Unconventional monetary policy: the cause of a new
financial crisis? An analysis of QE’s impact on
systemic risk in the US’ banking sector
By Marit Prins*
Supervised by prof. dr. L.H. Hoogduin January 8, 2020
Abstract
By performing a panel data analysis, this paper analyses the impact of unconventional monetary policy on systemic risk in the United States’ banking sector in the period 2005-2020. Particularly, it investigates the effect of QE on the expected capital shortfall of 53 US’ banks. Furthermore, this paper examines whether this effect is heterogeneous between Global Systemically Important Banks (GSIBs) and non-GSIBs. Statistical evidence is presented of a positive impact of QE on the expected capital shortfall of banks. Besides, a relatively more substantial impact of QE on non-GSIBs than GSIBs is confirmed.
Keywords: Unconcentional monetary policy – Quantitative Easing - Systemic risk - SRISK JEL codes: E44; E52; E58
2
1. Introduction
Over the years, banking has transformed significantly, with positive, as well as negative consequences. The latter include an enhanced risk-taking capacity, which has led to augmented systemic risk (Vives, 2016). Securitisation, for instance, is an example of misdirected financial innovation. Little doubt exists about the contribution of securitisation to economic growth, credit expansion, and financial markets’ advancement. Nevertheless, the Great Recession has exposed this innovation’s contribution to systemic risk, deteriorating financial stability. Likewise, globalisation of financial markets increased interconnectedness of financial institutions, fostering efficiency. However, it has also significantly amplified risk, as a single bank’s failure can severely impact the entire financial system (Engle, Jondeau & Rockinger, 2014). The Global Financial Crisis (GFC) revealed that the contagion of financial institutions can have a substantial impact on financial stability. Consequently, since the GFC, Central banks have a much greater focus on financial stability.
When a country is in recession, it can conduct expansionary monetary policy to increase the amount of money in circulation and stimulate the economy again. In this situation, the central bank can lower interest rates, loosen reserve limits, and purchase bonds from the open market with newly created money. The recent financial crisis proved that these conventional monetary policy (CMP) tools have limited usefulness in times of economic crisis or deep recession when nominal interest rates are at the effective lower bound. Consequently, unconventional monetary policy (UMP) tools were introduced, which effectively stabilised some sectors after the financial crisis (Chodorow-Reich, 2014). Following the GFC, the importance of financial stability has become more evident. During the pre-crisis period, central banks mainly focused on price stability and, as a result, failed to see the development of excessive credit creation.
3
The subsequent rise of asset bubbles threatened financial stability and ultimately resulted in the GFC (Joyce et al., 2012).
Consequently, some have proposed the central banks to include financial stability as an independent objective, in addition to the predominant mandate of price stability (Borio, 2014; Kapinos, 2017). Moreover, since the execution of UMP programs by the major central banks, policymakers and academics have increasingly focused on the potential trade-off between financial stability and price stability (Smets, 2014; Adrian & Liang, 2016). According to the Dutch Central Bank (2019), the prolonged period of UMP effectively keeps inflation low and at a constant level, warranting price stability. However, the risks of financial instability increase, due to increased risk-taking of financial investors and the possibility of asset bubble creation. This trade-off reveals the importance of powerful coordination between monetary and macroprudential policies2. To ensure this, a thorough understanding of the sources of financial stability and, accordingly, systemic risk, is of paramount importance. In recent years, a renewed interest in the (systemic) risk-taking channel of monetary policy rekindled. Even though the build-up of systemic risk was an essential contributor to the financial crisis, the empirical evidence on the effect of unconventional monetary policy on systemic risk, and, accordingly, the systemic risk-taking channel, is still relatively scarce.
Alessandri and Nelson (2015) propose that UMP induces bank risk-taking by flattening the term structure of interest rates. Aggregate risk and, in turn, systemic risk could increase if the risks of all banks are altered similarly. The former is more commonly referred to as the systemic risk-taking channel of monetary policy. Besides, the low interest rate environment, in which
4
UMP is conducted, enhances cross-holdings and asset commonality, which fosters default cascades and increases the likelihood of bank panics (Faia and Karau, 2018). This increases systemic risk in the financial sector. Additionally, due to the portfolio rebalancing channel of quantitative easing (QE), asset bubble creation and financial imbalances are likely to occur after the execution of such purchasing programs (Van den End, 2016). During the build-up, and even more during the burst of a bubble, banks’ systemic risk increases (Brunnermeier, Rother & Schnabel, 2020). While Verhelst (2017) and Deev and Hodula (2016) indeed find that UMP increases systemic risk in Europe, existing studies have not yet performed a thorough analysis of UMP’s impact on systemic risk in the US.
As a result of the financial crisis, the survival of the euro was feared by investors. On the 26th of July 2012, Mario Draghi held his famous ‘whatever it takes’ speech3, which restored trust
in the financial markets. These three magic words have been brought back into fashion by the present COVID-19 pandemic (The Economist, 2020). In March, the COVID-19 pandemic struck, which plunged international financial markets. Having in mind the success of the policies they implemented during the GFC, central banks quickly tried to stabilise the economy. More specifically, the Federal Reserve (Fed) and the European Central Bank (ECB) have announced, among other things, to execute extensive QE programs. Now more than ever, it is crucial to gain more knowledge about UMP’s economic consequences and, in particular, its effect on systemic risk, so that a next crisis might be prevented. Therefore, this paper analyses the impact of UMP on systemic risk in the banking sector of the United States. Particularly, it
3
The ECB announced to start buying sovereign bonds of distressed Eurozone countries, the so-called Outright Monetary Transactions (OMT). However, Draghi’s speech restored trust in the financial markets so successfully, that the ECB never needed to execute this program. His famous words were: ‘Within our mandate, the ECB is ready to do whatever it takes to preserve the euro. And believe me, it will be enough’.
5
investigates the effect of QE on the expected capital shortfall of US’ banks. Furthermore, as heterogeneous bank responses to monetary policy changes, resulting from dissimilar bank characteristics, influence monetary policy transmission (Brissimis & Delis, 2010), this paper analyses whether this UMP’s impact on systemic risk differs for Global Systemically Important Banks (GSIBs) and non-GSIBs.
This paper contributes to the existing literature in the following manner. In this paper, the effect of UMP on systemic risk in the US’ banking sector is analysed, whereas the existing literature mostly focuses on (a combination of) other countries. Only Faia and Karau (2018) and Kapinos (2017) conducted similar research for banks in the US. This paper differentiates itself from these papers in the research method adopted and the sample investigated. Faia and Karau (2018), Kapinos (2017), and almost all other papers investigating UMP’s effect on systemic risk, employ a VAR model. By contrast, this paper performs a panel data analysis. Lastly, this paper examines different sample than this literature; a different period and a larger number of banks is analysed.
The remainder of this paper is organised as follows. Chapter 2 discusses the unconventional monetary policies conducted by the Federal Reserve. Chapter 3 provides an overview of the existing literature on systemic risk and its relationship with UMP. Then, chapter 4 outlines the research methodology this paper employs. Furthermore, in chapter 5, the data is described, and several econometric problems are elaborated upon. In chapter 6, the results are presented, and robustness checks are performed. Lastly, chapter 7 concludes and discusses the limitations of this research and recommendations for further research.
