Does Monetary Policy Affect Bank Risk-taking? An Empirical
Study of U.S. Banks Following The 2008 Financial Crisis
Rajendra Marapin
(S2316994)
University of Groningen
Faculty of Economics and Business
MSc Economics & Finance
Supervisor: dr. C.G.F. van der Kwaak
Abstract
A recent set of literature has been debating whether the prolonged period of accommodative monetary policy in the U.S. after the 2008 financial crisis is influencing bank risk-taking incentives. Analyzing a sample of 6851 U.S. banks over the period 2004-2016, this paper presents strong empirical evidence that accommodative monetary policy leads to an increase in bank risk-taking, measured by the ratio of risk-weighted assets to total assets. Furthermore, this effect is diminished for banks with higher equity capital and amplified for larger banks. Overall, the findings suggest that the prolonged period of accommodative monetary policy in the U.S. might be harming financial stability.
Keywords: Accommodative monetary policy, Bank risk-taking, U.S. banks, Panel data
2 1. Introduction
In recent times, it has often been argued that the prolonged period of accommodative monetary policy during 2003 and 2005 was a crucial factor that led to the U.S. financial crisis of 2008 (e.g., Allen & Carletti, 2010; Maddaloni & Peydró, 2011; Taylor, 2009). In fact, Taylor (2009) argues that government interventions caused, prolonged and even worsened the crisis by setting interest rates extremely low compared to historical levels. Borio and Zhu (2008) were one of the first to analyze this element of the monetary policy transmission mechanism after the crisis and have named this concept “the risk-taking channel of monetary policy”.
3 Vissing-Jorgensen, 2011; Farhi & Tirole, 2012; Chodorow-Reich, 2014). Surprisingly, this claim has not yet been extensively studied for U.S. banks. Therefore, the research question this paper aims to answer is: What is the relationship between accommodative monetary policy and bank risk-taking in the U.S. after the 2008 financial crisis? To properly understand this relationship, it is important to specify what is meant by risk-taking. Here, risk-taking is characterized by the amount of risk a lender is prepared to hold in his/her portfolio. In the case of a bank, it concerns, among others, the ratio of risk-weighted assets to total assets.
This paper contributes to the existing literature as follows. As far as I am aware, this paper is one of, if not the first, to analyze the relationship between accommodative monetary policy and bank risk-taking in the U.S. following the 2008 financial crisis. Controlling for bank-specific and macroeconomic variables, I find that the prolonged period of accommodative monetary policy following the 2008 financial crisis has led to greater bank risk-taking in the U.S. This study analyzes the effects of expansionary monetary policy on bank risk-taking using a dynamic panel model, where the lagged dependent variable is included as an explanatory variable to account for the persistence of bank risk. The main measure for bank risk-taking is the ratio of risk assets to total assets, where risk assets are defined as risk-weighted assets (RWA). RWA is a measure of a bank’s assets, where the assets are weighted according to their risk. It is used by regulatory authorities to calculate minimum capital requirements for banks to reduce the risk of insolvency. For example, a loan that is secured by a letter of credit is considered to be riskier and consequently carries a larger weight than a mortgage loan that is secured with collateral. Naturally, an increase in the risk assets to total assets ratio reflects greater risk-taking.
4 addition, I find that the effects of accommodative monetary policy on the ratio of risk assets to total assets held by a bank is diminished for banks with higher capital and amplified for larger banks. Overall, the findings suggest that the Fed should carefully consider the fact that the prolonged period of accommodative monetary policy is causing banks to increase risk-taking, similar to the period before the crisis. If this is impacting financial stability negatively, it implies a tradeoff for the Fed between stimulating economic growth and avoiding vulnerabilities to develop within the financial system (Lamers, Mergaerts, Meuleman, & Vennet, 2016).
The paper consists of the following structure. Section 2 discusses the literature review, section 3 elaborates on the methodology of this paper, section 4 discusses the data and descriptive statistics, section 5 presents the main results and section 6 provides a discussion. Section 7 concludes.
2. Literature Review
This paper is related to a growing literature investigating the effects of accommodative monetary policy on bank risk-taking. In what follows, I discuss the two main transmission channels of monetary policy: the risk-taking channel and the risk-shifting channel. Afterwards, I discuss empirical evidence for the post-crisis period.
2.1 Risk-taking channel of monetary policy: theory
Historically, expansionary monetary policy have been considered an essential factor in boom-and-bust business cycles (Fisher, 1933; Hayek, 1939). An extended period of relatively low interest rates (i.e. below the levels of monetary policy suggested by historical evidence) can lead to financial imbalances via a lower risk perception of banks and investors. This element of the monetary transmission mechanism has been coined as the risk-taking channel of monetary policy and describes how monetary policy influences risk-taking behavior of financial intermediaries (Rajan, 2006; Borio & Zhu, 2008; Adrian & Shin, 2009). Borio and Zhu (2008) discuss 3 ways in which interest rates affect risk perception and tolerance.
Firstly, interest rates have an effect on valuations, incomes and cash flows that financial institutions often use as an input in risk management models.1 A lower policy rate increases asset
1 This is close in spirit to the familiar financial accelerator, where increases in collateral values lower borrowing
5 prices and collateral values as well as incomes and profits. In turn, this can reshape banks’ estimates of default probabilities, loss-given-default and volatilities. Specifically, rising asset prices together with decreasing asset price volatility lead to downsized risk estimations. This example can be applied to the universal use of value-at-risk (VaR) measures for economic and regulatory capital objectives (Danielsson, Shin, & Zigrand, 2004). Particularly, in rising financial markets with decreasing volatility and enhanced bank’s capital conditions, the use of VaR models tends to release risk budgets of banks and encourages higher risk positions. Similarly, Adrian and Shin (2009) emphasize that changes in estimated risk by banks are what drive adjustments in their balance sheets and leverage conditions, which, in turn, exacerbate business cycle movements. (Altunbas et al., 2014).
Secondly, interest rates affect risk perception and tolerance via an increase in the “search for yield” by banks (Rajan, 2006). Low interest rates can encourage financial institutions to increase risk-taking due to behavioral, institutional or contractual reasons. For example, to satisfy the demands of investors and shareholders, banks promise that they will commit to producing certain nominal rates of return. To the extent that low interest rates reduce bank profits, they make riskier assets more appealing, as banks look to increase their yields to maintain the trust of their investors and shareholders. Furthermore, investors often use short-term returns as a criteria to evaluate whether a manager is performing well, pushing managers to expose themselves to additional risk when interest rates are low (Altunbas et al., 2014). The strength of this effect increases with the prevalence of agency problems. Bank capital serves as a commitment device to limit risk-taking and mitigates agency problems. For example, weakly capitalized banks have a minimum amount of capital at stake in case of failure, tempting them to engage in excessive risk-taking as the potential benefits in case of success are high. In case of failure, debtholders incur relatively the most losses, thus these banks essentially transfer bank risk to debtholders. This suggests that less capitalized banks are expected to increase risk-taking more during periods of accommodative monetary policy (Lamers et al., 2016).
Thirdly, risk perception and tolerance might also be affected by the communication policies and reaction function of the central bank. For example, banks’ beliefs that the central bank will engage in expansionary monetary policy during periods of economic crisis could lower their
6 expectations of huge downside risks. This is a typical moral hazard problem, which arises as a result of this anticipated insurance effect (Altunbas et al., 2014).
