Determinants and Consequences of Credit Supply: Evidence from a Panel Vector Autoregression.

Determinants and Consequences of Credit Supply:

Evidence from a Panel Vector Autoregression.

Sebastiaan Pool+ Oktober 2013

Supervisors: Lammertjan Dam, Leo de Haan, Jan Jacobs

Abstract

This paper examines the determinants and consequences of credit supply using a theoretical model which is subsequently empirically estimated. The theoretical model embeds a banking sector with default risk in a standard macroeconomic framework. The paper’s main interest is to determine the causality between bank provisioning (approximating expected default risk) and credit supply. The model is estimated using a panel Vector Autoregressions approach to uncover the dynamic relationships that are common to all countries. The results show that: i) credit supply and provisioning are important determinants of economic growth, ii) bank provisioning decreases as credit supply increases, iii) credit supply is primary affected by the output gap.

Keywords: Monetary Policy, Transmission Mechanism, Provisioning, panel VAR JEL Code: E43, E44, E47

* _{This thesis was written during an internship at De Nederlandsche Bank. The views expressed are those of the author}

and do not necessarily reflect official positions of De Nederlandsche Bank or the University of Groningen. I thank Lammertjan Dam, Leo de Haan and Jan Jacobs for their helpful discussions and comments.

+ _{University of Groningen, Faculty of Economics and Business, P.O. Box 800, 9700 AV Groningen, The Netherlands.}

2

1 Introduction

Before the recent Global Financial Crisis liquidity in the global economy was abundant and credit supply awash (Adrian and Shin, 2009). Nonetheless, the Global Financial Crisis caused liquidity in the interbank money market to dry up constraining financial intermediaries to extend credit to firms and consumers. Central banks worldwide set historical low interest rates hoping to reactivate functioning of the interbank money market (Allen and Carletti, 2008). Despite these exceptional low short-term interest rates, the recovery of the interbank money market has been modest. This paper analyzes why, despite exceptional low short-term interest rates, liquidity constraints in the interbank money market were not eased enough to recover credit supply.

This paper introduces a simple macroeconomic framework with a banking sector and borrowers that are allowed to default. To keep the number of variables tractable, I simplify a banking sector a la Cournot for a representative bank. The representative bank maximizes its expected profits anticipating that a fraction of its existing loans will default in the future, see Greenbaum, Kanatas and Venezia (1989). I solve for the optimal levels of credit supply and the long-term interest rate (the price of credit), given the short-term interest rate (the cost of credit), and loan default risk. The equilibrium conditions for credit supply and the long-term interest rate are embedded in a standard closed economy macroeconomic framework often used to analyze the monetary transmission mechanism, see Svensson (1997). Hence, instead of assuming a perfect interest rate pass-through, default risk and market power in the banking sector determine an interest rate spread. Intuitively, the representative bank is exposed to default risk which imposes a potential cost for the bank. Consequently, an increase in the default risk increases the long-term interest and decreases credit supply. Although the theoretical framework is merely synthesis of some existing approaches in the literature, I believe that its properties can be helpful to improve our understanding of the monetary transmission mechanism.

3

results show that both credit supply and provisioning by the banking sector are important determinants of economic growth. The IRFs show that during turmoil times an increase in provisioning may decrease credit supply and economic activity despite expansionary monetary policy. In addition, the results indicate that if banks constrain credit supply they may worsen the economic slowdown. Second, it appears that banks decrease provisioning as a percentage of the total balance sheet as credit supply increases. Hence, during booms banks take on more risk by keeping less provisions. Third, the IRFs show that especially the output gap is important in determining credit supply while cost factors like provisioning and policy rates appear to be less important.

This study contributes to the existing literature in several respects. First of all, the theoretical model embeds a banking sector with loan default risk in a standard monetary policy framework instead of assuming a perfect interest rate pass-through. The recent Global Financial Crisis proved the importance of the financial sector while its impact on the real economy remains hard to quantify. It is therefore important to study the interaction between the financial sector and the real economy to enhance comprehension of the monetary transmission mechanism. Second, this study estimates a 12-country panel VAR instead of analyzing the monetary transmission mechanism in a single country or area. The panel VAR approach allows to uncover the dynamic relationships that are common to all countries while at the same time it improves the statistical inference.

The rest of the paper is structured as follows. Section 2 presents relevant prior work. Section 3 derives the theoretical model and Section 4 specifies the methodology. Section 5 describes the data. Section 6 presents the panel VAR estimations results and compares the theoretical predictions with the empirical results. Section 7 analyses the sensitivity of the empirical results and Section 8 concludes.

2 Literature review

2.1 Monetary transmission mechanism

4

Minetti (2008) which present evidence in favor of the credit channel. Boivin, Kiley and Mishkin (2010) provide an excellent review of the literature.

Allen and Carletti (2008) argue that for a well-functioning monetary transmission mechanism an important role is preserved for the interbank money market and its role of allocating liquidity from banks with a surplus to banks with a deficit. The recent Global Financial Crisis showed the vulnerability of the interbank money market to fulfill its role of allocating liquidity from banks with a surplus to banks with a deficit. Central banks worldwide introduced a wide range of measures hoping to restore the functioning of the interbank money market. Traditionally, changes in the short-term interest rate, i.e. the policy rate, are modeled as the main monetary policy instrument (Igan et al., 2013).1 Liquidity constrained banks can borrow from their central bank at this policy rate, as a consequence, expansionary monetary policy lowers a bank’s financing costs. Cornett et al. (2011) show that although policy rates have been historically low since the beginning of 2009 banks decreased credit supply.

Many theoretical solutions have been proposed to explain the traditional channels of monetary policy, however the empirical evidence on the strength of these channels differs substantially, see Bernanke and Gertler (1995) and references therein. For example Eickmeier, Hofmann and Worms (2009) argue that, on the one hand, expansionary monetary policy shocks may increase credit supply both directly and indirectly via demand and/or supply side effects. Lowering the interest rate directly decreases the costs of additional funding for banks which may increase credit supply. At the same time, lowering the interest rate also directly lowers the cost of financing investment and consumption expenditures which may increase the demand for credit, assuming that the interest rate reduction has been carried through. Expansionary monetary policy can also indirectly increase credit demand via its potentially expansionary effect on economic activity. On the other hand, households and firms may smooth the effect of a monetary policy shock on consumption and investment by altering the share of internal versus external financing, e.g. firms may increase the relative amount of internal financing when cash flows increase after expansionary monetary policy. For a formal discussion see Bernanke and Gertler (1995) and Friedman et al. (1993).

Peersman (2011) shows that expansionary monetary policy may affect the credit supply

1 _{Other policy instruments as Quantitative Easing, standing facilities, Long-Term Refinancing Operations, are not}

5

capacity by altering the risk taking channel in three different ways. First, standard economic textbook models state that expansionary monetary policy decreases the interest rate which inclines households to increase the relative amount of currency holdings and decrease the relative share of interest bearing deposits, e.g. Mishkin (2010), Walsh (2010). Consequently, the financial sector faces tighter liquidity constraints, limiting banks to supply additional credit. Second, expansionary monetary policy improves the quality and value of debt outstanding because the expected repayment flow has increased. Accordingly, the bank’s market-to-market value of equity increases which improves the bank’s balance sheet and inclines the bank to take more risk by increasing credit supply. Third, expansionary monetary policy may also stimulate a bank to increase credit supply because credit spreads increase which also raises a bank’s market-to-market value of equity. To conclude, theory is ambiguous about the net effect of monetary policy on credit supply. Traditionally the monetary transmission mechanism is studied using VAR models, see Bernanke and Gertler (1995), Christiano, Eichenbaum and Evans (1999) and Stock and Watson (2001) for a discussion of the literature. Panel VARs can be used to uncover the dynamic relationships that are common to all cross-sectional units, e.g. Love and Zicchino (2006) study the impact of financial factors on firm investment and de Haan and van den End (2013) study the impact of banks’ responses to market funding shocks.

2.2 Market power in the banking sector

6

Third, other studies report the contestability and regulatory restrictions as a source of market power, e.g. Demirgüç-Kunt, Laeven and Levine (2004). Several empirical papers confirm the presence of market power in the banking sector, see Berger et al. (2004) and Degryse and Ongena (2008) for a discussion of this literature.

