous risk persistence is the main driver for the evolution of systemic risk. This finding is supported by Fahlenbrach et al. (2012) who show that there is strong persistence in banks’ risk culture.
4.5
Additional analyses and discussions
This section discusses several issues that might impact the credibility of our main results in Section 4.4. First, we examine the reliability of the counterfactual obtained from the synthetic control group. Second, we discuss endogeneity concerns and examine if there is an anticipation effect. These tests are typically performed in the literature on policy evaluations, as summarized by Athey and Imbens (2017), to shed light on the credibility of the main analyses. Third, we examine if the DFA had greater impact on the six largest BHCs given that they play a significant role in the financial system.
4.5.1 Predictive power of the synthetic control group
In Section 4.4.2, we construct a control group using the SCM and show that the group has stronger in-sample predictive power than those groups that are simply asset-weighted averaged. However, one might wonder if such a group can really produce a counterfactual for systemic risk in the US banking system in the post-DFA period. That is to say, how about the out-of-sample predictive power of a group constructed using the synthetic control method?
To address this concern, we use part of pre-DFA data to reconstruct a SCG and another part as a comparison for predictive ability analysis, as suggsted by Abadie et al. (2010, 2015). We reassign the DFA to the middle of the pre-DFA period (i.e. April 2004) and use the data available between January 1998 and March 2004 to construct a new synthetic group. Figure 4.4 shows ∆CoVaR and MES of the newly constructed group (SCGApr04) and that of the US banking system in the pre-DFA period. Figure 4.4(a) shows that SCGApr04 ∆CoVaR closely fits the US ∆CoVaR in the whole period, which means that SCGApr04 ∆CoVaR predicts US ∆CoVaR very well both in-sample and out-of-sample. The in-sample RMSPE is 0.8261% while the out-of-sample RMSPE has almost the
4
same value (0.8304%), which means that the synthetic group’s out-of-sample predictive power is as strong as that in-sample. Surprisingly, the synthetic group’s ∆CoVaR still closely tracks US ∆CoVaR during the 2008 crisis period.
0% 10% 20% 30% 40%
Jan-98 Feb-00 Mar-02 Apr-04 May-06 Jun-08
(a) ΔCoVaR
US ΔCoVaR SCGApr04 ΔCoVaR 0% 4% 8% 12% 16%Jan-98 Feb-00 Mar-02 Apr-04 May-06 Jun-08
(b) MES
US MES SCGApr04 MES
US ∆CoVaR and US MES are copied from Figure 4.2. SCGApr04 ∆CoVaR and SCGApr04 MES are obtained from the synthetic control group constructed for the US banking system using information available before April 2004.
Figure 4.4. The synthetic control group’s out-of-sample predictive performance
For MES, it can be seen from Figure 4.4(b) that SCGApr04 MES closely fits the US MES before the 2008 crisis. The in-sample RMSPE is only 0.5836% while the out-of-sample RMSPE is 1.1813%. The larger out-of-sample RMSPE is mainly due to the shock of the 2008 financial crisis, where the synthetic group does not mimic the US MES very well. Without including the crisis period, the out-of-sample RMSPE between April 2004 and June 2008 is 0.4872%, even smaller than the in-sample RMSPE. After the crisis period, Figure 4.4(b) shows that US MES decreases sharply since September 2009 and synth MES closely tracks US MES again since March 2010. As shown in Figure 4.2(b) where we construct the synthetic group using data available up to June 2010, the synthetic group’s MES predicts US MES during the crisis period much better. It is only 2% lower than US MES between December 2008 and December 2009 while SCGApr04 MES is 6% lower than US MES. Therefore, we have reasons to believe that the SCM is able to predict US MES in the post-DFA period adequately when we incorporate the data during the crisis period to construct the synthetic group.
4
SCM has very good out-of-sample predictive ability in terms of systemic risk measures. As for the above examples, the out-of-sample predictive power re-mains good over a longer period (75 months). This performance of the synthetic group gives us confidence to believe that the synthetic group shown in Section 4.4.2 can provide counterfactual trends of US ∆CoVaR and US MES in the post-DFA period.
4.5.2 Endogeneity concerns and anticipation effect
One might be concerned that our research design is subject to endogeneity problems. For example, if the DFA imposes different regulation on small and large banks, using small banks as a comparison group for the group of large banks may be problematic. There is no such a concern in this chapter. The DFA indeed focuses more on BHCs with consolidated assets of $50 billion or more, which are (implicitly) deemed as systemically important. If we had adopted smaller BHCs in the US as a control group, the estimated treatment effect would be problematic because we can not rule out the effect of the DFA on small BHCs. Given that the DFA is passed to repair the whole US banking system, we do not adopt US small banks as a control group for large banks. Instead, we adopt EU large BHCs with consolidated assets of $50 billion or more as a candidate control group, which play similar roles in the EU financial system as those in the US banking system. The DFA, which is a US law, hardly affects EU BHCs. Another reason why endogeneity may be an issue is that banks may have responded to the DFA in advance and therefore we cannot observe the treatment effect of the DFA in reducing systemic risk in the US banking system in the post-DFA period. This is not a serious issue in this chapter. The DFA imposes stricter regulation on BHCs with consolidated assets of $50 billion or more. If these BHCs attempted to avoid the stricter regulation by responding to the DFA in advance, it seems that the only possible way is to shrink their assets to less than $50 billion. We examine the assets of these BHCs since 2009 and find no indication for such a reduction in assets. In addition, the two systemic risk indicators adopted in this chapter are constructed based on market data rather than financial data of banks. It is basically impossible for banks to manipulate market data and we do not see a reason for banks to try to do so. Therefore, our systemic risk indicators are not distorted and hence avoid the endogeneity concern.
