Gropp Vesala

Equity and Bond Market Signals as Leading Indicators of Bank
Fragility
Reint Gropp, Jukka Vesala, Giuseppe Vulpes
Journal of Money, Credit, and Banking, Volume 38, Number 2, March 2006,
pp. 399-428 (Article)
Published by The Ohio State University Press

DOI: https://doi.org/10.1353/mcb.2006.0032
For additional information about this article

https://muse.jhu.edu/article/196985
Access provided by Missouri @ St Louis, Univ of (24 Aug 2017 19:10 GMT)
REINT GROPP
JUKKA VESALA
GIUSEPPE VULPES
Equity and Bond Market Signals as Leading

Indicators of Bank Fragility
We analyse the ability of the distance to default and subordinated bond

spreads to signal bank fragility in a sample of EU banks. We find leading
properties for both indicators. The distance to default exhibits lead times
of 6–18 months. Spreads have signal value close to problems only. We
also find that implicit safety nets weaken the predictive power of spreads.
Further, the results suggest complementarity between both indicators. We
also examine the interaction of the indicators with other information and
find that their additional information content may be small but not insignifi-
cant. The results suggest that market indicators reduce type II errors relative
to predictions based on accounting information only.
JEL codes: G21, G12, C35, C41

Keywords: banking, bank fragility, market indicators, market discipline,
bankruptcy predictors, proportional hazard models.
From a supervisory perspective, the securities issued by

banks are interesting for two main reasons. One, market prices of debt and
equity may increase banks’ funding cost and, therefore, induce (direct) market
discipline. Second, supervisors are considering the use of market data to comple-
ment traditional accounting data for assessing bank fragility (indirect market
discipline). Market prices may efficiently summarise all available information in
Research assistance by Sandrine Corvoisier and Roberto Rossetti is gratefully acknowledged. The
views expressed in this paper are solely those of the authors and not those of the ECB. We thank Allen Berger,
Jürg Blum, Max Bruche, Mark Flannery (the editor), Vitor Gaspar, Christopher James, Myron Kwast,
Mark Levonian, Jose Lopez, Simone Manganelli, Adrian Pop, two anonymous referees, and seminar
participants at the European Central Bank, the Basel Committee Research Task Force Conference on
Applied Banking Research in Oslo, the third Joint Central Bank Research Conference on Risk Measure-
ment and Systemic Risk in Basel, the Federal Reserve Bank of Boston Conference “The Impact of
Economic Slowdowns on Financial Institutions and their Regulators” and the Columbia University/
Federal Reserve Bank of New York Conference “Beyond Pillar Three in International Banking Regulation:
Disclosure and Market Discipline of Financial Firms” for useful comments. All errors are the authors’.
Reint Gropp is at the European Central Bank (E-mail: Reint.Gropp얀ecb.int). Jukka
Vesala is at the European Central Bank. Giuseppe Vulpes is at UniCredit Banca
d’Impresa.
Received June 4, 2002; and accepted in revised form July 12, 2004.
Journal of Money, Credit, and Banking, Vol. 38, No. 2 (March 2006)
Copyright 2006 by The Ohio State University
400 : MONEY, CREDIT, AND BANKING
one convenient indicator. Moreover, market information is available at a very high

frequency relative to accounting information and market information is inherently
more forward looking than accounting data. Hence, it has been proposed that supervi-
sors use these signals as screening devices or inputs into early warning models
geared at identifying banks, which should be more closely scrutinised.1
This paper investigates the second question and aims to ascertain the quality (i.e.
predictive power and classification errors) of two market indicators as leading
indicators of bank fragility in Europe. Previous work has established that the prices
of banks’ securities reflect contemporaneous information about banks’ risks in the
U.S. and also in Europe.2 This paper builds on this finding by examining whether
market prices can predict future fragility. We focus on two indicators: the spread
on subordinated debt issued by the bank relative to the risk free rate and the equity
market-based distance to default (KMV 2003). We test whether these indicators are
useful in predicting a material weakening in banks’ condition. In the absence of
actual bank bankruptcies and lack of access to consistent supervisory ratings in
Europe, we measure banks’ financial condition through Fitch/IBCA individual rat-
ings. We use a downgrade to C or below as our measure of bank fragility and
document that this was almost always followed by a government or parent bank
intervention within one year’s time.
Employing two different econometric models–a logit model and a proportional
hazard model–we find support in favour of using both indicators as leading indicators
of bank fragility, regardless of the econometric specification. The logit models suggest
that the distance to default predicts downgrades between 6 and 18 months in advance,
while its predictive properties are weak closer to failure. In contrast, spreads’
predictive powers diminish beyond 12 months prior to a downgrade. We also argue
in the paper that a proportional hazard model may be quite suitable when predicting
downgrades. It measures the probability of being downgraded within a time period,
conditional on having survived (not being downgraded) up to the starting point of
that interval. The proportional hazard model highlights the time dimension of the
relationship between the indicators and the downgrade. It reveals that distances to
default have to be observed for relatively long periods before they yield useful
information. Combining distances to default with spreads in the same model reduces
this time period. In a forecasting context, proportional hazard models have the
additional advantage of yielding continuously improving forecasts as the time one
observes the indicator increases.
1. Supervisory early warning models combine a set of bank-level financial indicators (balance sheet,
income statement, and market indicators), as well as sometimes also other variables (e.g. macroeconomic
conditions), to make a prediction about the future state of a bank. A growing number of supervisory
agencies have been experimenting with this kind of models (see Gilbert et al. 1999).
2. There are numerous studies relating U.S. secondary bond and/or stock market data to banks’
risk: Hand et al. (1992), Flannery and Sorescu (1996), Docking et al. (2000), Jagtiani et al. (2000), and
Flannery (1998, 2001). U.S. primary bond market data: Morgan and Stiroh (2001). European primary
bond market data: Sironi (2003). European secondary bond and equity market data: Gropp and Richards
(2001). With few exceptions, all studies tend to find a significant relationship between market prices and
risk, while risk and market prices are measured in different ways and the methodologies may differ substan-
tially.
REINT GROPP, JUKKA VESALA, AND GIUSEPPE VULPES : 401
Both sets of results indicate that spreads are useful predictors only for banks,
which are not implicitly insured against default, while the public safety net does not
appear to affect the predictive power of the distance to default. Further, both
market indicators provide some additional information relative to accounting vari-
ables and we find that the two indicators together have more discriminatory power
in predicting failures than each alone. In particular adding market indicators to
accounting information and using both indicators in the same model reduces type
II errors (a sound bank classified as weak). Hence, our results give some support
to the use of market indicators as screening devices or inputs into supervisors’ early-
warning models.
A number of recent papers studying U.S. banks are closely related to our work.
Curry et al. (2001) find that stock prices exhibit a downward trend as much as two
years before a supervisory CAMEL rating downgrade to three, four or five. In
addition, adding market variables to standard models containing call report financial
data improves their predictive power, especially for banks in the greatest finan-
cial distress. Evanoff and Wall (2001) find that accounting information has almost
no predictive power for CAMEL and BOPEC supervisory rating downgrades, and
subordinated debt spreads perform only insignificantly better. In their paper, spreads
tended to classify many more banks as “bad” than a classification based on CAMEL
ratings only. Berger et al. (2000) examine the relationship between supervisory
information and a number of market indicators (rating changes, abnormal stock
returns, and the proportion of equity owned by institutional investors and bank insid-
ers). They find that supervisory assessments and bond ratings are able, at least
partially, to predict each other, whereas supervisory assessments and equity indicators
are not.
DeYoung et al. (2001) take the opposite approach and examine whether on-site
examinations produce information that influences market prices. They find that this
information is only gradually incorporated in banks’ bond spreads with particularly
poor supervisory assessments reducing spreads and vice versa. They suggest that
markets rely on supervisory discipline as a substitute for market discipline. Krainer
and Lopez (2003) investigate whether market prices contain additional information
over accounting variables in predicting BOPEC rating changes. They answer this
question in the affirmative, but caution that there is no improvement in out-of-
sample forecasts. Interestingly, they find, as we do, that debt market indicators have
predictive power close to a downgrade only, while equity prices react much earlier.
In the European context, Laviola et al. (2001) examine the forecasting ability of
accounting data for Italian banks. While they do not use market information, they
employ (like us) proportional hazard models for forecasting bank distress.
Bliss and Flannery (2002) argue that the fact that market prices reflect banks’
risks does not necessarily imply that they discipline manager’s behaviour (“monitor-
ing” versus “influencing”). For a sample of U.S. banks, they find mixed evidence
of significant influencing. If managers could perfectly counteract the shocks the bank
faces, an increase in e.g. the subordinated debt spread would result in a corrective
action, which, in theory, ultimately would take the spread back to its initial level.
This line of reasoning has implications for the use of market indicators as early
warning indicators, as it suggests that these indicators may be of use only to the
extent that “influencing” works only imperfectly (see also Rajan 2001).3
One of the main contributions of this paper is its focus on equity prices in addition
to spreads. Signals based on equity prices have been considered ill suited, because
equity-holders benefit from the upside gains that accrue from increased risk-taking
leading to increased asset volatility. The distance to default used in this paper avoids
this problem as it combines equity information with information about leverage and
asset volatility.4 There are several other aspects, which suggest that equity prices,
properly adjusted, may be attractive as monitoring devices. One, there is broad
consensus that the equity markets are quite efficient in processing available informa-
tion. Second, while bond spreads are conceptually simple, their implementation is
difficult. Hancock and Kwast (2001) find that different bonds issued by the same
U.S. bank may yield different estimates of the spread and Bliss (2001) demonstrates
that spreads may be of little use in predicting ratings, even though positive yield/
rating slopes are often cited as evidence of risk sensitivity. In the European context, the
construction of appropriate risk-free benchmarks, which is a necessary ingredient
to the calculation of spreads, may also be difficult, especially for smaller countries.
Further there may be time varying liquidity premia incorporated in spreads, which
may hamper their interpretation as signals of bank fragility.5 However, there are a
number of practical difficulties in the calculation of stock market indicators as well.
For example, the distance to default can be sensitive to trading irregularities in the
period close to default. Further, it may be sensitive to the actual measure of equity
volatility and the distributional assumptions about equity returns used in the calculation.
The remainder of the paper is organised as follows: Section 1 gives our main
empirical hypotheses. Section 2 defines our sample and the variables used in the
empirical study. Section 3 contains descriptive statistics of the variables of interest.
Section 4 reports our econometric specifications and results. Section 5 examines
the results in terms of classification errors. Section 6 concludes.
1. EMPIRICAL HYPOTHESES
In this paper we consider the distance to default, DD, and the spread on subordi-
nated debt (relative to risk free debt), S, as bank fragility indicators. Both indicators
have been discussed extensively in the previous literature (i.e. Board of Governors of
the Federal Reserve System, 1999, Hancock and Kwast, 2001, KMV, 2003). They
3. The same applies to supervisors using market indicators to trigger supervisory action. Namely,
if this were widespread practice and supervisors acted accordingly, we should not find any predictive
power in market indicators.
4. Gropp et al. (2004) show that the distance to default signals greater bank fragility if asset volatility
is increasing.
5. See e.g. Elton et al. (2001), who find that almost half of spreads on corporate bonds are unexplained
by default risk or taxes or Huang and Huang (2002) who find that credit risk typically accounts for less
than 20% of corporate-treasury yield spreads.
represent appropriate indicators of bank fragility in two respects. One, both indicators
reflect three crucial factors of bank fragility: (1) the market value of the firm’s
assets, which represents the discounted value of future revenues; (2) leverage, which
captures the obligations the firm has to meet; and (3) the volatility of assets. Second,
Gropp et al. (2004) show that they reflect these three sources of risk in the “correct”
way, namely both indicators suggest greater fragility for a lower market value of
assets, and for higher leverage and asset volatility.6 Hence, the central objective
of this paper is to empirically test for the following hypothesis:
Hypothesis 1: The distance to default and the spread on subordinated debt are
early warning indicators of a weakening in banks’ financial condition.
The indicators may differ with respect to the strength of their reaction to a shock
moving the bank closer to the default point. Based on the standard Black and Cox
(1976) model for valuing debt, the spread is a convex and decreasing function of
the market value of assets, as long as the bank’s assets are valued in excess of its
liabilities. Gropp et al. (2004) show that the spread therefore responds little to
changes in the market value of assets as long as the bank is far from default. This
is in contrast to the distance to default, which strongly reacts to adverse shocks also
far away from default. Hence, empirically, given that both indicators provide noisy
signals, unless we are close to default, only DD may yield a detectable signal, as
for S the signal to noise ratio may be quite low.
Hypothesis 2: The distance to default has better predictive ability compared to
the spread further away from default.
In case of fully insured debt (like insured deposits), the market value of the debt
equals the “no-default-risk” value and there is no signal of fragility obtainable from
spreads. However, even for debt not explicitly covered by the safety net, the spread
may be uninformative. The market may consider some banks as “too-big-to-fail,”
for example due to their systemic importance (e.g. Dewatripont and Tirole 1993).
However, useable signals from spreads may be obtained only as long as creditors
expect to be repaid in case of bank failure with some probability less than one.7
Similar arguments may also apply to the distance to default. However, in all
European countries the case for equity-holders to be bailed out in case of failure is
much weaker. Generally, they are not covered even in broad-based explicit safety
nets and tend to be explicitly excluded in most deposit insurance laws (Gropp and
Vesala 2004). At a minimum, the haircuts that equity holders take in case of failure
may be substantial even for very large systemic banks. Hence, we will test
6. This is in contrast to, say, stock returns, which tend to increase with increases in leverage and
volatility. For more discussion see Gropp et al. (2004).
7. Gropp and Vesala (2004) stress the importance of explicit limits in the safety net for market
discipline. They find that banks’ risk taking in Europe was reduced in response to the introduction of
explicit and restricted deposit insurance schemes. They also find evidence in favour of that a number
of banks are “too-big-to-fail.” In addition, in a sample of European banks, Gropp and Richards (2001)
find that banks’ bond spreads do not appear to react to rating announcements. Their findings could be
interpreted as evidence in favour of widespread “implicit safety nets.”
Hypothesis 3: The spread is a weaker leading indicator of bank fragility than