Freixas and Rochet (2008) model a banking sector by introducing utility maximizing consumers and profit maximizing firms and banks. The final solution is such that each agent in the economy behaves optimally and each market clears, i.e. the goods market, deposits market, and credit market. Recently, Gerali et al. (2010) analyzed the effect of shocks to the banking sector in a Dynamic Stochastic General Equilibrium (DSGE) model. They model market power in the banking sector by assuming a regime of monopolistic competition. Although these theoretical models are appealing, it is not efficient to estimate these models in a VAR framework. Adding variables creates complications because the number of VAR parameters increases with the number of variables squared (Stock and Watson, 2001). Instead the banking sector can also be modelled in a regime of oligopolistic competition by using the Monti-Klein (MK) (Monti, 1972) and (Klein, 1971). Greenbaum, Kanatas and Venezia (1989) extend the MK model by introducing loan default risk. The advantage of the MK model is that is allows for market power, but keeps the number of variables tractable.

3 The model

This section describes banking activities by a production function of loan services following the MK model of imperfect competition between a finite number of 𝑛 banks. The MK model assumes that banks are confronted with a demand curve for credit which is downward sloping in the interest rate. More general models explicitly model utility maximizing households and profit maximizing firms in order to derive equilibria levels for the credit, deposits, and interbank money market. I do not model these sectors explicitly, but proxy credit demand as a function of the long-term interest rate, the output gap and a risk premium expressing the probability of default, 𝐿𝑛𝑡(𝐼𝑡𝑙𝑟, 𝑌𝑡𝑔, 𝑅𝑡), where

𝐿𝑛_{𝑡 is credit demand, i.e. the bank’s decision variable, 𝐼}

𝑡𝑙𝑟 is the long-term interest rate, 𝑌𝑡𝑔 is the

7

profits, for this reasons it is more convenient to work with the inverse demand function 𝐼𝑡𝑙𝑟(𝑌𝑡𝑔, 𝐿𝑛𝑡, 𝑅𝑡).

Not explicitly modeling utility maximizing households and profit maximizing firms to find equilibria for the credit market and deposit market is a serious caveat of the model. These sectors will, however, be included implicitly by including the IS curve for the output gap which can be derived by maximizing utility in the life cycle model.

Each bank determines every period the amount of credit applications they will accept by maximizing their expected profit function. I abstract from any other (non-)interest income because the aim is to analyze what determines credit supply. In a Cournot equilibrium every bank maximizes each period its expected profit function taking the volume of deposits of the other banks as given 𝐸𝑡{𝜋𝑖,𝑡𝑝} = max_𝐿 𝑖,𝑡 𝑛 {(𝐸𝑡{𝑃𝑡}𝐼𝑡 𝑙𝑟_(𝑌 𝑡𝑔, 𝐿𝑛𝑡, 𝑅𝑡) − 𝑐𝑖,𝑡) 𝐿𝑛𝑖,𝑡 + 𝛽𝐸𝑡{𝜋𝑖,𝑡+1𝑝 }} , (1)

where 𝜋_𝑖,𝑡𝑝 denotes a bank’s profits indexed by 𝑖, 𝑃_𝑡 is the probability of paying interest, i.e. the probability that the loan does not default, 𝑐_𝑖,𝑡 are the marginal costs, 𝛽 is the discount rate, and 𝐸_𝑡 is the expectation operator. Equation (1) can be rewritten as, see Appendix A

𝐸_𝑡{𝜋_𝑖,𝑡𝑝} = max 𝐿𝑛_𝑖,𝑡 {∑(𝛽𝜆𝐸𝐸𝑡{𝑃𝜏}) 𝜏−𝑡_𝐸 𝑡{(𝑃𝜏𝐼𝑡𝑙𝑟(𝑌𝑡𝑔, 𝐿𝑛𝜏, 𝑅𝑡) − 𝑐𝑖,𝜏) 𝐿𝑛𝑖,𝜏} ∞ 𝜏=𝑡 } , (2)

where 𝜆_𝐸 is added to introduce that each period on average 1 − 𝜆_𝐸 percent of the loans matures. If a loan matures, the bank will no longer gain any interest income. The loan owner can apply for a new loan, which is subsequently considered by a bank, however, I assume that the chance of acceptance is independent of the borrowers previous loan contract.

8

credit was plentiful.” The statement suggests that until the Global Financial Crisis liquidity constraints were indeed not binding.

The bank has to finance new loans which is represented in Equation (2) by subtracting marginal costs, 𝑐𝑖,𝑡, from marginal revenues, 𝐼𝑡𝑙𝑟(𝑌𝑡𝑔, 𝐿𝑛𝜏, 𝑅𝑡), for each new loan contract. Since the

model abstracts from a deposits market in order to keep the number of variables tractable, one needs to assume a cost function independent of deposits. The literature often sets the deposit rate equal to the short-term interest and operating costs, see Freixas and Rochet (2008) and more recently Bolt et al. (2012). Besides deposits, the main sources of finance for a bank are the central bank and the interbank money market. For both sources of finance banks pay the short-term interest rate.2 Hence, abstracting from any operating costs and under the assumption of perfect elastic supply of funds, the short-term interest rate is the only cost for a bank to accept a loan application. For simplicity it is assumed that each bank has the same cost function

𝐶_𝑖,𝑡 = 𝐶_t = 𝑖_𝑡𝑠𝑟_𝐿 𝑖,𝑡 𝑛 _,

(3)

where 𝐶_𝑖,𝑡 is total costs and 𝑖_𝑡𝑠𝑟 _{is the short-term interest rate.}

The MK model assumes that the bank is confronted with a demand curve for credit downward sloping in the interest rate. I proxy credit demand by a functional form linear in the output gap, the long-term interest rate and the risk premium. Taking the inverse of the credit demand function yields

𝐼_𝑡𝑙𝑟 _{= 𝐴 − 𝜆}

𝑝𝑃𝑡− 𝜆𝑙𝐿𝑛𝑡 + 𝜆𝑦𝑌𝑡𝑔, (4)

where the risk premium 𝑅_𝑡, i.e. the probability that a loan defaults, is substituted for 1 − 𝑃_𝑡, i.e. one minus the probability of not defaulting, and 𝐴 is a constant which includes the 𝜆_𝑝 term.3 Subsequently, I substitute the marginal costs derived from Equation (3) and Equation (4) in Equation (2). After some rewriting, see Appendix A, and noticing that new loans can be

2 _{Assuming that the interbank market interest rate and the interest rate set by the Central Bank are equal.} 3 _{The interest rate depends positively on the risk premium, hence 𝐼}

𝑡𝑙𝑟= 𝑎 + 𝜆𝑝𝑅𝑡 where 𝑎 is a constant. Substituting

9 decomposed as 𝐿𝑛_{𝑡 = 𝐿} 1,𝑡 𝑛 _{+ 𝐿} 2,𝑡 𝑛 _{+ ⋯ + 𝐿} 𝑛,𝑡 𝑛 _{, Equation (2) becomes} 𝜋_𝑖,𝑡𝑝 = max 𝐿𝑛_𝑖,𝑡 {( 1 + 𝑟 𝑟 ) (𝑃𝑡[𝐴 − 𝜆𝑝𝑃𝑡− 𝜆𝑙 ∑ 𝐿𝑗,𝑡𝑛 𝑛 𝑗=1 + 𝜆𝑦𝑌𝑡 𝑔_{] − 𝑖} 𝑡𝑠𝑟) 𝐿𝑛𝑖,𝑡}, (5) where 𝑟 =1+𝑖𝑡+1𝑠𝑟

𝑃𝑡𝜆𝐸 − 1. The model is solved by maximizing profits with respect to loans and solving

for loans, assuming symmetry among banks. This yields the optimal solution for new loans at the bank and sector level

𝐿𝑛_𝑖,𝑡 = 1 𝜆𝑙(𝑛 + 1)(𝐴 − 𝜆𝑝𝑃𝑡− 𝑖_𝑡𝑠𝑟 𝑃𝑡 + 𝜆𝑦𝑌𝑡 𝑔_), (6) 𝐿𝑛𝑡 = 𝑛 𝜆𝑙(𝑛 + 1)(𝐴 − 𝜆𝑝𝑃𝑡− 𝑖_𝑡𝑠𝑟 𝑃𝑡 + 𝜆𝑦𝑌𝑡 𝑔_). (7)

The banks anticipate that a percentage 𝑃𝑡 of the loan ‘stock’ defaults at some future date.

