Credit standards and financial institutions leverage

Transcription

1 Credit standards and financial institutions leverage Gilles Dufrénot Benjamin Klaus Sheheryar Malik Alexandros P. Vardoulakis August 2012 Abstract We empirically test the interaction between the leverage of financial institutions and credit standards, which provide information about the future state of the real economy. Contrary to conventional wisdom that business cycles drive the decisions of financial institutions, we restrict the causality in the opposite direction and examine whether leverage cycles provide useful information about future credit conditions and thus the real economy. We find that a metric that captures both leverage and risk-taking by financial institutions can act as a leading indicator for credit conditions. On the contrary, measures based on market prices and prevailing expectations are not significant in predicting the variation in credit conditions. Keywords: Credit Standards, Leverage, Risk-taking, Procyclicality, Leading Indicators, Net tightening JEL Classification: E3, E44, G01, G21 We are grateful to the participants for their helpful comments at the Toulouse School of Economics- Banque de France seminar series, SAET conference and to Regis Breton, Dimitrios Tsomocos. All remaining errors are ours. The views expressed in this paper are those of the authors and do not necessarily represent those of the Banque de France, the European Central Bank or the Eurosystem. Aix-Marseille Université (AMSE), lopaduf@aol.com Banque de France, benjamin.klaus@banque-france.fr Banque de France, sheheryar.malik13@googl .com European Central Bank and Banque de France, alexandros.vardoulakis@ecb.int 1

2 1 Introduction Variations in the supply of credit have an effect on business activity through a lending and a reinforcing balance sheet channel (Bernanke (1983), Bernanke and Gertler (1989)). A measure of cyclically in credit is the index on banks lending standards produced by the Federal Reserve s Senior Loan Officer Opinion Survey, since it reflects the criteria according to which a bank makes its lending decisions. Credit conditions can be considered as a forerunner of future economic activity. Lown and Morgan (2006) find that recessions have often been preceded by a tightening in lending standards. In particular, changes in the Federal Reserve s lending standards index Granger-causes changes in output, loans and the federal funds rate. On the contrary, these macroeconomic variables are not successful in explaining variation in the lending standards index. 1 Related to the aforementioned studies, this paper takes a step back and attempts to empirically identify variables that can act as leading indicators for the cyclicality in lending standards. Given the results by Lown and Morgan (2006), such indicators should not be derived from real economic variables, such as GDP, or loans extended to the real economy. Lending standards reflect the subjective beliefs of loan officers about future economic conditions. We build our intuition for the selection of our leading indicators on the work of Geanakoplos (2003) and Bhattacharya et al. (2011). Both papers argue that the subjective beliefs of investors play a crucial role for their risk taking behavior and leverage decisions. In particular, Bhattacharya et al. (2011) show that periods of prosperity can invoke optimism about future economic conditions and result in financial institutions increasing their borrowing and shifting their portfolios towards riskier assets. Optimistic subjective beliefs both reduce the borrowing costs of financial institutions, since creditors demand lower premia, and ease the credit standards for contemporaneous credit extension and investment by financial institutions. However, this increases the vulnerability of the financial system to adverse shocks. A measure, which captures both risk-taking and increased leverage by financial institutions, should carry additional information about the cyclicality in credit standards over and above any information present in the current beliefs about credit conditions. In other words, changes in credit standards should not only depend on realized shocks, but also on the past optimizing behavior and risk-taking of financial institutions as the theories of endogenous 1 Asea and Blomberg (1998) also find empirical evidence for procyclicality in credit and conclude that cycles in bank lending standards are important in explaining aggregate economic activity. 2

3 leverage cycles advocate. These theories also predict an asymmetric response by financial institutions in good and bad times, which can be regarded as times of loose and tight credit standards, respectively. In good times, institutions are more optimistic about the future prospects of the economy, they increase their leverage and potentially invest in riskier projects, whose risk-return profile has improved. On average, this behavior can result in bigger losses in the future, a correction in expectations, which become less optimistic, and a subsequent tightening in lending standards. On the contrary, an increase in risktaking and leverage during bad times when credit standards are tight, signals a gradual improvement in expectations, and, thus, an improvement in future credit conditions. This asymmetric response motivates our choice of econometric model. On aggregate, credit conditions can be characterized either as improving or deteriorating. Hence, there is a need for a model which clearly distinguishes between these two regimes. Moreover, the degree of change in credit conditions depends on the level of previous conditions and the specific point in the cycle at which financial institutions evaluate the economic environment they face when deciding to extend credit. Any econometric model capable of tracking the leverage cycle and testing the aforementioned theories should allow for time-varying transition probabilities among different states of credit conditions. For that purpose, we implement the time-varying transition probability (TVTP) Markov switching model developed by Filardo (1994). This methodology allows for identification of not only credit conditions regimes, but also of leading indicators which carry statistically significant information about the cyclical behavior of credit standards. Our modeling approach has the advantage of testing the statistical significance of the chosen indicator in adding information about the cyclicality of credit conditions distinctly within each regime. An econometric model, which does not distinguish between regimes, can average out important information rendering the indicator insignificant overall. Alternatively, the significance of an indicator may result from its usefulness in only one regime, while its informational value may be limited in the other. There are a handful of variables that can potentially serve as leading indicators for credit conditions. The aforementioned studies suggest that such variables should capture elements of risk-taking behavior and leverage decisions by financial institutions. However, financial institutions differ in their behavior to take on risk and increase their leverage. The leading indicator that we propose follows closely Adrian and Shin (2010), who show that broker dealers adjust their leverage according to prevailing risk conditions more aggressively than commercial banks. This can create a painful deleveraging 3

