Skip to: [Printable Version (PDF)] [Bibliography] [Footnotes]

Finance and Economics Discussion Series: Screen Reader version ^♣

Non-Core Bank Liabilities and Financial Vulnerability*

Joon-Ho Hahm

Yonsei University

Hyun Song Shin

Princeton University

Kwanho Shin

Korea University

August 27, 2011

Abstract:

A lending boom is reflected in the composition of bank liabilities when traditional retail deposits (core liabilities) cannot keep pace with asset growth and banks turn to other funding sources (non-core liabilities) to finance their lending. We formulate a model of credit supply as the flip side of a credit risk model where a large stock of non-core liabilities serves as an indicator of the erosion of risk premiums and hence of vulnerability to a crisis. We find supporting empirical evidence in a panel probit study of emerging and developing economies.

1 Introduction

Banks are the most important financial intermediaries in emerging and developing economies. As intermediaries who borrow in order to lend, banks must raise funding in order to lend to their borrowers. In an economy with domestic savers, the main source of funding available to the bank is the retail deposits of the household sector. However, retail deposits grow in line with the size of the economy and the wealth of the household sector. When credit is growing faster than the pool of available retail deposits, the bank will turn to other sources of funding to support its credit growth. If we classify retail deposits as the core liabilities of the banking sector and label the other components of bank funding as the non-core liabilities, then the ratio of the non-core to core liabilities will reflect the underlying pace of credit growth relative to trend and may be expected to give a window on the risk premiums ruling in the economy.

Our paper investigates the role of non-core banking sector liabilities in signaling financial vulnerability. There are two parts to our inquiry. First, we formulate a model of credit supply as the flip side of a credit risk model where a bank maximizes profit subject to a Value-at-Risk (VaR) constraint. The bank maintains a large enough capital cushion to limit the probability of failure to a fixed threshold. When measured risks are low, the bank can expand lending without violating its VaR constraint, leading to higher credit supply to the economy, with consequent impact on the risk premium implicit in the price of credit. When core deposits are "sticky"and do not grow in line with credit supply, the liabilities side of banks' balance sheets will be filled with non-core funding from the capital market. In this way, a higher incidence of non-core funding will be associated with above-trend growth in credit and compressed risk premiums.

The second part of our paper is an empirical investigation where we put the main prediction of our model to the test. We conduct a panel probit study of the susceptibility of emerging and developing economies to a financial crisis using the non-core liabilities of the banking sector as the conditioning variable. We find evidence that various measures of non-core liabilities, and especially the liabilities to the foreign sector, serve as a good indicator of the vulnerability to a crisis, both of a collapse in the value of the currency as well as a credit crisis where lending rates rise sharply.

In formulating our model of credit supply as the flip side of a credit risk model, our approach rests on the corporate finance of bank balance sheet management. In textbook discussions of corporate financing decisions, the set of positive net present value (NPV) projects is often taken as being exogenously given, with the implication that the size of the balance sheet is fixed. Leverage increases by substituting equity for debt, such as through the an equity buyback financed by a debt issue, as depicted by the upper panel in Figure 1.

However, the upper panel in Figure 1 turns out not to be a good description of the way that the banking sector leverage varies over the financial cycle. The distinguishing feature of the banking sector leverage cycle is that leverage fluctuates through fluctuations in the total size of the balance sheet with equity being the pre-determined variable. Hence, leverage and total assets tend to move in lock-step, as depicted in the lower panel of Figure 1 ¹

Figure 1: Two Modes of Leveraging Up. In the upper panel, the firm keeps assets fixed but replaces equity with debt. In the lower panel, the firm keeps equity fixed and increases the size of its balance sheet.

Figure 1: Two Modes of Leveraging Up. This is 2 flow charts. In the upper panel, labeled (Increased Leverage with Assets Fixed) there are two squares of equal size. The first square, on the left, is split vertically down the center with the left portion shaded gray and labeled (Assets), with an (A) above its half of the square. The right portion has an L above its half of the square and is split horizontally into two sections , the top section labeled (Equity) and the bottom (Debt). The (Equity) portion is a bit larger than the (Debt) portion. There then is a green arrow pointing to another square of the same format, except that the Equity portion is much smaller. In the lower panel, labeled (Increased Leverage via Asset Growth) there is another square of the same format as the one on the upper left , except that (Equity) is shaded gray instead of (Assets). There is another green arrow, pointing to a rectangle of the same format, so that the height of the original square almost doubles and the areas of both the (Assets and Debt) portions are enlarged.

Banks and other financial intermediaries' lending depends on their "balance sheet capacity" . Balance sheet capacity, in turn, depends on two things - the amount of bank capital and the degree of "permitted leverage" as implied by the credit risk of the bank's portfolio and the amount of capital that the bank keeps to meet that credit risk. Bank lending expands to fill up any spare balance sheet capacity when measured risks are low. Since the balance sheet constraint binds all the time, lending expands in tranquil times so that the risk constraint binds in spite of the low measured risks.²

By addressing the up-phase of the financial cycle, and the potential for the compression of risk premiums during lending booms, our approach differs from models of leverage constraints or collateral constraints that bind only in the downturn. In such models, lending is always below the first best. As well as on the downturn, our focus is on the up-phase of the cycle when risk premiums become compressed, leaving the economy vulnerable to a potential reversal.

Our model is not sufficiently refined to address issues of the optimal level of risk premium or quantity of credit. However, the model delivers the feature that a large stock of non-core liabilities of the banking sector will be associated with compressed risk premiums in the market for bank credit - a feature that proves useful in our empirical investigation. We conduct a panel probit investigation for the incidence of financial crises in a large sample of emerging and developing economies and find that non-core bank liabilities do, indeed, have explanatory power for subsequent crises.

Figure 2: Lending Boom Financed by Non-Core Liabilities. This figure depicts the banking sector balance sheet before and after a credit boom. Increased lending during a credit boom is financed by non-core liabilities.

Figure 2: Lending Boom Financed by Non-Core Liabilities. This is a flow chart split into two sections. The upper panel is labeled (After Lending Boom) and starts on the right with two rectangles of equal size, the upper one shaded and labeled (Foreign Creditors) the lower one labeled (Domestic Depositors). Both are connected to gray arrows which point at a circle to the left. The circle is shaded gray, with a non-shaded smaller circle within it, tangent to the bottom of the outer circle, and labeled (Banking Sector). There are two more gray arrows coming out of the circle from the left, and pointing at two rectangles of equal size stacked on top of each other. The top rectangle is shaded and labeled (New Borrowers) the bottom is labeled (Borrowers). The lower panel is labeled (Before Lending Boom). It starts with one rectangle on the right labeled (Domestic Depositors) which is connected to a gray arrow pointing left to a circle labeled (Banking Sector). It is the same size as the smaller (Banking Sector) circle contained within the shaded circle in the upper panel. This circle is connected to another gray arrow which points left to a rectangle labeled (Borrowers).

Figure 2 is a schematic illustration of the build-up of vulnerabilities associated with the growth of non-core liabilites. The bottom panel is the banking sector before a credit boom, while the top panel illustrates the system after the boom. As traditional deposit funding does not keep up with the credit growth, the banking sector's expansion is funded by non-core liabilities (in this case, from foreign creditors), building up vulnerabilities to deleveraging by foreign creditors.

Figure 3: Non-Core Liabilites of Korean Banks. Panel on right plots six categories of non-core liabilities of Korean banks measured in Korean Won. Panel on the left plots the non-core series as a fraction of M2. Source: Bank of Korea and Shin and Shin (2010) Figure 3 is an illustration from Korea.