The expected probability of the interest payment is given by the expected percentage of loans that will pay interest

𝑃𝑡=

𝐿_𝑡− 𝐸_𝑡{𝐿𝐷_𝑡+1}

𝐿𝑡 ,

(8)

where 𝐿_𝑡 denotes the total loan stock and 𝐿𝐷_𝑡 denotes loan defaults. The loan stock carried over from one period to the next equals new loans minus loan defaults plus the part of the existing credit stock that does not mature

𝐿_𝑡 = 𝜆_𝐸𝐿_𝑡−1− 𝐿𝐷_{𝑡 + 𝐿} 𝑡

𝑛_{, 0 < 𝜆}

𝐸 < 1. (9)

10 𝐿𝐷_{𝑡 = (1 − 𝜌}

𝑙)𝜆𝐷𝐿𝑡−1+ 𝜌𝑙𝐿𝐷𝑡−1, 0 < 𝜌𝑙 < 1. (10)

Equation (10) assumes that the number of defaults is a fraction of the number of loans in the previous period and the number of defaults in the previous period, hence 𝜌 captures the persistence of defaults, and (1 − 𝜌) determines how fast the number of defaults returns to its equilibrium value after a shock. The advantage of this equation is that the AR(1) default process is stable.

The model is log-linearized around the steady state by means of a first order Taylor expansion to estimate the model using a linear estimation technique (Vector Autoregression). The derivations are in Appendix B. The log-linearized credit and default equations are4

𝑙_𝑡 = (1 + 𝜗1 𝜗₃𝜗₂) (𝜆𝐸𝑙𝑡−1− 𝜆𝐷𝑙𝑡𝐷+ 𝜗3 𝜗₁{𝜆𝑦𝑦𝑡𝑔− 𝑖𝑡𝑠𝑟 1 − 𝜆_𝐷+ 𝜗2𝑙𝑡𝐷}) + 𝜀𝑡𝐿, (11) 𝑙_𝑡𝐷 _{= (1 − 𝜌} 𝑙)𝑙𝑡−1+ 𝜌𝑙𝑙𝑡−1𝐷 + 𝜀𝑡𝐷, (12)

where 𝑙𝑡 and 𝑙𝑡𝐷 represent respectively the log-linearized versions of credit supply and loan

defaults, 𝑦_𝑡𝑔 ≡ 𝑦𝑡− 𝑦𝑡𝑝 is the log-linearized version of the output gap, 𝑦𝑡 and 𝑦𝑡𝑝 are respectively

the log-linearized version of actual and potential output, 𝜗1 ≡ 𝐴 − 𝜆𝑝(1 − 𝜆𝐷), 𝜗2 ≡ 𝜆𝑝𝜌𝑙𝜆𝐷 and

𝜗₃ ≡ (1 − 𝜆_𝐸+ 𝜆_𝐷). Hence, upper case letters denote level variables and lower case letters denote log-linearized variables.5 _{The term structure is given by log-linearizing Equation (4) and}

substituting the log-linearized solution for new loans

𝑖_𝑡𝑙𝑟 ₌ 𝑛 𝜗₁(1 − 𝜆𝐷)𝑖𝑡 𝑠𝑟₊𝜆𝑦 𝜗₁𝑦𝑡𝑔− 𝜗2 𝜗₁(𝑙𝑡− 𝑙𝑡𝐷), (13)

where 𝑖𝑡𝑙𝑟 denotes the log-linearized version of the long-term interest rate.

4 _{Adding a white noise term, 𝜀}

𝑡𝐷, to the default ratewithout violating the constraint that 𝑃𝑡 is bounded to the interval

(0,1) implies that the distribution of 𝜀𝑡𝐷 depends on 𝑃𝑡, see also Greenbaum, Kanatas and Venezia (1989) who assume

that 𝑃𝑡 follows a random walk process. 5 _{The short-term interest rate 𝑖}

𝑡𝑠𝑟 is not log-linearized, see Appendix B, because its steady state value is equal to zero.

11

To solve the model analytically one needs a definition for the short-term interest rate 𝑖𝑡𝑠𝑟

and the output gap 𝑦_𝑡𝑔. I assume a simple backward looking macro model often used to analyze monetary policy in VAR models. See for example Svensson and (1997), Giordani (2004), and Bernanke, Boivin and Eliasz (2005). The interest rate setting rule of the Central Bank is given by assuming that the monetary authority minimizes a loss function of the type:

𝐿𝑂𝑆𝑆_𝑡 = 𝐸_𝑡∑ 𝛽𝑖_[𝜃 𝑦(𝑦𝑡+𝑖𝑔 ) 2 + 𝜃_𝜋(𝜋_𝑡+𝑖)2_] ∞ 𝑖=0 , (14)

where the inflation target has been normalized to zero. The solution yields a standard Taylor rule Taylor (1993)

𝑖_𝑡𝑠𝑟 _{= 𝛾}

𝜋𝜋𝑡+ 𝛾𝑦𝑦𝑡𝑔+ 𝜀𝑡𝑀𝑃, (15)

where 𝜋𝑡 denotes the inflation rate. The IS curve is given by

𝑦_𝑡𝑔 = 𝛽_𝑦𝑦_𝑡−1𝑔 − 𝛽_𝑟(𝑖_𝑡−1𝑙𝑟 _{− 𝜋}

𝑡−1) + 𝜀𝑡𝐴𝐷, (16)

Equation (16) is derived by maximizing household utility with respect to consumption and leisure subject to the budget identity. The solution to this problem gives the well-known Euler equation see (Heijdra, 2009). The Euler equation is log-linearized around its steady state and after setting in the steady state actual output equal to consumption gives Equation (16). Potential output follows an AR(1) process

𝑦_𝑡𝑝 = 𝜌𝑦_𝑡−1𝑝 + 𝜀𝑡𝑇, 0 < 𝜌 < 1. (17)

Finally, the inflation rate is modeled as a Philips curve of the form

12

Equation (15)-(18) contain an error term which are introduced to shock the system. The shocks 𝜀_𝑡𝐿_,

𝜀_𝑡𝐷_{, ε} 𝑡 𝐴𝐷_{, ε} 𝑡 𝑇_{, ε} 𝑡 𝐶𝑃_{, ε} 𝑡 𝑀𝑃 _{𝑖𝑖𝑑~(0, 𝜎}

ε2𝑖) are called credit shock, provisioning shock, aggregate demand

(AD) shock, technology shock, cost push (CP) shock, and monetary policy (MP) shock, respectively.

The model can be summarized by the following reduced form VAR representation.

𝑍_𝑡 = 𝐴₁𝑍_𝑡−1+ 𝐴₀ε_𝑡, (19) where 𝑍_𝑡 = [𝑙_𝑡𝐷 _𝑦 𝑡𝑔 𝑦𝑡 𝜋𝑡 𝑖𝑡𝑠𝑟 𝑙𝑡] ′ , 𝑍𝑡−1 = [𝑙𝑡−1𝐷 𝑦𝑡−1𝑔 𝑦𝑡−1 𝜋𝑡−1 𝑖𝑡−1𝑠𝑟 𝑙𝑡−1]′, ε_𝑡 = [𝜀_𝑡𝐷 _𝜀 𝑡𝐴𝐷 𝜀𝑡𝑇 𝜀𝑡𝐶𝑃 𝜀𝑡𝑀𝑃 𝜀𝑡𝐿]′,

𝐴₀ and 𝐴₁ are specified in Appendix B.