6
2. Unconventional Monetary Policy in the US.
In November 2008, the Federal Reserve faced a financial crisis and deteriorating economy. The federal funds rate (FFR) had reached a historically low point; it had been reduced to virtually zero. The so-called zero lower bound was reached. Consequently, cuts in the FFR, known as conventional monetary policy (CMP), proved ineffective in stimulating aggregate demand during the financial crisis. This showed the need for new policies. Therefore, the Federal Reserve turned to unconventional monetary policies, which are targeted at altering long-term interest rates through asset purchases (i.e., quantitative easing) or central bank communication (i.e., forward guidance) (Bhattarai and Neely, 2016). The remainder of this paper will focus
exclusively on the consequences of QE.
During (the aftermath of) the Global Financial Crisis (GFC), four distinct asset purchasing programs have been executed: QE1, QE2, Operation Twist and QE34. Ultimately, these
purchases lead the Federal Reserves’ balance sheet to increase five-fold, from $900 billion to $4.5 trillion (Kuttner, 2018)5. With these QE programs, the Fed intended to combat weak
economic growth and low inflation, and loosen the monetary stance (Van den End, 2015). Nowadays, this form of UMP is conducted once more to counteract the economic consequences of the COVID-19 pandemic. In the remainder of this chapter, the different QE programs that the Federal Reserve has implemented will be elaborated upon.
4 Table A1 in the Appendix summarises the characteristics of these programs.
5 Figure A1 presents the system open market account (SOMA) over the period 2015-2020. All dollar-denominated assets obtained and sold through open market operations, and therefore the QE programs, are present in the SOMA.
7 QE1
On the 25th of November, 2008, the Fed initiated the first asset purchasing program by announcing the injection of $600 billion into the US economy. This type of stimulus, more commonly known as Large-scale Asset Purchases (LSAPs), consisted of $500 billion purchases of mortgage-backed securities (MBS) and $100 billion of agency debt. With this program, the Federal Reserve aimed to ‘reduce the cost and increase the availability of credit for the purchase of houses, which in turn should support housing markets and foster improved conditions in financial markets more generally’6. After this first attempt at economic stimulus,
the economy continued to contract. The Federal Open Market Committee (FOMC) expected inflation to remain lower than optimal for price stability and economic growth. Consequently, an expansion of this first asset program was announced on the 18th of March, 2009. The FOMC increased its purchases of MBS by $750 billion, and agency debt by $100 billion. Additionally, the FOMC started purchasing longer-term Treasury securities, worth $300 billion, to help improve conditions in private credit markets.
QE2
Even after QE1, the recovery in employment and output remained modest. The elevated unemployment rate and the low inflation rate prevailing in the US economy were inconsistent with the dual mandate7 of the Committee. The consumer price index (CPI) dropped to one per cent, a disinflationary trend that concerned many analysts (Bhattarai and Neely, 2016). To counter this trend, the Committee decided to increase its holdings of securities in November
6 See FOMC Press Release Nov 25, 2008.
7 The Federal Reserve has a dual mandate of price stability and maximum sustainable employment. To establish price stability, the Fed has an inflation target of 2 percent. The longer-run normal rate of unemployment, consistent with the employment mandate, is 4,1 percent.
8
2010. More specifically, it purchased an additional amount of $600 billion of longer-term Treasury securities, spread out over eight months.
QE2 received some criticism. Thomas M. Hoenig, a member of the FOMC, even voted against this new program, as he conjectured the risks of this policy to outweigh the benefits. According to Mr Hoenig, the continued expansion of the money supply raises the risks of future financial imbalances, which, over time, could ’cause an increase in long-term inflation expectations that could destabilise the economy’8. Moreover, as Swanson (2011) points out, several economists presumed that the Fed’s objective of fostering employment would not be achieved. Nevertheless, the members of the Fed decided to pursue QE2.
Operation Twist
In the summer of 2011, the financial stress index spiked and fears of recession renewed (Bhattarai & Neely, 2016). In response to these developments, on the 21st of September, the FOMC expanded the average maturity of its holdings of securities and reinvested maturing agency debt and MBS in MBS instead of Treasuries. This new program is more commonly known as the Maturity Extension Program (MEP) and Reinvestment Policy. It was intended to ‘put downward pressure on longer-term interest rates and help make broader financial conditions more accommodative’9. As long-term interest rates decline relative to short-term rates, a ‘twist’ of the yield curve, this program obtained the nickname ‘Operation Twist’10.
Operation Twist included the sale of $400 billion of short-term assets (1- to 3-year securities), accompanied by purchasing the same quantity of long-term assets (6- to 30-year Treasuries).
8 See FOMC Press Release Nov 3, 2010. 9 See FOMC Press Release Sep 21, 2011.
9
On the 20th of June, 2012, the Fed extended this program by an additional $267 billion. As opposed to the other three QE programs, the monetary base did not extend, as the sale of short term assets, rather than money creation, financed the new asset purchases.
QE3
In his speech during the Jackson Hole conference11, Fed’s chairman Ben Bernanke expressed his concern about the labour market’s stagnation12. Later, the Committee announced that ‘economic growth might not be strong enough to generate sustained improvement in labour market conditions’13. Therefore, the Committee deemed further policy accommodation necessary and executed a third round of QE on the 13th of September, 2012. It committed to purchasing $40 billion MBS per month, without setting a time limit when the program launched. Hence, the Committee committed to a pace of purchases instead of a given quantity (Kuttner, 2018). The purchase of additional MBS would continue until the outlook for the labour market improved significantly, in a context of price stability. This program launched to put downward pressure on longer-term interest rates, improve financial conditions, and support mortgage markets. On the 12th of December, 2012, the FOMC announced additional purchases of long-term Treasury securities, comprising a total amount of $45 billion per month. QE3 ultimately ended in October 2014, after a significant improvement of the economic conditions prevailing in the US, and the utilisation of labour resources in particular. With the end of this LSAP, the period of Quantitative Easing ultimately concluded.
11 The Jackson Hole conference is an annual event, where the Fed gives an update on how it assesses the economic conditions in the US and what future policy will entail.
12 See Bernanke, Ben S., Speech Aug 31, 2012, ‘Monetary Policy since the Onset of the Crisis’. 13 See FOMC Press Release Sep 13, 2012.
10 Policy Normalisation
After QE3 was terminated, the Federal Reserve’s policy comprised of reinvesting principal payments from its holdings of agency MBS and agency debt. As no new assets were purchased, the size of its balance sheet remained unchanged. However, due to the previous prolonged period of asset purchases, the balance sheet increased substantially. Therefore, to reduce its balance sheet, the Federal Reserve initiated a balance sheet normalisation program in September 201414. This program comprised two distinct approaches to monetary policy normalisation. The first approach encompassed targeting the federal funds rate at more normal levels, for the first time since December 200815. The economic conditions prevailing in the US and the favourable economic outlook warranted the Committee to ultimately employ this approach in December 2015. The reduction of its security holdings, the second approach that the FOMC proposed, took longer to adopt. Only in October 2017, this part of the program was executed. From this moment onwards, the Committee gradually reduced its purchased debt by ‘decreasing its reinvestment of the principal payments it receives from securities held in the System Open Market Account’. This reduction amounted $10 billion each month, of which $6 billion originated from the unloading of Treasury securities and $4 billion from the unloading of mortgage-backed securities and agency debt. These amounts were gradually increased to $30 billion and $20 billion per month in the subsequent year, respectively. In May 2019, the Fed reduced the unloading of Treasury securities from $30 billion to $15 billion per month. Eventually, the FOMC concluded the normalisation program in September 2019. The balance sheet was reduced to $3.7 trillion. Nevertheless, this was still approximately four times the amount it comprised prior to the GFC.
