An option theoretic model for

ultimate loss-given-default with systematic recovery risk

and stochastic returns on defaulted debt

Michael Jacobs, Jr1

1. Introduction and Summary

Loss-given-default (LGD),2 the loss severity on defaulted obligations, is a critical component
of risk management, pricing and portfolio models of credit. This is among the three primary
determinants of credit risk, the other two being the probability of default (PD) and exposure of
default (EAD). However, LGD has not been as extensively studied, and is considered a much
more daunting modeling challenge than other components, such as PD. Starting with the
seminal work by Altman (1968), and after many years of actuarial tabulation by rating
agencies, predictive modeling of PD is currently in a mature stage. The focus on PD is
understandable, as traditionally credit models have focused on systematic components of
credit risk which attract risk premia, and unlike PD, determinants of LGD have been ascribed
to idiosyncratic borrower specific factors. However, now there is an ongoing debate about
whether the risk premium on defaulted debt should reflect systematic risk, in particular
whether the intuition that LGDs should rise in worse states of the world is correct, and how
this could be refuted empirically given limited and noisy data (Carey and Gordy, 2007).
The recent heightened focus on LGD is evidenced by the flurry of research into this relatively
neglected area (Acharya et al [2007], Carey and Gordy [2007], Altman et al [2001, 2003,
2005], Altman [2006], Gupton et al [2000, 2005], Araten et al [2004], Frye [2000 a,b,c, 2003],
Jarrow [2001]). This has been motivated by the large number of defaults and near
simultaneous decline in recovery values observed at the trough of the last credit cycle circa
2000-2002, regulatory developments such as Basel II (BIS [2003, 2005, 2006], OCC et al
[2007]) and the growth in credit markets. However, obstacles to better understanding and
predicting LGD, including dearth of data and the lack of a coherent theoretical underpinning,
have continued to challenge researchers. In this paper, we hope to contribute to this effort by
synthesizing advances in financial theory to build a model of LGD that is consistent with a
priori expectations and stylized facts, internally consistent and amenable to rigorous
validation. In addition to answering the many questions that academics have, we further aim
to provide a practical tool for risk managers, traders and regulators in the field of credit.
LGD may be defined variously depending upon the institutional setting or modeling context,
or the type of instrument (traded bonds vs. bank loans) versus the credit risk model (pricing
debt instruments subject to the risk of default vs. expected losses or credit risk capital). In the

1
Senior Financial Economist, Credit Risk Analysis Division, Department of Economic and International Affairs,
Office of the Comptroller of the Currency, One Independence Square, Suite 3144, Washington, DC 20024,
202-874-4728, [email protected]. The views expressed herein are those of the author and do not
necessarily represent a position taken by of the Office of the Comptroller of the Currency or the U.S.
Department of the Treasury.
2
This is equivalent to one minus the recovery rate, or dollar recovery as a proportion of par, or EAD assuming
all debt becomes due at default. We will speak in terms of LGD as opposed to recoveries with a view toward
credit risk management applications.