the distance to default for banks covered by an implicit or explicit complete or
partial public guarantee.
2. EMPIRICAL IMPLEMENTATION
Our data set consists of monthly observations from January 1991 to March 2001.
We use monthly averages of daily data, in order to eliminate some of the noise in
high frequency equity and bond prices. The data set consists of those EU banks,
for which the necessary rating, equity, and bond market information is available.
In the sample selection process we started from 103 EU banks, which had an
individual rating from Fitch/IBCA. The sample size was then largely determined
by the availability of market data. The two sub-samples used in evaluating the
equity and bond market signals consist of 86 and 59 banks, respectively (see Table 1),
yielding about 5300 bank/month observations for the equity sample and 3600 bank/
month observations for the debt sample. All observations are used to estimate the
proportional hazard model.
We estimate logit models at x ⫽ 3, 6, 12, 18 and 24-month horizons, i.e. we use
the market indicator at time t ⫺ x to estimate the probability of a downgrade at
time t. We constructed the respective samples such that we combined each downgrade
TABLE 1
Composition of the Sample by Country
DD Spread
Belgium 3 (0) 1 (0)

Denmark 2 (0)
Germany 11 (1) 15 (4)
Greece 5 (1)
Spain 8 (2) 2 (0)
France 8 (5) 9 (5)
Ireland 4 (0) 2 (0)
Italy 20 (10) 7 (5)
Luxembourg 1 (0)
Netherlands 3 (0) 4 (0)
Austria 3 (3) 2 (2)
Portugal 4 (0) 1 (0)
Finland 1 (1) 2 (1)
Sweden 3 (2) 3 (2)
United Kingdom 11 (0) 10 (2)
Total 86 (25) 59 (21)
Note: Number of banks. Number of downgraded banks in parentheses.
with all non-downgraded observations for the same time period.8 The idea is to
reduce the potential for spurious results arising from differences in the mac-
roeconomic environment. As in the proportional hazard model, we have multiple
observations per bank and hence present all results with standard errors corrected
for clustering (i.e. correlated observations within banks). The sample size for the
logit model is then around 1000 bank/month observations for the equity sample and
350 for the debt sample. The samples contain banks from 14 (equity sample) and 13
(bond sample) EU countries (Table 1).
Formal bank bankruptcies have been extremely rare in Europe. Hence, we use a
proxy for banks’ seriously weakened financial condition, rather than outright default,
as the dependent variable. We use the month of the change in Fitch/IBCA’s individual
rating to C or below. Fitch/IBCA’s individual rating is designed to measure the true
condition of banks as such, without taking into account the possibility of public or
parent bank support. Further details on the rating are given in the Appendix.
We selected C or below as our cut-off point, because we found that such a downgrade
preceded all cases of serious bank problems in Europe on which public information
is available. The correspondence is almost exact both ways. Of the 31 downgrades
in our sample period 10 were associated with an injection of public funds, three
with a private injection of funds, two with a public or parent guarantee, and 12 with
a major restructuring of the banks’ operations typically initiated by the supervisor
(Table 2). In four cases, we were unable to ascertain what, if anything, happened
after the downgrade. Typically, the intervention took place within a year of the
downgrade. We were unable to find any interventions in banks rated by Fitch/IBCA
(our universe) that were not preceded by a downgrade to C or below.9 We argue
that if market indicators predicted such a weakening in banks’ condition, supervisors
could take early corrective action and perhaps avoid the use of public funds on the
basis of these signals. Essentially, we use the financial strength rating in a similar
way as previous studies have used supervisory ratings in the U.S., mainly because
in Europe internal supervisory ratings are either strictly confidential or do not exist.
The distance to default (DD) represents the number of standard deviations away
from the default point, where the default point is defined as the point when the
assets of the bank are just equal to its liabilities. We calculated the distance to
default DD for each bank in the sample and for each time period, t (i.e. month),
using that period’s equity market data and interpolated balance sheet data. To obtain
DD, we calculated
8. To illustrate, suppose we are interested in the 12 months horizon and suppose, for the sake of
simplicity, the sample contains only two banks which were downgraded to C or below and ten banks
in total. Suppose further that one of these downgrades took place in May of 2000 and the other in
October of 1998. Hence, we would select the market indicators for 12 months before May of 2000 (i.e.
May of 1999) for all banks for which data exist and similarly select the market indicators for October
of 1997. The sample would then consist of one observation for a downgraded bank in May 1999 plus
eight observations of non-downgraded banks in May 1999 plus one observation for the downgraded
bank in October 1997 plus nine observations of non-downgraded banks in October 1997, as the bank
which eventually is downgraded two years later would still be included as a non-downgrade. Sample
size in this illustration would be 1 ⫹ 8 ⫹ 1 ⫹ 9 ⫽ 19 observations.
9. Sixteen banks were downgraded to C, ten banks to C/D, four to D and one to D/E.
TABLE 2
Downgraded Banks
Bank Timing of the downgrade Type of intervention
a
Standard Chartered June 90 Unclear
HSBC Banka May 91 Unclear
Skandinaviska Enskilda Banken July 92 Injection of public funds
Svenska Handelsbanken Dec. 92 Injection of public funds
Banco Espanol de Creditob June 93 Injection of public funds
Credit Lyonnaisa June 94 Injection of public funds
Okobank Oct. 94 Injection of public funds
Banco di Napolib Jan. 95 Injection of public funds
Banco Zaragozanob March 95 Unclear
Banca Popolare di Novarab Oct. 95 Restructuring
Banca Popolare di Milanob Nov. 95 Restructuring
Bank Austria June 96 Restructuring
Natexis Banque Populaire Nov. 96 Restructuring
Banca di Roma Nov. 96 Injection of public funds
(to a bank acquired by BdR)
Creditanstaltb Jan. 97 Restructuring
Banca Nazionale del Lavoro June 97 Injection of public funds
CPRb Nov. 98 Injection of funds from parent
Westdeutsche Landesbanka Nov. 98 Injection of public funds
Commercial Bank of Greeceb Dec. 98 Restructuring
Bayerische Landesbanka Dec. 99 Public Guarantee
Entenial March 99 Injection of funds from parent
Bankgesellschaft Berlin June 99 Injection of public funds
Banque Worms Nov. 99 Restructuring
Erste Bank der Oesterreichischen Sparkassen Feb. 00 Injection of funds from the
savings bank system
Banca Popolare di Sondriob March 00 Unclear
Credit Foncier de France April 00 Guarantee from parent
Banca Commerciale Italiana June 00 Restructuring
Banca Popolare di Lodi June 00 Restructuring
Deutsche Genossenschaftsbanka Nov. 00 Restructuring
Banca Popolare di Intrab Feb. 01 Restructuring
Credito Valtellinese Feb. 01 Restructuring
Notes: The table shows those banks, which were downgraded to an individual rating of C or below by Fitch/IBCA and the subsequent
intervention. An intervention can take the form of an injection of public funds, an injection of funds or a guarantee from a parent organisation,
a public guarantee, a restructuring, generally associated with an exchange of management. Source: Fitch/IBCA. aOnly in the bond sample.
b
Only in the equity sample.
DD ≡
d
⫺ε⫽
ln ( ) (
VA
D
σ2
⫹ r⫺ A T
2 ) . (1)
σA√T σA√T
DD depends on VA (the value of assets), σA (asset volatility), and D (the face value
of debt liabilities), as well as on interest rate and time parameters. Though
unobservable, the first two parameters can be calculated from observable market
value of equity capital (VE) and the volatility of equity (σE) using the system of
equations below:
VE ⫽ VA N(d1) ⫺ De⫺rTN(d2) ,
σE ⫽ ( )
VA
VE
N(d1)σA ,
d1 ≡
ln ( ) (
VA
D
σ2
⫹ r⫹ A T
2 ) ,
σA√T
d2 ≡ d1 ⫺ σA√T .
Following Ronn and Verma (1989) we solved this system of equations by using
the generalised reduced gradient method. The necessary inputs are equity market
capitalisation, VE, equity volatility, σE, total debt liabilities, D, and the risk-free rate,
r. N denotes the normal distribution. We used monthly averages of VE, obtained
from Datastream. σE was estimated as the standard deviation of daily absolute equity
returns and, as proposed in Marcus and Shaked (1984), we took the six-month moving
average (backwards) to reduce noise. The presumption is that market participants do
not use the very volatile short term estimates, but more smoothed volatility mea-
sures.10 D was obtained from semi-annual published accounts and are interpolated
(using a cubic spline) to yield monthly observations. The time to the maturing of
the debt, T, was set to one year, which is the common benchmark assumption when
one has no particular information about the maturity structure. Finally, we used
short term government bond rates as risk free rates, r.11 Below we report results for
the negative distance to default, ⫺DD, rather than the distance to default, such that
an increase in the indicator signals greater fragility, in line with the spread. This is
intended to enhance the readability of the tables below.
The spread of subordinated debt relative to the risk free rate, S, is defined as
S ⫽ YTBit⫺YTCt , (2)
where YBT it denotes the yield of a subordinated bond Bi of bank i at time t with a
T
remaining term to maturity T and YCt denotes a corresponding government bond of
the bank’s country of incorporation C at time t with a remaining term to maturity
T. We used the standard Newton iterative method to calculate the bond yields to
maturity. In smaller countries the most liquid bank issues tended to be bonds
denominated in foreign currency (DM, euro, USD and in two cases, yen). Hence,
we matched the foreign currency denominated bank bonds with government bonds
issued in the same currency, again matching on all other criteria. We restricted the
sample to subordinated bonds with an issue size of more than euro 150 million, as
10. This is not an efficient procedure as it imposes the volatility to be constant. However, equity
volatility is accurately estimated for a specific time interval, as long as leverage does not change
substantially over that period (see for example Bongini et al. 2002).
11. In the calculation of DD we assume a log-normal distribution for the underlying asset values.
As pointed out by Bliss (2000), this assumption may not hold in practice. He argues that the normal
distribution does not take into account that closer to the default point it is likely that there are adjustments
in debt liabilities. It has also been pointed out that the assumption of a one-year term to maturity for