$Figure 3: Non-Core Liabilites of Korean Banks. This figure has two charts. The one on the left is labeled (Non-Core Liabilities as Fraction of M2). The y-axis measures fraction and ranges from .10 to .55. The x-axis ranges from Jan-91 to Jan-10 with a label each January. The series are plotted with small green circles, connected with a black line, although the points are regularly too close to see the line. The series increases from about .16 to a peak of .32 in Jan-98, which is labeled and the observation is a red circle instead of green. After Jan-98 the series experiences a sharp drop to a low of .15 in mid 1999-2000. The series starts to increase in 2001 and rises steadily to a peak of .50 in Jan-09, which is also labeled and marked in red. After Jan-09 it drops again, the last data point being around .35 in the beginning of 2010. The chart on the right is labeled (Non-Core Liabilities). It is a line chart that plots the six categories of non-core liabilities, measured in Trillion Korean Won ranging from 0 to 800 on the y-axis. The x-axis ranges from 1991 to the start of 2010 with labels at January every two years. The six categories are [Other] FX borrowing in red, (Lf) Debt Securities in yellow, (Lf) Repos in green, (M2) Promissory Note 2 in white, (M2) Promissory Note 1 in purple, and (M2) Certificate of Deposit in blue. The series increases from around 25 trillion won at the start of 1991 to about 175 won in 1998. It then decreases slightly but from Jan-01 steadily grows from 100 to a peak of just over 7-won in 2008, after which it sharply falls, ending at around 500 at the start of 2010. The red category fills under the series just described, and the rest of the colors follow in the order introduced. The red and yellow fill most of the area under the series, with the yellow following a similar trend series described, though with less dramatic peaks. The white and purple can barely be distinguished, and the green and blue remain relatively constant over the period, although the green is more present from 1998 to 2009, and the blue grows from 2003 to 2010.$

The right panel of Figure 3 plots six categories of non-core liabilities of the Korean banking sector, taken from Shin and Shin (2010). It is notable how the first peak in non-core liabilities coincides with the 1997 crisis. After a lull in the early 2000s, non-core liabilities increase rapidly in the run-up to the 2008 crisis.³ The left panel of Figure 3 is the plot of non-core liabilities as a fraction of M2, and highlights the highly procyclical nature of non-core liabilities. There is substantial variation in the ratio of non-core liabilities to M2, ranging from around 15% of M2 to a peak of 50% at the height of the 2008 crisis following the bankruptcy of Lehman Brothers.

There is an extensive literature on leading indicators of emerging market financial crises. Using a panel of over 100 developing countries from 1971 to 1992, Frankel and Rose (1996) find that currency crises tend to occur when output growth is sluggish, domestic credit growth is high, foreign interest rates are high, and the ratio of FDI to debt is low. Kaminsky and Reinhart (1999) explored the linkages between banking crises and currency crises, and found that financial liberalization and capital inflows, credit booms, and an overvalued currency often precede "twin crises" that combine banking and currency crises.⁴

Drawing on the earlier literature, Goldstein, Kaminsky and Reinhart (2000) conducted a comprehensive battery of empirical tests for the effectiveness of early warning systems that rely on macroeconomic (and some microeconomic) variables at various frequencies. Using the "signals" methodology of comparing Type I and Type II errors, they conclude that many of their in-sample leading indicators remain effective in out-of-sample analyses.

The recent global financial crisis has also stimulated renewed interest in measuring vulnerability. However, the fact that the crisis affected advanced and emerging economies alike, with outwardly disparate causes in the two groups, has meant that consistent indicators of vulnerability have been rare. Claessens et al. (2010) examine many candidate indicators of vulnerability but find support only for house price appreciation, current account deficits and bank credit growth. Using a Multiple Indicator Multiple Cause model based on 107 country data, Rose and Spiegel (2008, 2010) find that commonly cited causes of financial crises implicating a host of variables - macroeconomic, financial conditions, regulatory, and institutional - are in fact only weakly related to the incidence of crises, leading them to somewhat more skeptical conclusions on the usefulness of early warning systems.

Our objective differs from these earlier papers. Our motivation is primarily to draw attention to the role of the intermediary sector in driving fluctuations in risk premiums. For this reason, we employ only a small selection of variables motivated by the theory, and we do not attempt to maximize goodness of fit by employing a large number of explanatory variables from disparate categories.

Nevertheless, the empirical performance of non-core liabilities measures is encouraging and gives some cause for optimism that more elaborate versions of such models may be a useful input into early warning exercises. In any case, we note that previous research on forecasting crises did not focus explicitly on fluctuation of non-core bank liabilities as a potential indicator of financial vulnerability, focusing instead on the asset side of the banking sector balance sheet, such as on credit growth or credit to GDP ratios. Although our non-core liability measures are closely related to asset side measures, we show that they carry considerable information value over and above credit aggregates.

Liabilities of banks to the foreign sector constitute a major component of non-core bank liabilities in many emerging market countries as the domestic wholesale bank funding market is not sufficiently developed to support rapid bank lending growth. Earlier empirical studies cited above have examined the size and maturity structure of aggregate external debt positions - an example being the ratio of short-term external debt to official foreign exchange reserves. These ratios were employed as an indicator of vulnerability to foreign exchange liquidity shocks. Our contribution is to point to the banking sector as the likely engine of accumulating vulnerability.

Our study builds on Shin and Shin (2010), who laid out the conceptual distinction between core- and non-core banking sector liabilities, and how these aggregates relate to traditional monetary aggregates. Using Korean bank data, this earlier study finds that non-core bank liabilities as defined as the sum of foreign exchange liabilities and wholesale bank funding are associated with vulnerability to sharp depreciation of the Won and increased borrowing spreads. Hahm, Mishkin, Shin and Shin (2010) further elaborate on the role of non-core bank liabilities as an indicator of financial procyclicality. Using more disaggregated series by claim-holders of non-core liabilities in Korea, they find that, relative to core liabilities, non-core bank liabilities are more procyclical on various measures. Drawing on these earlier studies, the objective of our empirical analysis is to explore the potential usefulness of non-core bank liabilities as conditioning variables in a panel probit study of potential vulnerability of emerging economies to financial crises.

The outline of the paper is as follows. We begin in the next section by formulating our model of credit supply based on the Vasicek (2002) model of credit risk, and draw implications on the relationship between credit, non-core liabilities and risk premiums in the bank credit market. We then follow with our empirical investigation by conducting a panel probit study of financial crises in emerging and developing economies using the IMF's International Financial Statistics (IFS) data. In order to allow for persistent heterogeneity across countries in our sample, we use the random effects version of the panel probit model, and confirm the strong explanatory role of non-core banking sector liabilities in explaining crises.

2 Model

Our model is a static model of credit supply with two dates - dates 0 and . Loans are made at date 0 and repaid at date 1. A bank makes loans financed from three funding sources - the bank's equity , its deposits and its non-core liabilities, denoted by . The notation for the components of the bank's balance sheet is given as in Figure 4.

The bank's equity and total deposit funding are both fixed. Deposits are fully insured by the government, and so earn the risk-free rate of return, which we set to zero. Total lending satisfies the balance sheet identity:

$\begin{displaymath} L=E+D+N \end{displaymath}$

(1)

The bank has a well-diversified loan portfolio consisting of loans to many borrowers, and credit risk follows the Vasicek (2002) model, which is the basis for the Basel capital requirements (BCBS (2005)). Borrower repays the loan when $Z_{j}>0$ , where $Z_{j}$ is the random variable given by

$\begin{displaymath} Z_{j}=-\Phi ^{-1}\left( \varepsilon \right) +\sqrt{\rho }Y+\sqrt{1-\rho }% X_{j} \end{displaymath}$

(2)

where $\Phi \left( .\right)$ is the c.d.f. of the standard normal, $% \varepsilon$ is the probability of default on the loan and

and $% \left \{ X_{j}\right \}$ are mutually independent standard normal random variables.

is the common factor that drives credit risk while each $% X_{j}$ are the idiosyncratic component of credit risk for the particular borrower

. The parameter $\rho \in \left( 0,1\right)$ is the exposure of each loan to the common factor

. To verify that $\varepsilon$ is the probability of default, note that

$\begin{eqnarray*} \Pr \left( Z_{j}<0\right) &=&\Pr \left( \sqrt{\rho }Y+\sqrt{1-\rho }% X_{j}<\Phi ^{-1}\left( \varepsilon \right) \right) \ &=&\Phi \left( \Phi ^{-1}\left( \varepsilon \right) \right) =\varepsilon \end{eqnarray*}$

Conditional on the common factor , defaults are independent. Denote the loan interest rate as so that the notional value of assets (the amount due to the bank at date 1) is $\left( 1+r\right) L$ .