14 Table A1 summarises the characteristics of this program.
15 See FOMC Statement Dec 16, 2008. Three weeks after the deployment of the first asset purchasing program, the FOMC decided to reduce the target range of the FFR to between 0 and ¼ percent. After this approach was adopted, the FFR increased gradually until a value of around 2.40% in 2019.
11 QE in response to the COVID-19 pandemic
Since the eruption of the COVID-19 virus, which was declared a pandemic in March 2020, financial markets and businesses have been severely hit. Work-from-home policies, event cancellations and business closures have caused a significant economic downturn of which the duration is unknown. To mitigate the economic consequences of this new crisis, the Federal Reserve decided to take unprecedented steps16. It declared to ‘use its full range of tools to support the flow of credit to households and businesses and thereby promote its maximum employment and price stability goals’17. Most notably, the Federal Reserve executed a new
round of QE on the 15th of March. To initiate this program, the Fed announced to purchase at least $500 billion in Treasury securities and $200 billion in agency mortgage-backed securities ‘over the coming months’. As of the 23rd of March, the Fed extended this program by increasing its holdings of agency commercial mortgage-backed securities (CMBS). In the months following this decision, the economic conditions improved slightly, and, in turn, the pace of asset purchases declined. Subsequently, on the 10th of June, the Fed guaranteed to change the pace of its security holdings purchases to approximately $80 billion in Treasury securities, $40 billion agency MBS and between $250 and $500 million in agency CMBS every month. The Fed aimed to ‘sustain the smooth functioning of markets for these securities, thereby fostering effective transmission of monetary policy to broader financial conditions’18. At the moment of writing, this is still the tentative pace of purchases, except for the purchases of agency MBS, which increased to around and near $55 billion.
16 Table A2 presents the response of the Federal Reserve to the COVID-19 pandemic in more detail. 17 See FOMC Press Release Mar 15, 2020.
12
1. Literature review
Chapter two clarified the distinct QE programs introduced by the Federal Reserve since the Great Recession. These UMP measures have been implemented in times with the interest rates at the effective lower bound. The existing literature identifies several consequences of these low interest rates for the banks in the financial system. First of all, banks tend to reduce lending standards (Maddaloni & Peydro, 2011; Abbate & Thaler, 2019). Secondly, the profitability and margins of banks decline (Cerutti, 2016). Furthermore, banks raise the risk of their portfolios (Angeloni, Faia & Duca, 2015; Buch, Eickmeier, Prieto, 2014;Delis, Hasan, Mylonidis, 2012) and increase leverage (Angeloni et al., 2015; De Groot, 2014). Increased risk-taking of banks as a consequence of monetary policy has received ample attention in the existing literature. The risk-taking channel (RTC) of monetary policy is defined by Borio and Zue (2012) as: ‘the impact of policy rate changes on risk perception or risk tolerance and hence on the risk in portfolios, on the pricing of risk and on the terms by which funding is extended’. In their paper, Borio and Zue (2012) argue that the RTC might operate by three mechanisms. Firstly, cash flows, income and valuations are affected by changes in interest rates. This influences risk tolerance, risk perceptions, and, consequently, risk-taking. Secondly, agents’ behaviour is affected by a central bank’s commitment and accountability, which, therefore, may influence risk-taking. Lastly, when nominal rates-of-return targets are present, risk-taking may be induced by low interest rates. This latter mechanism was previously elaborated upon by Rajan (2006), who identified the ‘search for yield’. A financial institutions’ profitability reduces due to lower interest rates (see also Mamatzakis & Bermpei, 2016; Borio et al., 2017; Molyneux et al., 2019). In the presence of nominal rates-of-return commitments to stakeholders, these institutions are inclined to search for yield by increasing their investments in higher-risk assets, to compensate for this decline in profitability. Alessandri and Nelson (2015) propose that this search-for-yield mechanism exists when UMP is imposed, as UMP flattens the term-structure
13
of interest rates. Nonetheless, the existence of empirical evidence on the RTC of monetary policy does not necessarily imply a threat to financial stability. For this, evidence of risk-taking with systemic risk implications is necessary (Kabundi & de Simone, 2020). The so-called systemic risk-taking channel will be further elaborated upon below. First, the concept of systemic risk will be introduced.
Many academics and policymakers have tried to define systemic risk. The Financial Stability Board (FSB), International Monetary Fund (IMF) and Bank for International Settlements (BIS) (2009) define systemic risk as ’a risk of disruption to financial services that is caused by an
impairment of all or parts of the financial system and has the potential to have serious negative
consequences for the real economy’. Benoit, Colliard, Hurlin and Pérignon (2017) identify
three systemic risk sources: systemic risk-taking, contagion and amplification. They describe systemic risk-taking as high and correlated risk-taking by financial institutions. This correlated risk-taking could result from negative externalities a failing bank imposes on the surviving banks19. Banks are likely to invest in the same assets to minimise this externality, and thus fail or survive together (Acharya, 2009). Moreover, banks tend to be exposed to liquidity risk similarly. By redundantly investing in illiquid assets, banks expose the entire banking system to the risk of aggregate liquidity shortages. Next, contagion is specified as the spillover of the loss of one part of the financial system to another part. Contagion mechanisms include balance sheet contagion (Allen and Gale, 2000; Drehmann & Tarashev, 2011; Elsinger, Lehar, & Summer, 2006; Freixas, Parigi, & Rochet, 2000) and payment and clearing infrastructures (Rochet & Tirole, 1996; McAndrews & Potter, 2002). Furthermore, amplification is the situation where a small shock, if many institutions are affected, ends up having a substantially
19 For instance, a given bank’s default could lead creditors to believe that other banks may fail in the future (Acharya & Yorulmazer, 2008).
14
large impact. Amplification mechanisms include liquidity crises (Brunnermeier & Pedersen, 2009; Duarte & Eisenbach, 2015), market freezes (Afonso, Kovner, & Schoar, 2011; Flannery, 1996) and runs (Diamond & Dybvig, 1983; Iyer & Puri, 2012; Martin, Skeie, & Von Thadden, 2014). According to Benoit et al. (2017), these three sources are closely linked to each other: ‘the literature concerned with “systemic risk-taking” studies why financial institutions choose to be exposed to similar risks, thus reinforcing amplification mechanisms, and why they take large risk exposures, exposing themselves to default and their counterparts to contagion’.