4 when risk perceptions change, asset values fall and the creditworthiness of borrowers deteriorates. Adrian et al. (2010) show that innovations in broker dealers leverage are able to explain the bulk of variation in the cross-section of returns. Thus, broker dealers leverage is a priced factor, which is argued to correspond to changes in risk-taking behavior. We use the same interpretation, but we are interested in the evolution of broker dealers leverage, and hence overall risk-taking in the economy, over time and whether it can act as a leading indicator for future credit conditions. 2 Our approach puts emphasis on the relative risk shifting performed by financial institutions, which is funded by increased borrowing. Total financial borrowing can increase, but it arguably carries more information about future credit conditions when it is allocated to riskier activities. We construct our leading indicator therefore by multiplying total financial leverage, which is equal to the total assets of broker dealers and commercial banks divided by their total equity, with the relative allocation of borrowed funds by financial institutions, i.e. broker dealers total liabilities divided by commercial banks total liabilities. The second part of the indicator captures risk-shifting performed in the financial system, while the first one captures the total indebtedness of the financial sector. The assumption that broker dealers engage in riskier operations than commercial banks can be supported by the fact that the former are less strictly regulated and that they compete more intensively for positive yields on their investments (Hanson et al. (2010)). Higher leverage alone is not harmful if it is mainly allocated to safe investments, while risk-shifting introduces lower risks of deleveraging and bankruptcy when it is equity financed. We consider the case of the US from 1990Q1 to 2011Q2 for which we have quarterly data for credit standards and borrowing of financial institutions, taken from the flow of funds accounts. Our database includes three crisis periods, the Long-Term Capital Management (LTCM) collapse in 1998, the burst of the dot-com bubble in 2000 and the financial crisis starting in The analysis can be carried out for other countries as well. However, most other countries lack appropriate data for credit standards or it is difficult to construct a measure of risk-shifting due to the universal banking model which does not distinguish between broker dealers and commercial banking operations. 2 A related paper is that of Gorton and He (2008), which investigates how competition among banks affects their lending standards. First, they present a theoretical model which shows that banks competition for borrowers leads to periodic credit crunches, i.e. credit cycles can occur without changes in the macroeconomic environment. Second, they empirically test the model by using information on banks performance differences in terms of loan losses. They find that the relative bank performance (1) has predictive power for rival banks behaviour in the credit card market, (2) represents an autonomous source of macroeconomic fluctuations, (3) is a priced risk factor for both banks and non-financial firms. 4

5 Thus, regulatory data on investments by asset class are needed to perform this analysis for the UK or Europe. We find that the constructed indicator carries significant information for the evolution of credit conditions both in loose and tight credit regimes. Most importantly, total financial leverage or broker dealers to commercial banks borrowing are both insignificant by themselves in predicting variations in credit conditions. It is the combination of higher financial leverage with increased risk-taking that creates endogenous credit cycles. A higher value for the indicator, which can stem either from higher leverage, increased risk-taking or both, within a loose credit regime signals higher vulnerability in the financial system, since the probability of moving to a tightening regime increases. On the other hand, the probability of leaving a tightening regime increases with the value of the indicator, which signals that expectations are improving and a recovery is more probable. In order to check the robustness of our results, we bootstrap the statistic that enables us to test the informational content of the variables potentially causing variations in credit conditions. To our knowledge, this is the first time such a bootstrap analysis is proposed within the framework of time-varying transition probability Markov-switching models. The simulations suggest that our model does a good job of tracking the timevarying nature of the predictability of loose and tight credit regimes by our indicator. Moreover, we test additional variables which can presumably serve as leading indicators for credit conditions, such the VIX index, the TED spread and the spread between the yields of Moody s Baa and Aaa corporate bonds. Contrary to the quantity-based measure discussed above, the predictive power of price-based measures is limited. Although these measures are intended to signal risk in the financial system, they respond with a lag due to the fact that they are biased by optimistic expectations or, in other words, procyclicality. They can signal risk in bad regimes after subjective beliefs are updated, but their informational power for future credit conditions is limited due to the fact that credit standards already reflect the change in expectations. However, these high-frequency measures may carry useful information within tightening regimes, since they respond faster to the unraveling of risk, while credit standards only respond on a quarterly basis. The rest of the paper proceeds as follows. Section 2 describes the econometric model. Section 3 discusses the data used in the analysis and provides the economic rationale for the choice of variables. Section 4 outlines our conjectures and results, and presents the robustness check for our chosen leading indicator, while section 5 discusses alternative 5

