Journal of Banking & Finance 36 (2012) 923–933

Contents lists available at SciVerse ScienceDirect

Journal of Banking & Finance

journal homepage: www.elsevier.com/locate/jbf

The determinants of bank loan recovery rates

Hinh D. Khieu a,⇑, Donald J. Mullineaux b, Ha-Chin Yi c
a
College of Business, University of Southern Indiana, Evansville, IN 47725, United States
b
Gatton College of Business, University of Kentucky, Lexington, KY 40506, United States
c
McCoy College of Business Administration, Texas State University-San Marcos, San Marcos, TX 78666, United States

a r t i c l e i n f o a b s t r a c t

Article history: Using Moody’s Ultimate Recovery Database, we estimate a model for bank loan recoveries using variables
Received 13 May 2011 reflecting loan and borrower characteristics, industry and macroeconomic conditions, and several recov-
Accepted 6 October 2011 ery process variables. We find that loan characteristics are more significant determinants of recovery
Available online 18 October 2011
rates than are borrower characteristics prior to default. Industry and macroeconomic conditions are rel-
evant, as are prepackaged bankruptcy arrangements. We examine whether a commonly used proxy for
JEL classification: recovery rates, the 30-day post-default trading price of the loan, represents an efficient estimate of actual
G21
recoveries and find that such a proxy is biased and inefficient.
G28
Ó 2011 Elsevier B.V. All rights reserved.
Keywords:
Recovery rates
Ultimate recoveries
Loss given default
Credit risk

1. Introduction case studies on a single lender or combine bonds and loans in a

Recovery rates on defaulted loans play a critical role in credit risk
modeling. Under the 2004 Basel Accord (as well as the proposed
drafts of Basel III), some global banks are now permitted, and in
certain cases required, to use an ‘‘internal rating-based approach’’
to estimate ‘‘loss-given-default’’. The recovery rate is measured as
one minus loss-given-default. Prospective recovery is also a deter-
minant of the value of a credit default swap (CDS). The International
Swaps and Derivatives Association, Inc., reports that the global vol-
ume of credit default swaps grew from $631.5 billion in 2001 to
$54.6 trillion by mid-2008, yielding a growth rate of 8546%. A better
understanding of what drives loan recoveries is therefore essential.
Loan recovery is largely an unexplored area, presumably reflecting a
lack of adequate data. We use the Moody’s Ultimate Recovery Data-
base to analyze recovery rates on large syndicated bank loans origi-
nated in the United States during 1987–2007.
Most existing studies on loan recoveries appear in the practi-
tioner literature, which is dominated by ratings agencies such as
Standard & Poor’s and Moody’s. Studies in the academic literature
focus primarily on bond recovery rates. Only a few empirical aca-
demic studies of loan recoveries exist, and these either consist of
⇑ Corresponding author. Tel.: +1 812 465 7145; fax: +1 812 465 1044.
E-mail addresses: hdkhieu@usi.edu (H.D. Khieu), mullinea@uky.edu
(D.J. Mullineaux), hy11@txstate.edu (H-C. Yi).
single sample.
Although case studies are valuable, it is difficult to know
whether their results can be generalized. When bonds and loans
appear together in a single sample, the implicit assumption is that
the same factors influence recovery rates in both markets in iden-
tical ways. Yet loans and bonds have different characteristics, so
this assumption may not hold. For instance, bank loans are more
likely to be secured than bonds, and secured lenders recover more,
on average, than unsecured creditors (Acharya et al., 2007; Cantor
and Varma, 2004). Bank loans also are typically senior to bonds in a
company’s capital structure. The absolute priority rule, which
establishes the priority of claims in liquidation, implies bank lend-
ers should recover more than bondholders in default states. In
addition, banks are more involved in debt restructurings than
bondholders; consequently, banks are more likely to become major
shareholders, or assume director seats, or do both (Gilson, 1990). If
so, loans should have lower agency costs and higher expected
recoveries relative to bonds. Finally, bank loans are more likely
to be restructured privately than publicly (Gilson et al., 1990),
and private resolution tends to result in higher recoveries (Franks
and Torous, 1994). These differences between loans and bonds
warrant an investigation focused solely on loan recovery rates.
Recovery studies employ different measures of recovery rates.
One common measure is the trading price of the debt instrument
approximately 30 days after default (see, for example, Carty,
1998; Eberhart and Sweeney, 1992). This measure serves as a