Faia and Karau (2018) suggest that, in theory, three transmission channels exist through which UMP can affect systemic risk. One of these channels is the above-mentioned systemic risk-taking channel. UMP can influence the individual risk-risk-taking of a bank. Aggregate risk and, in turn, systemic risk could increase if the risks of all banks are altered similarly. Some studies investigate the existence of the systemic risk-taking channel (SRTC) (see, e.g. Colletaz, Levieuge, Popescu, 2018; Faia & Karau, 2018; Kabundi & De Simone, 2020). This empirical evidence points towards this channel’s existence, yet Colletaz et al. (2018) only validate the SRTC in the long run. The second channel operates through the low interest rate environment in which UMP is usually conducted. Due to these low interest rates, banks increasingly depend on other banks’ market funding, resulting in increased cross-holdings. When a financial system is highly interconnected, one entity’s distress can easily be transferred to more entities, which increases the likelihood of bank failures or stresses to co-occur (Corbacho and Peiris, 2018). Thus, due to increased interconnectedness, a bank’s idiosyncratic shock is more likely to spill over to other parts of the financial system. Hence, enhanced interconnectedness fosters default cascades. Correspondingly, the importance of adopting a macro-perspective, which complements the at that time conventional micro-perspective, was stressed by the IMF in 2013: ‘The traditional focus on idiosyncratic risks and the solidity of individual institutions needs to
15 be complemented by a system-wide perspective as both macro-financial linkages and
interconnections within the financial system can give rise to systemic risk’20. The final channel discussed by Faia and Karau (2018) is related to the search for yield behaviour of banks, also caused by low interest rates, discussed earlier in this chapter. This behaviour provokes banks to invest in identical, risky assets. Due to asset commonality and, accordingly, common exposures, a single shock could hit various banks simultaneously. Consequently, the likelihood of bank panics increases. This could increase systemic risk in the financial sector.
It is important to note that, as Brissimis and Delis (2010) point out, monetary policy transmission is affected by the heterogeneity of bank responses to monetary policy changes. The authors state that different balance sheet characteristics, namely market power, capitalisation and liquidity, cause this heterogeneous response in bank lending, profitability and risk-taking.
First of all, Brissimis and Delis (2010) find that, generally, monetary policy tightening negatively affects bank lending. Nevertheless, market power and high capitalisation tend to alleviate this effect due to higher buffers. Correspondingly, other papers state that monetary policy changes influence highly capitalised and large banks less than banks with lower capitalisation or smaller size, causing that the former’s lending capabilities are not affected (Cetorelli & Goldberg, 2008; Kashyap & Stein, 2000; Peek & Rosengren, 1995). Furthermore, Matthys et al. (2010) state that banks set different prices on loans, based on the borrower risk level. Following expansionary UMP implementation, less profitable, less capitalised, and smaller banks reduce spreads for safe firms more aggressively. Moreover, these weaker banks
16
raise spreads faster than stronger banks when firm risk increases. Second of all, Brissimis and Delis (2010) do not find a statistically significant effect of monetary policy on bank profitability. However, they observe a heterogeneous response and argue that contractionary monetary policy benefits banks with more market power or higher levels of equity. Similarly, Altavilla et al. (2018) conclude that the effect of monetary policy on bank profitability is affected by the heterogeneity of bank balance sheet characteristics. In relative terms, they find that banks with lower asset quality and higher operational efficiency benefit more from accommodative monetary policy. Lastly, Brissimis and Delis (2010) find that, in general, contractionary monetary policy negatively affects risk-taking. However, this effect seems to be almost zero, or even positive, for banks with higher capitalisation and liquidity. Thus, even though funding becomes more expensive following contractionary monetary policy, higher capitalisation and liquidity induce some banks to increase their portfolio risk. Given the aforementioned studies, which show that banks respond rather heterogeneously to monetary policy changes, one could question whether heterogeneity also affects the transmission of unconventional monetary policy to systemic risk.
While the forecited studies investigate the impact of UMP on banks, there is still limited empirical evidence on the impact of UMP on systemic risk. Verhelst (2017) analyses the effect of UMP on systemic risk in the European banking sector. The author concludes that an UMP shock, measured by the yield spread between German and Italian 10-year government bonds at the day of an ECB policy announcement, positively impacts systemic risk. The results preserved after controlling for heterogeneity between banks. Verhelst (2017) distinguished banks based on several balance sheet characteristics: the number of bank loans, deposits and capital, and the type of income. Similarly, Deev and Hodula (2016) analyse Eurozone countries and find that UMP shocks escalate financial instability, in the form of systemic risk in the
17
banking sector. On the contrary, Kapinos (2017) concludes that a single institutions’ contribution to systemic risk is not affected by UMP surprise shocks in the US.
Besides studies on the effects of UMP on systemic risk, some papers investigate the impact of QE in particular. Van den End (2016) points out that, by lowering risk spreads and interest rates, QE encourages risk-taking in financial markets. Increased risk-taking potentially increases systemic risk (Vives, 2016). Furthermore, Joyce et al. (2012) describe the portfolio rebalancing channel of QE, a channel through which central banks aim to boost the economy. The interest rates on the assets purchased by the central bank decline due to the increase in demand. Preferred habitat investors21, in turn, start buying assets that are a close substitute to the ones purchased by the central bank. Hence, the composition of investment portfolios changes due to central bank asset purchases. As a consequence, the price of the assets included, as well as not included in the purchasing programs, increases. The build-up of asset price bubbles and financial imbalances is likely in such market conditions (van den End, 2016). During the build-up, and even more during a bubble burst, banks’ systemic risk increases (Brunnermeier et al., 2020). Besides, Anh Nguyet Vu (2020) analyses the impact of QE on systemic risk in Japan. The results show that QE reduces systemic risk, but this effect may be neutralised by the rise in systemic risk due to the increased risk-taking in the low interest rate environment. Verhelst (2017) finds a positive impact of monetary policy shocks on systemic risk in Europe, following QE announcements. Similarly, Deev and Hodula (2016) conclude that QE raised systemic risk in Eurozone countries.
21 Preferred habitat investors are bond market investors with strong preferences to hold bonds of a specific maturity length. To make investments outside of their ‘preferred habitat’, they require a premium. If, as a result of QE, the interest rates decline on the preferred bonds, preferred habitat investors will switch to bonds that are close substitutes and yield a significantly higher return.
18
Except for Verhelst (2017), all these studies have in common that they employ a VAR or DSGE model. As explained in chapter two, the QE programs covered several months. When a VAR model is employed, UMP is perceived as an exogenous shock. Since the purchases made during the QE programs were spread out over time, considering QE as a shock seems erroneous. Therefore, this paper performs a panel data analysis. In chapter four, the research design will be further elaborated upon. Verhelst (2017) investigates UMP’s effect on systemic risk, measured by the Marginal Expected Shortfall (MES), in Eurozone countries from October 2008 until December 2015. This paper distinguishes itself from Verhelst (2017) in the period analysed, the systemic risk measure employed, the UMP measurement, and the countries investigated.
19
4. Research methodology
4.1 Systemic risk measure
Several academics have proposed a measure of a financial institutions’ contribution to the level of systemic risk in the financial sector. Girardi and Ergün (2013) introduce Value-at-Risk (VAR), which measures an individual institution’s risk in isolation. As VAR does not capture the entire system’s risk with non-linearities and feedback effects, it is not suitable to measure systemic risk (Kabundi & De Simone, 2020). Delta conditional VAR (∆CoVaR), introduced by Adrian and Brunnermeier (2011), considers the interconnectedness of the financial system. Adrian and Brunnermeier (2011) define ∆CoVaR as ‘the difference between CoVaR conditional on the institution being under distress and the CoVaR in the institution’s median state’. This measure has a few drawbacks. Firstly, as Acharya, Engle and Richardson (2012) indicate, CoVaR does not consider an institution’s volatility, whereas volatility contributes significantly to an institution’s riskiness. The measure only depends on this financial institution’s correlation with the market. Secondly, characteristics of financial institutions, such as size or leverage, do not explicitly affect CoVaR, even though these characteristics are important for systemic risk (Engle, Jondeau & Rockinger; 2014). Furthermore, spillover is measured by correlation, while only the former implies causation. Lastly, CoVaR cannot aggregate all individual institutions’ contributions to systemic risk and can only determine which institutions are systemically important (Kabundi & De Simone (2020)22.