Since 𝐴₀ is a lower triangular matrix I can estimate Equation (19) in reduced form and recover all structural shocks by a Cholesky decomposition of 𝐴₀𝐴₀′. 𝐴₀ and 𝐴₁ in Appendix B also show that a technology shock has no effect on the output gap, the inflation rate, the short-term interest rate, credit supply and defaults. For this reason it is common practice to eliminate the output variable from Equation (19) (Giordani, 2004).

Model properties

Provisioning, 𝑙_𝑡𝐷_{, is only contemporaneously affected by a provisioning shock and not by any other}

13

their loans in provisions. Hence, a positive credit shock results in an increase in provisioning during the next period and vice versa.

The output gap, 𝑦_𝑡𝑔, is only contemporaneously affected by an AD shock. There is considerable consensus in the literature that the output gap is only modestly affected by shocks of other variables, see for example Bernanke and Gertler (1995) and Christiano, Eichenbaum and Evans (1999). A Cholesky decomposition allows for a contemporaneous response of the output gap to a provisioning shock. The model predict that a provisioning shock results in an increase in the long-term interest rate which lowers the output gap during the next period, but predicts no contemporaneous effect of the output gap to a provisioning shock.

Prices, 𝜋𝑡, are only contemporaneously affected by a CP shock. The literature often

assumes that prices respond very sluggishly to shocks in other variables, see for example Bernanke and Gertler (1995) and Christiano, Eichenbaum and Evans (1999).

The short-term interest rate, 𝑖𝑡𝑠𝑟, is contemporaneously affected by an AD and CP shock

because the model assumes that the monetary authority reacts contemporaneously to a CP and AD shock as proposed by the Taylor rule. On the one hand this assumption is valid because the monetary authority is able to respond quickly to all information available. On the other hand, Christiano, Eichenbaum and Evans (1999) argue that this choice is certainly debatable because quarterly data on prices and output are typically only available with a lag. Nonetheless, presuming that monetary authorities typically have access to monthly data, Christiano, Eichenbaum and Evans (1999) argue that assuming that the monetary authority observes prices and the output gap when they set the interest rate is at least as credible as assuming that they do not. Besides, Leeper, Sims and Zha (1996) and more recently Sims and Zha (2006) show that the effects of a MP shock on output and prices is invariant to this assumption.

Credit supply, 𝑙𝑡, is contemporaneously affected by a provisioning, AD, CP, MP and credit

14

long-term interest rate through the risk premium which induces banks to increase credit supply. The net effect is ambiguous. The model also predicts that a positive AD shock contemporaneously increases credit supply because credit demand increases, but decreases credit supply because the short-term interest rate increases; the net effect is again ambiguous. Moreover, the model predicts that a positive CP shock contemporaneously decreases credit supply because short-term interest rate increases. Finally, the model predicts that a MP shock contemporaneously decreases credit supply. The increase in the short-term interest rate increases a bank’s financing costs which decreases credit supply.

4 Methodology

4.1 Panel VAR

To analyze the effects of monetary policy and credit supply on economic activity; I consider a reduced form panel VAR system. The advantage of using a VAR approach is that one does not need to make assumptions on the structure of the system (Sims, 1980). VAR models do not impose much structure ex-ante and treat all variables as endogenous. Since I am interested in the impulse responses functions after a shock rather than estimating the structural parameters itself, a VAR approach is attractive to analyze the effect of monetary policy on credit supply. I consider the following reduced form VAR

𝑋_𝑖𝑡 = 𝑐_𝑖+ Φ(𝐿)𝑋𝑖𝑡+ 𝑣𝑖𝑡, 𝐸{𝑣𝑖𝑡𝑣𝑖𝑡′} = 𝛴𝑣, (20)

where 𝑋_𝑖𝑡 is a vector of endogenous variables, 𝑖 is the country index, 𝑐_𝑖 is a vector of country-specific fixed effects, Φ(𝐿) is a lag polynomial Φ(𝐿) = 1 + Φ1𝐿 + Φ2𝐿2+ ⋯ + Φ𝑝𝐿𝑝, and 𝑣𝑖𝑡 is

a vector of reduced form residuals 𝑖𝑖𝑑~(0, 𝛴_𝑣). The vector 𝑋_𝑖𝑡 consists of the endogenous variables introduced in Section 3, 𝑋𝑖𝑡 = [𝑙𝑖𝑡𝑑, 𝑦𝑖𝑡𝑔, 𝜋𝑖𝑡, 𝑖𝑖𝑡𝑠𝑟, 𝑙𝑖𝑡]′.

15

that it implicitly imposes the same underlying structure to each country in the panel. See for a more in-depth discussion of cross-country heterogeneity in VAR models including credit variables Nikolay, Hülsewig and Wollmershäuser (2012). Gavin and Theodorou (2005) argue that the panel approach enables researchers to find the dynamic relationship that is common to all countries. Cross-country heterogeneity is allowed for by adding individual fixed effects; see Section 4.3.

4.2 Identification

It is not possible to estimate the IRFs of 𝑋𝑖𝑡 to the structural shocks, 𝜀𝑖𝑡, because the reduced form

disturbances, 𝑣_𝑖𝑡, represent the effect of all structural shocks in the economy. Therefore it is impossible to ascribe a particular structural shock in 𝜀𝑖𝑡, for example a MP shock, to 𝑣𝑖𝑡, see

Christiano, Eichenbaum and Evans (1999). To identify the structural shocks, it is common practice to assume first that the structural shocks are orthogonal, i.e. 𝛴_𝜀 is a diagonal matrix with the standard deviations on the diagonal. Second, one has to make an identification assumption to identify the relationship between the reduced form VAR disturbances, 𝑣_𝑖𝑡, and the structural shocks, 𝜀𝑖𝑡. At this point the model in Equation (19) can be used to identify this relationship by

assuming that 𝜀_𝑖𝑡 = 𝐴₀𝑣_𝑖𝑡~(0, 𝛴_𝜀 = 𝐴₀𝛴_𝑣𝐴₀′), where 𝐴₀ is the invertible square matrix in Equation (19) and 𝐸{𝜀𝑖𝑡𝜀𝑖𝑡′} = 𝛴𝜀. Since 𝐴0 is a lower triangular matrix the structural shocks in

Equation (19) can be identified by assuming a recursive system which imposes zero restrictions on all elements of 𝐴0 above the diagonal. The recursive system is also known as the Cholesky

Decomposition.

4.3 Estimation

I estimate Equation (20) using the Generalized Method of Moments (GMM).6 Love and Zicchino (2006) argue that the parameter structure most likely differs across countries. As a consequence, the restriction that the underlying structure is the same for each cross-sectional unit is almost certainly violated. One solution for incorrect parameter restrictions, implicitly imposed by using the panel VAR approach, is to include country fixed effects, 𝑐_𝑖. However, introducing fixed effects in a VAR model presents a second problem because the fixed effect terms are correlated with the

6 _{For more details about the estimation procure I refer to Love and Zicchino (2006), whose State code I thankfully}

16

regressors via the lags of the dependent variable. Common practice is to apply the mean-difference procedure which cancels out the fixed effect terms, but due to this correlation the coefficients are biased.

To overcome the correlation problem I follow Love and Zicchino (2006) and apply the forward mean-differencing procedure, i.e. Helmert procedure (Arellano and Bover, 1995). The Helmert procedure subtracts the mean of all future observations for each country and year. By applying this procedure, the orthogonality between the forward mean-differenced variables and the lagged dependent variables, is preserved. The lagged dependent variables can be used as instruments and the coefficients of the system can be estimated with GMM. Although the 𝐴₀ matrix in Equation (19) suggests over-identification restrictions, I estimate a ‘just identified’ version of the model using a Cholesky decomposition.7

5 Data

The sample ranges from 1970Q1:2011Q2 and includes 12 countries: Austria, Belgium, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, Spain, Sweden, United States. The data includes the following variables: output measured by real gross domestic product, inflation using the price deflator for private consumption, the short-term interest rate represented by the three month money market interest rate, credit supply measured as bank loans outstanding to the private sector, and provisions are measured by aggregate provisioning of the banking sector, for details see Table C.3 in the Appendix. The output gap is subsequently calculated as the difference between actual output and potential output the latter one is estimated using the Hodrick-Prescott filter on actual output.8

Expectations about loan defaults is a latent variable; I proxy this variable by bank provisioning. In general, provisions are an item on the balance sheet representing capital reserved to pay for anticipated future losses. Bank provisioning represent funds set aside to pay for anticipated loan defaults.9 The provisions data series starts, depending on the country, between

7 _{The software used to estimate Equation (19) does not allow over-identifying restrictions. It may, however, be}

interesting for future research to study whether the empirics confirm these over-identification restrictions.