6 leading indicators for credit conditions. Section 6 concludes. 2 Model We are interested in examining the cyclical features of credit conditions, y t, t = 1,...,T. Allowing for state-dependent means, the prototypical TVTP probability Markov switching model, provided in Filardo (1994), can be described as below (see also Filardo and Gordon (1998), Layton and Smith (2007)); y t = µ 0 + Φ(L)(y t 1 µ S t 1 ) + ǫ 0 t in state 0 (1) = µ 1 + Φ(L)(y t 1 µ S t 1 ) + ǫ 1 t in state 1 Here Φ(L) = δ 1 + δ 2 L δ d L d 1 is a lag polynomial, ǫ 0,1 t IIDN(0,σ 2 ), and S t {0, 1}; thus the state-dependent mean µ St = µ 0 + µ 1 S t. The state variable S t is governed by a hidden 2-state Markov-chain with transition probability matrix, P(S t = s t S t 1 = s t 1,v t ) = Π = [ p(v t ) 1 p(v t ) 1 q(v t ) q(v t ) ], (2) where v t = {v t,v t 1,...} is the history of the conditioning (leading indicator) variable(s) conjectured to be informative with regards to variation in credit conditions. The two states in our application hence correspond to what we refered to earlier as the tightening and easing regimes in credit conditions. The choice of functional form of q( ) and p( ) is typically probit or logistic type. For the purpose of this analysis essentially due to the simple form of its CDF we shall deal with the latter. 3 The relevant logistic function will take the form, and similarly, p(v t ) = exp(θ p0 + Σ M 1 m=1θ pm v t m ) 1 + exp(θ p0 + Σ M 1 m=1θ pm v t m ), q(v t ) = exp(θ q0 + Σ M 2 m=1θ qm v t m ) 1 + exp(θ q0 + Σ M 2 m=1θ qm v t m ). Here m = 1,..,M 1 denote the lags for the indicator variable contained in p(v t ), whereas m = 1,..,M 2 the lags corresponding to expression q(v t ). The model can be cast as a 3 In comparison the probit corresponds to a Normal CDF and thus involves an unevaulated integral which tends to be computationally more demanding; albeit marginally. In practice the results obtained from either functional form are indistiguishable. 6

7 conditional-joint density g( ) written as, g(y t y t 1,...,y t d,v t ) = 1 s t= s t d =0 ĝ(y t S t = s t,...,s t d = s t d,y t 1,...,y t d ) P(S t = s t S t 1 = s t 1,v t ) P(S t 1 = s t 1,...,S t d = s t d y t 1,...,y t d,v t 1 ), and correspondingly the log-likelihood is given by, (3) L(θ) = T ln[g(y t y t 1,...,y t d,v t ;θ)]. (4) t=1 Here θ denotes the parameter vector. g(y t y t 1,...,y t d,v t ) makes explicit the link between the conditioning variables contained in v t with regards to how they feature in the inferential procedure for a Markov switching model for series y t via the transition probabilities. Estimation is carried out via maximum likelihood (ML) methods adapted for mixtures of normals. This is naturally facilitated given the structure of the model as in (1) and (2) and functional form for the transition probabilities such that q(v t ),p(v t ) (0, 1) guarantees a well-defined log-likelihood function. The ML approach has advantages over competing estimation strategies for TVTP models. Notwithstanding its computational ease it may be preferable over estimation via EM algorithm put forth by Diebold et al. (1994) where it is generally difficult to implement the maximization step in the presence of AR dynamics. On the other hand Filardo and Gordon (1998) work with a Gibbs sampling approach which may be feasible in the context of this variety of models. Although a tractable approach, in practice we may require very tight priors for estimation, an aspect which we perceive as being perhaps more difficult to justify for purposes of a cross country-type analysis (see also Albert and Chib (1993)). Indeed if there is no statistically meaningful information with regards to evolution of the state of the economy contained in v t, then the specification tends to a fixed transition probability (FTP) model (see Hamilton (1989)); more specifically, when the restrictions on coefficients corresponding to the conditioning variables θ pi = θ qi = 0 for i 0 are upheld. Formally, under the null of non time-varying transition probabilities the likelihood ratio test statistic is given by, Ξ = 2 [L(θ) L R (θ)] χ 2 (M 1 +M 2 ),α (5) 7