2.2 Risk-taking channel of monetary policy: evidence
Several papers provide empirical evidence supporting the risk-taking channel. Delis and Kouretas (2011) analyze the relationship between low interest rates and bank risk-taking for Europe for the period 2000-2008. Using a 2SLS-IV and Generalized Method of Moments (GMM) model, they find unambiguous evidence that the low interest rate environment in Europe prior to the crisis significantly increased bank risk-taking. Additionally, they find that this relationship is weaker for banks with high levels of capitalization and stronger for banks that engage in non-traditional banking activities. Maddaloni and Peydró (2011) take a different approach and use evidence from a unique lending survey for the Euro-area and the U.S. to analyze whether monetary policy has an effect on the lending standards of banks.
7 quarterly U.S. data for the period 1980-2008. By means of a standard orthogonalized VAR model, they find that lowering policy rates increases bank riskiness, confirming the existence of the risk-taking channel.
2.3 Risk-shifting channel of monetary policy: theory and evidence
Another part of the literature concerns the risk-shifting channel (Dell’Ariccia et al., 2014). This channel functions through the liability side of a bank’s balance sheet and predicts a positive relationship between accommodative monetary policy and bank risk-taking. Everything else equal, accommodative monetary policy interventions which lower the cost of a bank’s liabilities will lead to higher bank profits when it succeeds, and thus encourages the bank to limit risk-taking to obtain these gains. In turn, this effect depends on the degree of limited liability protection the bank has. To understand why, consider a fully leveraged bank whose financing consists only out of debt/deposits. Under limited liability, this bank will not undergo any losses in case it collapses. Lower cost of liabilities will lead to higher net interest margins on all assets, by reducing the rate banks have to pay to depositors. The bank can maximize this effect by lowering the risk of its portfolio, opting for a safer portfolio with a higher likelihood that the bank will be required to repay depositors. On the contrary, for a fully capitalized bank, the effect of a reduction in funding costs will, everything else equal, raise net interest margins equally across portfolios and consequently have no impact on the bank’s risk choices. The more advantageous the limited liability protection is to a bank, the stronger the risk-shifting effect. Therefore, this risk-shifting effect is strongest for fully leveraged banks and lowest for banks with no leverage, as they do not benefit from limited liability.
8
2.4 Post-crisis empirical evidence
Chaudron (2016) uses standard fixed effects and bias-corrected least-squares dummy variable (LSDVC) methods to analyze interest rate risk-taking in a sustained low interest rate environment for the period after the 2008 financial crisis. Risk-taking is proxied using banks’ interest rate risk positions, i.e. the target duration of equity. The findings suggest that, in the case of the Netherlands, banks held modest interest rate risk positions over the period 2008 to 2015, indicating that low interest rates have not led to increased rate risk-taking by Dutch banks. Furthermore, Gehrig and Iannino (2017) investigate the evolution of resilience of European banks from 1987 to 2015 using 2 key measures of systemic risk. SRISK, a measure of a bank’s exposure to systemic risk and Delta CoVar, a measure of contagion risk emerging in a bank. Using an asymmetric GARCH and a correlation model, they conclude that low interest rates have (weak) destabilizing effects on European banks; they reduce bank resiliency.
Other papers provide evidence of a search-for-yield effect following the recent accommodative monetary policy in the U.S. (e.g., Becker & Ivashina, 2015; Di Maggio & Kacperczyk, 2017). However, these papers focus on nonbanks such as money market funds, mutual funds, and insurance companies. Finally, Lamers et al. (2016) investigate how monetary policy interventions by the European Central Bank and the Fed impacts bank systemic risk as perceived by stock markets. They identify monetary policy shocks using a structural VAR approach by analyzing how the volatility of these shocks changes on days of monetary policy announcements. Afterwards, they employ a panel regression analysis, where they examine the effect of these shocks on bank systemic risk. Here, bank systemic risk is defined as a bank’s contribution to the risk of the financial system. Ultimately, they measure bank risk using the stressed market value (SMV), i.e. the capital the bank expects to have after a large adverse shock. Their findings support the notion that accommodative monetary policy might stimulate risk-taking appetite of banks.
3. Methodology
The general model that will be used for empirical analysis takes the following form:
9 where bank risk-taking, denoted by r for bank i at time t, is defined as a function of the stance of monetary policy, ir; a set of bank-specific control variables, b; a set of macroeconomic control variables, c, which affects all banks, a dummy variable, D; and an interaction term involving the dummy variable and the monetary policy stance. This dummy variable is referred to as post-crisis and takes on the value of 1 after 2008, and 0 otherwise. It is included to account for the fact that a crisis occurred during the sample period, which might have altered the relationship between monetary policy and bank risk-taking; also known as a structural break in the parameters. Therefore, the interaction term captures how the relation between monetary policy and bank risk-taking has changed after the crisis. Further on in this paper, as the empirical analysis continues, Eq. (1) will be extended with theoretical and empirical components. However, I first discuss the dataset and the variables used in this paper.
Following the paper of Delis and Kouretas (2011), the focus of this paper lies in analyzing banks that take deposits. Consequently, investment banks are excluded as they do not take deposits and thus are not relevant to the theoretical discussion mentioned above. Additionally, I exclude cooperative banks due to persistent missing data for the sample period. Evidently, there might be survivorship bias present in the sample. Banks that originally existed but during the sample period went into bankruptcy or became dissolved due to a merger/acquisition are not (fully) reported because of this. Hence, this might lead to biased estimates as my sample only contains banks who survived and overlooks those who did not. For this paper, the focus lies on analyzing U.S. commercial and savings banks for the period 2004-2016. While I am interested in studying risk-taking behavior after the 2008 financial crisis, I extend the sample period back to 2004 to increase the statistical power of my empirical analysis, i.e. to obtain more accurate and reliable results.
10 data and conclude that using annual data is enough to explain the impact of monetary policy rates on bank lending and that the use of quarterly data does not add significant understanding to the results. Similarly, Delis and Kouretas (2011) compare their results obtained by both quarterly and annual data and concluded that the results were (surprisingly) similar. Hence, they argued that using annual instead of quarterly data to analyze the relationship between interest rates and bank risk-taking does not lead to a loss of information. Therefore, annual data is likely to be sufficient for this analysis. Furthermore, time-series data at a quarterly frequency has a considerable amount of missing data for these banks. Because of these theoretical and empirical arguments, I focus on annual data in my analysis.
Additionally, I filter the data such that only banks with at least 5 years of available data for the relevant variables are included. This ensures that all banks which are included in the sample have sufficient data available to properly analyze their risk-taking behavior over the sample period. Another potential issue with collecting data from Orbis Bank Focus is the problem of double counting bank statements, as Orbis Bank Focus sometimes includes both consolidated and unconsolidated bank statements from the same (mother) company. To mitigate the issue of double counting, I follow the guidelines set forward by Duprey and Lé (2016) and use data from unconsolidated accounts if available and otherwise from consolidated accounts. By following this method, I keep most banks at the disaggregated level and consequently maximize the sample size for my analysis of bank risk-taking. Moreover, I also exclude bank holding companies as it can occur that the consolidated balance sheet of bank holding companies reports the assets of its subsidiaries as well, leading to double counting issues. In what follows, I describe the dependent and explanatory variables used in this paper.
4. Data
4.1 Bank risk-taking
11 is riskier and requires more capital than a mortgage loan that is secured with collateral. Naturally, an increase in the risk assets to total assets ratio (denoted as risk assets) reflects greater risk-taking. The ratio of risk assets to total assets indicates the riskiness of a bank’s balance sheet and thus directly reflects bank risk-taking (Delis & Kouretas, 2011). The sample mean of risk assets equals 67.6%, with the lowest average value reported in 2012 (63.5%) and the highest in 2008 (71.3%), when the financial crisis unfolded. To ensure that results are robust, a second measure for bank risk-taking is implemented: the ratio of performing loans to total loans (denoted as
non-performing loans).