0378-4266/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jbankfin.2011.10.005
924 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933
proxy for the ultimate payoff creditors obtain when the defaulting
entity emerges from bankruptcy or is liquidated. The assumption is
that loan or bond traders rationally estimate the amount that even-
tually will be recovered at some future point and discount that va-
lue back to the trading date. The literature provides no strong
evidence on the effectiveness of trading prices as unbiased estima-
tors for ultimate recoveries, however. We will use an actual mea-
sure of recoveries in our analysis and examine how well trading
prices can proxy for ultimate recoveries.
We first investigate the factors that drive actual bank loan
recovery rates using a sample of loans originated by a great many
lending institutions. Ours is one of the first academic studies to use
Moody’s Ultimate Recovery Database (URD) for this type of analy-
sis.1 To summarize the results, we find that borrower leverage before
default negatively affects ultimate recoveries. Firm size also influ-
ences recoveries, but the impact differs by the type of loans. Recov-
eries from borrowers with a history of defaults are significantly
higher than those of first-time defaulters. A variety of loan contract
features are strongly related to the ultimate payoff to creditors, and
the use of prepackaged bankruptcy reorganizations increases actual
settlements. Loan recoveries are nonlinearly related to the length of
time to emergence. The probability of default measured at the time
of loan origination and the borrower’s operating cash flows are not
related to actual settlement values, however.
Second, we examine how well the market for bank loans antic-
ipates settlement values in recovery. We investigate whether the
trading price on loans roughly 30 days after default represents a
‘‘rational’’ forecast of actual recovery. A rational forecast, by defini-
tion, is unbiased, but is also ‘‘efficient’’ in the sense that the factors
that influence the forecast are identical to those that determine the
actual outcome. We find that the post-default trading price on
loans is not a rational estimate of actual recovery realizations.
The paper proceeds as follows. In Section 2, we review the re-
lated literature. In Section 3, we discuss our model. In Section 4,
we describe our data and present our econometric specification.
Sections 5–7 contain the empirical results. Section 8 concludes.
2. Related literature
Existing empirical studies of recovery rates primarily focus on
bonds. The lack of research on bank loan recoveries reflects the pau-
city of loan recovery data, given that loans are private debt con-
tracts.2 At a descriptive level, several empirical studies on defaulted
loans show that the recovery rates exhibit a bimodal distribution (Ara-
ten et al., 2004; Asarnow and Edwards, 1995; Schuermann, 2004).
That is, loan defaults typically result in full recovery or have a zero
or very low recovery rate. Dermine and Neto de Carvalho (2006) find
a similar distribution for loans in a sample drawn from one bank. How-
ever, other studies do not confirm bimodality, and instead show that
loan recovery rates are skewed to the right, while bond recoveries
are left skewed (Emery, 2007). The inconsistent results in these stud-
ies may reflect differences in datasets, time periods, or both.
In a practitioner study, Emery (2007) examines recovery rates
on bank loans using Moody’s database of ultimate recoveries and
finds a mean rate of 80% at resolution, compared with 65% for
bonds.3 Emery and Ou (2004) show that loss severity in the event
1
A study by Zhang (2009) also uses the Moody’s URD data.
2 sample of loans made by a single bank.
Frye (2000) attempts to explain the apparent doubling of loan loss severity over
the period 1970–1999, but notes that the data required for loan defaults are not
available. Instead, he fits his model to bond data.
3 4
Emery (2007) defines ultimate recoveries as ‘‘the recovery values that creditors
actually receive at the resolution to default’’. In Altman’s (2009) literature review,
‘‘ultimate recovery rates refer to the nominal or discounted value of bonds or loans
based on either the price of the security at the end of the reorganization period
(usually Chapter 11) or the value of the package of cash or securities upon emergence
from restructuring’’.
of default is about 2.25 times greater for bonds than loans, holding
ratings constant. Asarnow and Edwards (1995) present a univariate
analysis of bank loan default data on 831 commercial and industrial
(C&I) loans and 89 structured loans4 made by Citibank over 24 years
and find an average recovery of 65% for C&I loans and 87% for struc-
tured loans. The higher recovery rate on structured loans reflects the
fact that such loans are heavily collateralized and contain many
restrictive covenants. Acharya et al. (2007) report recovery rates of
81.12% for bank loans, 59% for senior secured bonds, 56% for senior
unsecured bonds, 34% for senior subordinated bonds, 27% for subor-
dinated bonds, and 18% for junior subordinated bonds, respectively,
for the period from 1982 to 1999.
Researchers have recently developed mathematical models that
deal directly with recovery rates. Guo et al. (2009) extend the re-
duced-form approach to credit-risk modeling to explicitly model
the recovery rate process in terms of the firm’s assets and liabili-
ties. Although risky debt is priced prior to default in existing mod-
els, Guo et al. (2009) present a regime-switching model and a jump
diffusion process to quantify defaulted debt prices and realized
recovery rates. Their recovery model differs from other reduced-
form approaches by delivering a pricing formula for zero-coupon
bond prices both before and after default, whereas other models
price such debt instruments only at or before default.
Bakshi et al. (2006) provide a multifactor model of recovery
rates. Their contributions are twofold: (1) credit spreads can be
separated into default and recovery components and (2) recoveries
as a fraction of the discounted par value fit a sample of BBB-rated
bonds better than recoveries of face value. The former are found to
reduce out-of-sample pricing errors.
Karoui (2007) also models recovery rates and derives pricing
formulas for risky debt instruments. His model is more tractable
than comparable models because recovery rates follow a discrete
– rather than continuous – time process. He estimates the model
using BBB and B Standard & Poor’s yield indices and shows that
the implied loss given default is approximately 27% for the BBB in-
dex and 74% for the B spreads.
Most early academic studies on credit risk assume that proba-
bilities of default (PD) and recovery rates (RR) are uncorrelated
(see, for example, Jarrow et al., 1997). There are reasons to doubt
this assumption, however. For instance, research on credit rating
transitions shows that recovery and default are each related to
conditions external to the firm such as macroeconomic, industrial,
geographic, and temporal factors. Altman (2009) notes that collat-
eral values, which affect bond and loan recovery rates, decline as
economic conditions deteriorate, whereas the number of defaults
increases in a weak economy. Altman et al. (2005) report a signif-
icant negative relation between aggregate default rates and recov-
ery rates on bonds over the period 1982–2002. They show that
previous studies, which ignore this correlation, understate both ex-
pected and unexpected losses. Jokivoulle and Peura (2000), Hu and
Perraudin (2002), and Das and Hanouna (2009) support Altman
et al.’s (2005) findings.
Our paper differs from the existing literature on recovery rates
in that we (1) use a direct measure of recoveries – discounted set-
tlement rates from Moody’s URD – rather than a trading price
proxy, (2) focus on defaulted bank loans rather than a combination
of bonds and loans, and (3) use a sample of loans originated by a
variety of financial institutions to many companies rather than a
Asarnow and Edwards (1995, p. 13) elaborate on the structured loans they study
as follows:  The loans are very closely monitored – the bank directly controls the
company’s cash receipts and disbursements.  The loans are highly structured and
contain many restrictive covenants.  The loans are highly collateralized, and lending
is done on a formula basis, for example, having a pre-determined advance rate against
customer receivables as collateral.’’
 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933
3. The model
We classify factors that should affect loan recoveries into four
groups: (1) loan characteristics, (2) recovery process characteristics,
(3) borrower characteristics, and (4) ‘‘external’’ factors, such as mac-
roeconomic conditions and industry characteristics. We also exam-
ine whether the probability of default is related to recovery. We
discuss the prospective roles of factors within each group, including
references to existing empirical evidence, where relevant.
3.1. Loan characteristics
3.1.1. Loan size
Loan size could influence recovery rates, although the existing
evidence is mixed. Dermine and Neto de Carvalho (2006) find, for
example, that loan size is negatively related to actual recovery rates
over the 12–48 months following default. The authors argue that
banks will delay foreclosure on larger loans because other bank cli-
ents who conduct business with a borrower in default would be ad-
versely affected by foreclosure and consequently could default on
their own loans. This prospective ‘‘spill-over effect’’ results in lower
recovery rates when these large loans eventually enter foreclosure.
In contrast, Acharya et al. (2007) find a positive relationship between
recovery rates and debt issue size, consistent with their hypothesis
that recoveries should increase with debt size because larger bor-
rowers have greater bargaining power in bankruptcy. Thorburn
(2000) studies Swedish small business bankruptcies and finds that
loan size is not a significant determinant of recovery.
We include loan size as a variable in our model, but are agnostic
with respect to the sign. Although the relationship aspect of lend-
ing suggests a prospective negative sign, the fact that large credi-
tors could exploit their bargaining power in resolving defaults,
especially in bankruptcy, argues for a positive coefficient between
loan size and recovery rates.
3.1.2. Loan types
Moody’s URD includes two types of loans: term loans and
revolvers. All are commercial credits. A term loan is normally used
to finance a long-term project, such as the purchase of fixed assets
or building a manufacturing plant. Compared with a term loan, a
revolving credit or a line of credit is typically short term, smaller,
and more often granted without collateral (Rose and Hudgins,
2008, p. 548; Strahan, 1999). No prior research examines whether
differences in loan type affect recovery rates differently.5 One char-
acteristic that distinguishes term loans from revolvers is maturity.6
Term loans are more likely to be secured than are revolvers, but their
longer maturities could enhance prospects for deterioration in col-
lateral value, thereby reducing recovery rates. The short duration
of revolvers also triggers frequent requests for renewals, which al-
lows banks to reevaluate default and recovery prospects by limiting
loan size or asking for more or better collateral.
We include a binary variable for the loan type in our model, but
do not hypothesize a sign. We also include an interaction term be-
tween the size of the borrower and the type of loan. If large firms
tend to rely more heavily on one type of loan, and loans to large
firms are more likely to fetch higher recoveries than small ones,
then the effect on recovery from loan types may also shift with
firm size. We discuss the potential impact of firm size on loan
recoveries below in the section on borrower characteristics.
5 10
Evidence shows that term loans are priced differently from revolvers, all else
equal (Gottesman and Roberts, 2004; Harjoto et al., 2006).
6
By definition, term loans have a maturity greater than one year. Revolvers can
have any maturity, but many have a maturity of less than a year. Loan commitments
of 364 days or fewer do not need to be backed by bank capital.
925
3.1.3. Collateral
Finance theory suggests that recovery rates should be higher on
secured than unsecured loans because creditors gain a legal right
to seize and sell certain designated assets in the event of default.
Both the practitioner and academic literature find collateral to be
significantly related to recoveries (see Altman and Kishore, 1996;
Araten et al., 2004; Van De Castle and Keisman, 1999).
Bank loans tend to be more highly collateralized than bonds at
origination.7 Banks, unlike bondholders, also have the option to seek
more or better collateral on existing loans when a borrower’s finan-
cial condition deteriorates.8 In contrast, bondholders cannot readily
revise the terms of a loan contract outside of bankruptcy, since the
Trust Indenture Act requires that the majority of the creditors must
assent to any contractual revisions.
We include various measures of collateral as dummy variables
in our model and expect a positive relation between the presence
of collateral and recovery rates on loans. Because we can observe
different types of collateral, we can also determine whether some
types of security yield higher recovery rates than others.
3.2. Recovery process characteristics
3.2.1. Prepackaged bankruptcy, traditional bankruptcy, and
out-of-court restructurings
A default can usually be resolved by restructuring (or reorgani-
zation) or liquidation. Restructuring or reorganization can be pub-
lic or private. A private reorganization is often referred to as a
private workout or out-of-court settlement. A public reorganiza-
tion involves a Chapter 11 bankruptcy filing. Liquidation can be
carried out through a Chapter 7 filing. The process debtors use to
resolve financial distress with various classes of claimholders is
likely to affect recovery rates. In the mid-1980s, a hybrid form of
financial reorganization emerged in which a debtor concurrently
files a bankruptcy petition and reorganization plan, known as a
prepackaged bankruptcy or prepack. Such a plan needs a confirming
vote from two-thirds of voting creditors in debt amounts and more
than 50% in number from each class of creditors, but the vote can
be held before or after the plan is filed.9 The pre-petition negotia-
tion and agreement result in lower legal costs, shorter time in bank-
ruptcy, and fewer bargaining frictions between shareholders and
creditors. This should translate into higher recovery rates on pre-
packaged bankruptcies.
Ryan (2008) shows that companies that prepackage appear to
be more sound operationally than those that do not. In addition,
firms that use distressed exchanges,10 which are informal and pri-
vate workouts, are normally more solvent at the time of reorganiza-
tion than bankrupt firms (Franks and Torous, 1994). As a result,
creditors to these firms may not have to accept large ‘‘haircuts’’ (dis-
counts) in the recovery process (Ryan, 2008). Franks and Torous
(1994) find that the average recovery rate is 80.1% in private work-
outs, compared with 50.9% in formal Chapter 11 procedures. McCon-
nell et al. (1996) document that the average recovery rate for debtors
that reorganized through prepacks from January 1980 to June 1993
7