Figure 4: Balance Sheet of Bank

By the law of large numbers, the realized value of the loan book at date 1 is the random variable $% w\left( Y\right)$ defined as:

$\displaystyle w\left( Y\right)$	$\textstyle \equiv$	$\displaystyle \left( 1+r\right) L\cdot \Pr \left( Z_{j}\geq 0\vert Y\right)$
	$\textstyle =$	$\displaystyle \left( 1+r\right) L\cdot \Pr \left( \sqrt{\rho }Y+\sqrt{1-\rho }X_{j}\geq \Phi ^{-1}\left( \varepsilon \right) \vert Y\right)$
	$\textstyle =$	$\displaystyle \left( 1+r\right) L\cdot \Phi \left( \tfrac{Y\sqrt{\rho }-\Phi ^{-1}\left( \varepsilon \right) }{\sqrt{1-\rho }}\right)$	(3)

The quantiles of the asset realizations can be derived as follows. The c.d.f. of the realized value of the loan portfolio at date 1 is given by

$\displaystyle F\left( z\right)$	$\textstyle =$	$\displaystyle \Pr \left( w\leq z\right)$
	$\textstyle =$	$\displaystyle \Pr \left( Y\leq w^{-1}\left( z\right) \right)$
	$\textstyle =$	$\displaystyle \Phi \left( w^{-1}\left( z\right) \right)$
	$\textstyle =$	$\displaystyle \Phi \left( \tfrac{\Phi ^{-1}\left( \varepsilon \right) +\sqrt{1-\rho }% \Phi ^{-1}\left( \frac{z}{\left( 1+r\right) L}\right) }{\sqrt{\rho }}\right)$	(4)

As prescribed by the Basel capital requirements (BCBS (2005))⁵, the bank follows the Value-at-Risk (VaR) rule of keeping enough equity to limit the insolvency probability of the bank to be some small $\alpha >0$ . We impose the condition that $\alpha <\varepsilon$ . That is, the bank defaults with a smaller probability than an individual borrower.⁶ The bank is risk-neutral otherwise. The bank's objective is to maximize expected profit subject only to its Value-at-Risk constaint.

Figure 5: Probability density of $w\left( Y\right)$

Figure 5: Probability density of w(Y). Both axes start at 0, y-axis labeled Density over repayments w(Y). A blue line moves along zero on the x-axis until about half way, and then slowly begins to rise, to a green vertical line, labeled under the axis (D+(1+f)N.) At this point the line rises more rapidly until it peaks at the top of the plot space, and then drops down to the end of the x-axis, which is labeled (1+r)L. The area under the blue line and left of the green line is labeled with an α.

The bank remains solvent as long as the realized value of $w\left( Y\right)$ is above its notional liabilities at date 1. Since the interest on deposits is zero while the funding rate on non-core liabilities⁷ is , the notional liability of the bank at date 1 is

$\begin{displaymath} D+\left( 1+f\right) N \end{displaymath}$

(5)

The optimal size of the loan book

for the bank keeps the insolvency probability at $\alpha$ , as illustrated in Figure 5. If

, then the shortfall in funding is made up by borrowing in the wholesale market. The bank's use of wholesale funding

and its loan supply

therefore satisfies:

$\begin{displaymath} \Pr \left( w<D+\left( 1+f\right) N\right) =\Phi \left( \tfrac{\Phi ^{-1}\left( \varepsilon \right) +\sqrt{1-\rho }\Phi ^{-1}\left( \frac{% D+\left( 1+f\right) N}{\left( 1+r\right) L}\right) }{\sqrt{\rho }}\right) =\alpha \end{displaymath}$

(6)

Re-arranging (6), we can derive an expression for the ratio of notional liabilities to notional assets.

$\begin{displaymath} \frac{\text{Notional liabilities}}{\text{Notional assets}}=\frac{D+\left( 1+f\right) N}{\left( 1+r\right) L}=\Phi \left( \dfrac{\sqrt{\rho }\Phi ^{-1}\left( \alpha \right) -\Phi ^{-1}\left( \varepsilon \right) }{\sqrt{% 1-\rho }}\right) \end{displaymath}$

(7)

We use the notational shorthand:

$\begin{displaymath} \varphi \left( \alpha ,\varepsilon ,\rho \right) \equiv \Phi \left( \tfrac{% \sqrt{\rho }\Phi ^{-1}\left( \alpha \right) -\Phi ^{-1}\left( \varepsilon \right) }{\sqrt{1-\rho }}\right) \end{displaymath}$

(8)

Clearly, $\varphi \in \left( 0,1\right)$ . Our condition that $\alpha <\varepsilon$ ensures that the expression inside $\Phi \left( .\right)$ in (8) flips sign from negative to positive as $\rho$ increases from zero to one. Figure 6 plots the notional debt to assets ratio $\varphi$ as a function of the common risk factor $\rho$ . The Value-at-Risk threshold level is fixed at $\alpha =0.1\%$ . The dark line is when the default probability $\varepsilon$ is 1%, while the light line is when $\varepsilon$ is 0.5%. We see that the debt to assets ratio is decreasing in both $\rho$ and $\varepsilon$ . Since the bank's leverage is monotonic in $\varphi$ , leverage declines in $\rho$ and $\varepsilon$ .

Figure 6: Plot of notional debt to assets ratio $\protect\varphi \left( \protect\alpha , \protect\varepsilon ,\protect\rho \right)$ . This chart plots $\varphi$ as a function of $\rho$ with $\protect\alpha =0.001$ . Dark line is when $\protect\varepsilon =0.01$ . Light line is when $\protect \varepsilon =0.005$ .

Figure 6: Plot of notional debt to assets ratio φ(α, ε, ρ). This is a line graph on a gridded plot. The y-axis ranges from 0.0 to 1.0 and there is a horizontal gridline at each .1 mark. The x-axis has the same labeling, with vertical gridlines at each .1. A dark line starts at the top left of the plot at (0.0, 1.0) and slopes down in a concave manner to about (0.9, .025). From that point the slope decreases and is constant until the end point at (1.0, 0.0). A light gray line starts at the same point at the top left (0.0, 1.0) and ends at (1.0, 0.0) and follows the same concave shape except it is to the right of the dark line, thus more distended from the origin.

From (7), we can solve for the bank's stock of non-core liabilities .

$\begin{displaymath} N=\frac{\varphi \left( 1+r\right) \left( E+D\right) -D}{1+f-\varphi \left( 1+r\right) } \end{displaymath}$

(9)

Using the balance sheet identity , we can also solve for the bank's loan supply function

$\begin{displaymath} L_{S}\left( r\right) =\frac{E+\frac{f}{1+f}\cdot D}{1-\frac{1+r}{1+f}\cdot \varphi } \end{displaymath}$

(10)

Note that the loan supply by the bank is increasing in and . Loan supply is well-defined only when $1+f>\varphi \left( 1+r\right)$ . Loan supply goes to infinity as the ratio $\left( 1+f\right) /(1+r)$ approaches $\varphi$ .

Figure 7: Loan Supply L_S(r)

Since the probability of default is $\varepsilon$ , the expected profit to the bank from one dollar's worth of loans is

$\begin{displaymath} \left( 1-\varepsilon \right) \left( 1+r\right) -1 \end{displaymath}$

(11)

Since the bank maximizes expected profit, its loan supply is zero if

falls below $\varepsilon /\left( 1-\varepsilon \right)$ . Otherwise, it will supply the full amount of loans constrained only by the VaR constraint (6). Figure 7 plots the loan supply curve of the bank as a function of the loan interest rate

. Note that the loan supply is zero if $r<\varepsilon /\left( 1-\varepsilon \right)$ , and goes to infinity as

approaches the asymptote $\left( \left( 1+f\right) /\varphi \right) -1$ from below. We summarize our results as follows.

Proposition 1 Non-core funding is increasing in $\varphi$ and decreasing in .

Corollary #.:

Bank credit supply

is increasing in $\varphi$ and decreasing in

Corollary 1 follows from the balance sheet identity and the fact that bank equity and deposit funding are fixed, so that total credit and non-core funding move together.

Credit market clearing determines the equilibrium loan rate , and hence the risk premium. Denoting loan demand as $L_{D}\left( r\right)$ , the equilibrium condition for the loan market is

$\begin{displaymath} L_{D}\left( r\right) =\frac{E+\frac{f}{1+f}\cdot D}{1-\frac{1+r}{1+f}\cdot \varphi } \end{displaymath}$

(12)

The market-clearing condition (12) determines the equilibrium loan rate . Since the default probability of loans is $\varepsilon$ , the risk premium in the credit market is given by

$\begin{displaymath} \pi \equiv \left( 1-\varepsilon \right) \left( 1+r\right) -1 \end{displaymath}$

(13)

When

is high, loan supply is high and hence the risk premium $\pi$ is low. For fixed $\varepsilon$ , the risk premium is monotonic in the lending rate

, so that the comparative statics of the risk premium inherit the comparative statics of the total credit supply given by Corollary 1.

Proposition 2 The risk premium $\pi$ is low when is high. The risk premium increases when the funding rate increases, or when $\varphi$ falls.

In a credit boom when the systematic risk factor $\rho$ is small, the measured risks in the loan portfolio is low, implying that less equity is needed to meet the bank's Value-at-Risk constraint, allowing the bank to increase its lending funded by an expansion in its wholesale funding . In Figure 5, a decrease in $\rho$ implies the shrinkage of the size of the left tail of the density of repayments, meaning that the bank can have a larger loan book for any given equity base .

Figure 8: Effect of rise in funding rate

Also, during a period of permissive funding conditions when the funding rate is low, the bank can maintain a larger stock of non-core liabilities .