8 where L R (θ) is the restricted log-likelihood, M 1 + M 2 are the number of restrictions on the test at significance level of α. This can be perceived as a test of the informativeness of the economic indicator variables in modelling and/or predicting business cycle turning points. Furthermore, the likelihood ratio test can also be implemented in order to choose between alternative lag specifications for the conditioning variables, i.e. to test if there is significant information contained in additional lags. The results of the test are reported in the tables by a statistic LTR and the corresponding p-value (marginal level of significance). The transition variable contains information about on the variations in credit conditions when the p-value lies under For purpose of robustness we bootstrapped the test statistics under the null on a Hamilton model (the interested reader can contact the authors for a detailed description of the methodology of bootstrap on this kind of models). In regards to being able to invoke classical likelihood theory, Kiefer (1978) demonstrated that in the case of an i.i.d. switching model the solution to the likelihood equations yields estimators which are consistent, asymptotically efficient and normal. Furthermore the inverse of the negative of the Hessian at the estimate is consistent estimator of the asymptotic variance-covariance matrix of parameter values. Under the assumption that functions of the restrictions are twice differentiable around the true parameters and the gradient of the function is of full rank in the neighbourhood of the true parameters, standard likelihood ratio tests of restrictions are valid. This rank condition can be violated due to the presence of a single data point representing a single regime. In such case we may attain a singularity in the likelihood; the variance may be zero, thus not fall within the permitted parameter space, yielding an unbounded likelihood and inconsistent parameter estimates. One particular way of circumventing this possibility is to model the variance to be constant across regimes; this is what was proposed in Hamilton (1989), followed in Filardo (1994) and what we shall maintain for the purpose of this analysis. 4 All the models we estimate incorporate autoregressive coefficients δ i, (i = 1,..,d) which are not state-dependent. The choice for the lags is one that satisfies the mispecification test on the expected residuals. 4 This assumption may not be inconsistent with the quarterly frequency of data we employ in this analysis. We refer the reader to Kiefer (1980), Phillips (1991) and Caudill and Acharya (1998) for further discussion about suggestions to deal with this issue. 8

9 3 Data The objective of our paper is to explain the variations in credit conditions by tracking the transition probabilities between easing and tightening regimes. We argue that these transition probabilities should depend on the level of risk and leverage in the financial system over time. Thus, we need to identify two series of data: one to proxy our dependent variable, i.e. the credit conditions, and another to act as a leading indicator, which proxies for risk taking behavior and financial leverage. We focus on the US financial system, since, compared to other countries, the available data is more extensive and covers a longer time span. However, our approach can be extended to other countries subject to available data, which can capture the evolution of credit conditions and risk-taking in the financial system. The data on credit conditions is obtained from the Federal Reserve s Senior Loan Officer Opinion Survey and covers the period from 1990Q2 until 2011Q3. In particular, we use the net percentage tightening, which is computed as the number of loan officers reporting tightening standards less the number reporting easing divided by the total number reporting, as our dependent variable. The data on leverage of commercial banks and security broker-dealers is obtained from the Federal Reserve s Flow of Funds Accounts and covers the period from 1990Q1 until 2011Q2. Both series correspond to quarterly data. The inconsistency between the data periods is due to the fact that lending standards are announced one quarter later, for example the survey of January 2001 (2001Q1) corresponds to the lending standards during the previous quarter (2000Q4). We construct two series using the flow of funds data. The first corresponds to the relative shift in riskiness in the financial system and is computed as the ratio of brokerdealers to commercial banks borrowing. We denote this index by RISKSHIFT. The second series signals the level of overall leverage in the financial system and is equal to the ratio of total assets to total equity of commercial banks and security broker-dealers. We denote this index by LEVFI. We argue that the combination of risk-taking behavior, which is signaled by RISKSHIFT, and leverage in the financial system, i.e. LEVFI, matters for future credit conditions. Higher risk-taking, which corresponds to a higher value for RISKSHIFT, should be less dangerous for the financial system and future credit conditions, if it is accompanied by lower leverage, and vice versa. Similarly, an improvement in credit conditions during a recovery should stem from both a recovery in leverage and an improvement in the 9

10 willingness to take risk. 5 Consistent with this, we show later that RISKSHIFT and LEVFI are not sufficient by themselves in capturing the variation in net percentage tightening. Figure 1 shows the evolution of RISKSHIFT and LEVFI for our sample period. Since the 1990s there has been a gradual increase in the size of the broker-dealers compared to commercial banking as measure by the RISKSHIFT index. According to our interpretation, this signals an increase in risk-taking over time. However, the index drops fast after the LTCM crisis in 1998, the burst of the dot-com bubble in 2000 and the subprime crisis in the summer of In particular, after the last crisis it quickly returns to the level observed in On the other hand, there has been a substantial deleveraging in the financial sector on aggregate after 1998 as shown by the evolution of the LEVFI index. Total borrowing in the financial sector is dominated by commercial banks, which started deleveraging to comply with the newly introduced Basel regulation and maintained a constant leverage ratio thereafter as documented by Adrian and Shin (2010). The relatively milder adverse effects of the dot-com bubble burst could be due to the sharp deleveraging in the financial sector before the bubble burst, though risk-taking was increasing, as measured by RISKSHIFT. On the other hand, overall leverage remained high before the subprime crisis despite the increase in risk-taking. Hence, we consider a leading indicator which incorporates both effects of risk-taking and leverage in the financial sector. We construct it by multiplying the two aforementioned series and denote it by RISKFI. Figure 2 presents the evolution of net percentage tightening and our leading indicator, which we denote by RISKFI. We argue that this indicator is not specific to a particular crisis event and can capture the cyclicality of credit conditions over a period of booms and busts with different underlying characteristics. In fact, we find that our leading indicator is highly significant in explaining the probability of transition from one regime to another during our sample period. The rest of the section discusses other variables, which could be used to proxy credit conditions or risk-taking by financial institutions, and provides the rationale for aforementioned choice of variables. 5 Higher risk taking under lower leverage should signal gambling to resurrect rather than a recovery, while an increase in leverage without the willingness to take risk means that at least some borrowers are still credit constrained. 10