Related studies use several other measures to measure bank risk-taking, such as the z-score and variance in bank profits. However, these measures are a better reflection of insolvency risk rather than “bank risk-taking” and therefore bear little relevance to the theoretical discussion put forward in this paper (Delis & Kouretas, 2011). The ratio of non-performing loans to total loans reflects the quality of bank assets, i.e. the potential negative effects of lower assets quality on earnings and asset values. An increase in non-performing loans entails an increase in bank risk-taking. The sample mean of non-performing loans equals 1.70%, with the lowest average value occurring in 2005 (0.80%) and the highest in 2010 (2.90%). Data for these variables are obtained from Orbis Bank Focus. It must be noted that I include the ratio of non-performing loans to total loans to proxy bank risk-taking purely to test the robustness of my results, as I derive my choice of bank risk-taking measures from Delis and Kouretas (2011). In fact, I later argue in the discussion of this paper that the ratio of non-performing loans to total loans is not an accurate measure to capture bank risk-taking.
4.2 Monetary policy stance
The monetary policy stance reflects whether monetary policy is accommodative; typically characterized by periods of decreasing federal funds rates, or contractionary; typically characterized by periods of increasing federal funds rates. Over the period 2009-2015, the federal funds rate has been kept close to the ZLB and has barely fluctuated. To capture a proper relationship between the monetary policy stance and bank risk-taking, implementing a measure which has fluctuated over the sample period is econometrically essential.
Thus, I opt to proxy the monetary policy stance using the Taylor rule Gap (denoted as
12 monetary policy. Monetary policy is argued to be endogenous as it depends on local economic conditions (Maddaloni & Peydró, 2011; Jiménez et al., 2014; Ioannidou et al., 2015). The Taylor rule gap is measured as the difference between the annual average of the nominal overnight interest rate and the annual average interest rate implied by the Taylor rule. The Taylor Rule is an interest rate model, developed by the economist John Taylor to guide policymakers. It prescribes what the federal funds rate should be, based on the output gap and current inflation. The output gap reflects the difference between potential and actual output. I use the Taylor rule as a standard benchmark to analyze monetary policy, as it does a good job at describing the monetary policy of the Fed over the last 30 years until 2007 (Castro, 2011).2 Consequently, the Taylor rule gap captures the stance of monetary policy, i.e. it indicates whether interest rates are “too high” or “too low” by comparing the current short-term nominal rate to the Taylor rule interest rate. A positive TGAP implies that the federal funds rate is higher than the Taylor rule interest rate; indicating tight monetary policy. The opposite holds for a negative TGAP, which indicates expansionary monetary policy.
Figure 1 depicts the nominal federal funds rate and the Taylor rule interest rate for the U.S. for the period 2004-2016. Between 2009 and 2015, the federal funds rate has continuously remained near the ZLB, reflecting the accommodative monetary policy by the Fed over the recent years. In contrast, the Taylor rule (and thus TGAP) has increased significantly since 2009. To ensure results are robust, I experiment with several specifications of the monetary policy stance. Besides the Taylor rule gap, I proxy the monetary policy stance using the annual change in the Fed’s assets as a % of nominal GDP (denoted as ΔFed assets3) and long-term interest rate (denoted
as long-term rate). The long-term rate is the annual average of the 10-year government bond yield. The significant expansion in the Fed’s balance sheet over the recent years reflects the Fed’s unconventional expansionary monetary policy activities since 2009, QE being the main one. An important objective of QE is to reduce long-term interest rates.
The Fed aims to achieve this by purchasing long-term assets from financial institutions, which increases the price and consequently reduces the yield of these long-term securities, ultimately lowering long-term rates. I interpret this variable as the following: a positive value for ΔFED Assets reflects an expansion of the balance sheet and is associated with expansionary
2 After 2007, the financial crisis unfolded and in 2009 the U.S. and the Fed purposely deviated from the Taylor rule
by setting nominal short-term interest rates near the ZLB.
3 I would like to thank professor Martien Lamers for his suggestion on using this measure to proxy unconventional
13 (unconventional) monetary policy. Therefore, a positive coefficient for this variable implies that accommodative monetary policy increases bank risk-taking. ΔFed assets has significantly increased after 2008, reflecting the start of the Fed’s unconventional monetary policy interventions, while the long-term interest rate has mostly been decreasing since 2004. Data for these proxies of the monetary policy stance I obtain directly from the FRED database of the Federal Reserve Bank of St Louis, including the Taylor rule interest rate.
Figure 1. Nominal federal funds rate and the Taylor rule interest rate for The U.S.
The figure depicts the annual average nominal fed funds rate and the annual average interest rate based on the Taylor rule for the U.S. for the period 2004-2016. Source: FRED database of the Federal Reserve Bank of St Louis.
bank-14 level lending rate (denoted as bank lending rate) by the ratio of interest income to total customer loans. This ratio indicates the average price that customers pay banks on their loans. Figure 2 illustrates a simple regression regarding the relationship between risk assets and the bank-level lending rate. The regression evidently reveals a negative relationship, presenting primary evidence that accommodative monetary policy is related to higher bank risk-taking. The question remains how this relationship will be altered when numerous control variables are added to the regression.
Figure 2. Bank-level interest rates and bank risk-taking
The figure reports the non-parametric regression between bank risk-taking, measured by the ratio of risk assets to total assets, and the bank-level lending rate, measured by the ratio of interest income to total customer loans.
4.3 Bank-specific control variables
15 2011). Hence, I control for this by means of the variable efficiency (denoted as efficiency), where I proxy efficiency using the ratio of operating revenues to operating expenses. Additionally, the profitability of a bank might also influence bank risk-taking. Intuitively, the increase in profits might be used to extend new loans the upcoming period, increasing the level of risk assets held by banks and thus bank risk-taking. I proxy profitability (denoted as profitability) using the ratio of profits before taxes to total assets.
Likewise, bank capitalization is an important determinant of bank risk-taking. Highly capitalized banks are better prepared to take on additional risk compared to weakly capitalized banks, since they have stronger capital buffers to shield itself against adverse events. Therefore, I expect highly capitalized banks to engage in greater bank risk-taking relative to less capitalized banks. However, it could also be the case that highly capitalized banks display prudent behavior and engage in less risky behavior, as they find it important to remain highly capitalized (Delis & Kouretas, 2011). Hence, the effect of capitalization on bank risk-taking is ambiguous. I control for bank capitalization using the ratio of equity capital to total assets (denoted as capitalization). Finally, bank size and non-traditional banking activities could be important determinants of bank risk-taking as well, thus I control for bank size (denoted as size) and off-balance sheet activities (denoted as off-balance sheet). For a comparable set of bank-specific control variables in risk equations, see e.g., Laeven and Levine (2009) and Demirgüç-Kunt, Detragiache, and Tressel (2008). I proxy size using the natural logarithm of real total assets and off-balance sheet using the ratio of off-balance sheet items to total assets. Data for these bank-level control variables I collect from Orbis Bank Focus.
4.4 Macroeconomic control variables
16 of failure is positively related with concentration. In turn, importance of banks is measured using the ratio of domestic credit provided by the banking to nominal GDP. The share of credit provided by banks indicates to what extent alternative sources of finance are available for companies and, thus also the level of competition and development of the banking system (Delis & Kouretas, 2011). I obtain data for these macroeconomic variables from the FRED database of the Federal Reserve Bank of St Louis and IMF International Financial Statistics.