However, after such a period of permissive financial conditions, risk premiums are low and the sector is vulnerable to a shock that reverses the permissive financial conditions. When eventually a shock arrives that either increases $\rho$ , or when the funding rate increases due to an overall deterioration of the wholesale funding market, there will be a sharp contraction in the stock of wholesale funding and in overall lending. Figure 8 shows the effect of a sharp increase in the funding rate . The increase in the funding rate shifts up the loan supply curve of the bank. For any given loan demand curve $L_{D}\left( r\right)$ , the shift upward in loan supply results in a sharp decrease in credit and in the use of non-core liabilities .

In open emerging economies, a substantial fraction of the non-core liabilities of the banks are foreign exchange-denominated liabilities, often short-term. Therefore, a sharp reduction in will be associated with capital outflows through the contraction of banking sector debt, and a depreciation of the domestic currency.

3 Evidence from Panel Probit

3.1 Data Description and Methodology

The primary data source for our study is the IMF's International Financial Statistics (IFS) database, focusing on the banking sector indicators at the country level. Although the country coverage of the IFS data is broad, the range of variables that can serve as the empirical counterpart of our non-core concept is somewhat limited. The IFS database lists 105 countries that have measures of banking sector liabilities to the foreign sector, 60 countries with liabilities of banks to non-bank financial sectors (with 50 that have both), and only 14 countries that list bonds issued by banking institutions. We also examine the difference M3 - M2 between two measures of broad money, M3 and M2 (reported by 64 countries) in order to get another fix on non-core liabilities of the banking sector.⁸ The sample period spans January 2000 to December 2010. All variables are monthly except for the credit to GDP ratio, which is annual. All missing values are replaced by using the linear interpolation method.

As measures of core liabilities, we sum demand deposits (reported by 121 countries), time, savings and foreign currency deposits (120 countries) and restricted deposits (80 countries). As an alternative, we use monetary aggregates M1 (reported by 120 countries) and M2 (120 countries). Eurozone countries do not report separate monetary aggregates and hence are excluded here. Use of the M3 measure in our study also reduces our sample. For instance, Korea does not figure in the regressions below, as it does not report M3 within IFS.

To investigate the predictive power of non-core bank liabilities for impending financial crises, we use two definitions of crises - currency crises and credit crises.

Currency crises are episodes where the value of the local currency drops abruptly and substantially. Following Frankel and Rose (1996) we define a currency crisis in terms of a currency depreciation of more than 25% in one year, and where the depreciation is at least 10% more than the depreciation in the previous year. That is

$\displaystyle \ln e_{t}-\ln e_{t-12}$	$\textstyle \geq$	$\displaystyle 0.25$	(14)
$\displaystyle \left( \ln e_{t}-\ln e_{t-12}\right) -\left( \ln e_{t-12}-\ln e_{t-24}\right)$	$\textstyle \geq$	$\displaystyle 0.10$	(15)

The second condition was introduced by Frankel and Rose (1996) to take account of countries that undergo rapid but steady depreciation due to high inflation.

The credit crisis definition uses two alternative measures of financial distress, associated with sharply higher market interest rates. In the first measure, we use the money market interest rate, and define a credit crisis as an episode where the money market rate reaches a level that is in the top 3% tail of the pooled in-sample distribution. We define such an episode as an instance of Credit Crisis I. Defining a crisis by means of the pooled in-sample tail rules out the case where no country undergoes a crisis during the sample period. However, since our sample period covers the turbulent times of the recent global financial crisis, our definition did not result in unreasonable classification upon a closer look at the data series.

Our second definition of a credit crisis appeals to the spread between the money market interest rate and the local treasury bill rate, and defines a credit crisis as an episode where the spread between the money market rate and the treasury rate lies in the top 3% of the pooled in-sample distribution across all countries. We call this variable Credit Crisis II.

A more standard measure of credit crisis would have been in terms of the spread between the local risk-free rate and the local rate on private liabilities, but data limitations due to the sample of countries examined in our study precludes the use of this more standard (and desirable) measure. ⁹

Once the crisis month is identified, we define a crisis episode by following the procedure used by Hausmann, Pritchett and Rodrik (2005)¹⁰, and assign the dummy value of 1 to the $\pm 6$ month period centered on the month of a crisis. That is, when the crisis happens at date , the crisis dummy equals to 1 at dates

$\begin{displaymath}t-6,t-5,\cdots ,t,t+1,\cdots ,t+6 \end{displaymath}$

We drop data for the six months before and after the crisis period so as to remove the ambiguity associated with the transition period when 1 or 0 may not be clearly assigned. The comparison group is the group of the countries that did not have a crisis in that same month.

Our definition of non-core bank liabilities follows the approach in Shin and Shin (2010). Non-core bank liabilities will be classified (in the first instance) broadly as claims on banks held by financial institutions and held by foreign creditors. In principle, non-core bank liabilities should include inter-bank liabilities, but data limitations for emerging economies prevent us from using interbank liabilities in gross terms. We adopt two alternative measures of non-core bank liabilities:

$\displaystyle Non-core 1 = Liability~~of~~banks to the foreign sector$
$\displaystyle + Liability of banks to the non-banking financial sector$			(16)

$\displaystyle Non-core 2 = Liability of banks to the foreign sector + (M3 - M2)$

(17)

Both measures of non-core bank liabilities include bank liabilities to the foreign sector, which constitutes an important source of non-deposit wholesale funding for banks in emerging and developing economies. In addition to foreign liabilities, non-core 1 adds bank liabilities to non-bank financial institutions such as insurance companies and pension funds, and non-core 2 adds M3 - M2 as additional components of non-core liabilities.

In actual estimations of the probit models below, we use various ratios of non-core to core. As a measure of core liabilities, we use three alternative measures - M1, M2 and core deposits. Core deposits are obtained by summing demand deposits, time and savings deposits, foreign currency deposits, and restricted deposits. Finally, to obtain the credit to GDP ratio, we use deposit-taking banks' claims on other residents as a measure of bank credit.

The appendix presents the full list of countries that experienced either a currency crisis or credit crisis as identified by the above, together with the crisis dates. The appendix also reports which countries have data on non-core bank liabilities and the credit to GDP ratio. 37 countries had currency crises during our sample period, and several countries had two or more currency crises according to our definition (Brazil, Colombia, Lesotho, Mozambique, Namibia, South Africa, Swaziland, Turkey and Zambia). As for credit crises, 18 countries underwent crises according to the money market interest rate criterion (credit crisis I), while 19 countries experienced crises on the interest rate spread criterion (credit crisis II). Dominican Republic, Indonesia, Latvia, Paraguay, and Ukraine experienced credit crisis I only, and Algeria, Bolivia, Pakistan, Papua New Guinea, and Poland experienced credit crisis II only.¹¹

Table 1: **Summary Statistics.** This table gives the summary statistics for the variables. Missing values are replaced by using a linear interpolation. The appendix contains the list of crisis episodes studied in this paper.
Variable	Obs	Mean	Std. Dev	Min	Max
Noncore1/M1	4239	0.70	0.93	0.00	9.13
Noncore1/M2	4250	0.28	0.47	0.00	5.10
Noncore1/Core (deposits)	4521	0.32	0.86	0.00	26.88
Foreign/M1	7594	0.85	1.43	0.00	22.90
Foreign/M2	7718	0.31	0.48	0.00	5.09
Foreign/Core (deposits)	8510	0.37	0.84	0.00	23.35
Nonbank/M1	4239	0.17	0.28	0.00	1.66
Nonbank/M2	4250	0.06	0.09	0.00	0.85
Nonbank/Core (deposits)	4521	0.06	0.14	0.00	3.53
Noncore2/M1	4494	1.41	1.38	0.09	10.10
Noncore2/M2	4598	0.66	0.64	0.02	5.98
Noncore2/Core (deposits)	4596	0.55	0.57	0.03	10.03
(M3-M2)/M1	4887	0.78	1.06	0.00	6.50
(M3-M2)/M2	5003	0.34	0.41	0.00	3.84
(M3-M2)/Core (deposits)	4596	0.26	0.27	0.00	1.95
Exchange rate growth	8767	0.01	0.12	-0.54	0.86
Interest rate	6605	6.96	9.17	0.00	400.27
Spread	3691	-0.85	6.79	-112.52	329.84
Currency crisis	8106	0.09	0.29	0.00	1.00
Credit crisis I	6454	0.07	0.26	0.00	1.00
Credit crisis II	3551	0.14	0.34	0.00	1.00
Credit/GDP (yearly)	548	0.45	0.38	0.00	3.19

Table 1 reports summary statistics for the variables used in the probit analysis. When non-core liability is defined as the sum of bank liabilities to the foreign sector and non-bank financial sector (Non-core I), the non-core liability is 70% of M1 and around 30% of M2 or core deposits. The currency and credit crisis variables are dummy variables with a value of 1 for the crisis period. Credit to GDP is an annual variable with a mean of 45% in our sample.