11 Figure 1: Risk-taking behavior and leverage in the financial sector RISKSHIFT LEVFI Figure 2: Net percentage tightening and our leading indicator NETTIGHT (lhs) RISKFI (rhs)

12 3.1 Credit Conditions Data Series We identify three candidate variables as proxies for the conditions under which banks extend credit to the economy; net tightening, non-performing loans, and charge-offs rates. The net tightening of credit standards for approving commercial and industrial loans or credit lines to large and medium-sized firms is the outcome of a quarterly survey among senior loan officers at approximately sixty large domestic banks and twenty-four branches of foreign banks across the United States. In aggregate, participating banks account for about 60% of all loans by U.S. banks and about 70% of all U.S. bank business loans. Compared to loan rates, credit standards capture non-price lending terms, such as standards of creditworthiness, collateral, loan limits, etc. which banks might apply to deal with adverse selection and moral hazard. As shown by Lown et al. (2000) and Lown and Morgan (2006), the availability of bank credit depends, in addition to loan rates, strongly on credit standards. There is also evidence that credit standards are linked to economic activity as they help to predict GDP even after controlling for past economic conditions. This is also the case during the 1990s, when capital markets were supposed to have eclipsed the role of banks in the economy. In addition, Lown and Morgan (2006) find that higher loan levels cause a tightening in credit standards and tighter standards are followed by lower loan levels, suggesting a kind of credit cycle. Given its advantages in capturing the credit cycle dynamics, we choose the net tightening index for our analysis. Alternatively, the ratio of non-performing loans to total loans could be used to draw inferences about the state of the economy. Non-performing loans are defined as loans that are 90 days or more past due or have nonaccrual status (Clair (1992)) and thus reflect the degree of deteriorating loan quality. Given that the overall loan quality deteriorates only when the economy is already in the bad state, i.e. when economic conditions have already deteriorated (Jimenez and Saurina (2006)), this measure cannot be considered as timely enough to capture the variation in agents expectations about the future state of the economy as contained in the net tightening of credit standards. A third candidate measure for credit conditions is the charge-off rate. When the quality of a loan is deteriorating, loans are first classified as non-performing and second, if the loans appear unlikely to be fully repaid, the loans or part of them is charged off. Thus, compared to the non-performing loan ratio, the charge-off rate is likely to signal deteriorating economic conditions even later. In addition, as the chargeoff rate is the ratio of charged-off loans extended in previous years to total current loans including those made only recently, the rate measures yesterday s mistakes relative to today s base. As a consequence, growth in total loans can distort this ratio such that 12

13 the charge-off rate would be lower than the actual default probability (Clair (1992)). Having in mind that credit reacts very quickly to changing economic conditions and expectations, we choose the net tightening index as our proxy, since it responds in a timelier manner compared to our other alternatives. Figure 3 presents the evolution of these three candidate series for credit conditions and shows the late response of non-performing loans and charge-off rates compared to net tightening. Figure 3: Net Tightening, Nonperforming Loans and Charge-off Rates NETTIGHT (lhs) NPL (rhs) COR (rhs) Risk-taking Data Series Candidate measures of risk-taking in the financial system fall broadly into two categories. The first and more exhaustive category consists of measures that depend on either prices, interest rates, and/or on their corresponding volatilities. The second, which is less populated, includes quantity based measures, which are broadly related to leverage. The latter measures where proposed as a response to an inherent procyclicality in the financial system. Adam and Marcet (2010) find evidence that investors expectations about asset returns in the US were highest at the peak of the internet boom period in early Such procyclicality is in the heart of Minsky s Financial Instability Hypothesis (1992). During good periods, risk perceptions go down and financial institutions increase their leverage to invest more heavily in assets that appeared more risky before. Although risk-taking may have increased, this may not be captured by 13