8 _{The smoothing term which represents the extent to which a deviation from the trend value is punished in the loss}

function is set equal to 1600, as is common for quarterly data.

9 _{The provisions variable also includes provisioning for other banking activities, e.g. derivatives and securities.}

17

1979 and 1988 and ends either in 2008 or 2009. Table 5.1 reports the descriptive statistics of the provisions variable converted to US Dollars. The provisions series of France does not start before 1988, while some other series start in 1979. Also interesting is the difference in the mean between countries, e.g. Sweden reports to have on average only -$234.97 million provisions while the United States has on average $42095.91 million provisions. The next paragraph will elaborate on these large cross-country differences.

All data series are reported on a quarterly basis except for provisions which are reported on an annual basis. In order to preserve quarterly estimation I interpolate provisions using the quadratic match average conversion.10 Section 7.3 will show that the main results hold if the model is estimated using annual data.

Table 5.1 - Distribution of provisions observations over country.

Countries Years Mean St. Dev. Median Negative

Values Austria 1989-2008 2432.06 3151.09 1938.62 4 Belgium 1981-2009 1853.64 2570.59 1197.14 5 Denmark 1979-2009 1487.93 2362.51 700.51 2 Finland 1979-2009 160.91 163.85 135.55 23 France 1988-2009 10145.18 6593.50 7651.35 2 Germany 1979-2009 15485.98 11744.73 14395.03 0 Italy 1984-2009 9085.65 5303.31 8117.81 0 Japan 1989-2008 23029.63 22206.44 15575.27 0 Netherlands 1979-2009 3046.58 4650.29 1685.46 0 Spain 1979-2009 6369.87 8208.00 3885.30 0 Sweden 1979-2009 -234.97 2240.35 333.34 50 United States 1980-2009 42095.91 52509.01 27798.60 0 Total sample - 9312.59 21210.29 2380.71 86

Provisions are in millions of US dollars

Variable Transformations

The model presented in Section 3 is log-linearized around its steady state in order to estimate the model using a linear estimation technique. Following the theoretical model I transform the credit,

10 _{The interpolation is done using eViews. The program fits a local quadratic polynomial (for 3 adjacent points) for}

18

inflation, actual output and potential output variable by taking the first difference of the natural logarithm. Subtracting subsequently potential output from actual output gives the output gap. The short-term interest rate is not log-linearized because its steady state value is equal to zero, see Appendix B. However, for stationarity concerns which are discussed below, the short-term interest rate is first differenced.

How to transform the provisions variable is less clear because two problems with respect to this variable occur if one would like to estimate equation (20) in a panel VAR setting. First, Section 3 log-linearizes the provisions variable, however provisions are sometimes negative and taking the natural logarithm of a negative value is not possible. In Table 5.1, especially Finland and Sweden, where there were banking crises in these periods, are showing many negative values. Table 5.1 also shows a second problem, i.e. provisioning differs substantively between countries, estimating a panel VAR would not make sense when differences are too large. Both problems can be solved using a relative simple transformation, i.e. one could divide aggregate provisioning by the aggregate bank balance sheet of that particular country. In addition, the large cross-country differences are no longer presents because a correction has been made for the size of the banking sector. For stationarity concerns, which are discussed next, provisioning as percentage of the balance sheet is first differenced.

Table 5.2 - Descriptive Statistics

Quarterly Annual

Variable Transformation Obs. Mean

St.

Dev. Median Obs. Mean

St. Dev. Median 𝑙_𝑡𝑑 _Δ𝐿 𝑡 𝑑_𝐵𝑆 𝑡−1 1320 0.000 0.001 0.000 321 0.000 0.003 0.000 𝑦_𝑡𝑔 Δln (𝑌𝑡 Δln (𝑌_𝑡𝑃₎ 2158 0.000 0.009 0.000 558 0.000 0.409 0.008 𝜋𝑡 Δln(Π𝑡) 2097 0.011 0.012 0.008 527 0.013 0.554 0.000 Δ𝑖𝑡𝑠𝑟 2142 -0.045 1.226 -0.010 555 -0.005 6.423 -0.400 𝑙𝑡 Δln(L𝑡) 2145 0.022 0.019 0.022 492 0.086 0.062 0.087

19

Table 5.2 shows the summary statistics for all variables after the transformation. These

statistics are not only presented at a quarterly basis but also at an annual basis, because in Section 7.3 the model is re-estimates using annual data to analyze whether results are sensitive to the frequency of the data used. As expected the quarterly data series have a lower standard deviation than the annual data series.

This paragraph describes why the transformations summarized in Table 5.2 are not only essential to estimate the model using a linear estimation technique but also necessary for stationarity concerns. To test whether the series contain unit roots, I perform the Levin-Lin-Chu (LLC) test (Levin, Lin, and Chu, 2002) which is a panel data unit root test.11 The LLC test for panel data assumes a balanced panel and requires that the number of time periods is larger than the number of panels. Even though not all series in the sample are balanced, the number of time periods is larger than the number of panels which makes the LLC test suitable.12 The results of the LLC test suggest that all variables, except provisions, are stationary if panel means are subtracted, see Table C.1 in the Appendix. The provisions variable is stationary only if panel means are not subtracted suggesting that this variable does not contain a country specific constant. Hence, applying the fixed effect estimator is appropriate for all but the provisions variable. Table C.1 in the Appendix also reports the LLC tests statistics for the short-term interest rate and the provisions variable before they are first differenced. These results suggest at the 1% level that to estimate a stationary panel VAR it is essential to take the first difference of these variables.13

Figure 5.1 shows the credit growth rate, the output gap and provisions divided by the total balance sheet for 8 countries in the sample. Credit growth rates are usually positive and quite stable over time. During the recent Global Financial Crisis the credit growth rate declined in most countries, but returned to pre-crises levels within a few years. The output gap shows a sharp decline during the early years of the crisis. Moreover, it appears that the output gap decline preceded the decline in the credit growth rate. Figure 5.1 shows that provisioning levels are highly volatile.

11 _{The Levin-Lin-Chu (2002) (LLC) test pools cross-section time series data to test the 𝐻}

0: each individual time series

contains a unit root against 𝐻𝑎: all individual time series are stationary. The LLC test assumes the same autoregressive

parameter for all panels. For specifics of the unit root test the reader is referred to Levin, Lin, and Chu (2002).

12 _{Alternative panel tests as for example the Harris-Tzavalis test are more suitable for data with a large number of}

panels and relatively few time periods.

13 _{The LLC test statistic for the short-term interest rate has a p-value of only 0.02 (both panel means and a trend}

20

Figure 5.1 - Descriptive Graphs

Left axis growth rates credit and the output gap; right axis provisions; horizontal axis time. Credit, output and provisions are respectively calculated as ln ( 𝐿𝑡

𝐿_𝑡−1), 𝑦𝑡 𝑔 , and 𝐿𝑡 𝑑 𝐵𝑆_𝑡. -0.005 0 0.005 0.01 0.015 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 Austria -0.002 0 0.002 0.004 0.006 0.008 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 France 0 0.002 0.004 0.006 0.008 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 Germany 0 0.002 0.004 0.006 0.008 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 Italy 0 0.002 0.004 0.006 0.008 0.01 0.012 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 Japan 0 0.002 0.004 0.006 0.008 0.01 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 The Netherlands 0 0.005 0.01 0.015 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 Spain 0 0.005 0.01 0.015 0.02 0.025 -0.1 -0.05 0 0.05 0.1 70 75 80 85 90 95 00 05 10 United States

21

Especially during the years pre-Global Financial Crisis provisioning levels were historically low for most countries while during the Global Financial Crisis provisioning levels started to rise tremendously.