4.5 Endogeneity of bank characteristics and monetary policy
According to Delis and Kouretas (2011), it is well-known in the banking literature that efficiency, profitability, bank capitalization and off-balance sheet items are endogenous variables in risk equations. For example, higher taking might lead to increased bank efficiency, if higher risk-taking is the source of higher operating revenues. Everything else equal, higher operating revenues leads to higher efficiency. Thus, it becomes clear that efficiency should be treated as an endogenous variable. Similarly, excessive bank risk-taking might lead to an increase in non-performing loans, which in turn harms profitability. As a result, I treat profitability as an endogenous variable and it enters the equation lagged once.4 Furthermore, regarding capitalization, banks are expected to tradeoff higher levels of bank capital for risk assets to increase their return on equity, therefore this relationship is evidently endogenous. Moreover, if a bank is currently engaging in a low level of bank risk-taking activity, it might decide to focus more on non-traditional banking activities instead. Hence, I treat this variable as endogenous.
Finally, I argue that bank size is predetermined, since banks are aware of their relative size when engaging in risk-taking. For example, the “too big to fail” theory argues that the failure of large financial institutions would be disastrous for the financial system. Hence, these institutions receive support from the government in case the threat of potential failure emerges. As a result, relatively large banks know they are “big”, take this into account and engage in greater risk-taking than they normally would. Consequently, I treat bank size as a predetermined variable, in line with Delis and Kouretas (2011). What is more, monetary policy is argued to be endogenous as well as it depends on local economic conditions (Maddaloni & Peydró, 2011; Jiménez et al., 2014;
17 Ioannidou et al., 2015). Since the Taylor rule gap reflects the difference between policy rates and a formula describing monetary policy rather than measuring it, this proxy is likely to not suffer from endogeneity. In addition, bank-level lending rates might be endogenous as well, considering that banks determine their lending rates by discounting the expected level of risk in their portfolios (Delis & Kouretas, 2011). Thus, ΔFED assets, long-term rate and bank lending rate I treat as endogenous.
4.6 Descriptive statistics and correlations
18 Table 1. Descriptive statistics for a sample of 6581 U.S. banks, 2004-2016
Variable Mean Standard deviation Min Max Observations
Risk assets 0.676 0.140 0.015 2.119 77,467
Non-performing loans .0170 0.027 0.000 1.000 77,432
TGAP -0.015 0.011 -0.032 0.009 85,631
ΔFED assets 0.017 0.031 -0.009 0.092 79,044
Long-term rate 0.032 0.010 0.018 0.048 85,631
Bank lending rate 0.110 3.231 0.000 566.315 77,398
Capitalization 0.113 0.055 -0.091 0.987 77,467 Lagged Profitability 0.009 0.030 -0.320 3.409 71,633 Size 18.983 1.344 14.643 28.365 77,467 Efficiency 1.515 1.533 -11.289 281.500 77,459 Off-balance sheet 0.231 3.075 0.000 400.314 77,467 Economic growth 0.018 0.016 -0.028 0.038 85,631 Inflation 0.021 0.012 -0.003 0.038 85,631 Importance of banks5 2.311 0.112 2.129 2.510 79,044 Concentration 0.382 0.036 0.333 0.430 85,631
The table reports descriptive statistics for the variables used in the empirical analysis. Risk assets is the ratio of risk assets to total assets, non-performing loans is the ratio of non-performing loans to total loans, TGAP is the Taylor rule gap, ΔFED assets is the absolute annual change in central bank assets as a % of nominal GDP, long-term rate is the annual average of the 10-year U.S. government bond yield, bank lending rate is the ratio of interest income to total customer loans, capitalization is the ratio of equity capital to total assets, lagged profitability is the ratio of profits at before taxes to total assets in year t-1, size is the natural logarithm of real total assets, efficiency is the ratio of operating revenues to operating expenses, off-balance sheet items is the ratio of off-balance sheet items to total assets, economic growth is real GDP growth, inflation is the annual average CPI inflation, importance of banks is the ratio of domestic credit provided by the banking sector to GDP, concentration is the 3-bank concentration ratio.
19 Table 2. Pairwise correlation matrix for a sample of U.S. banks, 2004-2016
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [1] Capitalization 1.00 [2] Lagged profitability 0.13 1.00 [3] Size -0.11 0.06 1.00 [4] Efficiency 0.11 0.14 0.08 1.00 [5] Off-balance sheet 0.02 0.04 0.04 0.04 1.00 [6] Economic growth 0.03 0.05 0.00 0.03 0.00 1.00 [7] Inflation -0.00 0.05 -0.07 0.02 -0.00 0.34 1.00 [8] Importance of banks 0.03 -0.01 0.09 -0.02 0.00 0.05 -0.59 1.00 [9] Concentration -0.02 0.03 -0.11 0.02 -0.00 -0.30 0.55 -0.67 1.00 [10] ΔFED assets -0.01 0.00 0.01 -0.02 0.00 -0.29 0.33 -0.04 0.04 1.000 [11] Long-term rate -0.01 0.05 -0.11 0.03 -0.00 0.00 0.55 -0.55 0.92 -0.09 1.00
[12] Bank lending rate 0.02 0.02 0.01 0.01 0.00 -0.01 0.00 -0.01 0.01 -0.00 0.01 1.00
[13] TGAP -0.03 -0.04 -0.04 -0.01 -0.00 -0.66 -0.26 -0.13 0.59 -0.15 0.45 0.01 1.00
20 5. Econometric analysis and results
5.1 Dynamic risk and endogenous controls
For simplicity, I start by considering a model in which there is no endogeneity present. The primary approach to estimating Eq. (1) is to estimate a panel model with random effects or fixed effects. The reasoning behind the use of fixed effects is to capture (possible) unobserved bank heterogeneity, i.e. unobserved bank-specific characteristics that I do not control for in my model which can contribute to explaining certain risk behavior of banks. For example, Maddaloni and Peydró (2011) find that high securitization activities by banks have amplified the effects of low interest rates on bank lending standards. Additionally, each bank has unique characteristics that might influence their risk-taking behavior, which is not easily measured or cannot be estimated. Furthermore, I verify that fixed effects are preferred over random effects using the Sargan test6. (This test reveals that the difference in coefficients between fixed and random effects is systematic, thus the fixed effects model is preferred).
However, besides the endogeneity problem, there is an essential concern in estimating Eq. (1) using a model with fixed effects. According to Delis and Kouretas (2011), bank risk is dynamic; it is a type of risk that persists over time and hence will deviate from its equilibrium value in the short run. They provide at least 4 theoretical arguments which explain the dynamic character of bank risk. Firstly, persistence can indicate the existence of strong competition, which tends to reduce bank risk-taking (e.g., Keeley, 1990; Cordella & Yeyati, 2002). Secondly, relationship-banking involving risky borrowers will have a longstanding effect on bank risk-taking, despite the improved efficiency from repeatedly working with the same customer. A similar process would occur in the presence of bank networks or if the banking sector is opaque. Thirdly, to the degree that bank risk is related to the stage of the business cycle, it could take time for banks to be able to smooth the effects of economic shocks. Additionally, risks might persist due to regulation. Notably, deposit guarantees or capital requirements can aggravate moral hazard issues, causing inefficient and risky investments to occur over a significant period of time. Finally, besides these theoretical arguments set forward, the accumulation of flows into stocks over time and a feedback
6 Commonly, the Hausman test is used to determine whether random effects or fixed effects is preferred. However,
21 effect between these two is mainly what leads to dynamic behavior. A stock variable indicates the quantity of a variable at one specific point in time (e.g., bank capital and risk assets), while a flow variable (e.g., economic growth and inflation), reflects the change of a variable over a period of time. Since both stock and flow variables are present in the model, the possible impact of stock variables on flow variables might be better captured by a dynamic model (Delis & Kouretas, 2011).