The Moody’s URD we have reveals that 938 of 1484 loans (63%) have all assets
pledged as collateral versus 134 of 2402 bonds (5.6%). Also, unsecured loans number
122 of 1484 (8.2%), whereas the proportion of unsecured bonds is 1825 of 2402 (76%).
The same holds true for all other types of collateral except for pledges of PP&E and
other equipment.
8
A necessary condition for a bank to ask for increased collateral is that the
borrower violates one or more covenants, resulting in ‘‘technical default.’’
9
A completely private workout before a bankruptcy filing would be successful only
if 100% of the creditors agreed.
According to Keenan (2000), a distressed exchange occurs (1) when the distressed
issuer makes a tender offer to pay cash for all or part of an outstanding debt, which is
usually at a price higher than the trading price but lower than the face amount; or (2)
the issuer makes an exchange offer to replace the outstanding debt securities with a
new package of securities such as cash, new bonds, stocks, or some combination.
926 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933
(72.9%) lies between those rates in the traditional (non-prepackaged)
Chapter 11 proceedings and private workouts. Izvorski (1997) shows
that average recoveries on 281 defaulted US corporate bonds for the
period 1983–1993 are 30.36% for debtors entering Chapter 11 and
36.23% for firms going through informal reorganizations.
To account for the effects of different forms of restructuring or
reorganization on recovery, we include a dummy variable for our
sample firms that went through Chapter 11 or Chapter 7 bank-
ruptcy procedures with a prepack, a dummy variable for those that
restructured outside of court, and a dummy variable for ‘‘other
types’’ of restructuring.11 The non-prepackaged bankruptcy dummy
variable serves as the reference group. We expect the coefficients of
the included dummies to have positive signs.
3.2.2. Time to emergence
The time it takes a defaulting borrower to emerge from bank-
ruptcy or a restructuring should also affect recovery rates. The
shorter this time horizon, the lower are costs generated in the
debtor’s legal battle to gain a new life. Also, a longer time to emer-
gence may imply that stakeholders cannot agree on the fair value
of the defaulting firm’s assets relative to its liabilities. Carty
(1998) points out that both secured and unsecured claimholders
may not be paid interest during the workout period, so neither
would prefer a long, drawn-out path to resolution. This should im-
ply lower recoveries as the time it takes to emerge from financial
distress increases. We consequently hypothesize a negative rela-
tionship between the observed time to emergence and recovery
rates.
We also interact the time to resolution with the existence of a
prepackaged bankruptcy. Because prepacks can reduce the time
to emerge, and both variables are included in the model, we are
interested in the marginal effect of prepacks on recovery rates,
holding the time to emergence constant.
Finally, Covitz et al. (2006) find that US bond recovery rates
from 1983 to 2002 are nonlinearly related to time in default. In
particular, they document that bond recovery rates increase with
time in default for the first one and a half years post-default, but
decrease afterward. To account for this prospect, we include a qua-
dratic term in our model for the time to emerge. Because our sam-
ple involves only bank loans, the estimated impact for this
quadratic term may be different from that observed in bond stud-
ies. Bank lenders are less dispersed than bondholders, so negotia-
tions with defaulting firms may proceed more quickly in the case
of loans.
3.3. Borrower characteristics
Borrower characteristics are commonly included as explanatory
variables in research involving bond recoveries (Acharya et al.,
2007; Covitz and Han, 2004). However, existing loan recovery
studies include only a limited number of firm characteristics.12
The question is: Can accounting information about borrowers near
the time of default help predict recovery rates to lenders?
11
Excluding bonds, Moody’s URD on loans alone contains 14 observations for firms
that defaulted and cured their default by paying back their obligations and 12
observations for other types of restructuring. We have no hypothesis for this type of
default, but still include the data in our sample for completeness in our binary
variable constructions.
12
Grunert and Weber’s (2009) bank loan recovery analysis includes firm size
measured as the natural logarithm of total assets. Dermine and Neto de Carvalho
(2006) employ only firm age as a firm characteristic. Thorburn (2000) studies debt
recovery rates on Swedish defaults without specifying whether the focus is on bonds
or loans and includes size and profit margin as firm characteristics.
3.3.1. Firm size
The impact of firm size on recoveries is ambiguous ex ante.
Large firm size may proxy for high bankruptcy costs, which may
in turn result in lower recovery rates. Yet larger firms presumably
present less severe information asymmetry problems to creditors.
If so, the restructuring process is likely to occur more quickly for
large firms than for small ones, which may in turn improve recov-
eries for lenders. Gilson et al. (1990) emphasize that creditors are
more willing to accept a restructuring plan from stockholders
who can provide more transparent information about firm value.
We include firm size as an explanatory variable in our model,
with no hypothesized sign for the coefficient. We measure firm size
as the natural logarithm of the market value of firm assets 1 year
before default, computed as the book value of debt plus the market
value of equity.
3.3.2. Cash flows, asset tangibility, and leverage
Like Acharya et al. (2007), we include cash flows, asset tangibility,
and firm leverage 1 year before default as borrower-related explan-
atory variables in the recovery model. Higher cash flows, measured
as EBITDA 1 year before default scaled by total book-
value assets, should positively affect recoveries. Acharya et al.
(2007) argue that more cash-rich firms can attract higher prices for
their assets from potential buyers, holding other variables constant.
Higher asset tangibility should potentially increase recovery va-
lue for creditors. Williamson (1988) argues that debtholders’
recovery depends on the degree to which the borrower’s assets
are not designated to specific uses or are re-deployable. We mea-
sure asset re-deployability or tangibility as a ratio of the sum of
net property, plant, and equipment to total assets 1 year before de-
fault. Conditional on the impact of any assets pledged as collateral
on recovery, this variable should have a positive coefficient, reflect-
ing the marginal effect of the amount of tangible assets on which
lenders hold a claim.
Two strands of literature express opposite views on the poten-
tial relation between recovery and leverage. An argument for a po-
sitive link claims that higher leverage prompts increased
monitoring of managers’ behavior, which in turn results in higher
asset quality. Therefore, higher leverage should vary positively
with recovery rates, when high-quality assets are liquidated.
However, some prior studies find a significant negative relation-
ship between leverage and recovery (Acharya et al., 2007; Cantor
and Varma, 2004).13 Acharya et al. (2007) argue that firms with
higher leverage have greater dispersion in debt ownership. This
complicates restructuring negotiations and results in lower recovery
rates. Carey and Gordy (2008) study firm-level recovery of firms
defaulting over the period 1987–2006 and find that recovery rates
are reduced by approximately 5 percentage points when the median
leverage ratio increases by one unit.
We measure leverage as the sum of total long-term debt and
debt in current liabilities scaled by the borrower’s book-value of
assets 1 year before default. We hypothesize a negative sign for
the impact of leverage on loan recoveries.
3.3.3. Prior defaults
Loans to borrowers with a history of prior defaults may produce
higher recovery rates, ceteris paribus. Such borrowers may be re-
quired to provide lenders with more collateral or be subject to
greater scrutiny in the loan review process. Prior research has ig-
nored this factor. We hypothesize that loans from borrowers with
prior defaults will have higher recovery rates. We include a binary
variable for this factor and expect a positive sign.
13
Their data include both bonds and loans, while our sample deals with bank loans
only.
 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933
3.4. Macroeconomic conditions, industry characteristics, and
probability of default
3.4.1. Macroeconomic conditions
Numerous studies show that recoveries decline during reces-
sions. Business cycles, not surprisingly, are also linked to the prob-
ability of default (Gersbach and Lipponer, 2003). Frye (2000) fits
his model to Moody’s data on US corporate bonds and estimates
that recoveries decline by 20–25% during downturns relative to
normal-year recoveries. Caselli et al. (2008) use a large sample of
11,649 loans to households and to small and medium-sized enter-
prises (SME) from the five largest Italian banks between 1990 and
2004 and find that LGD depends on various macroeconomic
variables.
We include a measure of real annual GDP growth as a proxy for
the state of the economy and expect a positive sign for this vari-
able’s estimate. Although there is a debate on whether GDP growth
is a good proxy for overall economic conditions, we include this
measure in our baseline model and examine alternative measures
in our robustness checks.
3.4.2. Industry characteristics
Two types of industry characteristics may be relevant to loan
recoveries: industry distress and cross-sectional variation across
industries themselves.
3.4.2.1. Industry distress. Theoretical and empirical studies exist on
how the asset values of defaulting firms change with industry con-
ditions. Shleifer and Vishny (1992) present a market-equilibrium
model focused on asset liquidation. Firm assets will sell at lower
values in default in distressed industries because the most likely
buyers are firms in the same industry that are also in distress.
Acharya et al. (2007) empirically confirm that industry distress is
negatively related to recoveries. We measure the presence of
industry distress with an indicator variable that takes a value of
one when a given firm is in an industry whose median stock return
in the year of default is 30% or less and zero otherwise. We expect
a negatively signed coefficient for this variable.
3.4.2.2. Cross-section variation across industries. Whether ‘‘industry
effects’’ exist on recoveries, as evidence shows in the case of de-
faults, is a matter of debate in empirical studies. Academic studies
(for example, Altman and Kishore, 1996; Dermine and Neto de
Carvalho, 2006) find cross-sectional variation in recoveries across
different industries, but industry’s impact on recoveries fails con-
firmation in practitioner studies (see Araten et al., 2004; Gupton
et al., 2000). We include binary variables for all industries using
the classifications employed by Moody’s. Where too few observa-
tions exist in any industry for meaningful regression analyses, we
aggregate and label them ‘‘Other Industry.’’
3.4.3. Probability of default
The probability of default should be an important factor in
determining the recovery rates for creditors. Numerous empirical
studies have established a direct or indirect link or both between
the probability of default (PD) and recovery rates (RR). Altman
et al. (2005) find a negative association between PD and RR using
defaulted bond data over the period 1982–2002. Hu and Perraudin
(2002) document a similar negative relationship from Moody’s de-
fault data covering the 22 years from 1971 to 2000.
We include a proxy for default prospects in our model and ex-
pect a negatively signed coefficient. The proxy is the all-in-spread
on the original and drawn loans that later default. The spread is
measured as the amount the borrower pays in basis points over
the London Inter-Bank Offered Rates (LIBOR) for each dollar drawn
down, including any annual fees, to the lender. The underlying idea
927
is that the spread captures the lender’s best estimate of the likeli-
hood of default at the time a loan is originated.
4. Sample data and econometric specification
4.1. Sample data
We use bank loan recovery and loan characteristics data from
Moody’s URD covering the period from 1987 to early 2007. The
borrowers are North American commercial and industrial firms
with total debt value at the time of default greater than $50 mil-
lion. Moody’s data sources include SEC filings, news, and bank-
ruptcy documents. We manually match these firms with
Standard & Poor’s Compustat firms by borrower names to obtain
firm characteristics. The US GDP data are from the US Department
of Commerce Bureau of Economic Analysis. The stock return data
are from the Center for Research in Security Prices (CRSP). The
all-in-spread for drawn funds is from Reuters Loan Pricing Corpo-
ration (LPC) DealScan database. Our final sample contains 1364
observations for settlement value recoveries.
Our dependent variable is the discounted settlement value re-
ported by Moody’s, which they define as ‘‘the nominal settlement
recovery amount discounted back from each settlement instru-
ment’s trading date to the last date cash was paid of the individual
defaulted instruments, using the defaulted instrument’s effective
interest rate,’’14 where the nominal settlement recovery is the value
of the settlement instruments received at or close to emergence and
expressed as a percentage of the principal amount at default.
4.2. Econometric specification
We estimate the relation between recovery rates (RR) and the
set of factors discussed above. Our initial specification is:
RR ¼ b0 þ b1  Loan Characteristics
þ b2  Recovery Process Characteristics
þ b3  Firm Characteristics
þ b4  Macroeconomic Conditions; Industry Characteristics; and PD þ e
ð1Þ
We apply two estimation methods to our models. Following
Acharya et al. (2007), we first apply OLS methods with Huber-
White (1980) heteroskedasticity-consistent standard errors,
assuming a linear relationship between recovery rates and the
explanatory variables. However, because our recovery data are
mostly bounded between 0 and 1 and a large concentration of
observations exist at the extreme boundary of 1, estimating a lin-
ear functional form for Eq. (1) may produce the same shortcoming
as using a linear probability model for binary data. That is, the im-
pact of any explanatory variable may not be constant throughout
its range of values and the predicted value is not guaranteed to
be bounded between 0 and 1. This method would also ignore
important nonlinearities that might exist with our recovery data.
According to Papke and Wooldridge (1996), using a linear model
for a log-odds function makes it difficult to retrieve the expected
value of the fractional dependent variable. One common method
to handle data with the characteristics of our mostly binary recov-
ery rates is the quasi-maximum likelihood estimation. The estima-
tion is obtained from the following maximization problem:
X
n
Max li ðh1 Þ
h1 2H
i¼1
where the Bernoulli log-likelihood function is given by:
li ðh1 Þ ¼ yi1 ln Gi þ ð1  yi1 Þ lnð1  Gi Þ
with the most common form for Gi being the logistic function.
14
From Moody’s URD’s technical specifications, p. 4.
928 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933