14 price-based measures, since (procyclical) optimistic beliefs can bias them downwards. More elaborate measures for risk-taking have been proposed, which, however, suffer from procyclicality as well given their dependence on price data. For instance, Altunbas et al. (2010) use the expected default frequency (EDF) computed by Moody s KMV on the basis of the Merton (1974) model. The EDF, which combines banks financial statements with stock market information and a default database, gives the probability that a company will default within a specified period of time (usually 12 months). However, as pointed out by the authors, this measure is likely to have underestimated banks risk in the pre-crisis period. Compared to purely price-based measures, this bias should be smaller since the EDF includes leverage which tends to rise during boom periods. However, the second major component of the EDF is asset volatility which declines during boom periods as agents lower their expectations about the realization of extreme negative events and thus the EDF will indicate lower risk in upswings (Borio et al. (2001)). In addition, the paper captures idiosyncratic bank risk through the residuals from a regression of bank returns on market-wide returns and through the individual bank volatility which is estimated using the approach proposed by Campbell et al. (2001). As both measures are based on stock prices that tend to underprice risk during boom periods (see e.g. Lowe (2002)); Borio (2003))) they are unlikely to capture banks risk-taking appropriately. The authors also use banks credit default swap (CDS) spreads to measure bank risk, which have been shown to have strongly underpriced credit risk before the financial crisis (Andersson and Vanini (2010))). Other studies, for example Boyd et al. (2006) and Laeven and Levine (2009), use the so-called Z-score, which equals the return on assets plus the capital-asset ratio divided by the standard deviation of asset returns. It gives the number of standard deviations below the mean by which profits must fall in order to eliminate equity capital and thus reflects the inverse of a bank s failure probability. The measure is computed using accounting data and should therefore be less affected by the underpricing of risk in the upswing than price-based measures, but is also based on realized asset returns which are backward looking by nature and has thus only a limited informational content with regard to future expectations. Another potential measure of banks risk-taking is the Value-at-Risk (VaR), which is an estimate of the expected loss that a financial institution is unlikely to exceed in a given period of time (usually 1 to 10 days) with a particular degree of confidence (usually 95% to 99%). It is one of the most widespread tools used by financial institutions to measure market risk. Interestingly, the evolution of the VaR of globally active 14

15 banks from 2003 until 2006 suggests that risk-taking increased at least mildly during the upswing (International Monetary Fund (2007), Committee on the Global Financial System (2010)). However, the VaR measure has several drawbacks as stressed by Daníelsson (2002). First, it provides no information about the expected losses in case a tail event occurs. Second, VaR is not a coherent risk measure, since the VaR of the sum of two portfolios can be higher than the sum of their individual VaRs. Third, the aggregation of risks across both institutions trading and banking books, and in particular the treatment of levered loans, is difficult, which can lead to an underestimation of risk reflected by the reported VaR (International Monetary Fund (2007)). 6. Most importantly, the statistical properties of market data are not the same in a crisis as during a benign market environment. Thus, the risk reflected by this measure is likely to be underestimated when using the most recent observations from a stable period for estimation (Daníelsson (2002)). We test for the informational context of three widely used price-based measures, the VIX index, the TED spread and the spread between Baa and Aaa bonds according to Moody s. We find that these indicators are not capable of providing valuable information for the evolution of credit conditions. Another important aspect is that the aforementioned measures focus mostly on the cross-sectional dimension of risk, i.e. they compare the risk across different financial institutions. However, as Borio et al. (2001) and others point out, in order to implement a macroprudential financial regulation, which is able to address procyclicality, candidate measures should capture the time dimension of risk appropriately. We, therefore, need a measure that reflects risk-taking behavior over time without being biased by prevailing expectations, which can be either optimistic or pessimistic. We look into quantity-based measures, such as financial imbalances and debt positions, for that purpose. According to Demirguc-Kunt and Detragiache (2005), credit booms are the best single-variable leading indicator of banking distress, and combinations of credit and asset price deviations from long-term trends are considered to be even better 6 VaR is a downside risk measure while VIX captures both parts of the distribution, i.e. also upside potential. This could be one reason why the VaR measure increased during the upswing while the VIX measure decreased. 15