Ultimately, if risk is truly persistent, a static model is biased. Thus, estimating Eq. (1) using a panel model with fixed effects produces invalid results. Nevertheless, for the readers who are interested, the results of the estimation of Eq. (1) using a panel model with bank fixed effects can be found in the Appendix; see Table A1. Instead, the use a dynamic empirical model is justified. Hence, Eq. (1) needs to be adjusted. This leads to the following estimation of Eq. (2):
𝑟𝑖,𝑡 = 𝛼 + 𝛿(𝑟𝑖,𝑡−1) + 𝛽𝑖𝑟𝑖,𝑡+ 𝜇𝑏𝑖,𝑡 + 𝛾𝑐𝑡+ 𝜆𝐷𝑡+ ø𝑖𝑟𝑖,𝑡∗ 𝐷𝑡+ 𝜀𝑖,𝑡 (2) where bank risk-taking, denoted by r for bank i at time t, is defined as a function of its 1-year lagged level r for bank i; the monetary policy stance, ir; a set of bank-specific control variables, b; a set of macroeconomic variables, c; a dummy variable, D; and an interaction term involving the dummy variable and the monetary policy stance. This dummy variable is referred to as post-crisis and takes on the value of 1 after 2008, and 0 otherwise. This equation can be estimated using the GMM model for dynamic panel data put forward by Arellano and Bover (1995) and Blundell and Bond (1998). The coefficient on the lagged dependent variable can be viewed as the speed of convergence to the equilibrium. A value of δ statistically equal to 0 implies that bank risk is characterized by a high speed of adjustment, while a value of δ statistically equal to 1 means that the adjustment is very slow. Values between 0 and 1 implies that risk persists, but returns to its normal level over time. Finally, δ takes implausible (negative) values if convergence to equilibrium cannot be achieved, which probably is a sign that there is a problem with the dataset (e.g., a very small time dimension of the panel).
22 variables as endogenous is found in section 4.5 of this paper. In econometric terms, endogeneity implies that these explanatory variables are correlated with current period as well as past period error terms of the regression equation, leading to biased and inconsistent estimates. In addition, bank size is treated as a predetermined variable, indicating that the value of this variable is determined prior to the current period.
Econometrically, this implies that the current period error term of the regression equation is uncorrelated with current and past values of the predetermined variable. However, this error term might be correlated with future values of the predetermined variable. For example, an unexpected shock (e.g., a financial crisis) can lead to bank assets (and thus bank size) decreasing in the future. In other words, bank size might be correlated to past period error terms. Furthermore, in Eq. (2) the lagged dependent variable included as explanatory variable is endogenous, since it is directly correlated with the error term. The GMM model uses lags of the endogenous variables and the predetermined variables as instruments to deal with the problem of endogeneity. Thus, the complete set of instruments is denoted by (ri1,…ri,t-n, xi1,…xi,t-n, zi1,…zi,t-n), where r is the dependent
variable, x is the set of remaining endogenous variables, z the predetermined variable and n the amount of lags.7
The findings are reported in Table 3 and, to save space, I only discuss the most important findings. For the detailed results, see Table A2 in the Appendix. By construction, the GMM model I implement to estimate Eq. (2) uses this equation both in levels and first-differences.8 First-order
autocorrelation could be expected in first differences of the error term, since the differenced error term (∆εi,t = εi,t− εi,t−1) and the lagged differenced error term (∆εi,t−1 = εi,t−1− εi,t−2) both contain the εi,t−1 term. Hence, autocorrelation in levels is identified by examining whether there is second order autocorrelation present in first differences. In fact, evidence of higher order autocorrelation would reveal that certain lags of the dependent variable are endogenous, making them bad instruments (Roodman, 2009). Evidently, I use the Arellano-Bond AR (2) test to detect if there is 2nd order autocorrelation present. This test has as null hypothesis that there is no 2nd order autocorrelation present and is applied to the differenced residuals. If this test is rejected however, second-order autocorrelation is detected, deeming instruments of lag length 2 invalid.
7 This discussion regarding instrumental variables is based, among others, on Arellano and Bover (1995), Blundell
and Bond (1998), and Bond (2002).
23 This appears to be the case for all the regressions in Table 4, except for regression VI. This issue is dealt with by using deeper lags, e.g., lags of the 3rd or 4th order of the dependent variable. After
doing so, I find that for regressions I, II and IV, the AR (3) test is not rejected. Similarly, for regressions III and V, the AR (4) test fails to be rejected. Thus, no evidence of third-order and fourth-order correlation is found for these regressions, deeming the lagged dependent variables as valid instruments for all the regressions. A limitation of the obtained results is that the Sargan test indicates that the model is over-identified. However, a major limitation of this test is that it requires a homoskedastic error term to be considered consistent, whereas the error term of this model is heteroskedastic (which is controlled for), causing this test to produce incorrect results. Hence, I presume the Sargan test to be invalid and do not make inferences from the results it produces; this is explained further in the discussion of this paper.
The findings illustrate that the coefficients on the lagged dependent variables are statistically significant for all estimated regressions, varying between 0.6 and 0.9, leading to the conclusion that bank risk-taking is highly persistent. I also experimented with a higher order of lags for the dependent variable and found no evidence of risk persistence beyond the first year. This implies that bank risk-taking persists for a while, although eventually it returns to its equilibrium level. This holds irrespective of the proxy used to measure bank risk-taking. The coefficient of the interaction term between monetary policy stance and post-crisis reflects how the effects of monetary policy on bank risk-taking has changed after the crisis. Therefore, the sum of the coefficients of the monetary policy stance and the interaction term reflects the overall effect of the monetary policy stance on bank risk-taking over the period analyzed. This effect is highly significant but changes sign, depending on the proxy used to measure bank risk-taking. The findings suggest a highly significant and positive impact of accommodative monetary policy on bank risk-taking (see columns I, II, III), while the coefficient of the bank-level lending rate is practically insignificant with a value of 0.000 (see column IV). A possible explanation for these findings is that the current low interest rate environment encourages banks to increase risk-taking, as they search for higher yields to satisfy investor demands (Rajan, 2006). However, these results hold only for the risk assets equation (see columns I-IV).
24 deposit rates via interest rate pass-through (Karagiannis et al., 2010), the risk-shifting channel argues that banks respond to this by decreasing bank risk-taking. Lower cost of liabilities leads to higher net interest margins on all assets, by reducing the rate banks must pay on its deposits. The bank can optimize this effect by engaging in less risk-taking, i.e. lowering the risk of its portfolio, with a higher likelihood that the bank will be required to repay depositors. Overall, the findings supporting the search for yield theory (columns I, II, III) are not robust to alternative specifications of bank risk-taking.
Furthermore, higher bank capital leads to an increase in the ratio of risk assets to total assets held by banks, irrespective of the variable used to proxy the monetary policy stance. In contrast, higher bank capital reduces the ratio of non-performing loans to total loans when the bank lending rate is used as proxy of the monetary policy stance (see column VI). Thus, the effect of bank capitalization on bank risk-taking is evidently ambiguous. On the one hand, I argue that highly capitalized banks are more likely to engage in bank risk-taking than weakly capitalized banks, as they have stronger capital buffers to shield itself against adverse events. On the other hand, Delis and Kouretas (2011) argue that a negative relation between bank capital and bank risk-taking is intuitive, since higher equity capital, through stricter capital requirements, reflects more prudent bank behavior.