Table 1
Descriptive statistics. Variable definitions are in Appendix A.

Variable N Mean Median Std. dev. Minimum Maximum

SETTLEMENT 1364 0.8414 1 0.2822 0 1.4021
LOANSIZE ($M) 1355 142 70 230 0 209
LOANTYPE 1364 0.4611 0 0.4987 0 1
ALLASSETCOLL 1364 0.6261 1 0.4840 0 1
INVENTRECEIVECOLL 1364 0.0242 0 0.1537 0 1
OTHERCOLL 1364 0.2427 0 0.4289 0 1
UNSECURED 1364 0.0770 0 0.2667 0 1
PREPACK 1364 0.1862 0 0.3894 0 1
RESTRUCTURE 1364 0.1246 0 0.3304 0 1
OTHERDEFAULT 1364 0.0191 0 0.1368 0 1
BANKRUPTCY 1364 0.6701 1 0.4704 0 1
TIMETOEMERGE 1364 13.8156 10.2667 14.8546 0 110.1667
FIRMSIZE ($M) 849 1176 16 10,302 1 157,459
FIRMPPE 838 0.3949 0.3703 0.2339 0 0.9586
FIRMCF 827 0.0529 0.0704 0.1678 3.0783 0.6880
FIRMLEV 836 0.8156 0.7183 0.4782 0 3.7418
EVERDEFAULTED 1364 0.1261 0 0.3321 0 1.0000
GDP 1364 0.0486 0.0469 0.0132 0.0317 0.0769
AIS 1364 0.0330 0.0275 0.0340 0 0.7500
INDDISTRESS 893 0.1680 0 0.3741 0 1