the debt does not hold in practice. We felt, however, that to explore alternatives, while extremely
interesting, would be far beyond the scope of this paper.
Hancock and Kwast (2001) argue that liquidity tends to be adequate for bonds of
this size. We limited the sample to “plain vanilla,” fixed rate, straight, subordinated
debt issues with the exception of five banks, for which we were unable to
find sufficiently liquid fixed rate bonds.12 In these five cases we used floating rate
bonds and matched on remaining term to maturity and on the base rate (LIBOR,
EURIBOR, etc.) of the adjustable coupon.13
We use the “support rating” of a bank issued by Fitch/IBCA to indicate the
likelihood of public support. Fitch/IBCA rates the likelihood of support on a scale of
one to five. A grade of one indicates existence of an assured legal guarantee for a
bank. A grade of two is assigned to a bank, for which in Fitch/IBCA’s opinion state
support would be forthcoming, even though this is not explicit. A rating of three to
five is given to banks where support may come from the parent organisation or
other owners, rather than from public sources. Complete definitions of the Fitch/
IBCA support ratings are given in the Appendix. The share of banks with a support
rating of one or two is quite high (around 65% in the equity sample and 80% in
the bond sample). This is not surprising, since we are considering banks with a
material securities market presence as an issuer. We considered the possibility that
there is some systematic correlation between the country of origin of the bank and the
Fitch/IBCA support rating, but this does not seem to be the case. Size does matter,
however, as banks with a support rating of one or two tend to be significantly
larger compared to those with a rating of three to five. Their average total debt
liabilities are roughly ten times higher.
How can we distinguish banks on the basis of the likelihood of being bailed out
as our definition of a major banking problem ultimately involves some intervention
at a later stage? Public coverage of different stakeholders may be incomplete and
different stakeholders may be wiped out altogether. This is especially true for
shareholders. In addition, it is important to distinguish between ex ante expectation
and ex post realisation. Even if ex post all banks survived and all stakeholders avoided
losses, market participants ex ante may still have assigned some non-zero probability
to not being bailed out. As we will see, for some (large) banks with Fitch/IBCA
support ratings of one and two, subordinated debt holders in fact assign a probability
of one to being bailed out and spreads are useless as a predictor of bank fragility. For
other banks and for equity holders this is not the case.
Recall that relative to the universe of 103 Fitch/IBCA rated banks,14 we were able
to obtain meaningful subordinated bond price data for only 59 banks. Sample selection
may be a problem, if the banks in sample differ in their probability of a downgrade
to C or below relative to those in the universe of banks. This is not the case. However,
the banks in the sample are larger than all rated banks (the universe), but the
12. All the bonds in our sample are “plain vanilla” issues in the sense that there are no bonds with
call or put options, nor are there any convertible bonds in the sample.
13. For example if the bank bond was variable coupon bond with LIBOR ⫹100bp, we matched it
with a variable rate government bond with a coupon also based on LIBOR.
14. This universe is substantially different from all EU banks. We would argue, however, that market
indicators are precisely of most use in case of large, complex financial institutions.
TABLE 3
Descriptive Statistics
Variable Definition N Mean Std.dev Min Max
(⫺DD) Monthly average 5450 ⫺4.97 9.16 ⫺0.71 ⫺639.41

of daily (⫺DD)
Spread (basis points) Monthly average 3677 83.94 116.69 ⫺81.89 662.44
of daily subordinated
debt spreads over
the risk-free yield
to maturity
Spread (UK banks only) 894 208.18 154.61 28.95 662.44
Spread (excluding 2783 44.03 61.35 ⫺81.89 587.87
UK banks)
High probability of Dummy variable equalling 3677 0.63 0.48 0 1
public support in one if Fitch/IBCA
case of failure support rating 1
(DSUPP) or 2 (zero otherwise)
SCORE (spread sample) Composite variable 2769 6.07 2.71 1 12
using the percentile
ranking of capital,
problem loans, cost/income
and RoA in each year
SCORE (DD sample) 4587 6.04 2.68 1 12
Note: Sample sizes determined by the data availability of time series secondary market data for subordinated bonds (spread), stock market
prices (⫺DD) and balance sheet information (SCORE). See text for further explanations.
difference is not statistically significant. A bias may also arise due to differences
in data availability for the banks that are in the sample. If banks that eventually are
downgraded remain in the sample only a relatively short period of time prior to
the downgrade, the proportional hazard model may overstate the predictive power
of indicators.15 However, the average time period in the sample for banks, which
eventually were downgraded, is 34 months. This should give us ample data to obtain
unbiased estimates.16 The sample, for which we were able to calculate DD consists
of 86 banks. The probability of problems in the sample is identical to that in the
universe. The average time of banks, which eventually were downgraded, is one month
longer in the equity than in the bond sample. Again, this should give us sufficient
data to estimate the model.
Variable definitions and descriptive statistics are given in Table 3.
3. DESCRIPTIVE EVIDENCE
The mean comparison tests reported in Table 4 are intended to provide a first
assessment of whether ⫺DD and S are able to distinguish weaker banks within our
15. Given that we chose a fixed starting point for our sample (1991) and given that we drop the
observations for all downgraded banks after the downgrade, the time period that non-downgraded banks
remain in the sample is longer. This by itself should not constitute a problem for the estimation. However,
if downgrades occur disproportionately at the beginning of the sample period, i.e. in 1991–94, this could
result in overstating the predictive power of our indicators in the proportional hazard model. Table 2
shows that this is not the case.
16. The average period in the sample of “non-failing” banks is, of course, longer with 76 months.
Notice that the maximum number of observations per bank is limited by our sampling period to 131 months.
TABLE 4
Ability of the Distance to Default and Spread to Predict Downgrades: Mean
Comparison Tests
x (months) Status Nobs Mean Std. error Differencea Difference ⬍ 0b
(⫺DDt⫺x)
3 0 1018 ⫺5.68 0.19 ⫺1.58 ⫺3.49***
1 25 ⫺4.10 0.41
6 0 1018 ⫺5.64 0.18 ⫺1.79 ⫺5.34***
1 25 ⫺3.85 0.28
12 0 1018 ⫺5.31 0.16 ⫺1.62 ⫺4.89***
1 22 ⫺3.69 0.29
18 0 1018 ⫺5.66 0.21 ⫺1.93 ⫺5.18***
1 21 ⫺3.72 0.31
24 0 1018 ⫺5.93 0.20 ⫺1.55 ⫺2.82***
1 18 ⫺4.38 0.51
St⫺x
3 0 457 0.88 0.05 ⫺0.19 ⫺0.68
1 21 1.07 0.27
6 0 454 0.86 0.05 ⫺0.18 ⫺0.55
1 20 1.04 0.32
12 0 438 0.79 0.05 ⫺0.10 ⫺0.37
1 19 0.89 0.26
18 0 417 0.74 0.05 ⫺0.12 ⫺0.43
1 15 0.86 0.27
24 0 393 0.70 0.05 ⫺0.03 ⫺0.13
1 14 0.73 0.26
St⫺xc
3 0 78 0.24 0.02 ⫺0.55 ⫺2.00**
1 5 0.79 0.28
6 0 78 0.20 0.02 ⫺0.38 ⫺3.04**
1 5 0.58 0.12
12 0 77 0.21 0.02 ⫺0.39 ⫺1.57*
1 4 0.60 0.25
Notes: Two sub-sample t-tests (unequal variances) are reported for the difference in mean values of (⫺DDt⫺x) and St⫺x in the sub-samples
of downgraded (SATUS ⫽ 1) and non-downgraded banks (STATUS ⫽ 0). *, **, *** indicate statistical significance at the 10%, 5%, and
1% levels, respectively. aMean (STATUS ⫽ 0) – Mean (SATUS ⫽ 1). bt-Statistics for testing the one-sided hypothesis that the difference
is negative. cBanks with a support rating of three to five and excluding UK banks.
data set. We examined whether the indicators lead downgrades by performing mean
comparison tests for various time leads (lead times of 3, 6, 12, 18 and 24 months).
Hence, one should read Table 4 as follows: we compare the mean of the market
indicators for banks downgraded at time t ⫺ x to the mean of the market indicators
of all other observations (where there was no downgrade at time t), where x ⫽ 3, 6,
12 …. We constructed the sample for the tests as for the logit model described above.
The results indicate that the banks that were downgraded subsequently had a
significantly higher mean value of (⫺DD) than those that did not. This is true up
to 24 months prior to the downgrade. We also find in the second panel that down-
graded banks had higher spreads and that spreads increase as the downgrade
approaches. However, the difference between the two groups of banks is never
statistically significant. As we argued above, the weaker predictive power of S may
be due to the government’s perceived willingness to absorb losses on behalf of
private creditors. Hence, we limit the sample to those banks with a support rating
of three or higher (i.e. banks that should generally not receive public support).
However, the results were virtually unchanged (output omitted). Only when we also
drop UK banks, the spread is significantly related to banking problems (third panel
of Table 4). The spreads of UK banks are significantly higher than those of
continental European banks in the sample.17 Consistent with this, Sironi (2003)
found that UK banks’ primary market subordinated debt spreads are much higher
than those of banks in the rest of Europe. He attributes this finding to differences
in the perceived safety net. However, if one supposes that the Fitch/IBCA support
rating correctly captures the likelihood of public support, this explanation is
not supported by our data.18 Alternative explanations could include the obligation
of UK pension funds to purchase government bonds of a certain maturity, which
in turn could artificially depress the yields of these bonds. Further, political economy
considerations relating to the structure of the financial system and the institutional
set up for bank failures could play a role (Gropp and Olters 2001). That paper shows
in a political economy model that the likelihood of a bail out may be a function of
the degree of independence from the political process of the authority deciding on
closure or bail out. It also shows that bailouts should be more likely in bank based
than in market-based financial systems. Both factors could make bail outs less likely
in the UK relative to continental Europe and, hence, result in higher spreads on UK
bank bonds. It is beyond the scope of this paper to explore the issue further. In the
following, when we examine spreads we will either control for UK banks using a
dummy variable or present results without UK banks.
4. EMPIRICAL ESTIMATION AND RESULTS
4.1 Estimation Methods