25 Table 3. Monetary policy and bank risk-taking: dynamic panel regressions
I II III IV V VI
Lagged risk assets 0.889*
(0.015) 0.832* (0.014) 0.853* (0.020) 0.824* (0.016) Lagged non-performing loans 0.568* (0.040) 0.797* (0.055) TGAP -0.596* (0.056) 0.112* (0.023) ΔFED assets 0.108* (0.015) Long-term rate -1.279* (0.133) Bank-level lending rate -0.000 (0.000) 0.001* (0.000) Monetary policy stance*post-crisis -0.251** (0.099) 0.097* (0.022) 0.423* (0.161) 0.000** (0.000) 0.056 (0.034) -0.000** (0.000) Capitalization 0.246* (0.083) 0.634* (0.103) 0.509* (0.097) 0.312* (0.099) -0.017 (0.025) -0.076* (0.025) Lagged profitability -0.368* (0.098) -0.781* (0.104) -1.024* (0.136) -0.304* (0.095) -0.614* (0.134) -0.120*** (0.067) Size 0.011* (0.002) 0.019* (0.002) 0.009* (0.002) 0.018* (0.002) 0.006* (0.001) 0.002* (0.001) Economic growth 0.039*** (0.021) 0.407* (0.028) 0.297* (0.022) 0.259* (0.023) -0.128* (0.009) -0.200* (0.006) Obs 65,029 65,029 65,029 64,990 64,995 64,969 Wald-test 24,705.22 19,830.80 17,760.53 21,665.79 8,184.53 282,109.20 p-value 0.000 0.000 0.000 0.000 0.000 0.000 AR2 0.256 AR3 0.168 0.233 0.183 AR4 0.466 0.565 Sargan 0.000 0.000 0.000 0.000 0.000 0.000
The table reports coefficients and standard errors (in parentheses). In regressions I-IV dependent variable is the ratio of risk assets to total assets and in regressions V-VI the ratio of non-performing loans to total loans. The explanatory variables are as follows: TGAP is the Taylor rule gap, ΔFED assets is the absolute annual change in central bank assets as a % of nominal GDP, long-term rate is the annual average of the 10-year U.S. government bond yield, bank-level lending rate is the ratio of interest income to total customer loans, post-crisis reflects a dummy variable which equals 1 after 2008 and 0 otherwise, monetary policy stance*post-crisis reflects the interaction term between monetary policy and the dummy variable post-crisis, capitalization is the ratio of equity capital to total assets, lagged profitability is the ratio of profits before taxes to total assets in year t-1, size is the natural logarithm of real total assets, and economic growth is real GDP growth. Obs is the number of observations, the Wald-test and its associated p-value denote the goodness of fit of the regressions and Sargan is the p-value of the Sargan Wald-test of overidentifying restrictions. AR2, AR3 and AR4 denote the p-values for the test for second, third and fourth order autocorrelation.
26 Finally, I find that in periods of economic growth, the ratio of risk assets to total assets that banks hold increases. This supports the theory of Delis and Kouretas (2011), who argue that during periods of prosperous economic conditions, banks tend to engage in more lending activities in search for higher yield. In contrast, the ratio of non-performing loans to total loans decreases, a finding that might be explained by the lagged nature of this measure, which I further elaborate on in the discussion of this paper. In short, I argue that a majority of these non-performing loans are loans extended in the past, where improving economic conditions since the financial crisis is improving borrowers’ ability to repay their loans. In turn, this reduces the probability that they will default on these loans.
5.2 Distributional effects of monetary policy due to bank characteristics
Following the analysis of the relationship between monetary policy and bank risk-taking, I examine how this relationship depends on bank characteristics. Specifically, keeping everything else constant, I analyze how bank size, bank capital and profits earned in the previous year influences the effects of monetary policy on bank risk-taking. Commonly, when estimating interaction terms between two continuous variables, variables are mean-centered, which consists of transforming the values of variables to deviations from their means. The reasoning behind centering variables is that it might make results more meaningful and interpretable. For example, if variables are not centered and I am analyzing the interaction between monetary policy and bank capitalization, the main effect of monetary policy would be the effects of monetary policy on a bank with no capital. In contrast, when variables are centered, the main effect would be the effects of monetary policy on banks with an average amount of capital. Hence, I ‘mean-center’ the following variables for this analysis: TGAP, bank-level lending rate, size, lagged profitability and
capitalization. To save space, I use only 2 proxies of the monetary policy stance in this analysis.
This leads to the estimation of Eq. (3):
𝑟𝑖,𝑡 = 𝛼 + 𝛿(𝑟𝑖,𝑡−1) + 𝛽𝑖𝑟𝑖,𝑡+ 𝜇𝑏𝑖,𝑡+ 𝛾𝑐𝑡+ 𝜆𝐷𝑡+ ø𝑖𝑟𝑖,𝑡∗ 𝐷𝑡+ 𝜎𝑖𝑟𝑖,𝑡∗ 𝑏𝑖,𝑡+ +𝜀𝑖,𝑡 (3)
27 dummy variable and the monetary policy stance, and an interaction term between monetary policy and bank characteristics. The dummy variable is referred to as post-crisis and takes on the value of 1 after 2008, and 0 otherwise. Estimations are carried out using the dynamic panel data method and, to save space, I only report the most relevant findings in Table 4. For the complete set of results, see Table A3 in the Appendix.
The results indicate that the effects of accommodative monetary policy on the ratio of risk assets to total assets is diminished for banks with higher equity capital and amplified for larger banks (see column I). Put differently, highly capitalized banks are able to absorb the effects of accommodative monetary policy on bank risk, while for larger banks this effect is stronger. These findings can be explained using the search for yield theory, which argues that less capitalized banks, via the increasing prevalence of agency problems, are expected to increase risk more in response to accommodative monetary policy (Lamers et al., 2016). In contrast, when
non-performing loans is used to proxy bank risk-taking, the main effect of TGAP on bank risk-taking
is positive. This effect is reduced for banks with higher equity capital and higher profits the previous year and magnified for larger banks. In other words, accommodative monetary policy leads to less bank risk-taking in this case, and this effect is weaker for highly capitalized and profitable banks and stronger for larger banks. These results can be explained by the risk-shifting theory (Dell’Ariccia et al., 2014), which argues that the effects of accommodative monetary policy on bank risk-taking is expected to be stronger for weakly capitalized banks, as they benefit more from reducing risk-taking via limited liability protection, compared to highly capitalized banks.
Analyzing these interaction terms allows me to examine whether the effects of monetary policy on bank risk-taking could change sign depending on bank characteristics. Interestingly, for
risk assets a negative and significant coefficient of TGAP is found and a positive and significant
effect for the interaction between TGAP and capitalization. This implies that for sufficiently capitalized banks, the impact of accommodative monetary policy turns positive, and consequently leads to a decrease in risk assets. On the other hand, for non-performing loans, the effect of TGAP on bank risk-taking is positive and significant, while the interaction between TGAP and
capitalization is negative and significant. Thus, for sufficiently capitalized banks the effects of
28
5.3 Robustness test: evidence from balanced panel data
In this paper, I adopt an unbalanced panel dataset for empirical analysis, implying that there is missing data for several banks over the sample period. As a final robustness test, it might be interesting to analyze whether the results obtained using the dynamic panel data method hold, when I use a balanced dataset instead. A balanced panel in this context implies that there is no missing data for any bank. I re-estimate Eq. (2) using the dynamic panel data method for the balanced panel dataset. To save space, I only report the most relevant findings in table 5. For the complete set of results, see Table A4 in the Appendix. The findings are similar to those obtained when estimating Eq. (2) using an unbalanced panel dataset. The sign and statistical significance of the coefficients of the monetary policy stance remains the same, except for the bank-level rate which becomes statistically insignificant for the non-performing loans equation. However, the impact of control variables slightly changes. The effect of capitalization on bank risk-taking becomes statistically insignificant when bank-level lending rates are used to proxy monetary policy (see columns IV and VI). In contrast, the impact of lagged profitability on a bank’s ratio of risk assets to total assets turns positive. This implies that higher profits earned in the previous year lead to higher bank risk-taking this year, which can be explained if one considers that part of these higher profits might be used to extend new loans. Overall, I conclude that while adopting a balanced panel dataset leads to slight changes in the results, no loss of information occurs with regards to the analysis of the relationship between monetary conditions and bank risk-taking. Hence, the main findings are robust to the use of balanced or unbalanced data.