BIS Papers No 58 257

case of bonds, one may look at the price of traded debt at either the initial credit event,3 the
market values of instruments received at the resolution of distress4 (Keisman et al, 2000;
Altman et al, 1996) or the actual cash-flows incurred during a workout.5 When looking at
loans that may not be traded, the eventual loss per dollar of outstanding balance at default is
relevant (Asarnow et al, 1995; Araten et al, 2004). There are two ways to measure the latter
– the accounting LGD refers to nominal loss per dollar outstanding at default,6 while the
economic LGD refers to the discounted cash flows to the time of default taking into
consideration when cash was received. The former is used in setting reserves or a loan loss
allowance, while the latter is an input into a regulatory or economic credit capital model.
In this study we develop various theoretical models for ultimate loss-given-default in the
Merton (1974) structural credit risk model framework. We consider an extension that allows
for differential seniority within the capital structure, an independent recovery rate process,
representing undiversifiable recovery risk, with stochastic drift. The comparative statics of
this model are analyzed and compared to a baseline model, all of these in a framework that
incorporates an optimal foreclosure threshold (Carey and Gordy, 2007). In the empirical
exercise, we calibrate alternative models for ultimate LGD on bonds and loans having both
trading prices at default and at resolution of default, utilizing an extensive sample of rated
defaulted firms in the period 1987-2008 (Moody’s Ultimate Recovery Database™ - URD™),
800 defaults (bankruptcies and out-of-court settlements of distress) that are largely
representative of the US large corporate loss experience, for which we have the complete
capital structures and can track the recoveries on all instruments to the time of default to the
time of resolution.
We find that parameter estimates vary significantly across models and recovery segments.
estimated volatilities of the recovery rate processes, as well as of their random drifts, are
found to increasing in seniority, in particular for bank loans as compared to bonds. We
interpret this as reflecting greater risk in the ultimate recovery for higher ranked instruments
having lower expected loss severities (or ELGDs). Analyzing the implications of our model for
the quantification of downturn LGD, we find the later to be declining in expected LGD, higher
for worse ranked instruments, increasing in the correlation between the process driving firm
default and recovery on collateral, and increasing in the volatility of the systematic factor
specific to the recovery rate process or the volatility of the drift in such. Finally, we validate
the leading model derived herein in an out-of-time and out-of-sample bootstrap exercise,
comparing it to a high-dimensional regression model, and to a non-parametric benchmark
based upon the same data, where we find our model to compare favorably. We conclude that
our model is worthy of consideration to risk managers, as well as supervisors concerned with
advanced IRB under the Basel II capital accord.
This paper is organized as follows. Section 2 reviews the literature, focusing on the treatment
of LGD in theoretical credit models, both academic and practitioner. Section 3 presents the
theoretical framework. Section 4 discusses comparative statics of the alternative models.
Section 5 describes the econometric framework. Section 6 describes the data used in this
study and presents the calibration analysis of structural model parameters. In Section 7 we

3
By default we mean either bankruptcy (Chapter 11) or other financial distress (payment default). In a banking
context, this defined as synonymous with respect to non-accrual on a discretionary or non-discretionary basis.
This is akin to the notion of default in Basel, but only proximate.
4
Note that this may be either the value of pre-petition instruments received valued at emergence from
bankruptcy, or the market values of new securities received in settlement of a bankruptcy proceeding or as the
result of a distressed restructuring.
5
Note that the former may viewed as a proxy to this, the pure economic notion.
6
In the context of bank loans, this is the cumulative net charge-off as a percent of book balance at default (the
net charge-off rate).

258 BIS Papers No 58

discuss the implications of our modeling framework for downturn LGD. In Section 8 we
perform an out-of-sample validation of our model and two alternative benchmarks. Finally,
Section 9 concludes and discusses directions for future research.