We used two different econometric models to investigate the signalling properties
of the market-based indicators of bank fragility. The first is a standard logit model
of the form:
Pr[STATUSt ⫽ 1] ⫽ ψ(α0 ⫹ α1 DIt⫺x ⫹ α2 DSUPPt⫺x* DIt⫺x) , (3)
where ψ( ) represents the cumulative logistic distribution, DIt⫺x the market indicator
at time t ⫺ x, and
STATUSt ⫽ {01 if bank was downgraded to C or below at time t

otherwise
.
We estimate the model for different horizons separately, i.e. we investigate the
predictive power of our two indicators 3, 6, 12, 18 and 24 months before the down-
grade. Significant and positive coefficients of the lagged market indicators would
support the use of ⫺DD or S as early indicators of bank fragility (Hypothesis 1).
17. The average spread of UK banks is 208 basis points (894 observations) and the average spread
of all other banks is 44 basis points (2783 observations). The difference is significant at the 1% level.
18. Sixty-one percent of UK banks have a support rating of one or two, while this share is 64% for
all other banks.
We created a dummy variable DSUPP, which equals one when the Fitch/IBCA
“support rating” is one or two in order to control for the perceived implicit safety
net. In order to test for the impact on the quality of the signal from market indicators,
we interacted the variable with the market indicators. Given a significantly positive
coefficient on S, an insignificant sum of the coefficients on S and DSUPP*S would
suggest that for banks covered by an implicit or explicit safety net, spreads are poor
indicators. In addition, a significantly positive coefficient on the sum of –DD and
DSUPP*⫺DD (given a significant coefficient on –DD), would be consistent with
Hypothesis 3, i.e. that –DD is a useful indicator even for banks implicitly or explicitly
insured. Since we use several observations for the same bank in case the bank is
not downgraded during our sample period, our observations are not independent
within banks, while they are independent across banks. Therefore, we adjusted the
standard errors for clustering using the generalised method based on Huber (1967).
Our second model is a Cox proportional hazard model of the form:
h(t,DI,X) ⫽ h0 (t) eβ1DI⫹B2X , (4)
where h(t,DI,X) represents the proportional hazard function, h0(t) the baseline hazard,
and X the control variables (see below). Again, we calculated robust standard errors, as
we had multiple observations per bank and used Lin and Wei’s (1989) adjustment
to allow for correlation of the residuals within banks. The model parameters were
estimated by maximising the partial log-likelihood function
ln L ⫽
D
{
兺兺 (β
j⫽1 r僆Dj
1 DIr ⫹ B2 Xr) ⫺ dj ln
[兺
i僆Rj
exp(β1DIi ⫹ B2 Xi)
]} , (5)
where j indexes the ordered times of a downgrade t(j) (j ⫽ 1,2,...,D). Dj is the set
of dj observations that are downgraded at t(j) and Rj is the set of observations that are
at risk at time t(j). The model allows for censoring in the sense that not all banks
were downgraded during the sample period.19
The two models provide a robustness check whether equity and bond market
indicators have signalling property with regard to downgrades. In addition, the
two models provide different insights. The logit model yields an estimate of the
unconditional probability of a downgrade at a pre-specified future point in time
(three months ahead, six months ahead, etc.). A hazard model allows us to
approximate the probability of being downgraded in the following month, conditional
on not being downgraded in the period up to that month. Hence, the proportional hazard
model not only uses the level of an indicator at a certain point of time, but also
how long the indicator remains at this level. This, conceptually, results in a more
efficient use of the available information and ultimately should lead to better discrimi-
natory power and improved forecasts.
19. For more details on estimating hazard models see Kalbfleisch and Prentice (1980) and for a more
recent treatment, Wooldridge (2002, chapter 20).
4.2 Estimation Results

Table 5A reports the results from estimating logit models with different time
leads using the distance to default as the indicator. A shorter distance to default
(i.e. an increase in ⫺DD) increases the likelihood of a downgrade. The respective
coefficient is significant at least at the 10% level for the 6, 12 and 18-month leads.
Hence, we find support for Hypothesis 1: -DD appears to have predictive properties of
an increased (unconditional) likelihood of a downgrade up to 18 months in advance.
The coefficient ceases to be significant for more than 18 months ahead of the event.
The coefficient of the three-month lead is also insignificant.20
We also find that the coefficient of DSUPP*(⫺DD)t⫺x is never statistically signifi-
cant from zero. Further, the sum of ⫺DDt⫺x and DSUPP*(⫺DD)t⫺x is significantly
different from zero at the 5% level for all lead times except for x ⫽ 24. Hence, the
safety net does not appear to be important for the predictive power of the distance to
default as an indicator of bank fragility. Equity holders seem to expect losses even
in case of default of systemically important banks.
The results for spreads, S, support Hypothesis 1 only for non-UK banks (see
Table 5B, output for the sample with UK banks omitted). The coefficients for lead
times of up to 18 months are significant at least at the 5% level. The results also
highlight that it is important to control for the expectation of public support in case
of spreads. The coefficient of DSUPP*St⫺x is significant and negative and a joint
hypothesis test reveals that the coefficient on the spread is zero for the banks with
a high expectation of public support. This is in contrast to the results for ⫺DD and
suggests that subordinated debt holders assign a probability of one to being bailed out
for some banks and, hence, bond prices do not adjust in line with risks for these
banks. Both S and ⫺DD explain relatively little of the variation in the dependent
variable: pseudo R2 is below 0.1 for all horizons.
The economic magnitudes of the effects are in line with this low explanatory
power. Evaluated at the mean, for ⫺DD an increase of one has the largest effect at
the 18 months horizon with a 1.1% increase in the probability of a downgrade. We
find the smallest effects at 3 and 24 months with 0.4%. In contrast, a one basis
point increase in spreads is associated with a 0.03% increase in the probability of
a downgrade 3 and 6 months before the downgrade and quickly diminishes to zero at
longer horizons.21 These patterns correspond closely to Hypothesis 2: the signal to
noise ratio for the spread is quite low further away from default.
20. This seems to be due to the erratic behaviour of asset volatility close to default, potentially
due to trading irregularities and/or non-linearity in the payoff to equity. Moreover, beyond some value,
the marginal impact of DD on the probability of failure becomes effectively zero. In part this motivates
the use of a dummy variable (specifying DD beyond some threshold), rather than DD itself in the
proportional hazard model estimated below.
21. In the logit model
∂E[y] eβ′X
⫽ Λ(β′X)(1 ⫺ Λ(β′X)), β ⫽ β,
∂x (1 ⫺ eβ′X)2
gives the marginal effects, with Λ being the logistic distribution.
TABLE 5A
Distance to Default: Logit Estimation
x ⫽ 3 months x ⫽ 6 months x ⫽ 12 months x ⫽ 18 months x ⫽ 24 months
Constant ⫺2.80*** (0.39) ⫺2.62*** (0.44) ⫺2.89*** (0.45) ⫺2.69*** (0.54) ⫺3.30*** (0.59)
(⫺DDt⫺x) 0.11 (1.12) 0.18* (0.10) 0.21** (0.11) 0.29* (0.15) 0.17 (0.13)
DSUPP*(⫺DDt⫺x) 0.16 (1.10) 0.11 (0.11) 0.02 (0.12) ⫺0.01 (0.13) ⫺0.03 (0.11)
Number of observations 1043 1043 1040 1039 1036
Number of banks 86 86 83 82 79
Number of downgrades 25 25 22 21 18
F-testa 5.22** 6.44** 3.78** 4.29** 1.11
Log likelihood ⫺114.4 ⫺114.0 ⫺104.0 ⫺102.7 ⫺89.3
Pseudo R2 0.03 0.03 0.03 0.03 0.02
Notes: All models are estimated using STATUS as the dependent variable. *, **, *** indicates statistical significance at the 10%, 5%, and 1% levels, respectively. Standard errors adjusted for clustering in parentheses.
a
F-test for the hypothesis that the sum of the coefficients of (⫺DDt⫺x) and DSUPP*(⫺DDt⫺x) is zero. χ2 values reported.
TABLE 5B
Spread: Logit Estimation
Constant ⫺3.36*** (0.39) ⫺3.50*** (0.42) ⫺3.42*** (0.40) ⫺3.53*** (0.44) ⫺3.43*** (0.47)
(St⫺x) 2.84*** (1.12) 4.07*** (1.56) 3.19** (1.31) 2.71** (1.11) 2.31 (2.28)
DSUPP*(St⫺x) ⫺2.55** (1.10) ⫺3.75*** (1.51) ⫺2.78** (1.29) ⫺2.40** (1.09) ⫺2.06 (2.19)
F-testa 1.41 1.86 2.09 0.67 0.29
Log likelihood ⫺69.9 ⫺66.5 ⫺64.4 ⫺52.3 ⫺50.0
Pseudo R2 0.06 0.07 0.05 0.04 0.02
Notes: Excluding UK banks. All models are estimated using STATUS as the dependent variable. *, **, *** indicates statistical significance at the 10%, 5%, and 1% levels, respectively. Standard errors adjusted for
clustering in parentheses. aF-test for the hypothesis that the sum of the coefficients of (St⫺x) and DSUPP*(St⫺x) is zero. χ2 values reported.
The results of discrete choice models may be quite sensitive to the underlying
distributional assumptions, in particular in cases where the distribution of the depen-
dent variable is as skewed as in this sample. Only around 4% of the samples are
downgraded observations. Hence, we estimated also the corresponding probit
models. We found essentially unchanged results, both in terms of magnitude and
significance.22
Next, we turn to the results from the proportional hazard model. Table 6 gives
the hazard ratios and corresponding P-values for five different models. In the model
with only ⫺DD, the hazard ratio is significant (at the 5% level) and has the expected
positive sign. The hazard ratio is increasing in the values of the fragility indicators,
which is consistent with the logit results. The table also shows results for tests of
the proportional hazard assumption (i.e. the zero-slope test), which amounts to testing,
whether the null hypothesis of a constant log hazard-function over time holds for
the individual covariates as well as globally. For ⫺DD, this assumption is violated.
The payoff from equity is equivalent to the payoff of a call option on the assets of
the firm with a strike price equal to the book value of debt. Hence, there is a non-
linear relationship between the distance to default and downgrades, which may
have been the reason for the rejection of the proportional hazard assumption. One
simple way of capturing this non-linearity is to use a dummy variable:
ddind ⫽ {01 if (⫺DD) ⬎ ⫺3.2

