Identifying the Risk-Taking Channel of Monetary Transmission and the Connection to Credit Spreads Over the Business Cycle

Transcription

1 Identifying the Risk-Taking Channel of Monetary Transmission and the Connection to Credit Spreads Over the Business Cycle Nimrod Segev February 12, 2017 Abstract I use loan-level data from the syndicated loan market in the U.S to investigate how monetary policy aects banks' sensitivity to risk. Using loan-level data and banks' sensitivity to risk enables me to identify the risk-taking channel and disentangle it from other monetary channels. I show that banks change their behavior toward risk following changes in monetary policy where loose monetary policy reduces banks' sensitivity to risk. Banks' sensitivity to risk is also negatively associated with real economic activity, suggesting a link between banks' risktaking behavior and the real economy. I then show that banks' sensitivity to risk has a real and predicted eect on credit spreads, specically on the Gilchrist and Zakraj²ek (2012) "excess bond premium", a leading indicator of real business cycles uctuations. The paper's primary contribution is in providing new loan level evidence for the existence of the risk-taking channel in the U.S, as well as a rst possible link between the risk-taking channel, credit spreads, and business cycle uctuations. Keywords: Monetary transmission, Risk-taking channel, Loan-level data, Credit spreads JEL Classication code: E44, E51, E52, G21 Fordham University, nsegev@fordham.edu 1

2 1 Introduction A recent line of research suggests that there is a risk-taking channel of monetary policy (Borio and Zhu 2012). According to this view, a reduction in the monetary policy rate causes nancial institutions to change their perception and tolerance of risk resulting in a lower risk premium, which in turn amplies the eect of the interest rate cut. The idea that banks take on excessive risk during periods of low policy rates has been getting increasing attention following the recent nancial crisis and the long period of extremely low rates that followed in the US and Europe. A number of economists have argued that the long period of relatively low policy rates and expansionary monetary policy was one of the main sources for the nancial system's fragile state prior to the great recession of The main idea is that persistent low levels of real interest rates induced nancial intermediaries to take excessive risk, which led to a fragile and unstable nancial system and had real negative eects on the economy (De Nicolò et al. 2010). There are at least three ways in which the risk-taking channel can operate: First, lower interest rates can aect cash ow, income, and collateral value, which can change banks' estimation of expected risk. For example, low interest rates can boost banks' asset value and reduce their volatility. This, in turn, can aect banks' estimates of probabilities of default and loss given default which could encourage banks' to take a riskier position and increase credit supply to riskier borrowers. This set is closely related to the credit channel of monetary policy, also known as the "nancial accelerator" (Bernanke, Gertler, and Gilchrist 1999). The main dierence between the two channels is that the risk-taking channel emphasizes banks' appetite to risk and willingness to supply credit, given borrowers risk, while the credit channel emphasizes changes in borrowers creditworthiness and the demand for credit. That is, the risktaking channel focuse on how interest rates aect the quality and not just the quantity of banks' credit. 1 Second, Assuming nancial intermediaries rate of return targets are "sticky", lower rates may cause a "reaching for yields" strategy. That is, reduction in interest rates can cause banks' and fund managers' to increase risky investments and supply riskier loans in order to meet expected high rates of return. Finally, monetary policy which increases economic activity may create an over-optimistic environment, which may then increase the aggregate level of risk-taking. During periods of good economic conditions, nancial intermediaries may have 1 See Borio and Zhu (2012) for more on treatment of risk in the nancial accelerator framework and Adrian and Shin (2010b) for an excellent discussion on the point of contact and dierence between the risk-taking channel and other channels of monetary policy. 2

3 unrealistically high positive expectations about future prots causing careless, and excessive risk taking. 2 Practically, the risk-taking channel can operate through both the nancial intermediary's assets and liabilities. On the liability side, a fall in policy rate could induce intermediaries to increase short-term liabilities, thus increasing their risk exposure. On the asset side, the search for higher yields or the change in the tolerance and perception of risk may cause asset managers to shift the composition of their investment towards riskier assets. Banks' overall risk level increases only if the riskiness of the assets and liability move in the same direction. That is, a "complete" risk-taking channel assumes that during periods of expansionary monetary policy, banks' take on more short-term debt on the liability side and supply credit to riskier borrowers on the asset side. This ensures that changes in the risk level on one side of the balance sheet are not oset by the reduced level of risk on the other side. This paper focuses on nding a risk-taking channel on the asset side of nancial intermediaries by focusing on banks' sensitivity to borrowers' levels of risk. Focusing on the riskiness of new assets builds on the work of Adrian and Shin (2010a; 2010b; 2014) who used data from banks in the U.S. to show that nancial intermediaries actively manage their leverage and balance sheet size according to economic conditions. They show that banks' leverage ratio is pro-cyclical and banks' generally expand their balance sheet size by taking on more short-term debt. Given that banks' risk exposure on the liability side is pro-cyclical, banks' overall exposure to risk increases only if the risk level on the asset side is also pro-cyclical. If the favorable economic and monetary conditions make banks' asset side safer, the increase in leverage does not necessarily imply an increase in banks' riskiness behavior. Since banks' tend to increase their short term debt when expending their balance sheet, Adrian and Shin (2010a) point to what could potentially be a good indicator of a risk-taking channel: "when balance sheets are expanding fast enough, even borrowers that do not have the means to repay are granted creditso intense is the urge to employ surplus capital " (p.436). Thus, this research complements their research by focusing on identifying banks' risk taking on the asset side. There are a number of challenges when trying to identify the eect of monetary policy on nancial intermediaries' level of risk exposure and tolerance of risk. First, the monetary policy rate may be endogenous to the level of risk and/or the stability of the nancial system. Second, 2 Others have suggested that low interest rates may also lead to more risk-taking by negativlay aecting banks' monitoring incentives (Dell'Ariccia, Laeven, and Marquez 2014) and by central banks' communication policies which could lead to moral hazard problems (Gambacorta 2009; Altunbasa, Gambacortab, and Marques-Ibanezc 2014) 3