Following Papke and Wooldridge (1996) and Dermine and Neto
de Carvalho (2006), we therefore apply the quasi-maximum likeli-
hood estimation (QMLE) method to estimate the following nonlin-
ear model:
0 1
b0 þ b1  Loan Characteristics
B C
B þb2  Recovery Process Characteristics
B C
EðRRjXÞ ¼ GB
B þb  Firm Characteristics
@ 3
þb4  Macroeconomic Conditions; Industry Characteristics; and PD
ð2Þ
e Xb
where G(  ) is the logistic function of the form: GðXbÞ ¼ 1þe Xb . The
attractiveness of this method is that the QMLE of b is simple for pffiffiffiffia
member of the linear exponential family, consistent, and N-
asymptotically normal, irrespective of the distribution of the RR
conditional on X(Papke and Wooldridge, 1996, p. 622; also see Gou-
rieroux et al., 1984).
5. Empirical results
5.1. Univariate analysis results
Descriptive statistics for the sample are in Table 1.15 The average
(median) settlement value recovery rate is 84.14% (100%). Our sam-
ple is split almost evenly between revolvers (54%) and term loans
(46%). About 19% of the recoveries have some type of a reorganiza-
tion plan that shareholders approved before or at the time of the
bankruptcy filing. The average length of time a sample firm stays
in default is 14 months, and the maximum is almost 10 years. All
of the firm characteristics in our baseline analyses are measured
1 year before default. The mean (median) cash flows relative to total
assets for the settlement sample firm are 5.29% (7.04%).
As shown in Table 2, most of the firms in the sample (67%) de-
faulted in a non-prepackaged bankruptcy, whereas a few went
through prepackaged bankruptcy (19%) and private workouts
(12%). The mean recovery rates for loans with a reorganization plan
lie between those for loans resolved in the traditional bankruptcy
and those for loans going through the other forms of default reso-
lution, which is consistent with McConnell et al. (1996). We also
15
Emery’s (2007) practitioner study uses Moody’s URD and provides simple
descriptive statistics. We use one more year of defaults than he employs, which
gives us 39 more observations. Our overall results are very similar to his.
find from our sample that debtors with a prepackaged negotiation
resolve their default in about four months, on average, compared
with 19 months for non-prepackaged bankrupt cases, which is
not reported for brevity. This difference is significant at the con-
ventional level.
C
5.2. Multiple regression results
C
C
A
Table 3 reports the multiple regression results. Model 1 shows
the OLS regression results with heteroskedasticity-consistent stan-
dard errors. Model 2 presents a quasi-maximum likelihood estimate
of recovery rates based on the logistic function. The negative sign of
the LOANSIZSE coefficient is consistent with the findings of Carty
and Lieberman (1996), Dermine and Neto de Carvalho (2006), and
Emery (2004). However, the coefficient is insignificant in the speci-
fication, corroborating the findings of Altman and Kishore (1996),
Asarnow and Edwards (1995), and Thorburn (2000). Acharya et al.
(2007) argue that the bargaining power of large creditors may
increase recoveries for large debts, while Dermine and Neto de
Carvalho (2006) suggest that delayed foreclosures on large loans
may decrease subsequent recoveries. Neither argument appears to
hold in the case of our loan-only models.
On average, a term loan recovers about 4 cents less on a dollar
than a revolver, when firm size is evaluated at its mean value.16
The result suggests that the short-term nature of revolving lines of
credit may subject borrowers to increased monitoring from lenders,
which in turn limits the loss in the event of default. This is consistent
with Rajan and Winton’s (1995) model, which argues that short-
term debt gives lenders greater flexibility and control to effectively
monitor managers. We also find a differential effect of firm size
across the two types of loans. In fact, recoveries increase with firm
size in the case of term loans.
Consistent with the prior literature on both bonds and loans,
our results show that loans supported by a pledge of any type of
assets have higher recovery rates than unsecured loans, other
things equal. We find loans collateralized by inventories or ac-
counts receivable (or both) provide about 30% more recovery than
non-collateralized credits, all else constant. This form of collateral
16
The marginal impact above was evaluated at the average FIRMSIZE of 3.2826. If
the average firm size is computed based on only those firms whose LOANTYPE = 1, the
magnitude of the marginal impact is similar. The partial derivative of RR with respect
to LOANTYPE, a binary variable, as shown above, is just a Taylor-series linear
approximation over the range between zero and one.
 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933
Table 2
Settlement recovery rates by types of default. Variable definitions are in Appendix A.
Prepackaged formal Non-prepackaged formal
bankruptcy bankruptcy
(1) (2)
Settlement Mean 90.00 79.83
recovery (Median) (100) (100)
N 1364 254 914
(%) (100) (19) (67)
yields greater protection against default than other types in our
estimations.
Turning to the recovery process variables, the coefficient esti-
mate of PREPACK is significant, albeit only at the 10% level, and
positively signed. The other recovery process variables are insignif-
icant. Restructuring with prepackaged negotiations yields higher
recoveries than non-prepackaged bankruptcies, other things equal.
Recoveries increase, on average, by about 11 cents on a dollar
when TIMETOEMERGE is evaluated at its average of 14 months.
The effect of PREPACK is significantly higher than that of RESTRUC-
TURE, which is inconsistent with McConnell et al.’s (1996) finding
that higher recoveries come from private workouts.
In our initial specification, the longer the time horizon of the de-
fault, the lower is the recovery rate. The quadratic and interaction
(with PREPACK) term coefficients are each positive and significant.
Like Covitz et al. (2006), who study bond recoveries, we find a non-
linear relationship between recoveries and time in bankruptcy.
However, although those authors find a concave relation between
recoveries and time to resolution, our relationship is convex. The
impact of TIMETOEMERGE on recoveries is not constant across
the two types of default resolution – one with a plan of restructur-
ing and one without it. Settlement recoveries decrease with bank-
ruptcy duration up to about 33 months in the case of prepackaged
bankruptcies and out to 14 months in those without prearranged
terms. By reducing the time in bankruptcy, defaulting firms with
an advance plan of reorganization produce higher recoveries for
their creditors.
Our findings for variables capturing borrower characteristics,
macroeconomic conditions, and industry factors are fairly consis-
tent with prior studies on bond and loan recoveries. For example,
the negative and significant LEVERAGE coefficient supports
Acharya et al.’s (2007) conjecture that higher leverage complicates
negotiations to resolve default, which in turn lowers recoveries.
Higher borrowers’ cash flows the year before default do not affect
recovery rates, however. Yost (2002) reports that firms with high
ratios of operating income to total assets are more likely to restruc-
ture out-of-court. By including a dummy variable for whether
firms restructure outside of court in the model, we may have made
the cash flow variable redundant.17 GDP growth is positively corre-
lated with ultimate recovery rates, a result similar to Frye (2000) and
Sabato and Schmid (2008). Industry distress negatively influences
them, a finding consistent with Acharya et al. (2007).
Strong evidence indicates that a prior history of defaults, rather
than first-time defaults, is associated with higher recovery rates.
The estimated coefficient on EVERDEFAULTED indicates that recov-
eries are 11 cents higher, on average, on loans where the borrower
has prior defaults. In these cases, lenders presumably negotiate at
origination for enhanced protection from default.
17
The correlation between the two variables is only 21.5%, however. We address the
prospect of multicollinearity by removing all dummies related to restructuring
methods and re-estimate the regression. The FIRMCF coefficient remains insignificant
(results not reported).
929
Out-of-court Other default and p-Value for mean
restructuring restructuring types difference
(3) (4) (1)–(2)
97.36 92.00 <.0001
(100) (100)
170 26
(12) (2)
In Model 2 of Table 3, we report estimates from a quasi-maxi-
mum likelihood estimation method proposed by Papke and
Wooldridge (1996) and recently used by Dermine and Neto de
Carvalho (2006).18 The signs, statistical significance, and economic
magnitude of most of the coefficients are the same as in the linear
model estimated with OLS methods, though the quasi-maximum
likelihood approach generally offers some advantages over the OLS
method as discussed earlier.
6. Do trading prices serve as viable proxies for ultimate
recoveries?
We next examine how well trading prices about 30 days after
default serve as proxies for the actual payoffs creditors receive.
In particular, we examine whether the trading price is an unbiased
and efficient predictor of ultimate recovery. Emery (2004), as one
example, uses the average bid prices on loans one month after de-
fault to proxy for the recovery rate. The assumption is that this
price reflects the expected present value of recovery or ‘‘the best
proxy for the ultimate realization of value for each particular debt
class’’ (Bonelli et al., 2001). Acharya et al. (2007) assume the dis-
counted price at default is an unbiased estimate of the recovery
at emergence. Eberhart and Sweeney (1992) find that the trading
price reasonably reflects the ultimate payoff to bondholders fol-
lowing the resolution of default. However, Emery (2007) finds that
ultimate recoveries and trading prices are not highly correlated.
Using trading prices as a proxy for ultimate recovery has several
shortcomings. First, the technique requires that the defaulted loans
trade in a market where prices can be observed. Not all loans trade.
Second, the observed price may not reflect recoveries on the origi-
nal principal because of partial loan prepayments before and dur-
ing bankruptcy (Bonelli et al., 2001). Third, the trading price may
be depressed by supply or demand fluctuations (or both) which
are unrelated to recovery in any precise way (Altman et al.,
2005). For example, Altman et al. (2005) argue that in a period of
tight credit,19 demand for defaulted loans or bonds will be low,
while supply is high. Fourth, the market value one month after the
default does not incorporate direct expenses, such as legal fees, in-
curred during the workout period, which will lower the ultimate
recovery. Finally, the 30-day horizon at which trading prices are col-
lected is arbitrary, but is necessitated by the scarcity of default loan
trading beyond that point. On average, the time to resolution of loan
18
A number of settlement recoveries exceeds 1 as shown in Table 1 and must
therefore be constrained to 1 so that RR is a fraction bounded between zero and one
for the quasi-maximum likelihood estimation method to work. The loss of observa-
tions is also due to firm characteristic variables obtained from Compustat for public
firms, while Moody’s URD contains some private firms. In addition, some firms that
disappeared after bankruptcy must be dropped from our regressions where firm
characteristics enter the models.
19
Altman et al. (2005) state: ‘‘For example, in a recession period with increasing
default rates, recovery rates would decrease, leading to higher credit losses. This, in
turn, would lead to higher capital requirements and, correspondingly, possibly to a
decrease in the supply of bank credit to the economy, thereby exacerbating the
recession.’’
930 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933