Additionally, Acharya, Pedersen, Philippon and Richardson (2010) propose Marginal Expected Shortfall (MES) to measure systemic risk. This measure captures the effect of a
22 A systemically important financial institution (SIFI) is financial institution that would pose a serious risk to the economy if it were to collapse. They are generally referred to as ‘too big to fail’.
20
marginal increase in a financial institution’s weight in the financial system on the system’s risk, measured by the expected shortfall23. Like CoVaR, MES does not consider a financial institution’s characteristics, such as leverage and size. Besides, MES does not correctly predict, in advance of a crisis, which institutions are more likely to suffer once the crisis has arisen (Idier and Lamé, 2014). In addition to and as an extension of MES, Acharya et al. (2010) propose Systemic Expected Shortfall (SES) as a systemic risk measure. SES measures a financial institution’s propensity to be undercapitalised when the system as a whole is undercapitalised. Acharya et al. (2012) calculate MES and SES using a financial institution’s equity returns. However, banks’ balance sheets do not account for the full transmission of micro risk-taking into systemic risk-taking. For this, a measure should account for contagion and interconnectedness in systemic risk developments (Faia & Karau, 2018). This renders both MES and SES to be inadequate measures of systemic risk. Furthermore, the distress insurance premium (DIP) is suggested by Huang, Zhou and Zhu (2009), which captures what is theoretically the price a bank would have to pay if it wanted to be insured against financial distress. This measure is computed using data on credit default swaps (CDS). However, Kabundi and De Simone (2020) argue that ‘as suggested by research on the risk-taking channel of monetary policy, systemic risk measures based on CDS are best viewed as indicators of risk perception than as indicators of risk’ (Kabundi & De Simone, 2020).
As an extension of MES, Acharya et al. (2012) and Adrian and Brunnermeier (2016) introduce SRISK, which is ‘the expected capital shortfall of a given financial institution, conditional on a crisis affecting the whole financial system’ (Benoit et al., 2017). If a financial institution faces
23The expected shortfall of a financial institution is its expected loss conditional on the loss exceeding the value-at-risk (VAR). The VAR is the loss level that will not be exceeded with a certain confidence level during a certain period of time. For instance, if a financial institution’s 5-day 95% VAR is 2 million USD, the chance this institution fails more than 2 million USD in 5 days is considered to be 1%. For a 95% confidence level, the expected shortfall is then calculated by taking the average of returns in the worst 5% of cases.
21
a capital shortage, it is unable to function. If this co-occurs with a financial constraint impeding the financial sector, the government might have to save this institution with taxpayer money, as the institution runs out of options to attract capital. Consequently, the institution’s failure has implications for the real and financial sectors (Acharya et al., 2010). Therefore, when an institution is likely to face a capital shortage simultaneously as the financial sector itself is weak, it can be considered systemically risky. An institution with a low SRISK has a small contribution to the entire system’s systemic risk and vice versa. This measure utilises data on a financial institution’s sensitivity of its return on equity to market shocks and market capitalisation and, most importantly, accounts for its size, financial leverage and interconnectedness. These characteristics are considered important determinants of an institution’s contribution to systemic risk (Engle et al., 2014). SRISK has proven ability to identify and rank SIFIs, besides predicting which institutions will suffer losses in a financial crisis (Kabundi & De Simone, 2020). This warrants ex-ante regulation of financial institutions that are likely to deteriorate the economy’s real and financial sectors when they fail.
This paper employs the measure SRISK to measure systemic risk in the financial sector in the US. This measure is constructed as follows:
(1) 𝑆𝑅𝐼𝑆𝐾𝑖,𝑡 = 𝑘𝐷𝑖,𝑡− (1 − 𝑘)𝐸𝑄𝑈𝐼𝑇𝑌𝑖,𝑡(1 − 𝐿𝑅𝑀𝐸𝑆𝑖,𝑡),
Where 𝑘 is a parameter for the prudential capital ratio, 𝐷𝑖,𝑡 denotes the book value of debt for
bank i at time t, 𝐸𝑄𝑈𝐼𝑇𝑌𝑖,𝑡 the market value of equity, and LRMES the Long-Run Marginal
Expected Shortfall. In this paper, the prudential capital ratio 𝑘 is set to 8%, following the total capital requirements of the Basel III Accords. LRMES captures the expected loss in terms of a financial institution’s equity level, conditional on a crisis. This crisis occurs when, over a six months horizon, the market were to fall by more than a certain threshold C. Acharya et al.
22
(2012) set C at 40%, and subsequently show that the LRMES can be calculated in the following way:
(2) 𝐿𝑅𝑀𝐸𝑆𝑖,𝑡 ≈ 1 − exp(−18 𝑥 𝑀𝐸𝑆𝑖,𝑡),
Where 𝑀𝐸𝑆𝑖,𝑡 represents the expected one-day loss in equity if market returns decline by more
than 2% 24. It is calculated as follows:
(3) 𝑀𝐸𝑆𝑖,𝑡 = 𝐸𝑡[−𝑅𝑖,𝑡|𝑅𝑀,𝑡 < −2],
Where 𝑅𝑖,𝑡 is the market return of bank i at time t and 𝑅𝑀,𝑡 the market return index for the
total market.
4.2 Research design
To investigate the effect of QE on a bank’s contribution to the systemic risk of the entire US’ financial sector, a model of the following form will be analysed:
(4) 𝑆𝑅𝐼𝑆𝐾𝑖𝑡 = 𝛽1 + 𝛽2 𝑄𝐸𝑡+ 𝛽3𝑉𝑡+ 𝛽4𝑟𝑡+ 𝛽5𝐷𝑡+ 𝑢𝑖𝑡,
where the systemic risk variable, 𝑆𝑅𝐼𝑆𝐾𝑖𝑡, of bank i at time t, is written as a function of a
constant 𝛽1; the variable 𝑄𝐸𝑡, which represents the quantitative easing programs of the Fed; a
variable for the overall volatility in the market 𝑉𝑡; a variable for the short term interest rate 𝑟𝑡;
a crisis dummy 𝐷𝑡; and an error term 𝑢𝑖𝑡.
As discussed in chapter three, by lowering risk spreads and interest rates, QE encourages risk-taking in financial markets. Increased risk-risk-taking leads to increased fragility of the banking
23
sector, increasing the risk of default. When a bank fails, this has external effects, potentially systemic (Vives, 2016). Moreover, QE likely increases asset price bubbles and financial imbalances (van den End, 2016). During the build-up, and even more during a bubble burst, banks’ systemic risk increases (Brunnermeier et al., 2020). This implies that QE increases systemic risk. Therefore, the first hypothesis this paper will test states:
Hypothesis 1: Quantitative Easing (QE) increases a bank’s contribution to the systemic risk of
the entire financial sector.
The coefficient 𝛽2 captures the impact of QE on systemic risk and is expected to be positive.