16 (Borio and Lowe (2002), Borio and Drehmann (2009)). 7 Nevertheless, higher credit or leverage growth should not necessarily signal procyclicality and increased risk taking as they may well reflect rapidly improving economic conditions and credit may be directed to safer investment and newly developed sectors, which enhance productivity. We argue that the evolution of the credit allocation between safer and riskier investment is a good signal of procyclicality and underlying beliefs. Financial institutions would not be allowed to persistently increase their leverage to invest in riskier assets if expectations remained constant through time, since they would be rationally penalized with higher borrowing costs and debt markets would eventually close down for them. In order to extract the probabilities governing the cyclical behavior of credit, we choose to look at the changes in the ratio of security broker dealers liabilities to commercial banking liabilities multiplied by the total leverage in the financial sector, as discussed before. We suggest that this variable can capture risk-taking at least in the US financial sector and that it is immune to changing expectations, which introduce biases in price-based measures. They are several additional aspects which motivate the choice of this variable apart from the ones already mentioned. First, security broker-dealers played a central role in the securitized banking of packaging and reselling loans (Gorton and Metrick (2009)). Shleifer and Vishny (2010) argue that this type of financial intermediation (as opposed to that of traditional banking ) transmits security market fluctuations into the real economy, thereby turning the volatility of sentiment into the volatility of real activity. Second, as shown by Adrian and Shin (2010), in contrast to commercial banks which appear to target a fixed leverage ratio, security broker-dealers adjust their leverage aggressively in response to changing economic conditions leading to high leverage in economic booms and low leverage in economic downturns. This implies that the leverage of security broker-dealers is pro-cyclical and might lead to an amplification of the financial cycle. In addition, Adrian et al. (2010) find that the leverage of security broker-dealers explains a substantial proportion of the cross-sectional variation of stock returns and argues that this due to their central role as active market participants. Third, regula- 7 An alternative quantity based measure is constructed by Delis and Kouretas (2011). The authors use the ratio of risk assets to total assets to measure banks risk-taking. Risk assets are defined to include all bank assets except cash, government securities at their market value and balances due from other banks. The paper reports that the mean value of this ratio for all banks in the sample increased by 5% from 2002 until 2006 and argues that this reflects an increase in banks risk-taking. Since these risk assets are assets whose value is subject to changes in market conditions, and given that the time period under consideration was a period of global economic growth, the increase in the ratio of risk assets to total assets is likely to reflect an increase in asset prices during a benign market environment rather than an increase in risk-taking. 16

17 tory reasons may justify why broker-dealer leverage is used for riskier investment than commercial banking leverage, since the latter institutions are more heavily regulated. Finally, non-regulated financial institutions may face more competition in the market that traditional commercial banks, and thus seek more aggressively for higher yields in the margin. This story has been put forward by Hanson et al. (2010). Considering that risk is underpriced during good times, these can increase the advantages of low cost leverage if they invest in assets that promise higher yields than they risk they are perceived to carry. Even if the aggregate level of leverage in the financial system remains constant over time, our index can predict changes in risk-taking and can be used to extract the degree of procyclicality in the financial system. 4 Results This section presents the results for our main indicator, which is computed as the product of RISKSHIFT and LEVBIB series. Its evolution is shown in figure 2. Additional tests for alternative leading indicators, which can potentially explain the variation in credit standards, are presented in section 5. Following the notation in section 2, the value of the indicator at time t is denoted by v t. Our objective is to examine whether this indicator contains valuable information for the variation in credit standards over time. As a proxy for prevailing credit conditions, we use the net tightening index, which is also presented in figure 2 and is denoted by y t. To illustrate the availability of the different data series, figure 4 presents the period that these data series cover and the timing at which they are released. For example, the Senior Loan Officer Opinion Survey dated April 2004 corresponds to officers decisions to tighten or loosen their credit standards over the preceding three-month period, i.e. 2004Q1. We want to examine whether our index can act as a leading indicator for credit standards one quarter ahead, i.e. how y t depends on v t 1. Hence, the risk-taking and leverage data we use to construct it are end of last quarter data, i.e. 2003Q4. Note that these data are not made public at the end of the quarter, but are rather released during the first quarter of Banks tighten credit conditions to avoid or lower losses on future loans in the expectation of deteriorating future economic conditions. Quoting from the Federal Reserve s Senior Loan Officer Opinion Survey, October 2008, [...] respondents pointed to a less favorable or more uncertain economic 17

18 Figure 4: Timing of the release of risk-taking and lending standards data FRB Loan Officer Survey April 2004 Changes in lending standards reflect the three months preceding the survey date 30/09/2003 (t 2) 31/12/2003 (t 1) 31/03/2004 (t) time Risk-taking data (t 1) becomes available during this period Risk-taking data 2003Q3 Risk-taking data 2003Q4 outlook as a reason for tightening their lending standards [...] If more loan officers report a tightening than an easing, i.e. net tightening takes a positive value, banks on aggregate can be considered to expect future economic conditions to deteriorate. This interpretation distinguishes between two regimes, one for positive values of net tightening and the other for negative. Our methodology should be able to identify these two regimes similar to Markov switching models, which endogenously pick the regimes, such as in Hamilton (1989). In technical terms, the estimated mean for a tightening regime is expected to be positive, while for an easing regime negative. Table 1 reports the estimation results (detailed estimation results are shown in table 4 in the appendix). The model we choose is able to differentiate between a tightening and an easing regime. An explanation for the fact that the tightening regime (regime 1) has a higher mean in absolute terms than the easing regime (µ 0 + µ 1 = versus µ 0 = 2.18) could be that loan officers respond more aggressively during bad times due to higher risk-aversion. Moreover, our sample includes three crisis periods - the LTCM crisis, the burst of the dot-com bubble and the financial crisis starting in during which there were substantial shocks in the net worth of the financial system. Thus, stricter borrowing constraints for banks (see for example Geanakoplos (2003) and Gromb and Vayanos (2002)) and an inability to raise new equity could be additional reasons for higher expected net tightening. 18