Table 3
Multiple regression results. Model (1) applies the OLS method. Model (2) applies the quasi-maximum likelihood estimation (QMLE) method. Variable definitions are in Appendix
A.

(1) (2)
Coefficients p-Value Coefficients p-Value
Loan characteristics
LOANSIZE 0.0007 (0.927) 0.0734 (0.410)
LOANTYPE 0.1126*** (0.002) 0.9357*** (0.000)
LOANTYPE  FIRMSIZE 0.0218*** (0.011) 0.1642*** (0.014)
ALLASSETCOLL 0.2179*** (<0.001) 1.8415*** (<0.001)
INVENTRECEIVECOLL 0.2994*** (<0.001) 3.7396*** (<0.001)
OTHERCOLL 0.2031*** (<0.001) 1.7203*** (<0.001)
Recovery process characteristics
PREPACK 0.0515* (0.094) 0.3068 (0.617)
RESTRUCTURE 0.0020 (0.964) 0.3612 (0.723)
OTHERDEFAULT 0.0136 (0.791) 1.2618 (0.165)
TIMETOEMERGE 0.0067*** (0.001) 0.0568* (0.079)
TIMETOEMERGESQ 0.0001*** (<0.001) 0.0009* (0.081)
PREPACK  TIMETOEMERGE 0.0039*** (0.008) 0.0794 (0.418)
Borrower characteristics
FIRMSIZE 0.0043 (0.623) 0.0880 (0.372)
FIRMCF 0.0082 (0.840) 0.3749 (0.706)
FIRMPPE 0.0706 (0.142) 0.8246 (0.237)
FIRMLEV 0.0626*** (0.006) 0.5358* (0.064)
EVERDEFAULTED 0.1075*** (0.001) 15.7309*** (<0.001)
Macroeconomic and industry conditions
GDP 2.4900*** (<0.001) 24.8111** (0.017)
INDDISTRESS 0.0765*** (0.011) 0.6263 (0.113)
Probability of default
AIS 0.3144 (0.398) 0.3846 (0.935)
INTERCEPT 0.3329* (0.079) 0.2167 (0.917)
Industry dummies Yes Yes
N 793 793
R-squared (Pseudo R-squared) 0.2510 0.3184
*
10% level of significance.
**
5% levels of significance.
***
1% levels of significance.

defaults, which ranges from 1.5 to 2.4 years (Altman and Hotchkiss,
2006; Araten et al., 2004; Gupton et al., 2000), is significantly longer
than 30 days.
Methodologically, we take the settlement recovery measure as
the benchmark. We first test whether the means of the settlement
values and trading prices are significantly different. Next, we re-
gress ultimate recoveries on trading prices.20 If the trading price
is an unbiased predictor of the settlement value, the intercept from
the regression should be zero and the coefficient of the trading price
should equal one. Finally, using the difference between the two val-
ues as the dependent variable, we estimate a multivariate regression
on the specification we estimated in our main results above. If the
secondary loan market is efficient, 30-day post-default trading prices
should incorporate all available information about anticipated recov-
eries of the sort reflected in our model.
The third test involves the joint hypotheses that the model of
bank loan recoveries is correctly specified and the forecast of the
actual recovery is rational. If the trading price is an efficient esti-
mator of settlement recovery, the trading price model should be
identical to the ultimate recovery model. In a regression of the dif-
ference between the two (which is simply the observed forecast er-
ror) on the ‘‘true’’ recovery model, none of the coefficients should
be significant. Observing significant variables implies that the fore-
cast error for settlement recovery could be improved by taking bet-
ter account of available information.
20
Emery (2007) conducts this test, but his sample consists of both bonds and loans,
whereas ours contains only loans.
Table 4 shows descriptive statistics of the two measures and
their mean and median differences. Trading prices are highly pos-
itively correlated with settlement recoveries (82%). However, the
trading price is, on average, 9 cents below the settlement recovery
dollar and the difference is statistically significant at the 1% level,
implying that the average loan is underpriced with respect to
recovery shortly after default.
When we regress settlement values on trading prices, the inter-
cept is not equal to zero and is strongly significant, whereas the
coefficient of the settlement measure fails a test that its value is
equal to one (results not shown). These results suggest that trading
prices are biased predictors of ultimate recoveries. When we next
estimate a regression of the difference between the two measures
of recovery on the explanatory variables used in our previous esti-
mations, a number of variable coefficients are significant. This im-
plies that these variables maintain predictive power in explaining
variation in the forecast error.21 In particular, firm leverage 1 year
before default, loan spread, industry distress, and dummies on
industry types are all statistically significant at the 10% or lower le-
vel. We conclude that the 30-day trading price, though correlated
with the settlement recovery, is not a rational predictor of settle-
ment value.
7. Econometric issues and alternative variable measurements
Our sample includes recovery rates on defaulted bank loans,
which can be observed only when borrowers default. To account
21
Results are omitted to save space, but are available on request.
 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933 931

Table 4
Descriptive statistics of ultimate recovery, trading prices, and their differences.

N Mean Median Std. dev. Min Max

Panel A
Settlement 1364 0.8414 1 0.2822 0 1.4021
Trading price 385 0.6467 0.7 0.2779 0.03 1.0550
Panel B
Difference = Settlement  Trading 329 0.0914 0.0682 0.1724 0.3577 0.7430

Correlation Settlement recovery Trading price

Panel C
Settlement 1
Trading price 0.8212 1

Table 5
Partial OLS regression results with alternative definitions of some variables used in Table 3.

(1) (2) (3) (4) (5) (6) (7)

TB3M 0.0870*
(0.087)
SPREADCORP 8.1511***
(0.000)
INDPRODCHG 7.8710*
(0.097)
INVINDCOVERAGE 0.0918*
(0.088)
ASSETVOL 0.0020
(0.818)
INVESTGRADE 0.0073
(0.895)
LSIZE-D 0.0059
(0.297)
N 742 742 742 787 782 389 792
R-squared 0.2350 0.2463 0.2392 0.2419 0.2579 0.3272 0.2767
*
10% level of significance.