4 banks' actions and risk level could be aected by changes in economic conditions as well as changes in the demand for, and supply of loans. Thus, one must disentangle the eects of monetary policy which work through the more traditional credit and lending channels from the direct eect of monetary policy on banks' risk tolerance and perception. The rst contribution of this paper is using loan-level data to apply a dierent identication strategy from the strategies that have been used so far in the literature when trying to nd empirical evidence to the existence of the risk taking channel. By combining a number of dierent data sources, I am able to obtain information on 7,575 loan deals from the syndicated loan market in the U.S from 1995 to I then use a two-step identication strategy similar to Kashyap and Stein (2000), Ashcraft and Campello (2007), and Aysun and Hepp (2013). In the rst step I estimate the sensitivity of banks' to borrowers' risk, in the second step I check the impact of monetary policy on that sensitivity. Using loan-level data with lender and borrowerspecic information along with this identication strategy allows me to disentangle the risk channel from other monetary channels while also mitigating the possibility that the results are driven by some unobservable macro shock. Using lender and borrower specic control variables in the rst stage regression (rst step) allows me to control for any balance sheet eect, liquidity eect and identify the independent eect of borrowers risk on the loan spreads. Using controls for the economic conditions, time trend, and seasonal eects in the second stage regression (second step), I can measure how monetary policy aects banks sensitivity to risk, providing a unique way to measure the risk-taking channel. To clarify the contribution, this paper is not the rst to use loan-level data from the syndicated loan market to identify the risk-taking channel. 3 However, to the best of my knowledge, this is the rst paper that uses the above two step identication approach, as well as banks' sensitivity to risk, in order to fully identify the risk-taking channel. Changes in banks' sensitivity to risk also provides an ex-ante measure of banks' asset side risk-taking which is based on banks' actual behavior. 4 The second contribution of this paper is in providing evidence to the connection between the risk-taking channel, credit spreads, and business cycle uctuations. While there is signicant empirical support for the fact that loose monetary policy increases bank risk-taking (De Nicolò et al. 2010; Gambacorta and Marques-Ibanez 2011; Buch, Eickmeier, and Prieto 2010; Delis and Kouretas 2011; Delis, Hasan, and Mylonidis 2012; Altunbasa, Gambacortab, and Marques- Ibanezc 2014; Jiménez et al. 2014; Ioannidou, Ongena, and Peydró 2015; Dell'Ariccia, Laeven, 3 See for example Delis, Hasan, and Mylonidis (2012) and Paligorova and Santos (2016). 4 Most of the existing literature use surveys or ex-post measures of banks' risk-taking. See section 2 for a review 4

5 and Suarez 2016; Paligorova and Santos 2016), the literature is silent with regards to the real eects of the risk-taking channel on the economy or on nancial stability. Most papers simply assume that the increase in banks' risk during periods of loose monetary policy contributes to nancial instability, economic distress, and other negative macroeconomic outcomes. However, since there are other channels through which monetary policy aects banks' stability, the increase in bank risk-taking does not necessarily imply "excessive" risk-taking that may lead to nancial instability and have real economic eects. This paper bridges this gap in the literature by providing evidence for the connection between banks' risk taking and real economic activity. Additionally, I suggest a possible mechanism where changes in banks' risk taking can cause or at least amplify business cycle uctuations. Specically, I show that banks' tolerance and risk behavior aects market capital spreads in a signicant and predicted way. The fact that market credit spreads are considered leading indicators for economic activity provides a possible link between banks' risk sensitivity and the real economy. 5 The main motivation for checking the eect of banks' sensitivity to risk on capital markets spreads comes from Barnea and Menashe (2015), who suggest that banks' strategies inuence capital market pricing due to asymmetric information between large banks and other smaller capital market lenders. Assuming nancial frictions in the form of asymmetric information, banks' sensitivity to risk provides signals to other less-informed capital market lenders, which in turn aects the pricing of loans in the capital market. I show that banks' sensitivity to risk has a real eect on the "excess bond premium" (EBP). EBP was developed rst by Gilchrist and Zakraj²ek (2012) as the component of the credit spread which is not related to the probability of default, and thus represents the cyclical changes in the connection between the default risk and the credit spread. Changes in banks' risk sensitivity can then signal to other credit market lenders to change their loan pricing for all borrowers, even those whose actual probability of default did not change. This change in capital markets pricing will show in the unexplained component of the credit spread, the EBP. The link between the risk-taking channel and the real economy then works as follows: Monetary policy aects banks' sensitivity and tolerance to risk, banks' sensitivity to risk then has a signicant eect on market credit spreads and specically on the EBP, changes in the excess bond premium translate to changes in the supply of credit, which has a real eect on 5 There is growing empirical and theoretical literature showing that credit spreads are extremely reliable leading indicators for future economic activity. See Gilchrist and Zakraj²ek (2012) and Bleaney, Mizen, and Veleanu (2016) for a review of the literature. 5