19 Table 1: Estimation results - information variable RISKFI Parameter Value t-statistic Autoregressive component µ ( 2.34) µ 0 + µ (9.40) δ (34.87) σ ǫ 7.04 (13.87) Transition function component θ p (32.18) θ p ( 17.07) θ q (29.88) θ q ( 16.75) The model also captures the transition from one regime to another. The intercepts in the transition probability functions, p(v t ) and q(v t ), are statistically significant and positive (θ p0 = 4.43 and θ q0 = 5.57). Abstracting from the information contained in our leading indicator, the values for θ p0 and θ q0 signal high persistence of regimes as the probability of remaining in the same regime in the future is very high. According to the arguments in this paper, the transition probability from one regime to another should not only depend on the current regime, but also on the point in the cycle, the risk-taking behavior and the leveraging decisions of financial institutions. Our indicator captures the dynamics of financial behavior over the cycle and is successful at predicting future variation in credit standards, reflected in θ p1 and θ q1 being statistically significant. The signs for θ p1 and θ q1 predict an asymmetric response of credit standards to our risk-taking indicator for the two regimes. A negative value for θ q1 implies that an increase in risk-taking within an easing regime increases the probability of moving to a tightening credit regime in the following quarter. Our result is in line with the prediction of endogenous leverage cycle theories, which argue that when times are good (loose credit standards in the context of our model) the more optimistic financial agents increase their leverage and their exposure to riskier assets. However, economic activity will contract if a bad shock realizes due to the fact that highly-leveraged institutions will see their capital being depleted and the marginal investor in the financial markets will become more pessimistic. To sum up, our model predicts a higher probability of deteriorating credit standards, which are a forerunner of economic activity, for an increase in our risk-taking indicator during regimes of easing credit standards. On the other hand, an increase in risk-taking within a tightening regime, which 19

20 should correspond to more pessimistic expectations about the future economic activity, can signal that the economy is starting to recover. This should correspond to a higher probability of moving out of the tightening regime, which is what our model predicts (θ p1 <0). It is important to highlight that our analysis does not argue that there is a threshold for our leading indicator after which the system moves from one regime to another. We do not compute turning points and, hence, we do not predict the event of a financial crisis. Nevertheless, our analysis has predictive power given that it computes the probability of moving from one regime to another depending on the current point in the leverage cycle. An increasing probability of moving to the tightening regime therefore signals the build-up of vulnerabilities in the financial system, which can lead to a sharp deterioration in credit standards and, thus, future GDP should a bad shock realize. Figure 5 presents the inferred probability of a tightening regime given the available data, P(S t = tightening ). When being close to 1, the measure provides strong evidence from the data that banks credit conditions were in a tightening regime; conversely, when being close to 0, there is evidence that credit conditions were in the easing regime. The graph reveals a relatively high correlation between the inferred probabilities and the Reinhart and Rogoff (2011) indicator for banking crises and stock market crashes which is indicated by the shaded areas. In contrast to that, plots of the transition probabilities p t and q t are difficult to interpret. For example, plotting p t = P(S t = tightening S t 1 = tightening,v t ) does not show fully the relevant contribution of the time variation because the variation is only relevant when S t 1 = tightening. The inferred probability starts declining at the end of 1990 signaling an improvement in credit conditions and remains close to 0 for the period leading to the LTCM crisis. The LTCM crisis was not connected with the leverage cycle in the US, but was rather caused by the exogenous Russian default event. The quick increase and subsequent decrease in the posterior probability illustrates the short-term nature of this specific crisis. On the contrary, our model predicts that the probability of entering a tightening regime increases sharply before the burst of the dot-com bubble in 2000 and especially prior to the financial crisis of It also signals a recovery in a timely manner. Our model suggests the possibility of a double-dip scenario in the years following the 2007 financial crisis. Although the probability of staying in a regime of tighter credit conditions decreases after the collapse of Lehman Brothers, our model signals that this recovery may be short-lived given that the probability of moving to a tightening regime increases sharply in the beginning of 2011, which has been associated 20

21 with the sovereign debt crisis in Europe. 1 Figure 5: Inferred probability of a tightening regime given the available data Alternative Indicators We have argued that the combination of risk-taking behavior together with leverage in the financial sector can explain the variation in future credit conditions. This section examines the robustness of this claim using alternative leading indicators. In particular, we check whether risk-taking, according to our definition, or financial institutions leverage are by themselves capable of providing valuable information about future credit conditions. The net tightening index is still our dependent variable. We consider two versions of the model, one with RISKSHIFT as our leading indicator, and the other with LEVFI. The leading indicators are lagged by one quarter. Table 2 presents the main results (detailed estimation results are shown in table 4 in the appendix). The model can still distinguish between a tightening and an easing regime. However, these two indicators alone do not provide valuable information for the evolution of credit conditions, as the coefficients in the transition function are not significant. We also test for the statistical significance of price based indicators, such as the VIX index, the TED-spread and the the spread between the yields of Moody s Baa and Aaa corporate bonds. We consider end of quarter data. The main results are summarized in 21