5% level of significance.
***
1% level of significance.

for prospective sample selection bias, we perform Heckman’s
(1979) two-step procedure. In the first stage, we estimate a model
for predicting default. The second stage estimation is the recovery
model and includes the inverse Mills’ ratio as an explanatory var-
iable. The coefficient estimate of this ratio is statistically insignifi-
cant (results not tabulated), suggesting that the hypothesis of no
selection bias cannot be rejected in the data. Therefore, least-
squares estimation of recovery rates, RRit = x0 b + uit, still yields con-
sistent estimators without lambda.
The standard process for writing loan contracts raises another
prospective econometric problem. Most loans contain cross-
default clauses such that if a firm defaults on any one loan to a
given lender, the firm is automatically in default on other debt
instruments specified in the loan agreement contract. This can
produce perfect correlation across certain observations on a given
default date. Because we use instrument-level recovery rates, our
sample data are clustered among the defaulting borrowers.
The settlement value sample includes 583 distinct borrowers
among the 1364 defaulted loans, and the number of clusters well
exceeds the loan units within each cluster. In addition, we can as-
sume independence across borrower clusters. Therefore, the OLS
method will produce consistent estimates and only the asymptotic
variance matrix needs to be corrected for clustering effects (Wood-
ridge, 2002, p. 135). The results accounting for clustering effects,
which are not tabulated for brevity, show no changes in the mag-
nitude or significance of the variable estimates we reported.
We also examine model re-specifications that apply alternative
measures for some of the variables. First, we account for time-
related effects on recoveries not fully captured by GDP growth by
adding year dummy variables to our base model. Although many
of the year dummies are significant, all of the other coefficient esti-
mates except PREPACK remain unchanged. Second, GDP growth is
used as a proxy for macroeconomic conditions in our main results.
Even though its coefficient estimate is positive and significant,
Sabato and Schmid (2008) find the variable to be a weak proxy
for the state of the economy in the context of recovery rates. The
main reason is a time lag in the effects of economic conditions
on recoveries.
Izvorski (1997) documents that yields on three-month Treasury
bills (TB3M) are negatively related to economic growth. He also
finds that the difference between the yield on an AAA corporate
bond and a 10-year Treasury bond (SPREADCORP) has a positive
correlation with economic conditions. Also, annual industrial pro-
duction growth (INDPRODCHG) should be positively correlated
with economic growth. In place of GDP growth, we separately em-
ploy TB3M, SPREADCORP, and INDPROCHG as three alternative
proxies for the state of the economy. The data used to construct
these variables are from the US Department of Commerce Bureau
of Economic Analysis and the Federal Reserve Board. The coeffi-
cient estimates on the three alternative proxies, as shown in
Table 5, are as expected, which confirms that higher recovery rates
are positively related to economic growth. The other coefficient
estimates remain qualitatively similar (not reported).
We also re-examine our models with alternative constructions
of other variables: industry financial distress, probability of de-
fault, and loan size. Finally, we retest our model using the natural
932 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933