otherwise
, (6)
where ⫺3.2 represents the top 25th percentile of the distribution of ⫺DD. Hence,
in this specification, we investigate whether banks with “short” distances to default
TABLE 6
Market Indicators: Proportional Hazard Estimation
Spread ⫹ DSUPP,
(⫺DD) only (⫺DD) ⬎ ⫺3.2 only Spread only UK banks excluded
(⫺DD) 0.73** (0.12)

Dummy for (⫺DD) ⬎ ⫺3.2 2.69*** (1.03)
Spread 1.00 (0.002) 1.02*** (0.01)
DSUPP*S ⫺0.99*** (0.01)
Number of observations 5365 5365 3604 2720
Number of banks 86 86 59 49
Number of downgrades 25 25 21 19
Wald χ2 4.08** 6.62*** 0.56 16.38***
Time at risk 5365 5365 3604 2720
Starting log likelihood ⫺100.5 ⫺100.5 ⫺69.8 ⫺61.5
Final log likelihood ⫺96.7 ⫺97.9 ⫺69.5 ⫺57.1
Zero-slope test 7.66*** 1.52 0.40 2.58
Note: Estimated using Cox proportional hazard regression. Partial Log-likelihood function given in the text. Dependent variable: duration
(in months) until downgrade or censoring. Standard errors are corrected for clustering using Wei and Lin’s (1989) method *, **, ***
indicate statistical significance at the 10%, 5%, and 1% levels. Hazard ratios reported.
22. The results are available from the authors upon request.
are more likely to be downgraded. We find that this indicator is significant at the
1% level and the proportional hazard assumption is no longer rejected.
As in the logit estimations, the spread alone is not significant, although it does
have a positive sign. When we exclude UK banks and control for the safety net, S
becomes significant at the 1% level. For ⫺DD the inclusion of the safety net dummy
or the UK dummy do not affect its coefficient, as in the logit specification (output
omitted). We will examine the economic magnitude of the coefficient in the context of
type I/type II classification errors in the following section.
So far we have examined the question whether market prices have any uncondi-
tional predictive power with respect to rating downgrades. Given that we answer this
question in the affirmative, we now turn to the question whether they add anything
to the information contained in other data. First, we examine whether the market
indicators contain information, which is not already summarised in the Fitch/IBCA
rating. Hence, we re-estimate the logit models controlling for the individual rating
at the time the market indicators were observed. The results given in Table 7A for
⫺DD are similar to those reported in Table 5A, albeit the significance of ⫺DD is
reduced. The rating itself is significant at least at the 10% level at all horizons, but
diminishes in magnitude as one moves away from the downgrade. The results
suggest that ⫺DD adds information to Fitch/IBCA ratings and this information
is more valuable further away from the downgrade. For spreads (Table 7B) the
magnitude and the size of the coefficients are unchanged relative to the specification
without initial ratings. The coefficient on the rating tends to be insignificant and
declines away from the downgrade.23
Next, we explore whether market indicators add information to that already
available from banks’ accounting data. Conceptually, this is obvious. Market-based
indicators should fully reflect past accounting information as well as forward looking
expectations about the prospects of the bank. The choice of which accounting variables
to use is somewhat arbitrary. We followed the previous literature (e.g. FDIC, 1994,
Sironi, 2003, and Flannery and Sorescu, 1996) and considered a set of accounting
indicators emulating the categories of CAMEL ratings (Capital adequacy, Asset
quality, Management, Earnings, and Liquidity). In order to maintain a sufficient
sample size in the set of downgraded banks, we had to consider only four out of
five indicators due to data limitations. Consequently, liquidity was dropped from
the indicator. The composite score is calculated as follows:
1. We calculated capital divided by total assets (C), problem loans/total loans

(A), operating costs/operating income (M), RoE (E) for each bank in each year.
2. We determined the percentile ranking of the bank in the distribution for every
indicator in each year (1991, 1992, 1993, …, 2001).
23. The initial rating enters linearly here. We also experimented with a more flexible form, in which
we used dummy variables for each rating level. Both the distance to default and the spread continue to
enter significantly in these regressions. The drawback of using a more flexible functional form is that
we cannot identify this model for banks with high ratings as they are never downgraded to C or below
and therefore lose a substantial number of observations.
TABLE 7A
Distance to Default: Logit Estimations, Controlling for the Fitch/IBCA Individual Rating
Constant ⫺4.14*** (0.72) ⫺3.89*** (0.71) ⫺3.95*** (0.66) ⫺3.43*** (0.75) ⫺3.93 (0.83)
(⫺DDt⫺x) 0.10 (0.10) 0.19 (0.12) 0.21* (0.12) 0.29* (0.16) 0.15 (0.13)
DSUPP*(⫺DDt⫺x) 0.14 (0.12) 0.09 (0.13) 0.022 (0.14) 0.02 (0.15) 0.02 (0.13)
RATINGt⫺x 0.36** (0.17) 0.36** (0.17) 0.32** (0.17) 0.28* (0.15) 0.24* (0.14)
F-testa 3.47* 4.52** 3.22* 4.25** 1.26
Log likelihood ⫺105.6 ⫺105.2 ⫺97.0 ⫺93.2 ⫺84.3
Pseudo R2 0.09 0.09 0.07 0.07 0.04
Notes: All models are estimated using STATUS as the dependent variable. *, **, *** indicate statistical significance at the 10%, 5%, and 1% levels. Standard errors adjusted for clustering in parentheses. aF-test for
the hypothesis that the sum of the coefficients of (⫺DDt⫺x) and DSUPP*(⫺DDt⫺x) is zero. χ2 values reported.
TABLE 7B
Spread: Logit Estimations, Controlling for the Fitch/IBCA Individual Rating
Constant ⫺9.66** (3.95) ⫺8.37*** (2.99) ⫺7.84*** (2.85) ⫺7.05*** (2.10) ⫺6.18*** (1.52)
(St⫺x) 2.28*** (0.71) 3.36** (1.56) 3.08*** (1.17) 2.93** (1.13) 2.82* (1.72)
DSUPP*(St⫺x) ⫺1.99*** (0.75) ⫺3.07** (1.46) ⫺2.79*** (1.09) ⫺2.89** (1.05) ⫺2.87* (1.59)
RATINGt⫺x 1.61 (1.02) 1.25 (0.81) 1.16 (0.81) 1.025* (0.61) 0.83* (0.43)
F-testa 1.48 1.38 0.62 0.01 0.00
Log likelihood ⫺36.6 ⫺36.5 ⫺35.3 ⫺31.9 ⫺28.3
Pseudo R2 0.36 0.32 0.28 0.23 0.17
Notes: Excluding UK banks. All models are estimated using STATUS as the dependent variable. *, **, *** indicate statistical significance at the 10%, 5%, and 1% levels. Standard errors adjusted for clustering in
parentheses. aF-test for the hypothesis that the sum of the coefficients of (St⫺x) and DSUPP*(St⫺x) is zero. χ2 values reported.
3. We divided the ranking distributions in four quartiles, and assigned a score