2. Review of the literature

In this section we will examine the way in which different types of theoretical credit risk
models have treated LGD – assumptions, implications for estimation and application. Credit
risk modeling was revolutionized by the approach of Merton (1974), who built a theoretical
model in the option pricing paradigm of Black and Scholes (1973), which has come known to
be the structural approach. Equity is modeled as a call option on the value of the firm, with
the face value of zero coupon debt serving as the strike price, which is equivalent to
shareholders buying a put option on the firm from creditors with this strike price. Given this
capital structure, log-normal dynamics of the firm value and the absence of arbitrage, closed
form solutions for the default probability and the spread on debt subject to default risk can be
derived. The LGD can be shown to depend upon the parameters of the firm value process as
is the PD, and moreover is directly related to the latter, in that the expected residual value to
claimants is increasing (decreasing) in firm value (asset volatility or the level of
indebtedness). Therefore, LGD is not independently modeled in this framework; this was
addressed in much more recent versions of the structural framework (Frye [2000 a,b], Dev
and Pykhtin [2002], Pykhtin [2003]).
Extensions of Merton (1974) relaxed many of the simplifying assumptions of the initial
structural approach. Complexity to the capital structure was added by Black and Cox (1976)
and Geske (1977), with subordinated and interest-paying debt, respectively. The distinction
between long- and short-term liabilities in Vasicek (1984) was the precursor to the KMV
model. However, these models had limited practical applicability, the standard example being
evidence of Jones, Mason and Rosenfeld (1984) that these models were unable to price
investment-grade debt any better than a naïve model with no default risk. Further, empirical
evidence in Franks and Touros (1989) showed that the adherence to absolute priority rules
(APR) assumed by these models are often violated in practice, which implies that the
mechanical negative relationship between expected asset value and LGD may not hold.
Longstaff & Schwartz (1995) incorporate into this framework a stochastic term structure with
a PD-interest rate correlation. Other extensions include Kim at al (1993) and Hull & White
(2002), who examine the effect of coupons and the influence of options markets,
respectively.
Partly in response to this, a series of extensions ensued, the so-called “second generation”
of structural form credit risk models (Altman, 2003). The distinguishing characteristic of this
class of models is the relaxation of the assumption that default can only occur at the maturity
of debt – now default occurs at any point between debt issuance and maturity when the firm
value process hits a threshold level. The implication is that LGD is exogenous relative to the
asset value process, defined by a fixed (or exogenous stochastic) fraction of outstanding
debt value. This approach can be traced to the barrier option framework as applied to risky
debt of Black and Cox (1976).
All structural models suffer from several common deficiencies. First, reliance upon an
unobservable asset value process makes calibration to market prices problematic, inviting
model risk. Second, the limitation of assuming a continuous diffusion for the state process
implies that the time of default is perfectly predictable (Duffie and Lando, 2001). Finally, the
inability to model spread or downgrade risk distorts the measurement of credit risk. This gave
rise to the reduced form approach to credit risk modeling (Duffie and Singleton, 1999), which
instead of conditioning on the dynamics of the firm, posit exogenous stochastic processes for
PD and LGD. These models include (to name a few) Litterman & Iben (1991), Madan & Unal

BIS Papers No 58 259

(1995), Jarrow & Turnbull (1995), Lando (1998) and Duffie (1998). The primitives
determining the price of credit risk are the term structure of interest rates (or short rate), and
a default intensity and an LGD process. The latter may be correlated with PD, but it is
exogenously specified, with the link of either of these to the asset value (or latent state
process) not formally specified. However, the available empirical evidence (Duffie and
Singleton, 1999) has revealed these models deficient in generating realistic term structures
of credit spreads for investment and speculative grade bonds simultaneously. A hybrid
reduced – structural form approach of Zhou (2001), which models firm value as a jump
diffusion process, has had more empirical success, especially in generating a realistic
negative relationship between LGD and PD (Altman et al, 2006).
The fundamental difference between reduced and structural form models is the
unpredictability of defaults: PD is non-zero over any finite time interval, and the default
intensity is typically a jump process (eg Poisson), so that default cannot be foretold given
information available the instant prior. However, these models can differ in how LGD is
treated. The recovery of treasury assumption of Jarrow & Turnbull (1995) assumes that an
exogenous fraction of an otherwise equivalent default-free bond is recovered at default.
Duffie and Singleton (1999) introduce the recovery of market value assumption, which
replaces the default-free bond by a defaultable bond of identical characteristics to the bond
that defaulted, so that LGD is a stochastically varying fraction of market value of such bond
the instant before default. This model yields closed form expressions for defaultable bond
prices and can accommodate the correlation between PD and LGD; in particular, these
stochastic parameters can be made to depend on common systematic or firm specific
factors. Finally, the recovery of face value assumption (Duffie [1998], Jarrow et al [1997])
assumes that LGD is a fixed (or seniority specific) fraction of par, which allows the use of
rating agency estimates of LGD and transition matrices to price risky bonds.
It is worth mentioning the treatment of LGD in credit models that attempt to quantify
unexpected losses analogously to the value-at-risk (VaR) market risk models, so-called
credit VaR models (Creditmetrics™ [Gupton et al, 1997], KMV CreditPortfolioManager™
[Vasicek, 1984], CreditRisk+™ [Credit Suisse Financial Products, 1997],
CreditPortfolioView™ [Wilson, 1998]). These models are widely employed by financial
institutions to determine expected credit losses as well as economic capital (or unexpected
losses) on credit portfolios. The main output of these models is a probability distribution
function for future credit losses over some given horizon, typically generated by simulation of
analytical approximations, as it is modeled as highly non-normal (asymmetrical and fat-
tailed). Characteristics of the credit portfolio serving as inputs are LGDs, PDs, EADs, default
correlations and rating transition probabilities. Such models can incorporate credit migrations
(mark-to-market mode - MTM), or consider the binary default vs. survival scenario (default
mode - DM), the principal difference being that in addition an estimated transition matrix
needs to be supplied in the former case. Similarly to the reduced form models of single name
default, LGD is exogenous, but potentially stochastic. While the marketed vendor models
may treat LGD as stochastic (eg a draw from a beta distribution that is parameterized by
expected moments of LGD), there are some more elaborate proprietary models that can
allow LGD to be correlated with PD.
We conclude our discussion of theoretical credit risk models and the treatment of LGD by
considering recent approaches, which are capable of capturing more realistic dynamics,
sometimes called “hybrid models”. These include Frye (2000a, 2000b), Jarrow (2001),
Bakshi et al (2001), Jarrow et al (2003), Pykhtin (2003) and Carey & Gordy (2007). Such
models are motivated by the conditional approach to credit risk modeling, credited to Finger
(1999) and Gordy (2000), in which a single systematic factor derives defaults. In this more
general setting, they share in common the feature that dependence upon a set of systematic
factors can induce an endogenous correlation between PD & LGD. In the model of Frye
(2000a, 2000b), the mechanism that induces this dependence is the influence of systematic
factors upon the value of loan collateral, leading to a lower recoveries (and higher loss