logarithm of the loan recoveries. The results, which are partially Appendix A (continued)
presented in Table 5 for brevity, are consistent with those reported
in the main results above. Variable definitions
LOANSIZE The dollar amount in millions of the
facility at the time of issuance
8. Conclusion LOANTYPE A dummy variable equal to one if the
loan is a term loan, and zero if it is a
Despite abundant literature on the determinants of default on revolver
loans and other debt instruments, relatively little is known about ALLASSETCOLL A dummy variable equal to one if the
the factors that influence bank loans recoveries following default. loan is secured by all firm assets, and
Yet, recovery is critical to bank performance as well as to the prop- zero otherwise
er measure of the capital needed to buffer against risk. The few INVENTRECEIVECOLL A dummy variable equal to one if the
academic studies on recovery rates on bank loans either include loan is secured by inventory, accounts
bonds and loans in a combined sample or rely on data from a single receivable, or both, and zero otherwise
lender, so robust evidence is lacking on what drives recoveries on OTHERCOLL A dummy variable equal to one if the
loans. In addition, researchers often rely on proxy values for recov- loan is secured differently from the
eries. In our paper, we focus on actual recoveries on loans gener- other types, and zero otherwise
ated by a large number of lenders. The source is Moody’s UNSECURED A dummy variable equal to one if the
Ultimate Recovery Database on North American loans in default, loan is unsecured, and zero otherwise.
covering the period 1987–2007. This variable serves as the reference
We find that, in general, loan characteristics are more signifi- group for collateral types
cant than borrower characteristic prior to default as determinants COLLATERAL A binary variable that equals 1 if the loan
of recovery rates. The results show that firm leverage before de- is collateralized, and zero otherwise
fault negatively affects ultimate recoveries, while borrower cash PREPACK A dummy equal to one if the
flow positions do not. A variety of loan contract features are bankruptcy is through a pre-packaged
strongly related to the ultimate payoff for creditors. Secured loans bankruptcy, and zero otherwise
have higher recoveries, especially when the collateral takes the RESTRUCTURE A dummy variable equal to one if
form of inventories and accounts receivable. Loans to borrowers default is resolved by out-of-court
with prior defaults yield higher recoveries than first-time defaults, restructuring, including distressed
and arranging a prepackaged bankruptcy increases recoveries. exchange offers, and zero otherwise
Loan recoveries vary significantly and nonlinearly with the length OTHERDEFAULT A dummy variable, equal to one if
of time to emerge. Macroeconomic conditions significantly affect default is resolved by other methods
recovery prospects, but the probability of default at the time of than an out-of-court restructuring,
loan origination is unrelated to ultimate recoveries. prepackaged formal bankruptcy, or
We also examine how well the secondary market for bank loans non-prepackaged formal bankruptcy,
anticipates settlement values for creditors. Our findings suggest and 0 otherwise
that 30-day post-default trading prices are highly correlated with BANKRUPTCY A dummy variable equal to one if
settlement recoveries, but that trading prices are biased and ineffi- default occurs only in the context of a
cient predictors of ultimate recovery. non-prepack formal bankruptcy filing,
and zero otherwise
Acknowledgements TIMETOEMERGE The length of time in months between
bankruptcy or restructuring and
We thank Ike Mathur, the editor, and an anonymous referee for emergence, often known as resolution
their valuable comments and suggestions. We also thank seminar time
participants at the Financial Management Association 2009 confer- FIRMSIZE The market value of firm-level assets
ence and the University of Southern Indiana’s College of Business. 1 year before default. The market
We thank John Butler, Paul Childs, John Garen, Kristine Hankins, value of firm-level assets is calculated
Mark Liu, and David Walker for their insightful comments and sug- as the book value of long-term and
gestions. We also thank Jackie Thompson for editorial assistance. short-term debt plus the number of
Hinh Khieu and Donald Mullineaux acknowledge financial support common shares outstanding
from University of Kentucky’s Gatton College of Business for par- multiplied by the price per share
tially funding the acquisition of Moody’s Ultimate Recovery FIRMPPE Firm asset tangibility, measured as net
Database. property, plant, and equipment over
total book assets 1 year before default
Appendix A FIRMCF Firm cash flows, measured as EBITDA
over total book assets 1 year before
default
FIRMLEV Firm leverage, measured as total long-
Variable definitions
term debt plus debt in current
SETTLEMENT The discounted recovery amount as a liabilities over total book assets 1 year
percentage of the principal amount at before default
default. The discount rate is the EVERDEFAULTED A dummy variable equal to one if the
defaulted instrument’s effective firm has defaulted before, and zero
interest rate otherwise
 H.D. Khieu et al. / Journal of Banking & Finance 36 (2012) 923–933
Appendix A (continued)
Variable definitions
INDDISTRESS A dummy variable equal to one if the
industry median stock returns in the
year of default is less than 30%, and
zero otherwise. The stock returns are
calculated without the defaulting
firms and the industry is defined
according to the three-digit SIC codes
GDP The annual GDP growth rate measured
1 year before the year when default
occurred
AIS The credit spread (in percent) at the
time of loan origination over LIBOR of
the drawn loan that defaulted
References
Acharya, V., Bharath, S., Srinivasan, A., 2007. Does industry-wide distress affect
defaulted firms? Evidence from creditor recoveries. Journal of Financial
Economics 85, 787–821.
Altman, E., 2009. Default recovery rates and LGD in credit risk modeling practice. In:
Jones, S., Hensler, D. (Eds.), Advances in Credit Risk Modeling and Corporate
Bankruptcy Protection. Cambridge University Press, Cambridge, England.
Altman, E., Brady, B., Resti, A., Sironi, A., 2005. The link between default and
recovery rates: theory, empirical evidence, and implications. Journal of Business
78 (6), 2203–2227.
Altman, E., Hotchkiss, E., 2006. Corporate Financial Distress and Bankruptcy: Predict
and Avoid Bankruptcy, Analyze and Invest in Distressed Debt. John Wiley &
Sons, Inc., New York, NY.
Altman, E., Kishore, V., 1996. Almost everything you wanted to know about
recoveries on defaulted bonds. Financial Analysts Journal (November/
December), 57–64.
Araten, M., Jacobs, M., Varshney, P., 2004. Measuring LGD on commercial loans: an
18-year internal study. RMA Journal (May), 28–35.
Asarnow, E., Edwards, D., 1995. Measuring loss on defaulted bank loans: a 24-year
study. Journal of Commercial Lending 10 (2), 11–23.
Bakshi, G., Madan, D., Zhang, F., 2006. Understanding the role of recovery in default
risk models: empirical comparisons and implied recovery rates. Working Paper,
University of Maryland.
Bonelli, S., Grossman, R., O’Shea, S., 2001. Bank Loan and Bond Recovery Study:
1997–2000. Fitch IBCA, Duff & Phelps Loan Products Special Report.
Cantor, R., Varma, P., 2004. Determinants of Recovery Rates on Defaulted Bonds and
Loans for North American Corporate Issuers. Moody’s Special Comment, pp. 1–24.
Carey, M., Gordy, M., 2008. The bank as grim reaper: Debt composition and
recoveries on defaulted debt. Working Paper, Federal Reserve Systems.
Carty, L., 1998. Bankrupt Bank Loan recoveries. Moody’s Investors Services, Global
Credit Research, Special Comment, pp. 12–16.
Carty, L., Lieberman, D., 1996. Defaulted Bank Loan Recoveries. Moody’s Investors
Services, Global Credit Research, Special Comment, pp. 12–19.
Caselli, S., Gatti, S., Querci, F., 2008. The sensitivity of the loss given default rate to
systematic risk: new empirical evidence on bank loans. Journal of Financial
Services Research 34, 1–34.
Covitz, D., Han, S., 2004. An empirical analysis of bond recovery rates: Exploring a
structural view of default. Working Paper, Federal Reserve Board.
Covitz, D., Han, S., Wilson, B.A., 2006. Are longer bankruptcies really more costly?
Working Paper, Federal Reserve Board.
Das, S., Hanouna, P., 2009. Implied recovery. Journal of Economic Dynamics and
Control 33 (11), 1837–1867.
Dermine, J., Neto de Carvalho, C., 2006. Bank loan losses-given-default: a case study.
Journal of Banking and Finance 30, 1219–1243.
Eberhart, A., Sweeney, R., 1992. Does the bond market predict bankruptcy
settlements? Journal of Finance 47 (3), 943–980.
Emery, K., 2004. Recovery Rates on North American Syndicated Bank Loans.
Moody’s Investors Service, Global Credit Research, Special Comment, pp. 12–16.
Emery, K., 2007. Moody’s Ultimate Recovery Database. Moody’s Investors Service,
Global Credit Research, Special Comment, pp. 12–16.
933
Emery, K., Ou, S., 2004. Credit Loss Rates on Similarly Rated Loans and Bonds.
Moody’s Investors Service, Global Credit Research, Special Comment, pp. 12–
12.
Franks, J., Torous, W., 1994. A comparison of financial recontracting in distressed
exchanges and Chapter 11 reorganizations. Journal of Financial Economics 35,
349–370.
Frye, J., 2000. Depressing Recoveries. Federal Reserve Bank of Chicago Policy
Studies.
Gersbach, H., Lipponer, A., 2003. Firm defaults and the correlation effect. European
Financial Management 9 (3), 361–377.
Gilson, S.C., 1990. Bankruptcy, boards, banks, and blockholders: evidence on
changes in corporate ownership and control when firms default. Journal of
Financial Economics 27 (2), 355–387.
Gilson, S.C., John, K., Lang, L., 1990. Troubled debt restructurings: an empirical study
of private reorganization of firms in default. Journal of Financial Economics 27
(2), 315–353.
Gourieroux, C., Monfort, A., Trognon, A., 1984. Pseudo maximum likelihood
methods: theory. Econometrica 52 (3), 681–700.
Gottesman, A., Roberts, G., 2004. Maturity and corporate loan pricing. Financial
Review 39, 55–77.
Grunert, J., Weber, M., 2009. Recovery rates of commercial lending: empirical
evidence for Germany. Journal of Banking and Finance 33, 505–513.
Guo, X., Jarrow, R., Zeng, Y., 2009. Modeling the recovery rate in a reduced form
model. Mathematical Finance 19, 73–97.
Gupton, G., Gates, D., Carty, L., 2000. Bank-loan loss given default. Moody’s Investors
Service, Global Credit Research. pp. 69–92.
Harjoto, M., Mullineaux, D., Yi, H-C., 2006. A comparison of syndicated loan pricing
at investment and commercial banks. Financial Management 35, 49–70.
Heckman, J., 1979. Sample selection bias as a specification error. Econometrica 47
(1), 153–161.
Hu, Y., Perraudin, W., 2002. The dependence of recovery rates and defaults. Working
Paper, Birkbeck College.
Izvorski, I., 1997. Recovery ratios and survival times for corporate bonds. Working
Paper, International Monetary Fund.
Jarrow, R., Lando, D., Turnbull, S., 1997. A Markov model for the term structure of
credit risk spreads. Review of Financial Studies 10, 481–523.
Jokivoulle, E., Peura, S., 2000. A model for estimating recovery rates and collateral
haircuts for bank loans. Bank of Finland Discussion Paper 2/2000.
Karoui, L., 2007. Measuring defautable securities with recovery risk. Working Paper,
McGill University.
Keenan, S., 2000. Moody’s Approach to Evaluating Distressed exchange. Moody’s
Investors Service, Global Credit Research’s Special Comment, pp. 12–26.
McConnell, J., Lease, R., Tashjian, E., 1996. An empirical analysis of prepackaged
bankruptcies. Journal of Financial Economics 40 (1), 135–162.
Papke, L., Wooldridge, J., 1996. Econometric methods for fractional response
variables with an application to the 401 (K) plan participation rates. Journal of
Applied Econometrics 11, 619–632.
Rajan, R., Winton, A., 1995. Covenants and collateral as incentives to monitor.
Journal of Finance 50, 1113–1146.
Rose, P., Hudgins, S., 2008. Bank Management and Financial Services, seventh ed.
McGraw Hill Companies, New York, NY.
Ryan, V., 2008. Mastering the turnaround. What it really takes to survive a corporate
bankruptcy. CFO Magazine (July/August), 63–68.
Sabato, G., Schmid, M., 2008. Estimating conservative loss given default. Working
Paper, Swiss Institute of Banking and Finance.
Schuermann, T., 2004. What do we know about loss given default? In: Shimko, D.
(Ed.), Credit Risk Models and Management. London, UK.
Shleifer, A., Vishny, R., 1992. Liquidation values and debt capacity: a market-
equilibrium approach. Journal of Finance 47, 1343–1366.
Strahan, P., 1999. Borrower risk and the price and nonprice terms of bank loans.
Working Paper, Boston College.
Thorburn, K., 2000. Bankruptcy auctions: costs, debt recovery, and firm survival.
Journal of Financial Economics 58, 337–368.
Van De Castle, K., Keisman, D., 1999. Recovering your money: insights into losses
from default. Standard & Poor’s CreditWeek, 29–34.
White, H., 1980. A heteroscedasticity-consistent covariance matrix and a direct test
for heteroscedasticity. Econometrica 27, 817–838.
Williamson, O., 1988. Corporate finance and corporate governance. Journal of
Finance 43 (3), 567–591.
Woodridge, J., 2002. Econometric Analysis of Cross Section and Panel Data. The MIT
Press, Cambridge, MA.
Yost, K., 2002. The choice among traditional Chapter 11, prepackaged bankruptcy,
and out-of-court restructuring. Working Paper, Auburn University.
Zhang, Z., 2009. Recovery rates and macroeconomic conditions: the role of loan
covenants. Working Paper, Boston College.