3.2 Probit Estimation Results

We estimate panel probit models to investigate the linkage between our crisis measures and the non-core bank liabilities constructed above. Under the probit model, the inverse standard normal c.d.f. of the probability of crisis is modeled as a linear function of the explanatory variables. We run separate probit regressions for each crisis definition, and use the random effects panel probit method to allow for country differences that persist over time. As a robustness check, we also ran all regressions using the pooled probit (no random effects) and the fixed effects logit method, and confirmed that the results to be reported below are qualitatively unchanged.¹² The panels are estimated by maximum likelihood, where the explanatory variables are detrended.In each probit regression, the binary outcome variable is the crisis dummy variable for either the currency crisis or credit crisis. All the regressors are lagged by six months in regressions with monthly data and by one year in regressions with annual data.

3.2.1 Currency Crisis

Table 2 presents the random effects panel probit regression results for currency crises.

Table 2: **Random Effects Panel Probit Regression for Currency Crisis: Monthly Data for Non-Core Sum.** The binary outcome variable is the currency crisis dummy. Regressors are six months-lagged values of the noncore-core ratios. Robust standard errors are in parentheses. Statistical significance at 10% ,5% and 1% level is denoted by *, ** and *** respectively.
	(1)	(2)	(3)	(4)	(5)	(6)
Noncore1/M1	2.80***
Noncore1/M1 Standard Error	(0.24)
Noncore1/M2		4.17***
Noncore1/M2 Standard Error		(0.50)
Noncore1/Core (deposits)			3.95***
Noncore1/Core (deposits) Standard Error			(0.55)
Noncore2/M1				0.93***
Noncore2/M1 Standard Error				(0.09)
Noncore2/M2					1.44***
Noncore2/M2 Standard Error					(0.17)
Noncore2/Core (deposits)						1.54***
Noncore2/Core (deposits) Standard Error						(0.22)
Pseudo $R^{2}$	0.15	0.10	0.04	0.07	0.05	0.05
Log-likelihood	-638.93	-681.89	-766.35	-932.94	-947.40	-947.40
Observations	3304	3310	3552	3482	3586	3581
Countries	38	38	40	41	42	42

As described above, we have two measures of noncore bank liabilities - non-core 1 (using liabilities to financial institutions) and non-core 2 (using M3 minus M2), and three proxies for core liabilities - M1, M2 and core deposits. Hence, we have six alternative ways of constructing the ratio of non-core to core liabilities. In Table 2, all non-core liability ratios are 6 months-lagged and detrended, and we report coefficient estimates along with robust standard errors in the parenthesis.

As can be seen in Table 2, for both non-core 1 and non-core 2 measures, and regardless of the form of core liabilities, all the non-core liability ratios have a positive and statistically significant coefficient with all but one at the 1% level. The results indicate that an increase in the non-core bank liability ratio is associated with an increase in the predicted probability of having a currency crisis. This finding is in line with the predictions from our theory section. Fluctuations in the non-core to core liability ratio can be interpreted as reflecting fluctuations in the changing degree of financial vulnerability to a crisis.

Table 3: **Random Effects Panel Probit Regression for Currency Crisis: Monthly Data for Separate Non-Core to Core Ratios.**
	(1)	(2)	(3)	(4)	(5)	(6)
Noncore1 Foreign/M1	4.39***
Noncore1 Foreign/M1 Standard Error	(0.33)
Noncore1 Nonbank/M1	0.89*
Noncore1 Nonbank/M1 Standard Error	(0.51)
Noncore1 Foreign/M2		7.08***
Noncore1 Foreign/M2 Standard Error		(0.70)
Noncore1 Nonbank/M2		5.45***
Noncore1 Nonbank/M2 Standard Error		(1.84)
Noncore1 Foreign/Core (deposits)			4.70***
Noncore1 Foreign/Core (deposits) Standard Error			(0.55)
Noncore1 Nonbank/Core (deposits)			8.57***
Noncore1 Nonbank/Core (deposits) Standard Error			(2.18)
Noncore2 Foreign/M1				3.40***
Noncore2 Foreign/M1 Standard Error				(0.24)
Noncore2 (M3-M2)/M1				-0.33**
Noncore2 (M3-M2)/M1 Standard Error				(0.15)
Noncore2 Foreign/M2					5.47***
Noncore2 Foreign/M2 Standard Error					(0.47)
Noncore2 (M3-M2)/M2					-0.85***
Noncore2 (M3-M2)/M2 Standard Error					(0.26)
Noncore2 Foreign/Core (deposits)						2.77***
Noncore2 Foreign/Core (deposits) Standard Error						(0.34)
Noncore2 (M3-M2)/Core (deposits)						-0.79
Noncore2 (M3-M2)/Core (deposits) Standard Error						(0.57)
Pseudo $R^{2}$	0.22	0.15	0.06	0.16	0.11	0.07
Log-likelihood	-588.95	-641.60	-744.29	-836.13	-891.69	-934.04
Observations	3304	3310	3552	3484	3487	3581
Countries	38	38	40	41	41	42

In Table 3, we present the panel probit regression results when we decompose the two non-core bank liability variables into their two respective separate components. As before, we introduce the non-core components as ratios of core liabilities.

The results reveal some insights on which components are relatively more important. When we use non-core 1, both foreign and non-bank components have a statistically significant positive effect. However, when we use non-core 2, only foreign liability terms are significantly positive. The coefficients on M3 - M2 are not statistically different from zero. This suggests that foreign liabilities play a more robust role as a predictor of currency crises in emerging economies. Nonetheless, there seems to be an additional and independent role of domestic non-core liabilities, although the non-core measure constructed from traditional monetary aggregates have little explanatory power.Monetary aggregates such as M2 and M3 are based on the legal form of the claim rather than on who holds the claim. Shin and Shin (2010) propose that classification by holder is more important for how "sticky" the claim is, rather than the legal form of the claim. Hahm, Mishkin, Shin and Shin (2010) provide evidence for this claim from disaggregated Korean banking sector data.

In Table 4, we check the robustness of our results by comparing our non-core measures to the much better known credit to GDP ratio. Borio and Lowe (2004) argued for the informativeness of credit aggregates in signalling financial excesses that expose an economy to potential crises, and have given prominence to the ratio of credit to GDP as an indicator. As the credit to GDP ratio is available at an annual frequency only, we run annual regressions. However, instead of re-identifying crisis episodes for annual data, we used the crisis episodes identified in the monthly data. Namely, the year in which the crisis occurs in the monthly data is identified as a crisis episode.

As can be seen in Table 4, we re-confirm the credit to GDP ratio as a significant indicator of an impending currency crisis. In every regression, it has a positive coefficient, and significant at the 1% level. Interestingly, however, note how non-core ratios still retain significance even in the presence of the credit to GDP ratio. In contrast with the monthly regressions above, non-core 2 measures seem to fare better than non-core 1 ratios when the credit to GDP ratio is included. This weak performance of non-core 1 may reflect the potential positive correlation between the credit to GDP ratio and the liability to non-bank financial institutions. The insignificant coefficient estimates of non-core 1 ratios may have also resulted from the weak power of test due to the loss of observations in the switch to annual data in these regressions.

Table 4: **Random Effects Panel Probit Regression for Currency Crisis: Annual Data with Credit to GDP Ratio Included.**
	(1)	(2)	(3)	(4)	(5)	(6)
Credit/GDP	13.91***	23.28***	23.03***	13.45***	17.98***	13.52***
Credit/GDP Standard Error	(3.06)	(7.05)	(7.16)	(3.83)	(5.58)	(3.61)
Noncore1/M2		1.97
Noncore1/M2 Standard Error		(1.75)
Foreign/M2			2.36
Foreign/M2 Standard Error			(1.95)
Nonbank/M2			1.37
Nonbank/M2 Standard Error			(5.47)
Noncore2/M2				2.12*
Noncore2/M2 Standard Error				(1.10)
Foreign/M2					11.84**	3.94**
Foreign/M2 Standard Error					(5.56)	(1.70)
(M3-M2)/M2					-2.33
(M3-M2)/M2 Standard Error					(1.93)
Pseudo $R^{2}$	0.11	0.21	0.21	0.12	0.17	0.15
Log-likelihood	-112.39	-53.49	-53.25	-71.97	-67.93	-99.87
Observations	426	211	211	235	235	364
Countries	63	32	32	36	36	55

It is noteworthy that the significance of the non-core 2 ratio relies heavily on foreign liabilities. We see this in the regression where we break out the non-core measures into their respective components. In column (5) only the foreign liability ratio remains significantly positive while the ratio using the M3 - M2 measure is insignificant. The predictive power of foreign bank liability ratio is again confirmed when it is included as the sole explanatory variable alongside the credit to GDP ratio in column (6).