260 BIS Papers No 58

severity) in periods where default rates rise (since asset values of obligors also depend upon
the same factors). In a reduced form setting, Jarrow (2001) introduced a model of
co-dependent LGD and PD implicit in debt and equity prices.7

3. Theoretical model
The model that we propose is an extension of Black and Cox (1976). The baseline mode
features perpetual corporate debt, a continuous and a positive foreclosure boundary. The
former assumption removes the time dependence of the value of debt, thereby simplifying
the solution and comparative statics. The latter assumption allows us to study the
endogenous determination of the foreclosure boundary by the bank, as in Carey and Gordy
(2007). We extend the latter model by allowing the coupon on the loan to follow a stochastic
process, accounting for the effect of illiquidity. Note that in this framework, we assume no
restriction on asset sales, so that we do not consider strategic bankruptcy, as in Leland
(1994) and Leland and Toft (1996).
Let us assume a firm financed by equity and debt, normalized such that the total value of
perpetual debt is 1, divided such that there is a single loan with face value  and a single
class of bonds with a face value of 1   . The loan is senior to that bond, and potentially has
covenants which permit foreclosure. The loan is entitled to a continuous coupon at a rate c,
which in the baseline model we take as a constant, but may evolve randomly. Equity
receives a continuous dividend, having a constant and a variable component, which we
denote as   Vt , where Vt is the value of the firm’s assets at time t. We impose the
restriction that 0    r  c , where r is the constant risk-free rate. The asset value of the firm,
net of coupons and dividends, follows a geometric Brownian motion with constant volatility  :

dVt  C
  r     dt   dZt
Vt  Vt 
(3.1)
Where in (3.1) we denote the fixed cash outflows per unit time as:
C  c   1      (3.2)

Where in (3.2),  and  are the continuous coupon rate on the bond and dividend yield on
equity, respectively. Default occurs at time t and is resolved after a fixed interval  , at which
point dividend payments cease, but the loan coupon continues to accrue through the
 

settlement period. At the point of emergence, loan holders receive  exp  c  ,Vt  , or the
minimum of the legal claim or the value of the firm at emergence. We can value the loan at
resolution, under risk neutral measure, using the standard Merton (1974) formula. Denote the
total legal claim at default by:
D   exp  c   1    (3.3)
.
This follows from the assumption that the coupon c on the loan with face value  continues
to accrue at the contractual rate throughout the resolution period  , whereas the bond with
face value 1   does not.