6 nancial conditions and macroeconomic outcomes. To the best of my knowledge, this paper is the rst to provide empirical evidence for a possible causal link between the risk-taking channel and the real economy. The rest of the paper is organized as follows: In section 2 I review the relevant literature. In section 3 and 4 I describe the empirical model and the data. Section 5 presents the main results. In section 6 I present a sensitivity analysis and Section 7 concludes. 2 Related Literature 2.1 Identication challenges and existing evidence Identifying the risk-taking channel is no easy task since it is so closely related to the credit channel of monetary policy. The main problem is to distinguish between changes in bank lending that are caused by the classic credit channel and changes that are caused by a risktaking channel. Ideally one could use a simple OLS model such as equation (1) to identify the eects of monetary policy on banks' risk. j Risk i,t = β 0,t + γ n,t MP t n + Controls i,t + e i,t (1) n=1 where bank i's level of overall risk at time t is a function of some lagged values of monetary policy (MP), and controlled for banks' xed eect, time eect, economic conditions, and any other bank specic variables. A number of challenges arise when trying to estimate equation (1). Assuming one has a reliable measure of banks' overall risk level (a strong assumption on its own), one must use control variables that are able to eliminate any eect of other monetary channels on banks' risk. That is, one must use the appropriate control variables to disentangle the eects of monetary policy on banks' risk level (which work by aecting loan supply and loan demand) from the eects on banks' risk, caused by actual changes in banks' perception and tolerance of risk. The second identication challenge is the endogenous problem of risk and monetary policy; in order for equation (1) to hold, one must assume that monetary policy is strictly exogenous to banks' risk, regardless of how it is measured. In the existing literature, a common measure of banks' risk is changes in lending standards obtained from lending surveys such as the Federal Reserve's Senior Loan Ocer Opinion Survey (SLOOS) (De Nicolò et al. 2010). In these surveys, the loosening of lending standards is 6

7 seen as an indication of an increase in banks' risk-tolerance, indicating a risk-taking channel. Dell'Ariccia, Laeven, and Suarez (2016) discuss a number of problems with using this measure. First, lending surveys only indicate the relative change in lending standards and not the absolute level. Additionally, the changes in lending standars may reect changes in the borrower's quantity and quality and banks' willingness to supply riskier lowns, which makes it impossible to directly measure the risk-taking channel. Dell'Ariccia, Laeven, and Suarez (2016) used a dierent survey, the Federal Reserve's Survey of Terms of Business Lending (STBL), which includes condential data on the internal rating of U.S. banks on loans to businesses. They show that ex-ante risk-taking by banks, measured by the risk rating of banks' new loans, is negatively related to increasing short-term rates. Other papers use rating measures of risk such as the Expected Default Frequency (EDF) (Gambacorta and Marques-Ibanez 2011; Altunbasa, Gambacortab, and Marques-Ibanezc 2014) or the ratio of risk assets to total assets (Delis and Kouretas 2011; Delis, Hasan, and Mylonidis 2012). Generally, these papers try to address the problem of distinguishing the risk-taking channel from other eects by using alternative micro-level control variables. Regarding the endogeneity problem, the solutions include using GMM (Delis and Kouretas 2011; Altunbasa, Gambacortab, and Marques-Ibanezc 2014), exogenous measures of monetary policy (Delis, Hasan, and Mylonidis 2012), and a sample split of the data (Dell'Ariccia, Laeven, and Suarez 2016). Most papers also argue that it is reasonable to assume that before the nancial crisis, price stability was considered a sucient target for monetary policy, so monetary policy was exogenous to nancial conditions before the crisis (Altunbasa, Gambacortab, and Marques- Ibanezc 2014; Dell'Ariccia, Laeven, and Suarez 2016). Two other papers use credit registries and measure borrowers' risk based on past default history and ex-post loan default rates. Jiménez et al. (2014) use a unique Spanish data set that includes information on credit applications, the rejection or granting decision of the loan, and other credit outcomes such as the probability of default and loan amount. Using this unique data set, they nd that low rates increase the probability that lowly-capitalized banks will grant a loan to ex-ante riskier borrowers. Ioannidou, Ongena, and Peydró (2015) use a unique Bolivian data set and nd similar results in Bolivia. The main problem with using the same identication strategy in the U.S. or other parts of the world is that, to the best of my knowledge, such data sets often do not exist or are not publicly accessible. In addition to the above micro-level studies, a number of papers use aggregate time series evidence to check for the eect of monetary policy on banks' risk. Angeloni, Faia, and Duca 7