The empirical results in Table 4 suggest that, independently from the credit to GDP ratio, the non-core liability ratio retains predictive power for currency crises in emerging and developing economies. This predictive power springs mainly from the foreign liabilities of banks, suggesting that liability side measures of vulnerability retain additional explanatory value that is not captured by the credit to GDP ratio.

The informativeness of liability side measures take on added significance when considering the more timely and higher frequency nature of such measures. Credit to GDP ratios are available at an annual frequency in most countries, while liability side aggregates are available more frequently, often monthly and sometimes even weekly. For purposes of real time surveillance exercises where timely identification of emerging vulnerabilities are important, the liability side aggregates identified in our paper may be promising as early warning indicators.

3.2.2 Credit Crisis

We now turn to consider credit crises. Credit crises are often associated with currency crises (as part of a "twin crisis"), but credit crises have occurred independently of currency crises, as is clear from the list of crisis episodes listed in the appendix to our paper.

As stated at the outset, we use two alternative definitions of a credit crisis episode. One is constructed from the money market interest rate alone (Credit Crisis I) and the second is defined by reference to the spread between money market rates and the local treasury rate (Credit Crisis II). Tables 5 to 7 report the random effects panel probit estimation results for credit crises using the money market rate definition. Tables 8 to 10 report analogous regression results for the second credit crisis definition using the spread definition.

Table 5: **Random Effects Panel Probit Regression for Credit Crisis I (Money Market rate only) with Monthly Data.** The binary outcome variable is the credit crisis dummy based on the money market interest rate. Regressors are six month-lagged values of noncore-to-core ratios. Robust standard errors are in parantheses. The statistical significance at the 10% ,5% and 1% level are indicated by by *, ** and *** respectively.
	(1)	(2)	(3)	(4)	(5)	(6)
Noncore1/M1	0.75***
Noncore1/M1 Standard Error	(0.12)
Noncore1/M2		1.59***
Noncore1/M2 Standard Error		(0.26)
Noncore1/Core (deposits)			10.88***
Noncore1/Core (deposits) Standard Error			(0.95)
Noncore2/M1				0.42***
Noncore2/M1 Standard Error				(0.09)
Noncore2/M2					1.07***
Noncore2/M2 Standard Error					(0.16)
Noncore2/Core (deposits)						0.35***
Noncore2/Core (deposits) Standard Error						(0.08)
Pseudo $R^{2}$	0.05	0.05	0.16	0.03	0.05	0.02
Log-likelihood	-492.37	-538.32	-482.70	-419.23	-456.01	-467.92
Observations	2389	2365	2679	2684	2753	2728
Countries	26	26	29	30	31	31

Table 6: **Random Effects Panel Probit Regression for Credit Crisis I (Money Market rate only) with Monthly Data for Separate Non-core to Core Ratios.**
	(1)	(2)	(3)	(4)	(5)	(6)
Noncore1 Foreign/M1	0.70***
Noncore1 Foreign/M1 Standard Error	(0.12)
Noncore1 Nonbank/M1	-0.13
Noncore1 Nonbank/M1 Standard Error	(0.35)
Noncore1 Foreign/M2		1.57***
Noncore1 Foreign/M2 Standard Error		(0.26)
Noncore1 Nonbank/M2		0.82
Noncore1 Nonbank/M2 Standard Error		(1.19)
Noncore1 Foreign/Core (deposits)			12.27***
Noncore1 Foreign/Core (deposits) Standard Error			(1.05)
Noncore1 Nonbank/Core (deposits)			3.02
Noncore1 Nonbank/Core (deposits) Standard Error			(2.59)
Noncore2 Foreign/M1				0.72***
Noncore2 Foreign/M1 Standard Error				(0.11)
Noncore2 (M3-M2)/M1				-0.83***
Noncore2 (M3-M2)/M1 Standard Error				(0.24)
Noncore2 Foreign/M2					1.19***
Noncore2 Foreign/M2 Standard Error					(0.19)
Noncore2 (M3-M2)/M2					-0.69
Noncore2 (M3-M2)/M2 Standard Error					(0.48)
Noncore2 Foreign/Core (deposits)						0.47**
Noncore2 Foreign/Core (deposits) Standard Error						(0.18)
Noncore2 (M3-M2)/Core (deposits)						-0.57
Noncore2 (M3-M2)/Core (deposits) Standard Error						(0.89)
Pseudo $R^{2}$	0.04	0.05	0.19	0.07	0.05	0.02
Log-likelihood	-496.88	-539.43	-467.52	-398.63	-407.53	-470.19
Observations	2389	2365	2679	2684	2684	2728
Countries	26	26	29	30	31	31

Table 7: **Random Effects Panel Probit Regression for Credit Crisis I (Money Market rate only): Annual Data with Credit to GDP Ratio Included.**
	(1)	(2)	(3)	(4)	(5)	(6)
Credit/GDP	4.88**	7.76	8.04	3.18	2.59	1.70
Credit/GDP Standard Error	(1.93)	(4.98)	(5.32)	(2.70)	(3.49)	(2.45)
Noncore1/M2		0.84
Noncore1/M2 Standard Error		(1.12)
Foreign/M2			0.66
Foreign/M2 Standard Error			(1.20)
Nonbank/M2			0.96
Nonbank/M2 Standard Error			(3.44)
Noncore2/M2				7.21***
Noncore2/M2 Standard Error				(1.97)
Foreign/M2					11.68***	4.37***
Foreign/M2 Standard Error					(3.54)	(1.56)
(M3-M2)/M2					1.85
(M3-M2)/M2 Standard Error					(2.60)
Pseudo $R^{2}$	0.01	0.04	0.08	0.07	0.26	0.31
Log-likelihood	-100.10	-45.15	-45.25	-36.63	-34.38	-85.13
Observations	435	124	124	176	176	369
Countries	56	20	20	26	26	50

The evidence from the credit crisis regressions indicate considerable explanatory role for the noncore measures. Consider the results shown in Table 5. All our non-core measures have positive coefficients and are significant at the 1% level.

In Table 6, as we did for the currency crisis regression, we break out the non-core liability aggregates into their respective components. Again, we see that liabilities to the foreign sector are an important explanatory variable for a credit crisis, as it was for a currency crisis. Note also how the foreign liability ratios remain statistically significant and of the correct sign even as the other components become insignificant in the regression.

Even more encouraging is the fact that for the credit crisis episodes, our liabilities side measures perform better than the better known credit to GDP ratio. In Table 7, although the credit to GDP ratio is significant when it enters as the sole explanatory variable, it is knocked out when our liabilities side variables are introduced. In columns (4) to (6), the non-core 2 ratio as well as the decomposed foreign liability ratio have a positive coefficient that is significant at the 1% level, even though the credit to GDP ratio becomes insignificant. These results confirm the intuition that liabilities to the foreign sector have implications for distress in the domestic financial market, also. By highlighting the explanatory role of banking sector liabilities, we train the spotlight on the behavior of the banking sector in the period preceding the crisis. Overall, these results for credit crises indicate that the non-core liability ratio may have an independent and even superior predictive power relative to measures of credit to GDP.

Finally, Tables 8 to 10 report the analogous regression results for credit crises using our second definition of credit crisis using the interest rate spread between money market and treasury rates.

We can see that the results in Tables 8 to 10 are qualitatively similar to the results using our first definition of crisis using the level of the money market rates only. In Table 8, all the non-core liability ratios have a positive coefficient and significant, with all but one at the 1% level. When we break out the non-core liability aggregates into their respective components, we again confirm the importance of liabilities to the foreign sector. In Table 9, we see that foreign sector liabilities are highly significant in all specifications, except for column (6). In contrast, the non-bank liability ratio and the non-core ratio using the M3 - M2 measure are either insignificant or have the "wrong" sign.

When we include the credit to GDP ratio in the regression in Table 10 , we see again that the non-core liability ratios fare better than the credit to GDP ratio. Indeed, the credit to GDP measure becomes insignificant in all cases. In contrast, both non-core 1 and non-core 2 ratios remain significantly positive in the presence of the credit to GDP ratio.