varying from zero (best) to three (worst) to the position of the bank in
the rankings.
We obtained the composite score by simply summing up the scores for each
indicator, yielding a variable ranging from zero (a bank in excellent condition) to
12 (a bank in very poor condition). SCORE was not available for all banks and all
time periods in the sample. In order to ensure comparability, we re-estimated the
model for S and ⫺DD only for those observations for which SCORE was available. In
order to save space, we report results for the 12 months horizon only.24
Comparing the results for the two different samples (Tables 5A and 5B with
Table 8) for the model with the market indicator only, we find that the results
are robust across the different samples. We also find that accounting variables have
some predictive power with respect to downgrades. The coefficient on SCORE is
positive and significant at the 1% level in all specifications. The R2 in the models with
SCORE alone is around 0.2. In the first three columns of Table 8 we see that the
distance to default adds some information to SCORE. The distance-to-default is
positive and significant at the 5% level, and the model fit, as measured by the pseudo
R2 increases from 0.17 to 0.19. In addition, the inclusion of (⫺DD) does not change
the magnitude or significance of SCORE and the inclusion of SCORE does not
change the significance of (⫺DD) while the magnitude is only slightly reduced. This
suggests that the distance to default and accounting information are complementary to
each other.
Empirical estimates from the same exercise for spreads are presented in the next
three columns of Table 8. They suggest that spreads also add some information
to that already available from accounting data. The model combining spreads with
SCORE has a better fit relative to the one containing SCORE only. The lower
significance of S and the reduction in its magnitude in the model with SCORE suggests
that there might be some substitutability between the information in spreads and
in accounting information.
In the seventh column of Table 8 we report results for a logit model with both
indicators together. While in such a case the sample size reduces quite significantly,
both ⫺DD and S enter significantly, and the fit of the model is better than for each
individual indicator alone. When we add SCORE to the model with both market
indicators (column 9), ⫺DD is not significant while S and SCORE are significant
at the 5% and 1% percent level, respectively. The R2 increases from 0.44 (SCORE
only) and 0.18 (market indicators only) to 0.49 when combining both market indica-
tors with SCORE. The combination of high explanatory power and low significance
of some coefficients suggests substantial collinearity and some substitutability
between the different indicators.25
24. In order to ensure comparability with the results reported in Tables 5A and 5B we tried to keep
the sample sizes as similar as the ones in the previous exercises. Hence, since SCORE is available only
on an annual basis, we repeated it for each DD and S observation belonging to the same year.
25. In this model, the coefficient for (⫺DD) has a negative sign, albeit not significant. This result
is entirely due to an outlier (Banca Nazionale del Lavoro with a ⫺DD of around ⫺4.5), which given
the small number of downgraded banks in the sample (nine banks) tends to influence the results rather
significantly. Without that bank, ⫺DD becomes positive, although remains insignificant.
TABLE 8
Market Indicators and SCORE: Logit Estimations, X ⫽ 12 Months
(⫺DD) onlya SCORE onlya (⫺DD) and SCOREa S onlyb SCORE onlyb S and SCOREb (⫺DD) and Sc SCORE onlyc S, (⫺DD) and SCOREc
Constant ⫺2.88*** ⫺8.06*** ⫺7.33*** ⫺3.41*** ⫺7.54*** ⫺7.58*** ⫺2.35*** ⫺13.69*** ⫺12.69
(0.44) (1.17) (1.18) (0.40) (1.51) (1.38) (0.67) (3.68) (3.60)
(⫺DDt⫺x) 0.21** 0.18** 0.29* ⫺0.05
(0.09) (0.08) (0.16) (0.20)
(St⫺x) 3.17** 2.84* 6.64*** 6.32**
(1.31) (1.67) (2.40) (2.82)
DSUPP*(St⫺x) ⫺2.748** ⫺2.39 ⫺6.56*** ⫺6.15**
(1.29) (1.65) (2.41) (2.74)
SCORE 0.58*** 0.59*** 0.58*** 0.54*** 1.22*** 1.06***
(0.13) (0.13) (0.17) (0.15) (0.35) (0.31)
Number of 975 975 975 344 344 344 171 171 171
observations
Number of banks 83 83 83 46 46 46 27 27 27
Number of 22 22 22 17 17 17 9 9 9
downgrades
F-testd 2.38 5.18** 0.03 0.0
Log likelihood ⫺102.8 ⫺87.4 ⫺84.7 ⫺64.1 ⫺54.3 ⫺52.0 ⫺28.9 ⫺19.6 ⫺18.0
Pseudo R2 0.02 0.17 0.19 0.05 0.20 0.23 0.18 0.44 0.49
Notes: All models are estimated using STATUS as the dependent variable. *, **, *** indicate statistical significance at the 10%, 5%, and 1% levels, respectively. SCORE is a synthetic variable summarising the
ranking of the bank with regard to four indicators representing respectively capital adequacy, asset quality, efficiency, and profitability. Standard errors adjusted for clustering in parentheses. aSample consists of banks
for which DD and SCORE are available. bExcluding UK banks; sample consists of banks for which spreads and SCORE are available. cSample consists of bank for which spreads, ⫺DD and SCORE are available.
UK banks are excluded. dF-test for the hypothesis that the sum of the coefficients of (St⫺x) and DSUPP*(St⫺x) are zero, respectively. χ2 values reported.
These findings are largely confirmed in the proportional hazard analysis (Table 9).
First, we are able to confirm the finding that the results seem to be quite robust
with respect to different samples, as the results reported in Table 6 and in Table 9
are consistent, both in terms of economic magnitude and statistical significance.
Adding ⫺DD to a model with SCORE yields statistically significant estimates for
both variables and improves the overall fit of the model. The proportional hazard
model also suggests that spreads may add some information to SCORE. In the
combined model S is significant at the 5% level and has retained its economic
magnitude. The improvement in model fit is quite substantial. Moreover, as in the
logit specification, the coefficient on DSUPP*S loses its significance when including
SCORE.26 In the final three columns, we–as before–examine models with both
market indicators included simultaneously. Without SCORE, both market indicators
are significant at the 5% level. ⫺DD’s coefficient is also much larger than in other
specifications, suggesting that it is beneficial to use a model with both indicators
simultaneously. Finally we estimate a model with both market indicators and SCORE.
As for the logit model, the statistical significance of most variables is reduced; only
⫺DD is significant at the 10% level. However, the Wald statistics suggests that
model fit significantly improves relative to models with market indicators or SCORE
alone. Again, the lack of statistical significance of individual coefficients and the
high joint significance suggests multicollinearity and some substitutability among
indicators.
5. IN-SAMPLE FORECASTING
Finally, we evaluate the potential usefulness of including market information in

supervisory early warning models of bank fragility. The standard approach to evaluat-
ing the performance of forecasting models is to consider classification errors (types
I and II). We follow convention and define type I errors as missed downgrades (i.e.
indicator classifies “no downgrade,” when the bank was in fact downgraded) and
type II errors as mis-classified non-downgrades (i.e. the indicator classifies “down-
grade” when the bank was in fact not downgraded). Based on the logit models in
Table 8, we report two exercises. One, we examine the changes in classification errors
of a model with only market indicators relative to models with market indicators and
SCORE. Second, we examine the classification errors of combining the two market
indicators and then adding SCORE. The results are presented in Table 10.
Looking at the market indicators individually, they show an overall classification
accuracy of 51% (distance to default) and 70% (spread). However, spreads do a
relatively poor job of picking up failures, as the type I error is 53%; while the
distance to default does better with a type I error of 27%. In case of the distance to
default this translates into predicting 16 of 22 downgrades and 484 of 953
26. One could take this as some indirect evidence of moral hazard, as a higher likelihood of support
seems to be correlated with weaker balance sheets.
TABLE 9
Market Indicators and SCORE: Proportional Hazard Estimation
(⫺DD) onlya SCORE onlya (⫺DD) and SCORE S onlyb SCORE onlyb S and SCORE (⫺DD) and Sc SCORE onlyc (⫺DD), S and SCORE
(⫺DD) ⬎ ⫺3.2 3.38*** 2.41** 10.46** 8.00*

(1.43) (1.08) (11.01) (8.96)
Spread 1.02*** 1.02** 1.01** 1.01
(0.01) (0.01) (0.01) (0.01)
DSUPP*S ⫺0.982*** ⫺0.989 ⫺0.99* ⫺0.99
(0.01) (0.01) (0.01) (0.01)
SCORE 1.43*** 1.38*** 1.44*** 1.35*** 1.36** 1.19
(0.10) (0.11) (0.16) (0.15) (0.17) (0.18)
Number of banks 83 83 83 46 46 46 33 33 33
Number of 22 22 22 17 17 17 10 10 10
downgrades
Number of 4546 4546 4546 2044 2044 2044 1307 1307 1307
observations
Wald χ2 8.32*** 24.86*** 33.91*** 12.71*** 11.02*** 19.97*** 7.02* 5.84** 12.48***
Time at risk 4546 4546 4546 2044 2044 2044 1307 1307 1307
Starting log ⫺82.7 ⫺82.7 ⫺82.7 ⫺45.3 ⫺45.3 ⫺45.3 ⫺30.9 ⫺30.9 ⫺30.9
likelihood
Final log likelihood ⫺79.2 ⫺73.5 ⫺71.6 ⫺41.5 ⫺40.3 ⫺38.0 ⫺25.7 ⫺28.3 ⫺25.1
Zero-slope test 0.06 0.00 0.04 2.49 0.01 0.93 0.17 1.17 2.06
Notes: Estimated using Cox regression. Partial Log-likelihood function given in the text. SCORE is a synthetic variable summarising the ranking of the bank with regard to four indicators representing respectively
capital adequacy, asset quality, efficiency, and profitability. Dependent variable: duration (in months) until downgrade or censoring. Standard errors corrected for clustering using Wei and Lin’s (1989) method are in
parenthesis. *, **, *** indicate statistical significance at the 10%, 5%, and 1% levels. Hazard ratios reported. aSample is the intersection of banks for which DD and SCORE are available. bSample consists of banks
for which spreads and SCORE are available. cSample consists of banks for which DD, spreads and SCORE are available.
TABLE 10
Classification Accuracy of Logit Models
(⫺DD) SCORE (⫺DD) and SCORE S and (⫺DD) and SCORE (⫺DD), S
onlya onlya SCORE S onlyb onlyb SCORE Sc onlyc and SCORE
Type I error 0.27 0.32 0.32 0.53 0.18 0.35 0.67 0.11 0.11
Type II error 0.49 0.24 0.23 0.28 0.30 0.26 0.25 0.21 0.15
Classification 0.51 0.75 0.77 0.70 0.71 0.74 0.73 0.80 0.85
accuracy
Notes: Based on logit models presented in Table 8. Results are for x ⫽ 12, i.e. one year before the downgrade. Type I error is a mis-
classified actual downgrade; type II error a mis-classified actual non-downgrade. aSample is the intersection of banks for which DD and
SCORE are available. bSample consists of banks for which spreads and SCORE are available. UK banks excluded cSample consists of
banks for which DD, spreads and SCORE are available. UK banks excluded.
non-downgrades accurately. Spreads predict six of 19 downgrades and 235 of 331