8 (2015) use VAR evidence to show the eect of monetary policy on bank funding, lending and overall banking risk, measured as the volatility in bank stock prices. They use quarterly U.S. data to show that lowering policy rates raise bank riskiness, particularly on the funding side. Buch, Eickmeier, and Prieto (2010) estimate an FAVAR on a large data set of banking sector variables and nd that an expansionary monetary shock induces small domestic banks to increase the fraction of risky loans while charging a lower risk premium. Bruno and Shin (2015) nd VAR evidence that contractionary monetary policy in the U.S has a cross-border eect on the risk taking of international banks. The paper most closely related to this is Paligorova and Santos (2016), who use evidence from the syndicated loan market to identify the risk-taking channel. They use a time series regression using the spread on newly-issued loans as the dependent variable, and measures of borrowers' risk and monetary policy as the independent variables. Their main variable of interest was the interaction term between borrowers' risk measure and monetary policy stance. This paper uses similar data but diers in the identication and main variable of interest. This paper is complementary to the above literature by providing new micro-level evidence with a dierent identication strategy that is able to mitigate the identication challenges highlighted above. By using banks' sensitivity to risk instead of direct measures of banks' overall risk I am able to better identify direct changes in bank perception and tolerance of risk. Additionally, the two-stage estimation approach explained in detail in the next section reduces the possible eect of endogeneity and unobserved variables. 2.2 Other related literature This paper is also closely related to the literature on the syndicated loan market. The main idea behind a syndicated loan is that it allows dierent nancial intermediaries to share the credit risk of granting a loan. In general, a syndicated loan allows banks to meet borrowers' demand for loans without having to bear all the market and credit risk by themselves. 6 The importance of the syndicated loan as a source of corporate nancing has increased signicantly since the early 90s. In the U.S. syndicated loans grew from $150 billion in 1987 to $1.7 trillion in 2006, making them the main source of corporate funding in the U.S (Ivashina 2009). A number of papers have studied the pricing and composition of a syndicated loan: 6 See Dennis and Mullineaux (2000) and Gadanecz (2004) for an in-depth review of the structure and development of a syndicated loan 8

9 Syndicated loans are smaller and more concentrated when there is little information about the borrower and when credit risk is high (Lee and Mullineaux 2004). The arranger share is smaller when borrowers are less risky and public information is available (Dennis and Mullineaux 2000; Lee and Mullineaux 2004), and spreads on a syndicated loan are smaller when there are more participants (Angbazo, Mei, and Saunders 1998; Focarelli, Pozzolo, and Casolaro 2008). 3 Empirical Model 3.1 Identication In order to minimize the bias in estimating equation (1) and test if banks are actively changing their perception and tolerance for risk, I build an empirical model which is closely related to the identication strategy used by Ashcraft and Campello (2007) and Aysun and Hepp (2013). The main idea is to use a two-step approach to check changes in bank sensitivity to risk in order to identify monetary transmission channels. To investigate the link between monetary policy and bank lending behavior I use the pricing of newly issued loans (i.e., the loan spread charged) as the main measure of banks lending behavior. In the rst stage regression of the model, I isolate the eect of risk on loan spreads. The second stage regression then checks how changes in monetary policy aect banks' sensitivity to borrowers' riskiness. The empirical model consists of two main parts. First, I use cross-section data to estimate how the riskiness of a loan is related to the loan spread. quarter from 1995 to 2007: This estimation is done for each LS i,j,k,t = β 0t + β t RISK i,t + ν t X i,t 1 + χ t Y j,t 1 + λ t Z k,t + e i,j,k,t (2) where subscripts i, j, k represent rm i, bank j, and loan k at quarter t. LS is the lending spread, RISK is a measure of the loan risk level. X, Y, and Z denote vectors of rm-specic, lender-specic and loan-specic control variables, respectively. The borrower control variables in equation (2) include borrower liquidity, and size. Lender-specic variables include liquidity, leverage, size, and lender return on assets ratio. The loan control variables include the maturity, loan amount, and the number of lenders participating in the syndicate. Additionally, I add a number of dummy variables which include a primary purpose dummy, term loan dummy, and a relationship dummy (see section 4 for a specication and description of all variables). In equation (2) the key coecient of interest is β t. It shows how the riskiness of the loan 9

10 relates to the lending spread. Signicant levels of β t indicate that changes in borrowers' risk have a signicant eect on loan spreads. That is, banks react to risk when issuing loans. In the second stage estimation, I estimate the impact of time-varying factors (U.S. output gap and monetary policy) on banks' sensitivity to risk. To do this I put the β t coecients that I estimate for each quarter in the rst stage estimation into a time series vector: Ψ RISK t. I then estimate the following time series regressions: Ψ RISK t = α + 4 γ k mp t k + k=1 4 ψ k y t k + k=1 3 θ k Q k + τt t + v t (3) where mp is the stance of monetary policy, y is the output gap, Q is quarter dummies, and T is the time trend. I use lagged values in order to account for possible persistent eects of monetary policy and economic conditions. 7 k=1 The risk-taking channel is captured by the coecient γ = 4 k=1 γ k. If γ is signicant, this implies that changes in the monetary policy stance are related to changes in banks' sensitivity to risk. That is, monetary policy induces banks to change their sensitivity and behavior towards borrowers' risk, which will indicate a risk-taking channel on the asset side of nancial intermediaries. There are three advantages of using this identication approach: First, using lender and borrower specic variables in the rst stage regression allows me to control for any eect of other monetary channels such as the lending and balance sheet channels. The lender-specic variables control for any eect of banks' ability to supply loans (the lending channel) while the borrower-specic variables control for borrowers balance sheet eects (the balance sheet channel). Thus, I am able to isolate the independent eect of borrows' risk. Second, using the cross-sectional sensitivity to risk in each quarter reduces the possibility of biased results due to time-varying unobserved variables. Since the rst stage regression measures the cross sectional sensitivity for risk for each quarter, the analysis accounts for any changes in the macro conditions between each quarter. The two-step approach then mitigates the possibility that the results are driven by unobserved macro conditions. Finally, the model is able to capture the ex-ante change in banks' tolerance (sensitivity) toward risk providing a new and more direct measure of the risk-taking channel. 7 The general practice in the empirical literature using a two-step approach is to use four or eight lags of monetary policy in the second stage regression. I use four in order reduce the degrees of freedom lose. Using eight does not signicantly change the results. 10