6 Empirical results

Section 6.1 presents the estimation results for the benchmark panel VAR, which consists of the following variables: the output gap, the inflation rate, the short-term interest rate and credit supply. Subsequently, in Section 6.2 the provisions variable is added to analyze the effects of a provisioning shock as specified in Equation (20). All shocks are labeled as specified in Section 3 and standardized to an economically interpretable size.14 The IRFs presented show the first 12 periods after the shock with the 5% and 95% confidence intervals generated by Monte-Carlo with 1000 iterations.15

The panel VAR is estimated including 3 lags based on the average of the Akaike and Schwarz information criteria for the individual time series; the results presented in Table C.1 of the Appendix suggest to include 1 to 5 lags.16 The model is estimated by including 3 lags which is in the middle of the range 1 to 5. There is no consensus in the literature which information criterion to use in selecting the appropriate lag length. Besides, the estimation results, which are not presented here but are available on request, prove to be rather robust to different lag length specifications. The general results are discussed first before I continue with the ones that are of my particular interest.

6.1 Benchmark model

Figure 6.1 presents the IRFs of the benchmark model. This section discusses the main consequences of an AD shock represented by a 0.01 percent point increase in the output gap, a CP

14 _{By default the shocks estimated, e.g. denoted by 𝜖}

𝑡, are one standard deviation shocks. The shocks and the

corresponding IRFs are divided by that variable’s standard deviation and multiplied by 𝑥 which is chosen to be of an easy interpretable size, i.e. 𝜀𝑡= 𝑥 (_𝜎𝜖𝑡2).

15 _{I experimented with a larger number of iterations and obtained similar results.}

16 _{Theoretical macroeconomist usually assume that there is at least some persistence in the shock term, hence they}

assume that 𝜀𝑡= 𝜌𝜀𝑡−1+ 𝜂𝑡, where 0 ≤ 𝜌 ≤ 1 is the autocorrelation parameter and 𝜂𝑡𝑖𝑖𝑑~(0, 𝜎η2). In equation (20)

𝜌 = 0 because Section 3 assumes that 𝜀𝑡𝑖𝑖𝑑~(0, 𝜎ε2). In order to introduce some persistence in the model after a shock

22

shock represented by a 0.1 percent point increase in the inflation rate, a MP shock represented by a 10 basis point increase in the short-term interest rate, and a credit shock represented by a 0.1 percent point increase in credit supply.

The consequences of a positive AD shock are a small increase in the inflation rate and a large increase in the short-term interest rate and credit supply. The output gap returns to its pre-shock level after 1 period while especially credit supply remains higher for more than 3 years. The result suggests that a temporary increase in the output gap has long lasting effects on credit supply. The main consequences of a positive CP shock are a large increase in the short-term interest rate and a small increase in credit supply. Nevertheless, the rise in credit supply should be interpreted with caution because the confidence intervals are wide and the increase is very small. The output gap appears to be unaffected by a CP shock.

The main consequences of a positive MP shock are a small decrease in the output gap and a small decrease in credit supply. The output gap appears to decrease in response to a MP shock and returns to its pre-shock level within a year suggesting that monetary policy can only affect output in the short run. Credit supply also decreases a bit and remains lower for more than 12 periods. The inflation rate rises, contrary to economic theory, during the first 3 periods after a MP shock. Economic theory argues that prices decline after a monetary contraction because the price of holding currency increases. The system produces the well-known price puzzle which is discussed in Section 7.1.

The consequences of a positive credit shock are an increase in the output gap, a long increase in the inflation rate, and a large increase in the short-term interest rate. The increase in the output gap after the credit shock is relatively persistent compared to the response of the output gap to a CP or MP shock. The result suggests that a credit shock could increase actual output relative to potential output for quite some time. The large increase in the short-term interest rate may be explained by the increase in the output gap and the inflation rate.

23

Figure 6.1 - Impulse response functions panel VAR: benchmark model.

Identification Cholesky decomposition; based on 3 lags; 5% and 95% confidence intervals on each side generated by Monte-Carlo with 1000 iterations; periods in quarters on the horizontal axis.

-0.05 0.00 0.05 0.10 0.15 0 3 6 9 12

Gap to Gap Shock

-0.02 -0.01 0.00 0.01 0.02 0 3 6 9 12

Gap to Inflation Shock

-0.002 -0.001 0.000 0.001

0 3 6 9 12

Gap to Interest Shock

-0.01 0.00 0.01 0.01

0 3 6 9 12

Gap to Credit Shock

-0.01 0.00 0.01 0.01 0.02 0 3 6 9 12

Inflation to Gap Shock

0.00 0.05 0.10 0.15

0 3 6 9 12

Inflation to Inflation Shock

-0.001 0.000 0.001 0.001

0 3 6 9 12

Inflation to Interest Shock

0.00 0.01 0.02

0 3 6 9 12

Inflation to Credit Shock

-2.00 0.00 2.00 4.00

0 3 6 9 12

Interest to Gap Shock

-2.00 0.00 2.00 4.00

0 3 6 9 12

Interest to Inflation Shock

-0.500 0.000 0.500 1.000 1.500 0 3 6 9 12

Interest to Interest Shock

-0.50 0.00 0.50 1.00

0 3 6 9 12

Interest to Credit Shock

0.00 0.02 0.04

0 3 6 9 12

Credit to Gap to Shock

0.00 0.01 0.02 0.03

0 3 6 9 12

Credit to Inflation Shock

-0.002 0.000 0.002

0 3 6 9 12

Credit to Interest Shock

0.00 0.05 0.10 0.15

0 3 6 9 12

24

6.2 Including provisioning for bad loans

Figure 6.2 presents the most important IRFs after adding provisioning as endogenous variable to the benchmark model. The provisioning shock is set equal to a 0.002 percent point increase in provisioning (rescaled by the aggregate bank balance sheet). Table 5.2 shows that the sample standard deviation is 0.001 percent point. Figure 6.2 shows that many countries experienced a large increase in provisioning at the beginning of the recent Global Financial Crisis. Figure C.1 in the Appendix shows for each country the remaining provisioning residuals after estimating Equation (20). Especially during the start of the recent Global Financial Crisis many countries experienced provisioning shocks close to 0.002. For these reasons the provisioning shock is set equal to a 0.002 percent point increase. This Section compares the IRFs of Figure 6.1 with the IRFs of Figure 6.2 and subsequently Section 6.3 compares the IRFs of Figure 6.2 with the theoretical model predictions.

The IRFs presented in Figure 6.2 are quite similar to the corresponding IRFs presented in Figure 6.1. Nevertheless, there are some clear differences. First, credit supply is not affected by a CP shock while in the benchmark model credit supply rises after a CP shock. Second, the inflation rate does not increase after a MP shock, while in the benchmark model the inflation rate rises after a MP shock (the price puzzle is no longer significant). Third, credit supply does not decrease after a MP shock as is the case in Figure 6.1. Fourth, the output gap is not significantly affected by a credit shock while in Figure 6.1 the output gap increases after a credit shock.

6.3 Comparison to theoretical IRFs

25

the long-term and short-term interest rate not incorporated in the credit demand equation.17 If a direct link exists a monetary contraction may increase the long-term interest rate independent of the behavior of an individual bank, this would give an incentive to supply more credit.

Figure 6.2 shows that credit supply increases after an AD shock. The model predicts that a positive AD shock increases loan demand which increases loan supply ceteris paribus. The monetary authority responds to a positive AD shock by increasing the short-term interest rate which increases a bank’s financing costs and decreases credit supply. The model is ambiguous about the net effect. The IRF in Figure 6.2 shows indeed that the short-term interest rate rises after an AD shock, while the previous paragraph discussed that credit supply is barely affected by a MP shock. It is therefore not surprising that credit supply increases after a positive AD shock. Apparently, especially the output gap, and therefore credit demand, is an important determined of credit supply while the costs of financing credit (the short-term interest rate and provisioning) appears to be less important.

The main consequences of a positive provisioning shock are a decrease in the output gap and a decrease in credit supply on impact. The model predicts a decrease in the output gap after a provisioning shock. A positive provisioning shock increases the long-term interest rate. The IS curve states that the output gap depends negatively on the long-term interest rate, so a provisioning shock decreases the output gap. The output gap indeed declines after a positive provisioning shock for more than 6 periods suggesting that provisioning shocks are important for economic growth.