The control variable for volatility is included to correct for the volatility already present in the market. In this way, fluctuations in systemic risk will not be incorrectly awarded to the policy shock, while they are in fact regular movements in the market (Faia & Karau, 2019). 𝛽3 is
expected to be positive; the more volatile the market, the higher a single bank’s contribution to the financial market’s overall systemic risk. Furthermore, as the macroeconomic conditions prevailing in a country influence bank risk (see, e.g. Laeven & Levine, 2009; Barrell et al., 2010), the control variable for the interest rate is incorporated. As discussed in chapter three, interest rates could influence risk tolerance, risk perceptions, and risk-taking. Besides, low interest rates may induce risk-taking through the ‘search for yield’ mechanism, potentially increasing systemic risk. Hence, 𝛽4 is expected to be negative. Lastly, the crisis dummy’s
inclusion warrants that a few extreme observations during the crisis do not contaminate the estimated coefficients. The expectation is that in a crisis period, the expected capital shortfall is higher. Thus, 𝛽5 is expected to be positive.
24
As indicated in chapter three, some of the existing literature point out that monetary policy transmission is affected by the heterogeneity of bank responses to monetary policy changes (resulting from different bank characteristics). Hence, this paper analyses whether bank heterogeneity also affects the transmission of UMP to systemic risk. Therefore, in addition to an analysis covering all banks, separate analyses are performed for Global Systemically Important Banks (GSIBs) and non-GSIBs. Whether a bank is a global systemically important bank is based on its size, substitutability, interconnectedness, cross-jurisdictional activity and complexity25. The larger a bank, for instance, the sooner it will be perceived as global systemically important. By performing the separate analyses, it can be determined whether having these characteristics influences QE’s effect on a bank’s contribution to the systemic risk of the entire financial sector. Monetary policy changes influence large banks less than banks of a smaller size (Cetorelli & Goldberg, 2008; Kashyap & Stein, 1995, 2000; Peek & Rosengren, 1995). Therefore, the impact of QE is expected to be relatively larger for non-GSBIs than non-GSBIs. Hence, the second hypothesis is as follows:
Hypothesis 2: The impact of QE is relatively larger on non-GSIBs than GSIBs.
25 See BIS Publication, December 18, 2020, ‘Global systemically important banks: Assessment methodology and the additional loss absorbency requirement’.
25
5. Data
5.1 Variable description
This paper analyses the effect of UMP on an individual bank’s contribution to the total systemic risk in the US’ financial sector for the period 2005 until 2020. Hence, the sample period covers pre-crisis, crisis and post-crisis periods. The analysis is performed from 2005 onwards, as from this moment the SRISK data is available for a significantly larger number of banks than before this date. The sample period ends in September 2020, the most recent moment on which data is available. In this paper, a sample of 53 US’ banks is analysed. Appendix A3 shows a list of these banks, of which nine are considered to be global systemically important according to the Financial Stability Board (FSB) and the Basel Committee on Banking Supervision (BCBS). These banks are indicated in bold. Banks for which the available data does not cover the whole sample period were excluded from the sample.
SRISK data is retrieved from Stern-NTU’s V-LAB. This institute presents daily calculations of SRISK of financial institutions worldwide. Unconventional monetary policy is measured by the SOMA holdings of the Federal Reserve. The FOMC has appointed the Federal Reserve Bank of New York to perform transactions for the SOMA. Hence, this data is retrieved from the website of the Federal Reserve Bank of New York. All dollar-denominated assets obtained and sold through open market operations, and therefore the QE programs, are present in the SOMA. Data on the SOMA is available every five days on average. The value on a given day is then used as the value of the subsequent days for which the data is not available.
26
Furthermore, data on the market volatility, inflation and short term interest rates stem from the Federal Reserve Economic Data (FRED). To measure the market’s volatility, daily data on the CBOE Volatility Index (VIX) is used. This index displays the market’s expectation of a 30-day forward-looking volatility of the S&P 500 index26 and represents investors’ market risk perception. This variable is transformed using the logarithm function, to make its distribution less skewed. Following Faia and Karau (2018), the short term interest rate is measured by the 3-month Treasury Bill rate. This data consists of daily values. Lastly, the analysis includes a crisis dummy, which takes value one from December 2007 until June 200927. The value of this dummy is zero in all other periods.
Table 1: Summary statistics
Variable Count Mean Std. Dev. Min Max
Srisk 211.896 4.509,11 20.359,20 -82.832,28 161.529,20
SOMA 212.053 2.741.813 1.567.120 473.300 6.595.300
Interest rate 210.887 1,25 1,59 -0,05 11.63
VIX 212.053 19,07 9,47 9,14 82,96
Crisis 212.265 0.10 0,30 0 1
Note: SRISK and SOMA are in millions of USD, interest rate in %, and VIX is an index.
The summary statistics of the above-mentioned variables are presented in Table 1. The dataset is strongly balanced, with 211.877 observations per variable on average. The variable SRISK
26 This is a stock market index that measures the stock performance of 500 large companies listed on stock exchanges in the United States.
27 The U.S. National Bureau of Economic Research (the body deciding on the begin and end of economic downturns) concluded that the start of the recession was in December 2007 and the end in June 2009.
27
contains negative as well as positive values. This variable ranges from high expected capital surpluses to high expected capital shortfalls. The interest rate also takes positive and negative values. Additionally, the data is checked for outliers and normality.
5.2 Econometric problems
Endogeneity
The model could suffer from endogeneity problems. Firstly, this could be the case since the included regressors may be correlated with explanatory variables relevant but omitted from the model. This causes the estimated coefficients to be biased and inconsistent. As the model does not include bank-specific characteristics, for instance, this omitted variable bias is likely to be present. To solve this bias, the model is run with a fixed effect (FE) estimator. FE estimation allows for correlation between the regressors and unobserved bank-specific characteristics. Hence, heterogeneity across banks is taken into account (Baltagi, 2008). An F-test on the presence of unobserved individual effects reveals that the model is indeed influenced by bank heterogeneity. Secondly, endogeneity issues might arise because, while QE is expected to influence an individual bank’s contribution to systemic risk, the reverse might also be true. This suggests a potential simultaneity problem. The GFC has revealed the importance of addressing financial stability in macroeconomic policy. Nevertheless, the Fed’s primary monetary policy objectives are still maximum sustainable employment and price stability. Macroprudential policy, on the other hand, is defined as: ‘the use of primarily prudential tools to limit systemic risk’28. Thus, financial stability concerns are mostly tackled in macroprudential policy.This reduces the probability of a simultaneity problem. Additionally, due to the usage of daily data, the concern for simultaneity is partly alleviated. Monetary policy
28
is generally influenced by changes in economic conditions taking place over a more extended period. If, from one day to the next, an individual banks’ contribution to the entire financial sector’s systemic risk increases, this likely does not immediately influence the Fed’s monetary policy. Longer-term developments in a bank’s measure of systemic risk are expected to influence monetary policy. Hence, whereas the policy decisions and implementations could have a daily impact on a bank’s contribution to systemic risk, a daily impact of the reverse is not expected to be present. Next, some specification tests are performed.
Non-stationarity
First of all, the Fisher unit-root test has been performed to test for stationarity of the data. When variables are non-stationary, the regression could be spurious. The Fisher unit-root test verifies that all variables are stationary, except for the variables SOMA and interest rate. These variables are non-stationary of order I(1). Therefore, both variables are transformed using the logarithm function. Moreover, first differences are applied to all control variables. After these alterations, the Fisher unit root test is performed once more. All variables appear stationary. The results are presented in Table A5.