11 It is important to note that using a generated variable as a dependent variable does not cause the same endogeneity problems as using it as a regressor. The only shortcoming with using a generated dependent variable is that measurement errors may give an additional error term which reduces the test power. However, the coecients, in this case, are still consistently estimated. 3.2 Banks sensitivity to risk and economic activity After identifying the risk-taking channel, the second contribution of this paper is to try to relate it to the real economy. In this section I will rst check if the risk-taking channel, operating by aecting banks' sensitivity to risk, is signicantly correlated with economic activity. To do this, I rst use a simple time series regression to check the connection between banks' sensitivity to risk and real economic activity. I then use a simple vector auto-regression (VAR) framework to investigate the relationship between risk sensitivity and macro-aggregates more formally. Letting Y t be a quarter indicator of economic activity I estimate the following equation: Y t = α + δβ RISK t + χ 1 Z 1 t 1 + χ 2 Z 2 t 1 + ρ Y t 1 + ɛ t (4) Where Y t is the annual growth rate of the economic indicator at quarter t. β RISK t is banks sensitivity to risk at every quarter obtained from the rst stage regression (equation (2)), Z 1 is a vector of economic controls and Z 2 is a vector of nancial controls. Z 1 includes lagged values of ination and the federal funds rate. Along with the lagged value of the economic indicator ( Y t 1 ), these variables control for any eects of past macroeconomic conditions. Z 2 includes lagged values of the "excess bond premium" EBP, equity market volatility, term spread and credit spreads. The EBP has been used in many recent papers as a measure of nancial risk and strains in the nancial markets. The EBP and the three other measures are used since they potentially obtain useful information in the nancial markets that may aect real economic activity. I use industrial production as the main measure of economic activity since we can expect that it will be signicantly eected by the credit availability. 8 Thus, to the extent that banks' sensitivity to risk has a signicant eect on the supply of credit, it may potentially aect industrial production and the real economy. I then consider the following four variable structural VAR: 8 I also check the robustness of the results by using other economic indicators such as the growth in real GDP and change in Unemployment. 11

12 AZ t = φz t 1 + ɛ t The endogenous variables of interest are collected in Z t = [IP t, CP I t, F F R t, RISK t ] and include the log of industrial production, log of the CPI index, the eective federal funds rate and my measure of risk sensitivity. I estimate the reduced-form VAR: Z t = BZ t 1 + u t where B denotes A 1 φ and u t denotes A 1 ɛ. To identify the system I use the standard Cholesky decomposition to impose six restrictions on matrix A, making it a lower-triangular. The identication assumption in the ordering of the variables is that monetary authority's react to prices and real economic conditions and that banks react to all three. The VAR is estimated with two lags, according to the Akaike criterion. 3.3 Banks sensitivity to risk and credit spreads In this section I will explore one specic mechanism through which changes in banks' sensitivity to risk can aect the real economy. Specically, I explore a possible causal relation where banks' sensitivity to risk amplify business cycle uctuation by aection market credit spreads. The main hypothesis is that banks' sensitivity to risk (captured by β from the rst stage identication model) has a real eect on market credit spreads. In order to check this hypothesis I use the Gilchrist and Zakraj²ek (2012) "excess bond premium" (EBP), which is a leading indicator of real business cycle uctuations. Gilchrist and Zakraj²ek (2012) (GZ) show that an unanticipated increase of 100 basis points in the EBP cause a signicant reduction in real GDP for a number of quarters. If banks' sensitivity to risk has a real eect on the EBP that could potentially explain the connection between changes in banks' behavior and real economic activity. My main assumption is asymmetric information between large banks and other capital market lenders. Given the asymmetric information, changes in large banks' sensitivity to risk may send signals to other capital market lenders to adjust their risk premium, even if borrowers' quality did not change. This may eect the unexplained component of the market credit spreads - the EBP. Since changes in the EBP are good indicators for changes in investments and economic activity, this channel suggests a causal link between banks' sensitivity to risk, the pricing credit, and economic activity. These assumptions diverge from the classical Modigliani and Miller (1958) view where all sources of funding are the same, and adopt the Bernanke and Blinder (1992) "credit channel" 12