Table 8: **Random Effects Panel Probit Regression for Credit Crisis II (Spread Measure) with Monthly Data for Non-Core Sum.** The binary outcome variable is the credit crisis dummy based on the money market interest rate. Regressors are six month-lagged values of noncore-to-core ratios. Robust standard errors are in parantheses. The statistical significance at the 10% ,5% and 1% level are indicated by by *, ** and *** respectively.
	(1)	(2)	(3)	(4)	(5)	(6)
Noncore1/M1	0.67***
Noncore1/M1 Standard Error	(0.13)
Noncore1/M2		1.10***
Noncore1/M2 Standard Error		(0.24)
Noncore1/Core (deposits)			5.46**
Noncore1/Core (deposits) Standard Error			(0.90)
Noncore2/M1				0.18*
Noncore2/M1 Standard Error				(0.10)
Noncore2/M2					0.56***
Noncore2/M2 Standard Error					(0.17)
Noncore2/Core (deposits)						0.38***
Noncore2/Core (deposits) Standard Error						(0.08)
Pseudo $R^{2}$	0.04	0.03	0.05	0.01	0.01	0.03
Log-likelihood	-393.86	-425.54	-420.94	-331.95	-356.62	-351.59 Observations
Countries	16	17	18	20	21	21

Table 9: **Random Effects Panel Probit Regression for Credit Crisis II (Spread Measure) with Monthly Data for Separate Non-core to Core Ratios.**
	(1)	(2)	(3)	(4)	(5)	(6)
Noncore1 Foreign/M1	0.97***
Noncore1 Foreign/M1 Standard Error	(0.17)
Noncore1 Nonbank/M1	-0.97***
Noncore1 Nonbank/M1 Standard Error	(0.35)
Noncore1 Foreign/M2		1.58***
Noncore1 Foreign/M2 Standard Error		(0.29)
Noncore1 Nonbank/M2		-2.05**
Noncore1 Nonbank/M2 Standard Error		(0.82)
Noncore1 Foreign/Core (deposits)			6.18***
Noncore1 Foreign/Core (deposits) Standard Error			(0.99)
Noncore1 Nonbank/Core (deposits)			-1.95
Noncore1 Nonbank/Core (deposits) Standard Error			(2.25)
Noncore2 Foreign/M1				0.61***
Noncore2 Foreign/M1 Standard Error				(0.12)
Noncore2 (M3-M2)/M1				-2.42***
Noncore2 (M3-M2)/M1 Standard Error				(0.37)
Noncore2 Foreign/M2					0.95***
Noncore2 Foreign/M2 Standard Error					(0.22)
Noncore2 (M3-M2)/M2					-6.10***
Noncore2 (M3-M2)/M2 Standard Error					(0.94)
Noncore2 Foreign/Core (deposits)						0.89***
Noncore2 Foreign/Core (deposits) Standard Error						(0.26)
Noncore2 (M3-M2)/Core (deposits)						-2.45*
Noncore2 (M3-M2)/Core (deposits) Standard Error						(1.32)
Pseudo $R^{2}$	0.06	0.05	0.05	0.14	0.13	0.03
Log-likelihood	-383.72	-416.47	-419.07	-286.01	-290.40	-349.51
Observations	1398	1437	1545	1531	1531	1547
Countries	16	17	18	20	20	21

Table 10: **Random Effects Panel Probit Regression for Credit Crisis II (Spread Measure): Annual Data with Credit to GDP Ratio Included.**
	(1)	(2)	(3)	(4)	(5)	(6)
Credit/GDP	3.44*	-1.76	-6.74	2.10	1.50	1.02
Credit/GDP Standard Error	(2.03)	(6.23)	(7.32)	(2.39)	(3.62)	(2.75)
Noncore1/M2		2.08*
Noncore1/M2 Standard Error		(1.20)
Foreign/M2			6.16*
Foreign/M2 Standard Error			(3.38)
Nonbank/M2			-4.41
Nonbank/M2 Standard Error			(3.98)
Noncore2/M2				2.62*
Noncore2/M2 Standard Error				(1.56)
Foreign/M2					5.71**	5.20***
Foreign/M2 Standard Error					(2.34)	(1.92)
(M3-M2)/M2					-8.53*
(M3-M2)/M2 Standard Error					(4.57)
Pseudo $R^{2}$	0.02	0.06	0.13	0.07	0.20	0.11
Log-likelihood	-86.96	-31.30	-29.09	-37.86	-32.72	-69.55
Observations	248	66	66	100	100	217
Countries	36	12	12	17	17	33

3.3 Summary of Findings

Our probit estimation results could be summarized as follows. First, our measures of the non-core bank liability ratio has significant predictive power both for currency crises and credit crises. Second, most of the predictive power of the non-core liability ratio stems from the information contained in the banking sector's liabilities to the foreign sector. Third, the non-core bank liability ratio has independent predictive power over the much better-known and debated credit to GDP ratio. Moreover, the non-core liability ratio remains significant in the presence of the credit to GDP ratio in the currency crisis regressions, and the non-core liability ratio seems to fare better than the credit to GDP ratio in the credit crisis regressions.

In sum, the empirical evidence strongly suggests that non-core liability measures contain considerable information value for financial vulnerability, and provides an additional support for the "excess elasticity" hypothesis of Borio and Disyatat (2011), as formalized in our theory section. Our findings suggest that, at least in emerging and developing economies, non-core bank liabilities may be usefully monitored as a complementary measure to the credit to GDP ratio in gauging the stage of financial cycles and the build up of financial risk.

4 Concluding Remarks

The discussion in our paper serves to focus attention on the banking sector and its role in the fluctuations in financial conditions. Through its pivotal role in the determination of risk premiums and overall financial conditions, the banking sector can be seen to lie at the heart of issues of financial stability.

Our discussion has focused mainly on the theoretical underpinnings of the possible information value of non-core banking sector liabilities and their empirical properties in forecasting crisis episodes. Although much more work is necessary in refining the results, there are some lessons for the broader application of balance sheet aggregates in questions of financial regulation and the mitigation of financial vulnerability.

Our empirical result that non-core liability aggregates perform at least as well as the credit to GDP ratio (and sometimes much better) has far-reaching implications for the choice of aggregate measures to be used in surveillance and regulatory frameworks. As part of the Basel III overhaul of bank regulation, the Basel Committee on Banking Supervision (BCBS) has agreed on a countercyclical capital buffer. In operational terms, the countercyclical buffer will make use of the credit to GDP ratio as the indicator of procyclicality that triggers increased capital requirements on banks (BCBS 2010). However, as seen in our results above, the simple credit to GDP ratio may be a somewhat coarse indicator, as well as being available only at a low frequency.

One of the main reasons for the lack of universal support for the cyclical capital surcharge under Basel III has been the concern that the simple credit to GDP aggregate may be too noisy in capturing finer institutional details that are relevant when considering harmonized international standards for bank capital regulation. Our results concerning the informativeness of liabilities side measures raise the prospect of having possibly more finely distinguished measures of financial vulnerability that tie up better with the underlying cyclical trends in the banking sector. Further research will be illuminating in shedding more light on these questions.

References

Adrian, Tobias and Hyun Song Shin (2008) "Procyclical Leverage and Value-at-Risk"Federal Reserve Bank of New York Staff Report 338, http://www.newyorkfed.org/research/staff_reports/sr338.html

Adrian, Tobias and Hyun Song Shin (2010) "Liquidity and Leverage,"Journal of Financial Intermediation, 19, 418-437

Basel Committee on Banking Supervision (2005) "International Convergence of Capital Measurement and Capital Standards: A Revised Framework", Bank for International Settlements, November 2005 http://www.bis.org/publ/bcbs118.pdf

Basel Committee on Banking Supervision (2010) , "Guidance for National Authorities Operating the Countercyclical Capital Buffer,"BIS, December.

Berg, Andrew and Catherine Pattillo (1999) "Predicting Currency Crises: The Indicator Approach and an Alternative"Journal of International Money and Finance, 18, 561 - 586

Borio, Claudio and Piti Disyatat (2011) "Global imbalances and the financial crisis: Link or no link?"BIS Working Papers 346 http://www.bis.org/publ/work346.pdf

Borio, Claudio and Philip Lowe (2004) "Securing Sustainable Price Stability: Should Credit Come Back From the Wilderness?"BIS Working Paper, No.157.

Bruno, Valentina and Hyun Song Shin (2011) "Capital Flows, Cross-Border Banking and Global Liquidity"working paper.

Claessens, Stijn, Giovanni Dell'Ariccia, Deniz Igan and Luc Laeven (2010) "Cross-Country Experiences and Policy Implications from the Global Financial Crisis"Economic Policy, 62, 267-293

Frankel, Jeffrey, and Andrew Rose (1996) "Currency Crashes in Emerging Markets: An Empirical Treatment,"Journal of International Economics, Vol. 41, pp.351-66.