Credit supply appears to decrease on impact after a positive provisioning shock. The model is ambiguous about the response of credit supply to a provisioning shock. On the one hand, a bank immediately writes off some of its outstanding loans which decreases credit supply, on the other hand, the long-term interest rate rises through the risk premium which increases credit supply. Figure 6.2 shows that credit supply decreases during the first 2 periods after the positive provisioning shock suggesting that on impact the negative effect is larger; after more than 2 periods the effect becomes insignificant.

17 _{This paper disconnects the short-term from the long-term interest rate while the conventional approach assumes a}

26

Figure 6.2 - Impulse response functions panel VAR: including defaults.

Identification Cholesky decomposition; based on 3 lags; 5% and 95% confidence intervals on each side generated by Monte-Carlo with 1000 iterations; periods in quarters on the horizontal axis.

Note: Prov. are Provisions in millions, Infl. is inflation, Intr. is the interest rate.

-0.001 0.000 0.001 0.002 0.003 0 3 6 9 12

Prov. to Prov. Shock

-0.002 -0.001 0.000 0.001

0 3 6 9 12

Prov. to Gap Shock

-0.002 0.000 0.002

0 3 6 9 12

Prov. Infl. Shock

-0.0002 -0.0001 0.0000 0.0001 0.0002 0 3 6 9 12

Prov. to Intr. Shock

-0.001 0.000 0.001

0 3 6 9 12

Prov. to Credit Shock

-0.004 -0.002 0.000 0.002

0 3 6 9 12

Gap to Prov. Shock

-0.05 0.00 0.05 0.10 0.15 0 3 6 9 12

Gap to Gap Shock

-0.04 -0.02 0.00 0.02 0.04 0 3 6 9 12

Gap to Infl. Shock

-0.002 -0.001 0.000 0.001 0.002 0 3 6 9 12

Gap to Intr. Shock

-0.01 0.00 0.01

0 3 6 9 12

Gap to Credit Shock

-0.002 -0.001 0.000 0.001

0 3 6 9 12

Infl to Prov. Shock

-0.01 0.00 0.01 0.02

0 3 6 9 12

Infl. to Gap Shock

0.00 0.05 0.10 0.15

0 3 6 9 12

Infl. to Infl. Shock

-0.001 -0.001 0.000 0.001 0.001 0 3 6 9 12

Infl. to Intr. Shock

-0.005 0.000 0.005 0.010

0 3 6 9 12

27

Figure 6.2 continued - Impulse response functions panel VAR: including defaults.

Identification Cholesky decomposition; based on 3 lags; 5% and 95% confidence intervals on each side generated by Monte-Carlo with 1000 iterations; periods in quarters on the horizontal axis.

Note: Prov. are Provisions in millions, Infl. is inflation, Intr. is the interest rate.

. -0.40 -0.20 0.00 0.20 0.40 0 3 6 9 12

Intr. to Prov. Shock

-1 0 1 2 3 0 3 6 9 12

Intr. to Gap Shock

-2.00 0.00 2.00 4.00

0 3 6 9 12

Intr. to Infl. Shock

-0.500 0.000 0.500 1.000 1.500 0 3 6 9 12

Intr. to Intr. Shock

-0.50 0.00 0.50 1.00

0 3 6 9 12

Intr. to Credit Shock

-0.004 -0.002 0.000 0.002 0.004 0 3 6 9 12

Credit to Prov. Shock

0.00 0.02 0.04 0.06

0 3 6 9 12

Credit to Gap to Shock

-0.02 0.00 0.02 0.04

0 3 6 9 12

Credit to Infl. Shock

-0.002 0.000 0.002

0 3 6 9 12

Credit to Intr. Shock

0.00 0.05 0.10 0.15

0 3 6 9 12

28

Provisioning appears to be unaffected by an AD, CP and MP shock but decreases after a credit shock. The model suggests, by construction, that provisioning increases after a positive credit shock because 𝜆_𝐷 > 0 meaning that banks hold a fixed percentage of their loans in provisions. The IRF suggests that provisioning decreases after a positive credit shock. A positive credit shock increases a bank’s balance sheet which decreases the ratio provisions over the total balances sheet. Hence, the result suggests that banks do not adjust their level of provisions after a positive credit shock. The finding is in line with exiting evidence in the literature, Cavallo and Majnoni (2001) find for non-G10 countries a negative correlations between pre-provisioning income and provisioning and Laeven and Majnoni (2003) present evidence that banks delay provisioning in good times until it is too late.

6.4 Variance decomposition

Table 7.1 shows the variance decomposition of the IRFs, i.e. the proportion of the forecast error variance that can be attributed to a particular shock 4, 8 and 12 periods ahead. In the benchmark model the variation in credit supply is explained for about 90% by a credit shock. Only an AD shock appears to play a significant role explaining about 5% of the variation in credit supply. Less than 2% of the variance in credit supply can be attributed to a MP shock. In the model that includes provisions an AD shock appears to explain about 10% of the variance in credit supply. A provisioning shock does not explain much of the variation in credit supply, however it appears that the consequence of excluding the provisions variable is an overestimations of the variance caused by a CP and MP shock and an underestimations of the variance caused by an AD shock. The variance in provisioning and the variance in output gap is primary caused by a provisioning shock and an AD shock respectively.

7 Robustness

29

Table 7.1

Variance Decompositions

Benchmark model Model including provisions

4 periods ahead Variables 𝑦𝑔 𝜋 𝑖𝑠𝑟 𝑙 𝑙𝑑 𝑦𝑔 𝜋 𝑖𝑠𝑟 𝑙 𝑙𝑑 _{99.0 0.3 0.1 0.4 0.3} 𝑦𝑔 _{97.7 0.4 1.5 0.3 0.6 97.3 0.9 1.1 0.0} 𝜋 1.0 96.4 0.3 2.4 1.0 2.3 94.8 0.1 1.8 𝑖𝑠𝑟 _{6.7 1.1 91.6 0.7 1.9 9.8 0.9 87.0 0.5} 𝑙 4.9 0.7 1.2 93.3 0.3 7.4 0.4 0.4 91.6 8 periods ahead Variables 𝑦𝑔 𝜋 𝑖𝑠𝑟 𝑙 𝑙𝑑 𝑦𝑔 𝜋 𝑖𝑠𝑟 𝑙 𝑙𝑑 _{98.1 0.3 0.2 0.5 1.0} 𝑦𝑔 _{97.5 0.6 1.6 0.3 1.0 96.7 1.2 1.2 0.0} 𝜋 2.2 88.3 0.2 9.3 0.8 2.8 89.8 0.1 6.5 𝑖𝑠𝑟 _{6.7 1.1 91.0 1.2 2.7 9.9 1.0 85.7 0.7} 𝑙 5.9 1.4 1.7 91.0 0.3 10.1 0.4 0.4 88.8 12 periods ahead Variables 𝑦𝑔 𝜋 𝑖𝑠𝑟 𝑙 𝑙𝑑 𝑦𝑔 𝜋 𝑖𝑠𝑟 𝑙 𝑙𝑑 _{97.6 0.3 0.2 0.5 1.3} 𝑦𝑔 _{97.4 0.7 1.6 0.3 1.0 96.5 1.3 1.1 0.1} 𝜋 3.1 80.5 0.3 16.1 0.8 3.4 85.1 0.1 10.7 𝑖𝑠𝑟 _{6.7 1.1 90.8 1.4 2.7 9.9 1.0 85.6 0.8} 𝑙 6.3 2.0 1.9 89.8 0.4 10.9 0.3 0.5 87.9

Notes: percentage variation of row variable explained by column variable 4, 8 and 12 periods ahead

7.1 Prize Puzzle

30

𝑡 + 1, but not inflation at time 𝑡. The monetary authority increases the short-term interest rate 𝑖𝑡𝑠𝑟

to decrease future inflation. Consequently at time 𝑡 𝜋_𝑡 > 0 and 𝑖_𝑡𝑠𝑟 _{> 0, if 𝐸}

𝑡{𝜋𝑡+1} is omitted than

𝑐𝑜𝑟𝑟(𝑖_𝑡𝑠𝑟_{, 𝜋}

𝑡) > 0, i.e. conditional correlation arises.