7
Jarrow (2001) also has the advantage of isolating the liquidity premium embedded in defaultable bond
spreads.

BIS Papers No 58 261

Thus far we have assumed that the senior bank creditors foreclose on the bank when the
value of assets is Vt, where t is the time of default. However, this is not realistic, as firm value
fluctuates throughout the bankruptcy or workout period, and we can think that there will be
some foreclosure boundary (denoted  ) below which foreclosure is effectuated.
Furthermore, in most cases there exists a covenant boundary, above which foreclosure
cannot occur, but below which it may occur as the borrower is in violation of a contractual
provision. For the time being, let us ignore the latter complication, and focus on the optimal
choice of  by the bank. In the general case of time dependency in the loan valuation
equation F Vt |  ,  , r ,  , following Black and Cox (1976), we have to solve a second-order
partial differential equation. Following Carey and Gordy (2007), we modify this such that the
value of the loan at the threshold is not a constant, but simply equal to the recovery value of
the loan at the default time. Second, we remove the time dependency in the value of the
perpetual debt. It is shown in Carey and Gordy (2007) that under these assumptions, so long
as there are positive and fixed cash flows to claimants other than the bank,  1     0 or
  0 , then there exists a finite and positive solution  * , the optimal foreclosure boundary.
We model undiversifiable recovery risk by introducing a separate process for recovery on
debt, Rt . This can be interpreted as the state of collateral underlying the loan or bond. Rt is a
geometric Brownian process that depends upon the Brownian motion that drives the return
on the firm’s assets Z t , an independent Brownian motion Wt and a random instantaneous
mean  t :

dRt
  t dt   dZ t   dWt (3.6)
Rt

d t      t  dt   dBt (3.7)

Where the volatility parameter  represents the sensitivity of recovery to the source of
uncertainty driving asset returns (or the “systematic factor”), implying that the instantaneous
1  dA dR 
correlation between asset returns and recovery is given by Corrt  t  t    . On
dt  At Rt 
the other hand, the volatility parameter  represents the sensitivity of recovery to a source of
uncertainty that is particular to the return on collateral, also considered a “systematic factor”,
but independent of the asset return process. The third source of recovery uncertainty is given
by (3.7), where we model the instantaneous drift on the recovery by an Orhnstein-Uhlenbeck
mean-reverting process, with  the speed of mean-reversion,  the long-run mean,  the
constant diffusion term, and Bt is a standard Weiner process having instantaneous
correlation with the source of randomness in the recovery process, given heuristically by
1
 Corrt  dBt , dWt  . The motivation behind this specification is the overwhelming
dt
evidence that the mean LGD is stochastic.
Economic LGD on the loan is given by following expectation under physical measure:
LGDP  Rt ,  t |  , c,  , ,  , ,  ,  

exp  c      2  2   
 1 Et  min  exp  c  , Rt exp    t     Z t    Wt    
     2    
 

262 BIS Papers No 58

 exp   t  c  
 1 B  Rt ,  t |  exp  c  , ˆ ,  t ,  (3.8)

Where the modified option theoretic function B    is given by:

B  Rt ,  t |  exp  c  , ˆ ,  t ,   Rt    d '   exp   c   t     d '  (3.9)

z u2
1 
having arguments to the Gaussian distribution function   z    e 2
du :
2 u 

1   Rt   1 2 
d '   log        ˆ   (3.10)
ˆ     exp  c    2  
t

A well-known result (Bjerksund, 1991) is that the maturity-dependent volatility ˆ is given by:
2
     2    1  
ˆ    2   2   2  2   2      2   2   2    1  e     1  e 2
         2    
   

(3.11)
The recovery to the bondholders is the expectation of the minimum of the positive part of the

difference in the recovery and face value of the loan  Rt    exp  c   and the face value of
the bond B, which is structurally identical to a compound option valuation problem (Geske,
1977):
LGDBP Vt , Rt ,  t |  , c,  ,  , ,  , ,  ,   

exp   ,        2  2    
 1 
Et min  B, max  Rt exp    t       Z t   Wt     exp  c,   , 0   
B     2     