Multicollinearity and correlation
Second, multicollinearity could exist in the model, which indicates an approximate or exact linear relationship between the explanatory variables. Strong multicollinearity causes estimates to have an unexpected magnitude or sign and become unreliable with high standard errors. The variance inflation factor (VIF) of the explanatory variables is analysed to test for potential multicollinearity. The results are presented in Table A6. A VIF of 10 or higher is generally seen as an indicator of problematic multicollinearity (Alin, 2010; Bowerman & O’connell,
29
1990). To indicate the presence of multicollinearity, the mean of the VIF values could also be investigated. If the is mean considerably larger than one, this indicates multicollinearity. As the VIF of all variables is around 2, and the mean of these values does not significantly exceed one, there is no strong indication that the estimators cannot be trusted. Besides, the correlation between the explanatory variables is analysed. Table A7 presents the correlation values. As all values are close to or below 0.5, there is no indication of high collinearity between the variables.
Serial correlation and heteroskedasticity
Lastly, the Wald test for groupwise heteroskedasticity and the Wooldridge test for serial correlation are performed. Table A8 presents the results of these tests. When heteroskedasticity arises, different error terms do not have the same variance. The Wald test discloses the presence of heteroskedasticity. Serial correlation is the presence of correlation between different error terms. The null hypothesis of no serial correlation is rejected. This indicates that first-order autocorrelation is present in the model. As the error is estimated for the same bank over time, this bank’s observations are correlated. This could explain the presence of autocorrelation in the model. In the presence of heteroskedasticity and autocorrelation, the estimated coefficients are unbiased. However, the standard errors cannot be trusted anymore. If the model is run with a random effects estimator, the autocorrelation problem could be solved. However, due to omitted variable bias, the fixed effects estimation is still the preferred method. To eliminate the effect of heteroskedasticity and autocorrelation, cluster robust standard errors are employed.
30
6. Results
6.1 Regression analysis
In the first column of Table 2, the results of the analysis with all banks are presented. All estimated coefficients are statistically significant, yet at different significance levels. The SOMA coefficient is statistically significant at the 1% significance level. Thus, a significant relationship is found between an increase in the SOMA holdings and a single bank’s contribution to systemic risk. If the change in the SOMA increases by 1%, a bank’s expected capital shortfall increases by 274.9 million USD, keeping all other variables constant. The coefficients of the interest rate, volatility index and crisis dummy are significant at the 5%, 1% and 1% significance level, respectively. If the interest rate change increases by 1 percentage point, a single bank’s expected capital shortfall decreases by 1,167 million USD. Besides, if the change of the volatility index increases by 1 %, SRISK increases with 9.90 million USD. Lastly, during the crisis period, SRISK was 5,863 million USD lower than in non-crisis periods.
Next, in the second column of Table 2, the results of the analysis with the sample of only GSIBs are summarised. A positive, significant relationship is found between an increase in the SOMA holdings and a single bank’s contribution to systemic risk. An increase of 1% in the change in the SOMA holdings increases SRISK by 1,226.47 million USD. The coefficients of the interest rate, volatility index and crisis dummy are significant at the 5%, 5% and 10% significance level, respectively. If the interest rate change increases by 1 percentage point, SRISK decreases with 6,231 million USD. In addition, if the change in the VIX increases by 1%, a single bank’s expected capital shortfall increases by 38,03 million USD. Lastly, for GSIBs, SRISK was 27,408 million USD lower in the crisis period than in the non-crisis period.
31 Table 2: Fixed effects estimation
(1) (2) (3)
VARIABLES Total sample GSIBs Non-GSIBs
SOMA 27,491*** 122,647** 15,541*** (9,024) (47,845) (5,615) Interest rate -1,167** -6,231** -432.1** (438.2) (2,237) (183.5) VIX 990.0*** 3,803** 652.6*** (251.2) (1,192) (194.9) Crisis -5,863*** -27,408* -2,956** (2,143) (11,759) (1,110) Constant 4,481*** 32,287*** 698.9*** (8.590) (47.83) (4.812) Observations 164,456 24,818 139,640 R-squared 0.001 0.003 0.001 Number of Banks 53 8 45
Note: all values are in millions of dollars and SOMA and VIX are log transformed.
Column three of table 2 shows the results of the analysis of the sample of non-GSIBs. All estimated coefficients are statistically significant. The SOMA coefficient is statistically significant at the 1% level. Hence, a positive, significant relationship is found between an increase in the SOMA holdings and a single bank’s contribution to systemic risk. If the change in the SOMA holdings increases by 1%, a single’s bank expected capital shortfall increases by 155.41 million USD. The coefficient of the volatility index is also statistically significant at the 1% significance level. If the change in the VIX increases by 1%, SRISK increases by 38.03 million USD. The coefficient of the interest rate and crisis period are significant at the 5% level. An increase in the change of the interest rate of 1 percentage point decreases a bank’s
32
expected capital shortfall by 6.23 million USD. Lastly, in the crisis period, a firm’s expected capital shortfall was 2,956 million USD lower than in pre- or post-crisis periods.
Comparison GSBIs and non-GSIBs
Next, the results of the second and third analysis are compared. The estimated coefficients of all variables, resulting from the analysis of the sample with only GSIBs, are much higher than those of from the sample with non-GSIBs. For instance, an increase of 1% in the SOMA holdings increases a non-GSIB’s expected capital shortfall by 155.4 million USD, whereas this is 1,226.47 million USD for GSIBs. This implies that GSIBs have higher expected capital shortfall resulting from QE. Furthermore, the signs of all estimated coefficients are the same. For both samples, all estimated coefficients are statistically significant, yet some at different significance levels. The coefficients of the variables SOMA, VIX, and risis are significant at a higher significance level for non-GSIBs than GSIBs. Moreover, the difference between the expected capital shortfall during a crisis period and crisis periods is slightly larger for non-GSIBS, indicating that a crisis affects the expected capital shortfall more for these type of banks. The significance of the interest rate coefficient is the same for both samples.
Comparison with hypotheses
A significant relationship between the SOMA holdings and a single bank’s expected capital shortfall is found in all analyses performed. The analyses confirm a positive relationship between QE and a bank’s contribution to the total systemic risk of the US’ financial sector. This confirms the first hypothesis of this paper. Furthermore, a significant relationship between the SOMA holdings and a single bank’s expected capital shortfall is found in both the analysis of the GSBIs and that of the non-GSIBs. GSIBs have higher expected capital shortfall resulting
33
from QE than non-GSBIs. Nevertheless, the results confirm the second hypothesis of this paper; QE has a more relatively substantial impact on non-GSIBs than GSIBs. By taking the expected capital shortfall of these two types of banks separately and dividing it by the joint total assets of their type29, QE’s impact on these banks can be calculated. Whereas the expected capital shortfall is 0.15% of total assets for non-GSIBs, this is 0.08% for GSIBs30. Hence, it is
found that QE’s influence is relatively larger on non-GSIBs than GSIBs.