13 view, where banks' loans play a special role in the transmission of monetary policy. The main dierence between the "risk-taking channel" and the classical "credit channel" is that the risk channel aects the lender's willingness to take on risk. The risk-taking channel operates by changing banks' tolerance to risk, which aects the risk premium in the loan market as well as in the bonds market thus shifting the supply of credit in the economy. To empirically check the above hypothesis, I use the following empirical model: 3 3 EBP t = β 0,t + δ k βt k RISK + η t Z t 1 + θ k Q k + τt t + v t (5) k=1 where EBP is the excesses bond premium in the U.S at quarter t, β RISK t k k=1 are banks' risk sensitivity coecients from the rst stage regression, and Z is a vector of control variables. Z includes the VIX index, which is the quarterly Chicago Board Options Exchange (CBOE) volatility index obtained from the Federal Reserve Economic Data (FRED), ROA is the return on assets in the U.S. nancial corporate sector (in percent, annualized), and xmktret is a measure of the excess (value-weighted) market return (in percent, annualized). The control variables account for exogenous changes in risk perception in the economy and the eects of the nancial system conditions. Regression (5) is estimated by OLS with the lag of the risk sensitivity determined by the Akaike Information Criterion (AIC). There are two main reasons for using the EBP when checking the eects of banks' sensitivity to risk on market credit spreads. First, while the syndicated loan market is dominated by mostly large banks, the corporate bond market, which is used to measure the EBP is dominated by institutional investors, such as insurance companies, pension funds, hedge funds, and other highly-leveraged lenders. Thus, we can assume that there is asymmetric information between the main lenders in the syndicated loan market and the main lenders in the corporate bond market. Second, the EBP by construction is orthogonal to the default risk of borrowers, reducing the possibility of endogeneity in equation (5). To see this, we need to rst understand how the EBP is constructed. EBP in each period is dened by GZ as: where S GZ t of the credit spread. EBP t = S GZ t is the observed credit spread and ŜGZ t Ŝ GZ t ŜGZ t (6) is the predicted or explained component is dened as the average of the predicted level of spread for each bond at every time period. To calculate the predicted level of spread for each bond k, of rm i at every time period t, GZ use: Ŝ i,t,k = exp[ ˆβDD i,t + ˆγZ i,t,k + σ2 2 ] (7) 13

14 Where DD is rm's i Distance to Default at time t, ( ˆβ, ˆγ) are OLS estimated parameters and Z is a vector of bond-specic variables. Equations (6) and (7) show that the EBP is the unexplained component of the credit spread, which is not related to borrowers' creditworthiness. It is then reasonable to assume that any variable which aects banks' sensitivity to risk in the rst stage regression (equation (2)) will aect the explained component of the credit spread, the ŜGZ t, and not the EBP. Thus, when using banks sensitivity to risk from equation (2) as regressors on EBP in equation (5) there is no risk that some unobserved variable eecting β RISK and EBP, making β RISK correlated with the error term and biasing the result. Using lagged values of banks sensitivity to risk, and given that by construction equation (5) does not suer from endogeneity, suggests that equation (5) can test a possible causal link between banks risk sensitivity and EBP. In the sensitivity analysis, I use the Hausman test to formally check that there is no endogeneity problem in equation (5). Additionally, I check the robustness of my results using seven other risk sensitivity measures to ensure the results are not aected by borrowers' risk measurement errors. 4 Data In order to compute the data set, I combined a number of dierent data sources. My main source of loan information is from the Thomson Reuters DealScan database, which I obtained through the Wharton Research Data Services. DealScan provides detailed information on 152,144 commercial loans given between 1995 to I then used the Compustat data set to obtain detailed lender and borrower information. Since the two data sets do not have a common identier for borrowers and lenders, I used the link le provided by Chava and Roberts (2008), which matches the two data sources' identication codes. My third source of data is the Center for Research in Security Prices (CRSP). I use the CRSP data set to obtain data on equity return and volatility for borrowers. I used the ticker of the borrowing company provided in DealScan to match the market information from CRSP. I am able to match 30,340 facilities to the lender, borrower, and market information collected between 1995 to Finally, I use the data provided by Gilchrist and Zakraj²ek (2012) for the EBP information and the controls used in equation (5). 9 While the dierent data sets include information on foreign loans, I limit my sample to include only loans where the borrower is an American institution and the loan was syndicated in the U.S. 14

15 I start in the second quarter of 1995 due primarily to the availability of data. 10 An additional reason to start the analysis in the mid-1990s is the strong increase in the amount and impotance of the syndicated loan market as a source of funding at that time (Gadanecz 2004). I chose to end the sample before 2007 for two main reasons. The rst reason is that I wish to prevent any eects of the nancial distress caused by the nancial crisis on bank lending decisions. The second reason is the zero interest rate policy (Zero Lower Bond) which prevailed in the U.S from the end of 2008 and induced monetary authorities to conduct unconventional monetary policy. 4.1 Measuring risk A key feature of this study is the measure of loan riskiness. In general measures of default risk try to provide investors and institutions with a measured assessment of rms' ability to pay back its debt and obligations. The results of such measures are that rms generally pay a spread that is correlated with their risk level (or default probability) to compensate lenders for the extra risk they are willing to take. There are generally two main sources of available information when trying to assess rms' default probability: nancial statements and market prices. While nancial statements generally report rms' past performances, market prices represent the market view of rms' future performance. Market prices are therefore the best objective view of the rms' ability to meet its future obligations given all available information. For the purpose of this study the question of whether markets are ecient and are able to correctly predict rms' future performance is actually quite minor. Since I model the ex-ante change in banks' behavior to risk, the only relevant measurement is a measurement of what banks think the riskiness of a borrower is. Thus, the ideal measure of risk will use market-based information and nancial statements to model what we could anticipate as the risk perception of nancial institutions when granting the loan. My main measure of risk is, therefore, a version of the Distance to Default (DD) measure developed rst by Merton (1974). Equation (8) presents the general equation to calculate distance to default used by Bharath and Shumway (2008). DD = ln( V F ) + (µ 0.5σ2 V )T σ V T (8) 10 While I am able to nd loan information dating back to 1988 in DealScan, the Compustat data set which I have only provided information starting in