(3.12)

  2  2  
where Rt   Rt exp    t      Z t     Wt  ,   is the value of recovery on the
 2  
collateral at the time of resolution. We can easily write down the closed-form solution for the
LGD on the bond according to the well-known formula for a compound option, where here
the “outer option” is a put, and the “inner option” is a call. Let R* be the critical level of
recovery such that the holder of the loan is just breaking even:

 exp  c    1  LGDP  R* ,  t |  , c,  , ,   , ,  ,   (3.13)

where   is the time-to-resolution for the loan, which we assume to be prior to that for the
bond,     B . Then the solution is given by:

exp   b 
LGDBP  Rt ,  t |  , c,  ,  , ,  , ,  ,  , B   1    Rt ,  t |  , c,  ,  , ,  ,  , ,  , B 
B
(3.14)
  Rt ,  t |  , c,  ,  , ,   , ,  ,  , B  

BIS Papers No 58 263

      
 B exp  t     2   a , b ;     Rt  2   a , b ;      exp  c      a  (3.15)
 B   B 

1   Rt   1 2 
a   log  *       t  ˆ   (3.16)
ˆ    R   2 

1   Rt   1 2 
b   log     B   t  ˆ   (3.17)
ˆ B   B   2 

Where  2  X , Y ;  XY  is the bivariate normal distribution function for Brownian increments

TX
the correlation parameter is given by  XY  for respective “expiry times” TX and TY for
TY
X and Y, respectively. Note that this assumption, which is realistic in that we observe in the
data that on average earlier default on the bond even if the emerges from bankruptcy or
resolve a default at a single time (which in addition is random), is matter of necessity in the

log-normal setting in that the bivariate normal distribution is not defined for  XY   1 in

the case that TX  TY  

We can extend this framework to arbitrary tranches of debt, such as for a subordinated issue,
in which case we follow the same procedure in order to arrive at an expression that involves
trivariate cumulative normal distributions. In general, a debt issue that is subordinated to the
dth degree results in a pricing formula that is a linear combination of d+1 variate Gaussian
distributions. These formulae become cumbersome very quickly, so for the sake of brevity
we refer the interested reader to Haug (2006) for further details.

4. Comparitive statics
In this section we discuss and analyze the sensitivity of ultimate LGD in to various key
parameters. In Figures 1 through 5 we examine the sensitivity of the ultimate LGD in the two-
factor model of Section 3, incorporating the optimal foreclosure boundary. In Figure 1, we
look at the ultimate LGD on the bond and the loan for three different settings of the factor
loading of the recovery rate process on the systematic factor in the firm value processes
(  = 0.05, 0.45 and 0.9), while fixing other parameters at “reasonable” values motivated by
prior literature (drift in recovery  = 0.08, face value of loan  = 0.5, coupon rate on loan
c = 0.06, LGD side volatility  = 0.3, volatility of recovery return drift process  = 0.5, speed
of mean reversion in LGD return  = 0.5, correlation between LGD side systematic factor
and random factor in recovery rate drift  =0.3, and time-to-resolution  =1). We observe that
ultimate LGD is monotonically decreasing at increasing rate in value of the firm at default,
that this increasing in the correlation between the PD and LGD side systematic factors, and
that this is also uniformly higher for bonds than for loans. In Figure 2 we show the ultimate
LGD as a function of the volatility in the recovery rate process attributable to the LGD side
systematic factor  , fixing firm value at default at Vt = 0.5. We observe that ultimate LGD
increases at an increasing rate in this parameter, that for higher correlation between firm
asset value and recovery value return the LGD is higher and increases at a faster rate, and
that for bonds these curves lie above and increase at a faster rate. In Figure 3 we show the
ultimate LGD as a function of the volatility  in the recovery rate process attributable to the