non-downgrades accurately. Considering SCORE individually yields a classification
accuracy of 75% and 71% in the ⫺DD sample and the spread sample, respectively.
The addition of market indicators to SCORE brings about an improvement in the
classification accuracy of the model which stems from a reduction in the type II
errors: the overall classification accuracy increases to 77% in the ⫺DD sample and
to 74% for the spread.
Combining the two market indicators yields an improvement in the classification
accuracy of the model, again largely due to a reduction in type II errors. More
than half of all failures continue to be missed. Nevertheless, the results suggest that
combining the two indicators reduces type II errors quite significantly. What about the
additional information of both market indicators relative to accounting information
(“SCORE”) alone? In the (smaller) sample, SCORE alone provides an overall
classification accuracy of 80%. Adding both market indicators yields an improvement
to 85%. All of the improvement comes through a reduction in type II errors as
before.27 While these improvements are far from dramatic, they do suggest some
useful role for market information in early warning models.
These points are re-enforced, when evaluating the predictive performance of the
proportional hazard model. Table 11 gives the probabilities of not experiencing a
downgrade (“survival”) based on Kaplan–Meier survivor functions. The figures
represent the proportion of banks “surviving” for a certain amount of time with a
certain characteristic. For example, consider the model with ⫺DD only. After 36
months, 68.2% of banks with –DD ⬎ ⫺3.2 are not downgraded to C or below,
compared to 85.2% of those banks with ⫺DD ≤ ⫺3.2. In terms of classification
errors this means that 68.2% of banks are classified as good even though they were
downgraded (type I error) and 14.8% were classified as bad even though they had
not been downgraded (type II error). Using ⫺DD ⬎ ⫺3.2 as the explanatory variable
gives us total discriminatory power of 17% after 36 months. In the final column,
27. Accounting data alone classify seven of 31 non-downgrades as downgrades, accounting data plus
(⫺DD) six of 31 and both indicators plus accounting information five of 31.
TABLE 11
Classification Accuracy of the Proportional Hazard Model
Time observed 6 months 12 months 24 months 36 months 48 months Log Rank testd
Distance to defaulta
Probability of survival (low DD) 100% 95.2% 87.9% 68.2% 48.7%
Probability of survival (high DD) 97.0% 93.8% 89.2% 85.2% 78.8% 5.80**
Discriminatory power ⫺3.0% ⫺1.4% 1.3% 17.0% 30.1%
SCOREa
Probability of survival (high SCORE) 100% 98.3% 90.7% 76.8% 67.6%
Probability of survival (low SCORE) 100% 100% 100% 100% 100% 9.84***
Discriminatory power 0.0% 1.7% 9.3% 23.2% 32.4%
Distance to default and SCOREa
Probability of survival (low DD and high SCORE) 100% 100% 80.8% 65.0% 36.2%
Probability of survival (high DD and low SCORE) 100% 98.5% 96.8% 90.7% 88.3% 17.01***
Discriminatory power 0.0% ⫺1.5% 16.0% 25.7% 52.1%
Spreadsb
Probability of survival (high spread) 94.1% 86.9% 70.2% 52.0% 52.0%
Probability of survival (low spread) 100% 100% 92.7% 92.7% 92.7% 3.69*
SCOREb
Probability of survival (high SCORE) 96.9% 90.4% 80.5% 70.4% 70.4%
Probability of survival (low SCORE) 100% 100% 100% 100% 100% 3.47*
Spreads and SCOREb
Probability of survival (high spread and high SCORE) 100% 83.5% 66.7% 32.0% 32.0%
Probability of survival (low spread and low SCORE) 97.1% 94.4% 88.5% 88.5% 88.5% 8.33***
Discriminatory power ⫺2.9% 10.9% 21.8% 53.5% 53.5%
Spreads and distance to defaultc
Probability of survival (low DD and high spread) 94.7% 90.0% 76.7% 62.8% 55.8%
Probability of survival (high DD and low spread) 95.5% 95.5% 89.1% 89.1% 89.1% 3.79**
SCOREc
Probability of survival (high SCORE) 100% 92.9% 70.8% 53.1% 44.3%
Probability of survival (low SCORE) 94.1% 88.2% 88.2% 88.2% 88.2% 5.23**
Discriminatory power ⫺5.9% ⫺4.7% 17.4% 35.1% 43.9%
DD, Spread and SCOREc
Probability of survival (low DD, high spread, high SCORE) 100% 87.5% 62.5% 39.1% 26.0%
Probability of survival (high DD, low spread, low SCORE) 93.4% 90.2% 86.6% 86.6% 86.6% 10.69***
Discriminatory power ⫺6.6% 2.7% 24.1% 47.5% 60.6%
Notes: Based on proportional hazard models presented in Table 9. Kaplan–Meier survivor functions. “Survival” means not being downgraded. a(⫺DD) ⬎ ⫺3.2 and SCORE ⬎ 7; Sample consists of banks for which
DD and SCORE are available. bS ⬎ 56.9 and SCORE ⬎ 7; Sample consists of banks for which S and SCORE are available c((⫺DD) ⬎ ⫺3.2 or (S ⬎ 56.9 and DSUPP ⫽ 0)) and SCORE ⬎ 7; Sample consists of
banks for which S, DD and SCORE are available. dTests equality of survivor functions; *,**,*** denote significance at the 10%, 5%, and 1% level, respectively.
we report the log rank test for the equality of the survivor functions. For ⫺DD ⬎
⫺3.2 the difference is statistically significant at the 5% level.
Again let us examine the two questions of interest: does using both market indicators
give us more discriminatory power than one alone? And do market indicators add
information to SCORE? In general the answer to both questions is yes. ⫺DD alone
after observing the indicator for 48 months has a discriminatory power of 30%;
S alone after being observed for 48 months has a discriminatory power of 41%. Both
indicators together have a discriminatory power of 43%. We also find that ⫺DD
has to be observed for some time in order to yield discriminatory power. We attribute
this to its relatively high volatility. On the other hand, observing spreads for more than
24 months yields no additional information. These two results correspond closely
to the patterns of predictive power we found in case of the logit models. As to the
second question, adding either ⫺DD, S, or both, to a model with SCORE only
yields about 20% points in additional discriminatory power. As before, the added
discriminatory power comes from a reduction in type II errors. Adding ⫺DD to
SCORE yields a reduction in type II errors of 31% points, adding S to SCORE a
reduction of 38% points and adding both a reduction of 18% points. Note that these
figures are not strictly comparable, as the sample is smaller when we analyse both
indicators simultaneously.28 Finally, the significance of the log rank test increases
in all cases when combining indicators relative to models with either market informa-
tion or accounting information only.
6. CONCLUSION
In this paper, we analyse market price based measures as early indicators of

bank fragility. We find support for using both indicators as leading indicators of bank
fragility, where bank fragility is measured as a downgrade of the Fitch/IBCA
individual rating to C or below. We document that such a rating was generally
associated with an intervention at the bank, either through public authorities or
through parent banks. Further, we find that while the distance to default predicts
downgrades between 6 and 18 months in advance, its predictive properties are quite
poor closer to the downgrade. In contrast, spreads’ predictive powers diminish
beyond 12 months prior to the downgrade. The results further suggest that spreads
are useful predictors only for banks, which are not (implicitly or explicitly) publicly
insured against default, while the public safety net does not appear to affect the
predictive power of the distance to default. We find, as in Sironi (2003), that
the spreads of UK banks are significantly higher in a given ratings class than the
spreads of continental European banks.
28. Note that type I errors are actually slightly increased by adding market indicators, which, however,
is largely due to the precise definition of the cut-off point we used to define the indicator variable. We
attempted to maximise discriminatory power (i.e. total error) here; alternatively we could have designed
indicator variables which would have minimised type I or type II errors. Qualitatively, the results are
unaffected by this choice.
The indicators provide some additional information relative to accounting data

alone. Comparing the performance of logit models to proportional hazard models we
find that proportional hazard models have the inherent advantage to yield continually
improving forecasts as the indicators are observed through time. Finally, we find
support for the notion that the two indicators together have more discriminatory
power in predicting downgrades than each alone. We perform an analysis of the
classification errors of the indicators and find that the main use of market information
may be to reduce type II errors. Overall, we interpret our findings in a way to
suggest that supervisors (and possibly the literature) may want to devote more
attention to the equity market when considering the use of the information embedded
in the market prices of the securities issued by banks. Equity market data could provide
supervisors with useful information, which is complementary to accounting data
and subordinated debenture spreads.
This recommendation should be qualified by the usual argument that applies to
any variable used in forecasting. The mere fact that an indicator is found not to enter
significantly in a forecasting regression does not necessarily mean that authorities
should be advised not to pay attention to the variable. For the absence of forecasting
power may simply imply that the variable is already being used. Conversely, the
mere fact that an indicator is found to be useful in forecasting bank fragility does
not tell us much about the desirability of a policy that involves feedback from that
indicator. The ability of the indicator to signal the underlying sources of financial
fragility that one wants to respond to may be impaired by the very fact that the
authorities responds to it. Specifically, if it became apparent that supervisors
use market prices as indicators of bank fragility and if supervisors conditioned their
intervention on market prices, market participants may reduce their monitoring effort
and, hence, stop providing useful signals. Market participants may learn to rely on
regulatory discipline as a substitute for market discipline, as shown for example
in DeYoung et al. (2001).
It is important to highlight one caveat in the interpretation of the results. The finding
that market indicators tend to add information to a model with accounting data
should only be taken as suggestive evidence that supervisors should include market
information in their supervisory models. The reason is that the accounting data used in
this paper were obtained from Bankscope, a publicly available database. Supervisors
may have access to accounting data of higher quality and at a higher frequency.
Hence, we would suggest interpreting our findings as a necessary rather than a
sufficient condition for the use of market indicators in early warning models.
Finally, it is important to stress that there might be considerable practical difficult-
ies in using either of the indicators proposed in this paper. For example, the distance
to default, apart from its relative computational complexity, may be sensitive to
shifts in derived asset volatility. This, in turn, may be due to irregularities in the
equity trading in the period closer to default. Further, the measure is quite sensitive to
the measure of equity volatility used and distributional assumptions about equity
returns. Similarly, the calculation of bond spreads may be difficult in practice,
because of relatively illiquid bond markets, resulting in noisy price data for bank bonds
and the lack of reliable risk-free benchmarks (especially in smaller countries).
APPENDIX
Fitch/IBCA Ratings Definitions

Fitch/IBCA’s Individual Ratings attempt to assess how a bank would be viewed
if it were entirely independent, and could not rely on external support. These ratings
are designed to assess a bank’s exposure to, appetite for, and management of
risk, and thus represent the view on the likelihood that it would run into significant
difficulties. The principal factors analysed to evaluate the bank and determine these
ratings include profitability and balance sheet integrity, franchise, management,
operating environment, and prospects.
Fitch-IBCA distinguishes among the following categories:
A. A very strong bank. Characteristics may include outstanding profitability
and balance sheet integrity, franchise, management, operating environment,
or prospects.
B. A strong bank. There are no major concerns regarding the bank. Characteristics
may include strong profitability and balance sheet integrity, franchise, manage-
ment, operating environment or prospects.
C. An adequate bank which, however, possesses one or more troublesome aspects.
There may be some concerns regarding its profitability and balance sheet
integrity, franchise, management, operating environment or prospects.
D. A bank which has weaknesses of internal and/or external origin. There
are concerns regarding its profitability and balance sheet integrity, franchise,
management, operating environment or prospects.
E. A bank with very serious problems which either requires or is likely to require
external support.
Note that, in addition, there are gradations among these five rating categories,
i.e. A/B, B/C, C/D, and D/E.
The Support Ratings do not assess the quality of a bank. Rather, they are Fitch-
IBCA’s assessment of whether the bank would receive support should this be
necessary.
1. A bank for which there is a clear legal guarantee on the part of the State, or
a bank of such importance both internationally and domestically that, in Fitch-
IBCA’s opinion, support from the State would be forthcoming, if necessary. The
State in question must clearly be prepared and able to support its principal banks.
2. A bank for which, in our opinion, state support would be forthcoming, even
in the absence of a legal guarantee. This could be, for example, because of the
bank’s importance to the economy or its historic relationship with the authorities.
3. A bank or bank holding company which has institutional owners of sufficient
reputation and possessing such resources that, in our opinion, support would
be forthcoming, if necessary.
4. A bank for which support is likely but not certain.