16 where V is the rm value, F is the face value of the rm's debt, µ is the expected annual return of the rm's assets, σ V is the expected annual volatility of the rm's assets and T is the forecast horizon. The main challenge when computing the DD is that V and σ V are not directly observable and must be estimated. The rst assumption when estimating rm value and volatility is that the total value of the rm follows a stochastic process: dv = µ V V dt + σ V V dw (9) where µ V is the expected return on V, σ V is the rms' asset value volatility, and dw is the increment of the standard Weiner process. The model then assumes that there are only two types of liabilities, one class of debt and one class of equity. Given that F is the book value of the debt which is due at time T, the market value of equity and the market value of assets then follow the Black-Scholes-Merton option pricing framework: E = V N(d 1 ) e rt F N(d 2 ) (10) where E is the value of the rms' equity, r is the risk free rate, N is the cumulative density function of the standard normal distribution and d 1, d 2 are dened as: d 1 = V F + (r + 0.5σ2 V )T σ V T (11) d 2 = d 1 + σ V T (12) I use the Black-Scholes-Merton framework to relate the volatility of the rm assets and its equity using the formula: σ E = V F N(d 1)σ V (13) I then estimate each rm's distance to default (DD) using the following data specication: I dene F to be equal to rms' short-term debt (Compustat dlcq) plus one-half of the long-term debt (Compustat dlttq), the forecast horizon (T) is one year, the risk-free rate is measured as the three-month T-bill rate, σ E is the annualized percent standard deviation of return estimated from the previous year's stock return data for each quarter (CRSP ret), and E is the market value of each rm's equity calculated as the product of share price and the number of shares 16

17 outstanding (Compustat prccq and cshoq). Since the excessive volatility of V/F in equation (13) causes large swings in the estimated volatility of σ V, I follow the iterative procedure described by Bharath and Shumway (2008). In general, a higher value of DD implies safer and more stable rms. As we can see, the main inputs in the DD model are the market value of equity, the face value of debt and equity volatility. We can see that in order for the model to produce good measures of default probability markets must be ecient and well-informed. For the purpose of this research, the DD model is a good method of calculating risk since it represents the best available market perception of the rm risk level, given all available information. In the sensitivity analysis, I use other measures of default risk such as a dierent measure of distance to default, the "naive-dd", and the S&P credit rating score. 4.2 Measuring monetary policy My main measure of monetary policy stance is the real federal funds rate. The real federal funds rate is computed as the dierence between the eective (nominal) federal funds rate and the consumer price index ination rate (De Nicolò et al. 2010). The main reason for using the federal funds rate is that during my sample period it was the primary tool used for implementing monetary policy. Since banks rely on short-term funding, which is highly correlated with the federal funds rate, we can expect that it will have a signicant eect on banks lending and risk behavior (Delis, Hasan, and Mylonidis 2012). 11 The main problem in using the federal fund's rate is that it may be endogenous to other macroeconomic variables, such as the output gap. Thus, I also used a number of measures that focus on exogenous shocks to the measurement of monetary policy. I follow Aysun and Hepp (2013) and use four alternative measures of monetary policy shocks: A measure of orthogonal 11 While most of the literature uses the federal funds rate as the main measure of monetary policy stance, there is no consensus as to whether one should use the real or the nominal rate. Dell'Ariccia, Laeven, and Suarez (2016) state that it is better to use the real rate as long as the two are highly correlated. In my sample, the correlation between the nominal and real federal funds rate is very high at Additionally, some papers use the change in the federal funds rate and not its absolute level (see for example Paligorova and Santos (2016)). However, since the risk-taking channel builds on the idea that banks change their willingness to take on risk during periods of loose monetary policy, using absolute levels will give a better indication of the monetary condition versus quarter-by-quarter changes. In any case, I obtained results similar to the main specication when using the nominal federal funds rate or changes in the federal funds rate instead of the real federal funds rate. 17

18 shocks to borrowed reserves (BR), the Christiano and Eichenbaum (1992) index, the Strongin (1995) index, and the Bernanke and Mihov (1998) index. I obtain all measures from Aysun and Hepp (2013) who extended these measures to the end of Variables description In this section, I describe in greater detail all the other variables used in equation (2). For the dependent variable, I use the Compustat AllinDrawn as my measure of loan spread. AllinDrawn is dened as the total cost including all fees which are paid by the borrower over LIBOR. I control for any borrower and lender factors that might aect loan spreads given borrowers' risk level. I control for borrower and lender size and liquidity since arguably, larger and more liquid borrowers and lenders will have better information and access to funds, which will lower the loan spread. Borrower and lender size is dened as the logarithm of total assets (Compustat atc). Liquidity is dened as the ratio of liquid assets which include cash and short-term investments (Compustat cheq) to total assets. I also control for lenders' leverage, since it may aect banks' ability to supply funds. Leverage is dened as total debt which includes long-term debt and debt in current liabilities (Compustat dlttq + dlcq) to total assets. I also include lenders' return on assets ratio, since it may provide information on the lender eciency, nancial position, and capital regulation eect (Altunbasa, Gambacortab, and Marques-Ibanezc 2014). Return on assets ratio (ROA) is de- ned as the lender net income (Compustat niq) divided by total assets. I use the rm and lender specic variables reported one period prior to the loan deal in order to minimize the risk of reverse causality. Additionally, I control for any loan-specic factors that could aect the loan spread. The loan specic variables include the maturity and loan amount, since both may signicantly aect the pricing of the loan (Ivashina 2005). Due to the fact that a number of studies have suggested that the purpose of the loan could have signicant implications on the pricing of the loan, I add a dummy variable which takes the value of one if the main purpose of the loan is designated to nance asset acquisition lines, debt repayment, LBO, spino, and takeover. 12 To control for relationship lending and asymmetric information between dierent lenders I add a dummy variable equal to one if the lender and borrower headquarters are in the same city, because we could assume that some long-term relationship may exist between the two 12 See Rutherford and Chakravarty (2016) for a review of the relevant literature. 18