6. Discussion
29 crisis unfolded; it reached its peak in 2010. Yet, it is highly unlikely that after the crisis, bank taking was still increasing. Therefore, this measure reflects past rather than current bank risk-taking behavior. Additionally, I argue that for the sample period considered, this measure captures bank risk exposure rather than bank risk-taking, as it is heavily influenced by economic conditions, beyond the control of banks. Hence, I include this variable solely because I am following the choices of Delis and Kouretas (2011) with regards to bank risk-taking measures.
Secondly, for the empirical analysis of this paper I use a GMM model for dynamic panel data, put forward by Arellano and Bover (1995) and Blundell and Bond (1998). One of the virtues of this model is that it accounts for endogeneity of monetary policy and bank characteristics by using lagged levels of these variables as instruments. However, the Sargan test, which examines the validity of these instruments, is rejected for effectively all the regressions estimated using the GMM model. This suggests that the instruments are invalid. However, a major limitation of this test is that it is built on the assumption of a homoskedastic error term. In contrast, heteroskedasticity is present in my data, which I control for. A heteroskedastic error term causes the Sargan test to produce inconsistent results (Roodman, 2009). In fact, Arellano and Bond (1991) show that this test over-rejects in the presence of heteroskedasticity. Additionally, Roodman (2009) argues that the Sargan test should not be relied upon too faithfully, as it is prone to weakness, with regards to the consistency of its results. The test is in fact weaker the more instruments there are present in the model.
30 Table 4. Monetary policy and bank risk-taking: distributional effects of monetary policy
due to bank characteristics
The table reports coefficients and standard errors (in parentheses). In regressions I-IV dependent variable is the ratio of risk assets to total assets and in regressions V-VI the ratio of non-performing loans to total loans. The explanatory variables are as follows: TGAP is the Taylor rule gap, bank-level lending rate is the ratio of interest income to total customer loans, post-crisis reflects a dummy variable which equals 1 after 2008 and 0 otherwise, monetary policy stance*post-crisis reflects the interaction term between monetary policy and the dummy variable post-crisis, capitalization is the ratio of equity capital to total assets, lagged profitability is the ratio of profits before taxes to total assets in year t-1, and size is the natural logarithm of real total assets. Furthermore, the coefficients of the interactions between monetary policy and capitalization, size, and lagged profitability are reported. Obs is the number of observations, the Wald-test and its associated p-value denote the goodness of fit of the regressions and Sargan is the p-value of the Sargan test of overidentifying restrictions. AR2 and AR3 denote the p-values for the test for second and third order autocorrelation.
* Statistical significance at the 1% level. ** Statistical significance at the 5% level. *** Statistical significance at the 10% level.
I II III IV
TGAP -0.551*
(0.066)
0.273* (0.022)
Bank-level lending rate -0.001
(0.002)
-0.000 (0.001)
Monetary policy stance*post-crisis -0.283**
(0.117) -0.001 (0.003) -0.142* (0.044) -0.001 (0.001) TGAP*Capitalization 7.854* (2.241) -0.896*** (0.499) TGAP*Size -0.149* (0.023) 0.091* (0.007) TGAP*Lagged profitability -4.792 (4.782) -10.102* (1.871)
Bank-level lending rate*Capitalization 0.013
(0.042)
0.011 (0.010)
Bank-level lending rate*size 0.001
(0.003)
0.001 (0.001) Bank-level lending rate*Lagged
31 Table 5. Monetary policy and bank risk-taking: balanced panel dataset
I II III IV V VI
Lagged risk assets 0.839*
(0.023) 0.802* (0.020) 0.815* (0.024) 0.814* (0.019) Lagged non-performing loans 0.682* (0.015) 0.692* (0.014) TGAP -0.534* (0.055) 0.158* (0.013) ΔFED assets 0.126* (0.014) Long-term rate -1.348* (0.116) Bank-level lending rate -0.004* (0.002) -0.000 (0.000) Monetary policy stance*post-crisis -0.117 (0.098) 0.080* (0.020) 1.093* (0.145) -0.008 (0.007) 0.042*** (0.022) 0.000 (0.001) Capitalization 0.199** (0.089) 0.317* (0.090) 0.217** (0.087) 0.084 (0.106) 0.008 (0.016) 0.017 (0.017) Lagged profitability 0.456* (0.149) 0.679* (0.134) 0.693* (0.138) 0.512* (0.155) -0.220* (0.051) -0.220* (0.049) Size 0.013* (0.002) 0.016* (0.002) 0.014* (0.002) 0.015* (0.002) 0.002* (0.000) 0.002* (0.000) Economic growth 0.037*** (0.021) 0.390* (0.031) 0.303* (0.025) 0.247* (0.024) -0.090* (0.007) -0.159* (0.005) Obs 53,757 53,757 53,757 53,757 53,757 53,757 Wald-test 15,379.4 13,521.13 14,676.47 16,914.18 9,968.80 9,935.75 p-value 0.000 0.000 0.000 0.000 0.000 0.000 AR2 0.219 0.332 AR3 0.522 0.505 0.297 0.815 Sargan 0.000 0.000 0.000 0.000 0.000 0.000
The table reports coefficients and standard errors (in parentheses). In regressions I-IV dependent variable is the ratio of risk assets to total assets and in regressions V-VI the ratio of non-performing loans to total loans. The explanatory variables are as follows: TGAP is the Taylor rule gap, ΔFED assets is the absolute annual change in central bank asset as a % of nominal GDP, long-term rate is the annual average of the 10-year U.S. government bond yield, bank-level lending rate is the ratio of interest income to total customer loans, post-crisis reflects a dummy variable which equals 1 after 2008 and 0 otherwise, monetary policy stance*post-crisis reflects the interaction term between monetary policy and the dummy variable post-crisis, capitalization is the ratio of equity capital to total assets, lagged profitability is the ratio of profits before taxes to total assets in year t-1, size is the natural logarithm of real total assets, and economic growth is real GDP growth. Obs is the number of observations, the Wald-test and its associated p-value denote the goodness of fit of the regressions and Sargan is the p-value of the Sargan Wald-test of overidentifying restrictions. AR2 and AR3 denote the p-values for the test for second and third order autocorrelation.
32 Figure 3. Median non-performing loans to total loans ratio for the sample of U.S. banks
The figure illustrates the median ratio of non-performing loans to total loans of the sample of U.S. banks for the period 2004-2016.
Lastly, my sample only contains banks who survived until the end of the sample period and overlooks those that did not. Consequently, there is survivorship bias present in my data. However, I argue that the results are therefore more striking, since the majority of the banks that did not survive were likely the ones who were engaging in excessive risk-taking in the period leading up to the crisis. This ultimately led to many loan defaults, causing these banks to collapse. Indeed, there is empirical evidence supporting this claim (e.g., Wang & Cox, 2013). This implies that the sample effectively underreports the amount of bank risk-taking during this period. Thus, I contend that the concluding remarks in this paper regarding the existence of a risk-taking channel in the U.S. after the 2008 financial crisis still hold, as it is not significantly affected by this bias.
7. Conclusion
33 current expansionary monetary policy interventions in the U.S. are setting the stage for the next financial crisis (e.g., Krishnamurthy & Vissing-Jorgensen, 2011; Farhi & Tirole, 2012; Chodorow-Reich, 2014). As this has not been studied extensively yet for U.S. banks, this paper aims to investigate this gap in the literature.