5. A bank or bank holding company, for which support, although possible, cannot
be relied upon.
LITERATURE CITED
Berger, Allen, Sally Davies, and Mark Flannery (2000). “Comparing Market and Supervisory
Assessments of Bank Performance: Who Knows What When?” Journal of Money, Credit
and Banking 32, 641–667.
Black, Fischer, and John Cox (1976). “Valuing Corporate Securities: Some Effects of Bond
Indenture Provisions.” Journal of Finance 2, 351–367.
Bliss, Robert (2000). “The Pitfalls in Inferring Risk from Financial Market Data.” Mimeo,
Federal Reserve Bank of Chicago, December.
Bliss, Robert (2001). “Market Discipline and Subordinated Debt: A Review of some Salient
Issues.” Federal Reserve Bank of Chicago Economic Review, First Quarter, 24–45.
Bliss, Robert, and Mark Flannery (2002). “Market Discipline in the Governance of U.S. Bank
Holding Companies: Monitoring vs. Influencing.” European Finance Review 6, 363–395.
Board of the Governors of the Federal Reserve System (1999). “Using Subordinated Debt
as an Instrument of Market Discipline”, Staff Study 172, Board of the Governors of the
Federal Reserve System, December.
Bongini, Paola, Luc Laeven, and Giovanni Majnoni (2002). “How Good is the Market at
Assessing Bank Fragility? A Horse Race between Different Indicators.” Journal of Banking
and Finance 26, 1011–1028.
Curry, Timothy, Peter Elmer, and Gary Fissel (2001). “Regulator Use of Market Data to
Improve the Identification of Bank Financial Distress.” FDIC Working Paper No 2001/
01, Federal Deposit Insurance Corporation.
Dewatripont, Mathias, and Jean Tirole (1993). The Prudential Regulation of Banks. Cambridge
MA: MIT Press.
DeYoung, Robert, Mark Flannery, William Lang, and Sorin Sorescu (2001). “The Information
Content of Bank Exam Ratings and Subordinated Debt Prices.” Journal of Money, Credit
and Banking 33, 900–925.
Docking, Dianne, Mark Hirschey, and Elaine Jones (2000). “Reaction of Bank Stock Prices
to Loan-Loss Reserve Announcements.” Review of Quantitative Finance and Accounting
15, 277–297.
Elton, Edwin, Martin Gruber, Deepak Agrawal, and Christopher Mann (2001). “Explaining
the Rate Spread on Corporate Bonds.” Journal of Finance 56, 247–277.
Evanoff, Douglas, and Larry Wall (2001). “Sub-Debt Yields as Bank Risk Measures.” Journal
of Financial Services Research 20, 121–145.
Federal Deposit Insurance Corporation (1994). EMS/CAMEL User Manual. Washington D.C.:
FDIC Division of Supervision.
Flannery, Mark (1998). “Using Market Information in Prudential Banking Supervision. A
Review of US Evidence.” Journal of Money, Credit and Banking 30, 273–305.
Flannery, Mark (2001). “The Faces of Market Discipline.” Journal of Financial Services
Research 20, 107–119.
Flannery, Mark, and Sorin Sorescu (1996). “Evidence of Bank Market Discipline in Subordi-
nated Debenture Yields: 1983–1991.” Journal of Finance 51, 1347–1377.
Gilbert, R. Alton, Andrew Meyer, and Mark Vaughan (1999). “The Role of Supervisory
Screens and Econometric Models in Off-Site Surveillance.” Federal Reserve Bank of St.
Louis Review 31–56.
Gropp, Reint, and Jan-Peter Olters (2001). “The Political Economy of Bank Bail-Outs.”
Mimeo, European Central Bank, June.
Gropp, Reint, and Anthony Richards (2001). “Rating Agency Actions and the Pricing of
Debt and Equity of European Banks: What Can We Infer About Private Sector Monitoring
of Bank Soundness?” Economics Notes 30, 373–398.
Gropp, Reint, and Jukka Vesala (2004). “Deposit Insurance, Moral Hazard and Market
Monitoring.” Review of Finance 8, 571–602.
Gropp, Reint, Jukka Vesala, and Giuseppe Vulpes. “Market Indicators, Bank Fragility, and Indirect
Market Discipline.” Federal Reserve Bank of New York Economic Policy Review 10, 53–62.
Hancock, Diana, and Myron Kwast (2001). “Using Subordinated Debt to Monitor Bank
Holding Companies: Is It Feasible?” Journal of Financial Services Research 20, 147–187.
Hand, John, Robert Holthausen, and Richard Leftwich (1992). “The Effect of Bond Rating
Agency Announcements on Bond and Stock Prices.” Journal of Finance 57, 733–752.
Huang, Jing-Zhi, and Ming Huang (2003). “How Much of the Corporate-Treasury Yield
Spread is Due to Credit Risk?” Stanford University Working Paper, May 2003.
Huber, Peter (1967). “The Behaviour of Maximum Likelihood Estimates under Non-Standard
Conditions,” In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics
and Probability, pp. 221–233. Berkeley, CA: University of California Press.
Jagtiani, Julapa, George Kaufman, and Catharine Lemieux (2000). “Do Markets Discipline
Banks and Bank Holding Companies? Evidence from Debt Pricing.” Federal Reserve Bank
of Chicago Emerging Issues, Series 3R.
Kalbfleisch, John, and Ross Prentice (1980). The Statistical Analysis of Failure Time Data. New
York: Wiley Series in Probability and Mathematical Statistics, John Wiley and Sons.
KMV Corporation (2003). Modelling Risk. San Francisco: KMV Corporation.
Krainer, John, and Jose Lopez (2003). “Forecasting Bank Supervisory Ratings Using Securities
Market Information.” Proceedings of the Federal Reserve Bank of Chicago Conferences on
Bank Structure and Competition, pp. 278–289. Chicago: Federal Reserve Bank of Chicago.
Laviola, Sebastiano, Paolo Reedtz, and Maurizio Trapanese (2001). “Forecasting Bank Fragil-
ity.” Mimeo, Banca d’Italia.
Lin, D. and L. Wei (1989). “The Robust Inference for the Cox Proportional Hazard Model.”
Journal of the American Statistical Association 84, 1074–1078.
Marcus, Alan, and Israel Shaked (1984). “The Valuation of FDIC Deposit Insurance Using
Option-Pricing Estimates.” Journal of Money, Credit and Banking 16, 446–460.
Morgan, Donald, and Kevin Stiroh (2001). “Market Discipline of Banks: The Asset Test.”
Journal of Financial Services Research 20, 195–208.
Rajan, Raguran (2001). “Comment on: Market Discipline in the Governance of US Bank
Holding Companies: Monitoring versus Influencing.” In Prudential Supervision: What
Works and What Doesn’t, edited by F. Mishkin, pp. 143–145. Chicago and London: NBER
Conference Report Series.
Ronn, Ehud, and Avinash Verma (1989). “Risk-Based Capital Adequacy Standards for a
Sample of 43 Major Banks.” Journal of Banking and Finance 13, 21–29.
Sironi, Andrea (2003). “Testing for Market Discipline in the European Banking Industry:
Evidence from Subordinated Debt Issues.” Journal of Money, Credit and Banking 35,
443–472.
Wooldridge, Jeffrey (2002). Econometric Analysis of Cross Section and Panel Data.
Cambridge, MA: MIT Press.

Gropp Vesala

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Gropp Vesala

Hochgeladen von

Copyright:

Verfügbare Formate

Equity and Bond Market Signals as Leading Indicators of Bank

Reint Gropp, Jukka Vesala, Giuseppe Vulpes

Published by The Ohio State University Press

For additional information about this article

Equity and Bond Market Signals as Leading

We analyse the ability of the distance to default and subordinated bond

JEL codes: G21, G12, C35, C41

From a supervisory perspective, the securities issued by

one convenient indicator. Moreover, market information is available at a very high

Hypothesis 3: The spread is a weaker leading indicator of bank fragility than

Belgium 3 (0) 1 (0)

Bank Timing of the downgrade Type of intervention

Variable Definition N Mean Std.dev Min Max

(⫺DD) Monthly average 5450 ⫺4.97 9.16 ⫺0.71 ⫺639.41

x (months) Status Nobs Mean Std. error Differencea Difference ⬍ 0b

4. EMPIRICAL ESTIMATION AND RESULTS

4.1 Estimation Methods

STATUSt ⫽ {01 if bank was downgraded to C or below at time t

h(t,DI,X) ⫽ h0 (t) eβ1DI⫹B2X , (4)

4.2 Estimation Results

x ⫽ 3 months x ⫽ 6 months x ⫽ 12 months x ⫽ 18 months x ⫽ 24 months

x ⫽ 3 months x ⫽ 6 months x ⫽ 12 months x ⫽ 18 months x ⫽ 24 months

ddind ⫽ {01 if (⫺DD) ⬎ ⫺3.2

(⫺DD) 0.73** (0.12)

1. We calculated capital divided by total assets (C), problem loans/total loans

x ⫽ 3 months x ⫽ 6 months x ⫽ 12 months x ⫽ 18 months x ⫽ 24 months

x ⫽ 3 months x ⫽ 6 months x ⫽ 12 months x ⫽ 18 months x ⫽ 24 months

3. We divided the ranking distributions in four quartiles, and assigned a score

Finally, we evaluate the potential usefulness of including market information in

(⫺DD) ⬎ ⫺3.2 3.38*** 2.41** 10.46** 8.00*

non-downgrades accurately. Spreads predict six of 19 downgrades and 235 of 331

In this paper, we analyse market price based measures as early indicators of

The indicators provide some additional information relative to accounting data

Fitch/IBCA Ratings Definitions

4. A bank for which support is likely but not certain.

Das könnte Ihnen auch gefallen

(⫺DD) ⬎ ⫺3.2 3.38* 2.41 10.46** 8.00*