19 institutions (Jones, Lang, and Nigro 2005). I also add the number of banks participating in the syndicate, since larger loan facilities with more members tends to have lower spreads (Lim, Minton, and Weisbach 2014). Since the vast majority of the facilities have more than one lender (only 12.9% have a single lender) I use the Compustat bankallocation to control for the share of the loan that each lender contributed to the total facility amount. I exclude from my data any observation where any of the above lender or borrower variables are missing. Additionally, I exclude any implausible observation with a negative lender and borrower liquidity or leverage. I follow Ivashina (2009) and exclude all rms belonging to nancial rms or regulated utilities (Compustat SIC code and ). Overall, these corrections reduce the nal sample to 7,575 observations. Table 1 provides some descriptive statistics for the main variables I use in my baseline regression and in computing DD. The risk measure is right-skewed, which is consistent with other studies using these measures (Bhagat, Bolton, and Lu 2015; Bharath and Shumway 2008). All values are consistent with other studies using these variables, the syndicated loan market and the DD measure. 13 Table 2 presents the correlation between the main variable used in equation (2) including the three alternative measures of risk. As expected, the distance to default and the naive version are highly correlated. The third risk measure, the S&P credit rating, is also relatively highly correlated with the two distance-to-default measures. 5 Results Table 3 presents the results from the rst stage regression. 44 of the 47 quarters have negative β coecients, indicating a negative relation between loan spreads and the distance-to-default measure. The vast majority of negative coecients are signicant at 5% or lower signicant level (only two are not signicant and three are signicant at 10%), while only one of the positive coecients is signicant. Since a higher DD represents safer or more credit-reliable borrowers, we can see that in the vast majority of sample quarters, higher risk levels (lower DD) are correlated with higher spreads. This indicates that banks are sensitive to the borrower's risk level and charge riskier lenders higher spreads as expected. 13 See Lee and Mullineaux (2004), Focarelli, Pozzolo, and Casolaro (2008), Ivashina (2009), and Lim, Minton, and Weisbach (2014) for the syndicated loan market and Bharath and Shumway (2008), Campbell, Hilscher, and Szilagyi (2008), and Bhagat, Bolton, and Lu (2015) for measures of DD. 19

20 Table 4 presents the results from the second stage regression using all measures of monetary policy. All γ coecients are negative and signicant at 5% or lower. The negative coecients indicate that during periods of tight monetary policy, banks' sensitivity to risk increases. That is, during periods of high rates, banks' reaction to risk (measured as an increase in loan spread) is stronger compared to periods of low rates. The results show that monetary policy has a real eect on how banks value and react to risk. The negative γ coecients indicate that during periods of tight monetary policy, the β in the rst stage regression will be more negative and borrowers will be charged higher relative spreads for each increase in their level of riskiness (lower DD). The main implication of Table 4 is that during periods of high rates, banks are more sensitive to borrowers' riskiness level, which increases the risk premium borrowers must pay to obtain capital. On the other hand when monetary policy is expansionary not only does the rate on loans go down, but also banks' sensitivity to risk and therefore the premium the borrowers need to pay for their risk level is lower. The fact that monetary policy has a signicant eect on lenders' sensitivity to borrowers' risk serves as direct evidence of the existence of a risk-taking channel of monetary policy. That is, contractionary (expansionary) monetary policy has a signicant positive (negative) eect on banks' sensitivity to risk. Table 5 reports the results for equation (4) investigating the connection between banks' risk sensitivity and real economic activity. Lagged values of Banks' risk sensitivity are signicantly correlated with industrial production. Since higher levels of my measure of risk sensitivity (β RISK ) imply that banks' are less sensitive to risk, the positive coecients imply that banks' relaxing their risk sensitivity is associated with an increase in economic activity and specically industrial production. 14 The impulse response function computed from the VAR analysis is presented in gure 1. The gure shows that the response of industrial production to a negative shock to banks' risk sensitivity (higher β RISK ) is positive and persistent. The results suggest that when nancial intermediaries become less sensitive to borrowers' risk the result is an increase in the supply of credit which could potentially cause or amplify the increase in economic activity. The above results do not, of course, imply any casual relationship between banks' sensitivity to risk and aggregate economic activity. Despite the simple suggestive nature of these results, 14 When using GDP growth or change in unemployment as economic indicator all 6 coecients are also positive and 3 are signicant at 10% level and one is signicant at 5% level. 20