Goodness-of-fit
The R-squared reported in Table 2 represents the goodness-of-fit of the model. Hence, it explains how much of the SRISK variable’s variance is predictable from the explanatory variables. A change in the SOMA probably has a lasting effect on a bank’s contribution to the systemic risk of the entire financial system. Moreover, there is a certain persistency in the magnitude of this effect; a bank’s contribution presumably does not change considerably from one day to another. Therefore, the initial analysis is conducted once more, including the one year lagged values of SRISK as an explanatory variable. This significantly improves the goodness-of-fit of the model, to approximately 0,99 in the analysis of all three samples. The signs of the SOMA and interest rate coefficients change. Hence, an increase in the change of the SOMA or interest rate is found to decline a bank’s contribution to the entire financial system’s systemic risk. As an innumerable amount of studies have proved the reverse, the results seem strange. Besides, it is not plausible that a bank’s contribution to systemic risk is almost entirely determined by external factors instead of internal, bank-specific factors.
29 The value of these banks’ total assets on March 31, 2020 is used. See National Information Center, March 31, 2020, ‘Large Holding Companies’ and Federal Reserve Statistical Release, March 31, 2020, ‘Large Commercial Banks’.
30 For non-GSIBs, this is calculated as (1.554,10/102.950)*100% = 0,15%. For GSIBs, this becomes (1.226,47/1.602.630)*100% = 0,08%.
34
Previous studies have shown that other factors, such as interconnectedness and bank characteristics, are important determinants of systemic risk. Therefore, it is unlikely that a model not including these factors as variables can predict 99% of SRISK’s variance. Hence, this model specification seems unreliable.
6.2 Robustness checks
Pre-, post- and crisis periods
The previous analyses show that the presence of a crisis period significantly affects the expected capital shortfall of a bank. Therefore a robustness check containing a separate analysis for pre-crisis, crisis and post-crisis periods is performed. The results of this analysis are summarised in Table 3. It appears that results are not entirely robust to dividing the total sample into pre-crisis, crisis and post-crisis periods.
First, the pre-crisis period is considered. In all three samples, most of the estimated coefficients and their significance change. Moreover, a change in the signs of the coefficients can be observed. Hence, a negative relationship between the SOMA holdings and a bank’s contribution to the total systemic risk in the financial sector is found. All estimated coefficients are much larger than those resulting from the initial analysis. If all banks are considered, and the change in the SOMA holdings increases by 1%, a bank’s expected capital shortfall decreases by 3,867.66 million USD. This was 274.91 million USD in the initial analysis. For GSIBs, an increase in the change of the SOMA holdings of 1% decreases a bank’s expected capital shortfall by 21,650.42 million USD. In the initial analysis, this was 1,226.47 million USD. For non-GSIBs, if the change in the SOMA holdings increases by 1%, SRISK decreases
35
by 1,109.46 million USD, as opposed to 155.41 million USD in the initial analysis. Besides, the standard errors increase considerably. The standard error measures how precisely the coefficient's unknown value is estimated by the model. Hence, the higher standard errors reveal that the estimates of this analysis are less precise than those of the initial analysis. Before the crisis, the soma account existed, but QE only started during the financial crisis (on the 25th of November, 2008). Therefore, before this date, the numbers of the SOMA do not represent QE, but other refinancing operations. Hence, the results of the pre-crisis period do not represent a correlation between QE and a bank’s expected capital shortfall. Still, the substantially high, negative coefficients could impact the size of the coefficients resulting from the initial analysis. However, the pre-crisis period only covers about one and a half years and, therefore, the impact of this on the overall results is expected to be limited.
Second, the crisis period is evaluated. All estimated coefficients are somewhat larger than those of the initial analysis. Moreover, the standard errors of all coefficients are higher, indicating that this analysis’ estimates of the regression coefficients are less precise than those of the initial analysis. The SOMA coefficients’ sign remains the same; a positive relationship is found between an increase in QE and a single bank’s contribution to the total systemic risk in the financial sector. The analysis of the total sample and the sample with non-GSIBS confirm the existence of this positive relationship. The estimated coefficients are significant at the 5% level. This is contrary to the results of the main analysis, in which case both estimated coefficients were significant at the 1% level. The results do not confirm a positive relationship if only GSIBs are considered. This is not in line with the results of the initial analysis. The coefficient of the volatility index is insignificant for all samples. Additionally, the interest rate coefficient is not significant for GSIBs. For non-GSBIS, on the other hand, the significance of this coefficient increased. Besides, compared to the initial analysis, the signs of the coefficients of
36 Table 3: Robustness check; pre-, post- and crisis periods.
(1) (2) (3) (4) (5) (6) (7) (8) (9) VARIABLES Pre-crisis total Pre-crisis GSIBs Pre-crisis non-GSIBs Crisis total Crisis GSIBs Crisis non-GSIBs Post-crisis total Post-crisis GSBIs Post-crisis non-GSBIs SOMA -386,766** -2,165,042* -110,946** 30,622** 133,947 19,744** 15,083*** 62,160** 9,149*** (184,430) (1,056,351) (52,687) (13,277) (80,224) (8,302) (4,332) (21,837) (2,930) Interest rate -340.6* -2,014 -84.64* 2,343** 9,721 1,418*** -14,390*** -67,074** -7,394*** (181.6) (1,066) (47.08) (898.0) (5,271) (473.0) (4,348) (20,344) (2,738) VIX 1,951*** 9,453** 866.8*** -415.4 -2,247 -308.0 1,500*** 6,069*** 960.6*** (644.7) (3,203) (301.9) (701.5) (4,559) (351.4) (381.0) (1,697) (312.5) Constant -1,228*** 3,495*** -1,664*** 9,309*** 56,982*** 2,776*** 5,363*** 36,702*** 1,048*** (24.32) (137.9) (7.240) (8.841) (57.80) (6.668) (4.659) (24.80) (2.740) Observations 31,216 4,712 26,504 16,218 2,448 13,770 117,764 17,770 99,996 R-squared 0.003 0.015 0.003 0.006 0.017 0.007 0.001 0.004 0.001 Number of Banks 53 8 45 53 8 45 53 8 45
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
37
the interest rate and volatility index change. This does not match the findings of the existing literature. As explained in chapter four, SRISK is conditional on a crisis affecting the whole financial system, measured by a 40% fall in the market over a six-month horizon. However, during a crisis, this market fall has already taken place. Subsequently, SRISK measures the expected capital shortfall of a financial institution, conditional on an additional fall of the market of 40%. Therefore, SRISK is presumably not applicable as a measure for systemic risk during a crisis period. This could explain the unexpected results of the analysis of the crisis period.
Lastly, the post-crisis period is considered. The signs of all estimated coefficients are the same as those of the initial analysis. The results suggest a positive relationship between QE and a bank’s contribution to systemic risk. Also, all estimated SOMA coefficients are smaller than those of the initial analysis. Besides, the standard errors of all SOMA coefficients decline. Hence, this analysis’ estimates of the regression coefficients are more precise than those of the initial analysis. The coefficients resulting from the analysis of all banks and only non-GSIBs are still significant at the 1% significance level. For GSIBs, the estimated coefficient is still significant at the 5% level. Furthermore, the interest rate coefficients increase considerably. For instance, if the total sample is considered, a 1% increase in the interest rate change decreases a bank’s capital shortfall by 14,390 instead of 1,167 million USD. The coefficients resulting from the analysis of all banks and only non-GSIBs are significant at the 1% level, instead of the initial 5%. Furthermore, the coefficients of the volatility index increase. For GSIBs, the coefficient is significant at the 1% level, instead of the initial 5%. Additionally, the coefficients resulting from the analysis of all banks or only non-GSIBs for GSIBs remain significant at the initial 1% level. The coefficients of the interest rate and volatility index have larger standard