The conventional approach to solve the problem is to include commodity prices (Sims, 1992) because commodity prices are useful in forecasting inflation. Also other solutions have been proposed to solve the price puzzle. Pagan (2013) argues that the monetary base might be the omitted variable since the VAR estimated in Section 6 contains a money supply equation, the Taylor rule, but no money demand equation. Giordani (2004) states that including output in lieu of the output gap may cause prices to rise after a monetary contraction. I experimented with these suggestions, the results are not presented here but are available on request. First, including the commodity prices variable as endogenous variable to the panel VAR specified in Equation (20) does not change the main results presented in Section 6; hence this approach does not solve the price puzzle. Second, the estimated panel VAR as specified in Equation (20), contains the output gap and not actual output. Hence, the solution proposed by Giordani (2004) is also not able to solve the price puzzle. Third, the euro area countries share a common monetary base and 8 out of the 12 countries in the sample are euro area countries; for this reason it does not make sense to include the monetary base. It appears that none of the traditional solutions suffices to solve the price puzzle.

7.2 Sample period

31 for ‘normal’ times than for turmoil times.

7.3 Variables and frequency

This section discusses whether the choice of variables has any effect on the results presented. The analyses presents consecutively the choice of quarterly data, the output gap measure and the definition of the provisions variable. For conciseness only the most important IRFs are reported in Appendix C.

Annual data

In order to estimate the model using quarterly data the annual provisions series is interpolated. However, standard errors decrease with the number of observations; hence, t-statistics increase with the number of observations. To ensure that the interpolation does not drive the results presented in Section 6, the model is re-estimate using annual data. Figure C.3 in the Appendix shows the most important IRFs resulting from the panel VAR estimation using annual data.18 The IRFs show that the main results presented in Section 6 remain intact. Nevertheless, there are some differences. First, the decline in the output gap after a positive provisioning shock is about 10 times as large as the corresponding decline in Figure 6.2. Subsequently, it appears that 2 years after the positive provisioning shock the output gap rises above its pre-shock level suggesting that the medium-term effect of a positive provisioning shock on the output gap is positive. Second, the output gap increases after a credit shock and seems to remain at a higher level for many years which is similar to the results presented in Section 7.2.

Output gap

The Hodrick-Prescott (HP) filter has been criticized for structural reasons, e.g. Cogley and Nason (1995) show that the HP filter can generate business cycle dynamics even if the original data series does not contain any business cycle dynamics. Figure C.4 shows the most important results after substituting the output gap for actual output. Results are quite similar to those presented in Section 6, however, in line with the results presented in Section 7.2 and 7.3, output increases after a positive credit shock.

32

This section substitutes the HP filter output gap for a more sophisticated output gap measure of the OECD which is only available on an annual basis. Figure C.5 in the Appendix shows the estimation results. Again the IRFs are very similar to IRFs presented in Figure 6.2, however, the IRFs also show some noticeable differences. Similar to the IRF presented in Figure 6.2, the output gap declines after a positive provisioning shock, however the decline is much stronger. A 0.002 percent point increase in provisioning decreases the output gap for more than 5 years by approximately 0.25 percent point. Moreover, it appears that the output gap responds very strongly to a credit shock. A positive credit shock of 0.1 percent point increases the output gap up to 2 percent point for more than 6 years.19 The results should, however, be interpreted with caution since the confidence intervals are wide and not always different from zero.

Provisions

Section 6 uses provisioning to proxy the expected default rate; this paragraph analyzes whether the choice to proxy expected defaults by provisions has any effect on the results. Figure C.6 in the Appendix shows the estimating results after substituting provisions for non-performing loans as percentage of total loans, see Table C.3 in the Appendix. Note however, that the number of observations drops to 98 because non-performing loans data are only available at an annual basis and the series begins in 2009. Nevertheless, the results obtained after substituting provisions for non-performing loans are similar to the results presented in Section 6.

8 Conclusion

This paper presented a simple theoretical model to analyze the determinants and consequences of credit supply. The theoretical model introduced a banking sector with borrowers that were allowed to default. The equilibrium solutions have been embedded in a standard macroeconomic framework to be able to analyze the role of credit supply in the real economy. Subsequently, the model was summarized by a reduced form VAR representation and estimated using a panel VAR approach.

19 _{A 0.1 percent point credit shock is not an exceptionally large shock, Table 5.2 shows that the annual standard}

33

Overall the empirical results are in line with the theoretical model. First, the results show that a negative credit shock or a positive provisioning shock decreases the output gap. In addition, during turmoil times a positive provisioning shock may decrease credit supply and economic activity despite expansionary monetary policy. The results also show that if banks constrain credit supply, the output gap may decrease even more. Second, the empirical results show that a positive credit shock may decrease provisioning. Hence it appears that banks do not keep a fixed percentage of their credit in provisions, in contrast banks tend to take on more risk by provisioning less during economic booms. Third, the IRFs provide strong evidence that credit supply is primary determined by the output gap and therefore credit demand, while a MP or a provisioning shock, affecting a bank’s costs, appears to be less important.

The sensitivity analysis showed that results proved to be rather robust. Besides, the estimates with annual data showed a much stronger responsiveness of the output gap to a credit and a provisioning shock than the quarterly data estimates. Especially the more sophisticated OECD output gap measure showed that a negative credit shock and a positive provisioning shock may have severe consequences on economic growth.

The monetary authority may play an important role during turmoil times when provisioning is high, but credit supply and economic growth are low. Since the results presented show limited responsiveness of provisioning, the output gap, and credit supply to conventional measures, other channels may be worthwhile considering. It seems that an important role is reserved for unconventional monetary policy. Measures such as buying up bad loans as done by the Federal Reserve or offering warrants for future loan defaults may stimulate credit supply by directly lowering provisioning.

34

Finally, future research may extend this panel VAR analyses by allowing for over identification restriction and restrictions on the parameters. In addition a panel VAR approach using sign restrictions to identify the structural shocks may be an interesting avenue for future research to study whether zero restrictions affect the results presented here.

References

Adrian, T., & Shin, H. S. (2009). Money, liquidity, and monetary policy. American Economic

Review 99(2), 600-605.

Allen, F., & Carletti, E. (2008). The role of liquidity in financial crises. 2008 Jackson Hole Conference Proceedings, Federal Reserve Bank of Kansas City. 379-412.

Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of error-components models. Journal of Econometrics 68(1), 29-51.

Berger, A. N., Demirgüç-Kunt, A., Levine, R., & Haubrich, J. G. (2004). Bank concentration and competition: An evolution in the making. Journal of Money, Credit and Banking 36(3), 433-451.

Bernanke, B. S., & Gertler, M. (1995). Inside the black box: the credit channel of monetary policy transmission. Journal of Economic Perspectives 9(2), 27-48.

Bernanke, B. S., Boivin, J., & Eliasz, P. (2005). Measuring the effects of monetary policy: A factor-augmented vector autoregressie (FAVAR) approach. The Quarterly Journal of

Economics 120(2), 387-422.

Boivin, J., Kiley, M. T., & Mishkin, F. S. (2010). How has the monetary transmission mechanism evolved over time? Handbook of Monetary Economics forthcoming. Bolt, W., de Haan, L., Hoeberichts, M., Oordt, M. R., & Swank, J. (2012). Bank profitability

during recessions. Journal of Banking & Finance 36(9), 2552-2564.

Cavallo, M., & Majnoni, G. (2001). Do banks provision for bad loans in good times? The World

Bank Policy Research Paper 2619.

Chirinko, R. S., de Haan, L., & Sterken, E. (2008). Asset price shocks, real expenditures, and financial structure: a multi-country analysis. CESifo working paper No. 2342.

Christiano, L. J., Eichenbaum, M., & Evans, C. L. (1999). Monetary policy shocks: What have we learned and to what end? Handbook of Macroeconomics 1, 65-148.