Using a panel of 6851 U.S. banks, I find that expansionary monetary policy positively influences risk-taking, measured by the ratio of risk assets to total assets held by banks. This implies that the prolonged period of accommodative monetary policy interventions after the crisis has led to greater bank risk-taking in the U.S. Thus, the findings are in line with the existing empirical evidence for the risk-taking channel in the U.S. pre-crisis period (e.g., De Nicolò et al., 2010; Maddaloni & Peydró, 2011; Dell'Ariccia et al., 2017; Altunbas et al., 2014). A potential explanation for these findings is that through the search for yield effect (Rajan, 2006), banks engage in higher risk-taking when interest rates are low to increase their yields, to satisfy the demands of investors and shareholders. Additionally, I find that the effects of accommodative monetary policy on bank risk-taking is weaker (however still significant) for banks with higher capitalization and stronger for larger banks. The former supports the theory set forward by Dell’Ariccia et al. (2014), who argue that the effects of monetary policy on bank risk-taking depends on the capital structure of banks. These results hold for several specifications of the monetary policy stance; however, they are not robust to alternative measures of bank risk-taking. Ultimately, the results suggest that the Fed should carefully consider how accommodative monetary policy is causing greater risk-taking in the banking sector, similar to the pre-crisis period. As Lamers et al. (2016) put it, if accommodative monetary policy is indeed impairing financial stability, this implies a tradeoff for the Fed between stimulating economic growth and preventing vulnerabilities from developing within the financial system.
34 Appendix
Table A1. Monetary policy and bank risk-taking: fixed effects regressions.
I II III IV V VI TGAP 0.588* (0.058) 0.307* (0.016) ΔFED assets 0.051* (0.018) Long-term rate -0.846* (0.126)
Bank-level lending rate -0.000*
(0.000) 0.002*** (0.001) Post-crisis -0.050* (0.002) -0.058* (0.002) -0.147* (0.005) -0.058* (0.002) 0.018* (0.001) 0.017* (0.000)
35 I II III IV V VI Economic growth -0.105* (0.019) -0.302* (0.025) -0.084* (0.019) -0.332* (0.019) -0.136* (0.009) -0.216* (0.007) Importance of banks -0.023* (0.004) -0.014* (0.005) -0.007 (0.004) -0.026* (0.003) -0.022* (0.001) -0.018* (0.001) Obs 65,029 65,029 65,029 64,990 65,016 64,990 Wald-test 329.57 334.04 335.73 321.97 269.25 2,564.33 p-value 0.000 0.000 0.000 0.000 0.000 0.000 Within- R2 0.13 0.13 0.14 0.13 0.17 0.21 Sargan 0.000 0.000 0.000 0.000 0.000 0.000
The table reports coefficients and standard errors (in parentheses). In regressions I-IV dependent variable is the ratio of risk assets to total assets and in regressions V-VI the ratio of non-performing loans to total loans. The explanatory variables are as follows: TGAP is the Taylor rule gap, ΔFED assets is the absolute annual change in central bank asset as a % of nominal GDP, long-term rate is the annual average of the 10-year U.S. government bond yield, bank-level lending rate is the ratio of interest income to total customer loans, post-crisis reflects a dummy variable which equals 1 after 2008 and 0 otherwise, monetary policy stance*post-crisis reflects the interaction term between monetary policy and the dummy variable post-crisis, capitalization is the ratio of equity capital to total assets, lagged profitability is the ratio of profits before taxes to total assets in year t-1, size is the natural logarithm of real total assets, efficiency is the ratio of operating revenues to operating expenses, off-balance sheet items is the ratio of off-balance sheet items to total assets, inflation is the annual average CPI inflation, economic growth is real GDP growth and importance of banks is the ratio of domestic credit provided by the banking sector to nominal GDP. Obs is the number of observations, the within-R2,the Wald-test and its associated p-value denote the goodness of fit of the regressions and Sargan is the p-value of the Sargan-Hansen test for the test of fixed vs random effects model.
* Statistical significance at the 1% level. ** Statistical significance at the 5% level. *** Statistical significance at the 10% level.
Table A2. Monetary policy and bank risk-taking: dynamic panel regressions
I II III IV V VI
Lagged risk assets 0.889*
(0.015) 0.832* (0.013) 0.853* (0.020) 0.824* (0.016)
Lagged non-performing loans 0.568*
(0.040) 0.797* (0.055) TGAP -0.596* (0.056) 0.112* (0.023) ΔFED assets 0.108* (0.015) Long-term rate -1.279* (0.133)
Bank-level lending rate -0.000
(0.000)
0.001*
36 I II III IV V VI Post-crisis -0.053* (0.001) -0.051* (0.002) -0.072* (0.002) -0.047* (0.002) 0.002** (0.001) -0.001 (0.001)
Monetary policy stance*post-crisis -0.251** (0.099) 0.097* (0.022) 0.423* (0.161) 0.000** (0.000) 0.056 (0.034) -0.000** (0.000) Capitalization 0.246* (0.083) 0.634* (0.103) 0.509* (0.097) 0.312* (0.099) -0.017 (0.025) -0.076* (0.025) Lagged profitability -0.368* (0.098) -0.781* (0.104) -1.024* (0.136) -0.304* (0.095) -0.614* (0.134) -0.120*** (0.067) Size 0.011* (0.002) 0.019* (0.002) 0.009* (0.002) 0.018* (0.002) 0.006* (0.001) 0.002* (0.001) Efficiency -0.002 (0.002) -0.002 (0.003) -0.001 (0.003) -0.003 (0.002) 0.001 (0.001) 0.000 (0.000) Off-balance sheet -0.000 (0.000) -0.000*** (0.000) 0.000 (0.000) -0.000*** (0.000) 0.000 (0.000) -0.000 (0.000) Inflation -1.088* (0.052) -1.123* (0.038) -0.880* (0.034) -0.736* (0.030) 0.115* (0.016) -0.002 (0.018) Economic growth 0.039*** (0.021) 0.407* (0.028) 0.297* (0.022) 0.259* (0.023) -0.128* (0.009) -0.200* (0.006) Importance of banks 0.046* (0.004) 0.019* (0.005) 0.066* (0.004) 0.046* (0.004) -0.011* (0.002) -0.016* (0.001) Obs 65,029 65029 65029 64990 64995 64969 Wald-test 24,705.22 19,830.80 17,760.53 21,665.79 8,184.53 28,2109.20 p-value 0.000 0.000 0.000 0.000 0.000 0.000 AR1 0.000 0.000 0.000 0.000 0.000 0.000 AR2 0.256 AR3 0.168 0.233 0.183 AR4 0.466 0.565 Sargan 0.000 0.000 0.000 0.000 0.000 0.000
The table reports coefficients and standard errors (in parentheses). In regressions I-IV dependent variable is the ratio of risk assets to total assets and in regressions V-VI the ratio of non-performing loans to total loans. The explanatory variables are as follows: TGAP is the Taylor rule gap, ΔFED assets is the absolute annual change in central bank asset as a % of nominal GDP, long-term rate is the annual average of the 10-year U.S. government bond yield, bank-level lending rate is the ratio of interest income to total customer loans, post-crisis reflects a dummy variable which equals 1 after 2008 and 0 otherwise, monetary policy stance*post-crisis reflects the interaction term between monetary policy and the dummy variable post-crisis, capitalization is the ratio of equity capital to total assets, lagged profitability is the ratio of profits before taxes to total assets in year t-1, size is the natural logarithm of real total assets, efficiency is the ratio of operating revenues to operating expenses, off-balance sheet items is the ratio of off-balance sheet items to total assets, inflation is the annual average CPI inflation, economic growth is real GDP growth and importance of banks is the ratio of domestic credit provided by the banking sector to nominal GDP. Obs is the number of observations, the Wald-test and its associated p-value denote the goodness of fit of the regressions and Sargan is the p-value of the Sargan test of overidentifying restrictions. AR1, AR2 AR3 and AR4 denote the p-values for the test for first, second, third and fourth order autocorrelation.