Frankel, Jeffrey, and George Saravelos (2010) "Are Leading Indicators of Financial Crises Useful for Assessing Country Vulnerability?"NBER working paper No. 16047

Goldstein, Morris, Kaminsky, Graciela, and Carmen Reinhart (2000) Assessing Financial Vulnerability: An Early Warning System for Emerging Markets, Institute for International Economics, Washington DC.

Hahm, Joon-Ho, Mishkin, Frederic S., Shin, Hyun Song and Kwanho Shin (2010) "Macroprudential Policies in Open Emerging Economies,"paper prepared for the Bank of Korea.

Kaminsky, Graciela, and Carmen Reinhart (1999) "The Twin Crises: The Causes of Banking and Balance of Payments Problems,"American Economic Review, Vol. 89, No. 3, pp. 473-500.

Lane, Philip R. and Gian Maria Milesi-Ferretti (2010) "The Cross-Country Incidence of the Global Crisis", working paper

Rose, Andrew and Mark Spiegel (2008) "Cross-Country Causes and Consequences of the 2008 Crisis: Early Warning,"NBER working paper 15357, forthcoming in Global Journal of Economics

Rose, Andrew and Mark Spiegel (2010) "The Causes and Consequences of the 2008 Crisis: An Update", forthcoming in the European Economic Review, 2011.

Shin, Hyun Song and Kwanho Shin (2010) "Macroprudential Policy and Monetary Aggregates,"a paper prepared for the 2010 Bank of Korea International Conference.

Vasicek, Oldrich (2002) "The Distribution of Loan Portfolio Value"Risk, December 2002, http://www.moodyskmv.com/conf04/pdf/papers/dist_loan_port_val.pdf

Wooldridge, Jeffrey M. (2010) Econometric Analysis of Cross Section and Panel Data, second edition, MIT Press, Cambridge, MA.

Appendix: Crisis Episodes

This appendix lists the crisis episodes that quallify as crises according to the criteria for a currency crisis or a credit crisis. See the text for the methodology to identify the two types of crises: currency crises and credit crises. Credit Crisis I uses the money market interest rate only whereas Credit Crisis II uses the interest rate spread, defined as the difference between the money market interest rate and the Treasury bill rate. Crisis countries for which noncore1, noncore2 and credit/GDP ratio data are available are marked with "o".

Country Name	Non-Core1	Non-Core2	Credit/GDP	Currency Crisis	Credit Crisis I	Credit Crisis II
Algeria	O		O			07m10-10m09
Armenia			O		00m2-02m8	00m7-02m8
Belarus	O	O	O	08m8-10m6
Bolivia	O	O	O			08m6-09m6
Botswana	O	O	O	08m4-09m7
Brazil	O	O	O	02m6-03m10	02m6-04m2	02m12-04m2
Brazil cont'd				08m6-09m10
Burundi		O	O	02m10-03m10
Chile	O	O	O	08m5-09m10
Colombia	O	O	O	02m8-03m10
Colombia cont'd				08m8-09m10
Czech		O	O	08m8-09m9
Republic
Dominican	O		O	02m7-04m11	02m9-05m6
Republic
Egypt	O		O	02m11-04m7
Eritrea		O		01m7-03m2
Georgia	O	O	O		00m1-00m8	01m7-02m7
Georgia cont'd					01m2-03m7
Ghana	O				02m10-04m5	06m10-07m10
Haiti		O	O	02m5-04m4
Hungary		O	O	08m8-10m1
Iceland		O	O	08m1-10m2	00m8-02m8	00m8-02m8
Indonesia	O		O	08m8-09m9	00m8-01m8
Jamaica	O		O		02m8-04m8	00m1-04m8
Jamaica cont'd					08m6-09m6
Latvia		O	O		08m12-09m12
Lesotho			O	01m12-02m10
Lesotho cont'd				08m4-09m9
Lithuania		O	O
Malta		O	O	07m7-08m7
Mauritius	O		O	08m10-09m10
Mexico	O	O		08m8-10m2
Moldova	O	O	O		00m1-00m11	00m1-00m9
Moldova cont'd						03m9-06m8
Moldova cont'd						09m6-10m9
Mongolia			O	08m9-09m9
Mozambique			O	05m5-06m8	00m5-02m11	01m2-02m2
Mozambique cont'd				09m12-10m9
Namibia	O		O	01m7-02m10
Namibia cont'd				08m4-09m9
Nigeria	O		O	09m2-10m3
Pakistan			O	08m4-09m9		00m1-00m12
Pakistan cont'd						08m3-09m4
Papua NG			O			01m5-02m5
Paraguay	O	O	O	01m7-03m10	03m1-04m3
Paraguay cont'd					08m7-09m9
Philippines			O			06m7-08m1
Poland		O	O	08m7-10m2		00m8-02m7
Romania		O	O	08m8-10m1	01m12-03m5	02m1-03m11
Romania cont'd						08m1-09m8
Russian Federation			O	08m7-10m2	01m6-02m6	01m6-02m6
Russian Federation cont'd						03m3-04m11
Serbia		O	O		05m7-07m2	05m5-09m6
Seychelles	O		O	07m4-10m3
Solomon Is		O	O	01m12-03m7
South	O	O		01m7-02m10
Africa				08m4-10m2
Swaziland	O		O	01m7-02m10
Swaziland cont'd				08m4-09m9
Sweden		O	O	08m8-09m12
Turkey	O	O	O	01m7-02m8	00m1-05m2	00m1-01m8
Turkey cont'd				08m5-09m9		02m10-04m4
Uganda	O	O	O	08m10-10m1
Ukraine	O	O	O	08m5-10m5	00m1-02m2
Ukraine cont'd					08m5-09m9
Uruguay	O		O	01m11-03m12	00m3-03m11	01m12-02m12
Uruguay cont'd					08m4-09m4	08m4-09m4
Zambia	O	O	O	02m3-03m3
Zambia cont'd				06m8-07m10
Zambia cont'd				08m7-10m3

Footnotes

* Preliminary draft of paper for the Federal Reserve Board and JMCB conference on "Regulation of Systemic Risk", Washington, DC, September 15-16, 2011. Return to Text

1. See Adrian and Shin (2008, 2010) for a discusion of the evidence from the US. Return to Text

2. Borio and Disyatat (2011) have coined the term "excess elasticity"to describe the tendency of the banking system to expand when financial constraints are relaxed. Return to Text

3. The peaks in the series occur some weeks after the start of the crisis, as the non-core series are measured in Korean Won and the Won depreciated sharply during the 1997 and 2008 crises, increasing the Won value of foreign exchange-denominated liabilities. Return to Text

4. See Berg and Pattillo (1999) for a survey of the early literature and comparison of methodologies. Return to Text

5. The regulatory requirement was intended to emulate private sector best practice. See Adrian and Shin (2008) for a possible derivation of the VaR rule in a contracting setting. Return to Text

6. This conditions is useful in our comparative statics results that follow. It ensures that increasing $\rho$ (and hence greater systematic risk in the loan portfolio) leads to lower leverage. Return to Text

7. The funding rate

is fixed and determined outside our model. See Bruno and Shin (2011) for a model of credit supply that endogenizes

by modeling the global banking sector. Return to Text

8. Although the detailed breakdown of M2 and M3 categories differs across financial systems, the US is a useful benchmark (although the Federal Reserve no longer reports M3). For the US, the difference between M3 and M2 is given by large time deposits, institutional money market mutual funds, and repurchase agreements. In this respect it captures some aspects of wholesale bank funding. Return to Text

9. Another issue to bear in mind with our definition is that the maturity of the money market instrument is, in many cases, considerably shorter than the treasury rate, so that sometimes the money market rate is lower than the Treasury rate. Return to Text

10. Hausmann, Pritchett and Rodrik (2005) used a probit model to identify factors in growth accelerations. Return to Text

11. Except for Latvia, countries that experienced Debt Crisis I are not recorded as having experienced Credit Crisis II due to the lack of data on the Treasury bill rate. Return to Text

12. The fixed effects logit model has the advantage of being robust to potential correlation between cross-section country heterogeneity and the error term, but we lose sample observations of countries that did not have a crisis (when the dependent variable is constant at 0). The robustness of our results to the choice of regression method is a case in favor of the random effects model used here. See Wooldridge (2010, ch.15) for a discussion of relative advantages of probit and logit. Return to Text

^♣ This version is optimized for use by screen readers. Descriptions for all mathematical expressions are provided in LaTex format. A printable pdf version is available. Return to Text