264 BIS Papers No 58

PD side systematic factor, fixing LGD side volatility  = 0.5, for different firm values at default
at Vt   0.3, 0.5, 0.8  . We observe that ultimate LGD increases at an increasing rate in this
parameter, that for lower firm asset values the LGD is higher but increases at a slower rate,
and that for bonds these curves lie above and increase at a lower rate. In Figure 4 we show
the ultimate LGD as a function of the volatility  of the stochastic drift in the recovery rate
process, for three different settings of the factor loading of the recovery rate process on the
systematic factor in the firm value processes (  = 0.05, 0.45 and 0.9). We observe that
ultimate LGD in this parameter decreases at a decreasing rate, although the sensitivity is not
great (especially for loans), and that as expected the curves lie above for greater PD-LGD
correlation and for bonds as compared to loans. Finally, in Figure 5 we fix   0.3 and vary
 , the coefficient of mean reversion in the drift process for the recovery rate, and observe
that ultimate LGD is increasing in this parameter, at a decreasing rate and having a
discontinuity for these parameter settings; and as expected, for higher levels of default and
recovery correlation, or for bonds as compared to loans, the curves lie everywhere above.
LGD(V|cor(R,V)={.05,.15,.45},alpha=.08,lam=.5,c=.06,eta=.3,beta=.5,nu=.5,kap=.5,kce=.3,tau=1)

Figure 1: Ultimate Loss-Given-Default vs. Value of Firm at Default

1.0

Loan-cor(R,V)=0.15
Loan-cor(R,V)=0.05
Loan-cor(R,V)=0.45
Bond-cor(R,V)=0.15
Bond-cor(R,V)=0.05
0.5

Bond-cor(R,V)=0.45
0.0
-0.5
-1.0

0.0 0.2 0.4 0.6 0.8 1.0

V_t
Stochastic Collateral & Drift Merton Model
LGD(nu|cor(R,V)={.05,.15,.45},V=.5,alpha=.08,lam=.5,c=.06,eta=.3,beta=.5,kap=.5,kce=.3,tau=1)

Figure 2: Ultimate Loss-Given-Default vs. Sensitivity of Recovery Process to LGD Side Systematic Factor
1.0

Loan-cor(R,V)=0.15
Loan-cor(R,V)=0.05
Loan-cor(R,V)=0.45
Bond-cor(R,V)=0.15
0.8

Bond-cor(R,V)=0.05
Bond-cor(R,V)=0.45
0.6
0.4
0.2
0.0

0.5 0.6 0.7 0.8 0.9 1.0

nu
Stochastic Collateral & Drift Merton Model

BIS Papers No 58 265

 Figure 3: Ultimate Loss-Given-Default vs. Sensitivity of Recovery Process to PD Side Systematic Factor
LGD(beta|v={.3,.5,.8}.alpha=.08,lam=.5,c=.06,eta=.3,kap=.5,kce=.3,tau=1)

1.0
0.5
0.0

Loan-V=0.3
-0.5

Loan-V=0.5
Loan-V=0.8
Bond-V=0.3
Bond-V=0.5
-1.0

Bond-V=0.8

0.0 0.2 0.4 0.6 0.8 1.0

beta
Stochastic Collateral & Drift Merton Model
LGD(eta|cor(R,V)={.05,.15,.45},V=.5,alpha=.08,lam=.5,c=.06,beta=.5,kap=.5,nu=.5,kce=.3,tau=1)

Figure 4: Ultimate Loss-Given-Default vs. Volatility in the Drift of the Recovery Rate Process
1.0
0.8
0.6
0.4

Loan-cor(R,V)=0.15
Loan-cor(R,V)=0.05
0.2

Loan-cor(R,V)=0.45
Bond-cor(R,V)=0.15
Bond-cor(R,V)=0.05
0.0

Bond-cor(R,V)=0.45

0.0 0.1 0.2 0.3 0.4

eta
Stochastic Collateral & Drift Merton Model
LGD(kap|cor(R,V)={.05,.15,.45},V=.5,alpha=.08,lam=.5,c=.06,eta=.3,beta=.5,nu=.5,kce=.3,tau=1)

Figure 5: Ultimate Loss-Given-Default vs. Mean-Reversion Recovery Drift Process

1.0
0.5
0.0

Loan-cor(R,V)=0.15
-0.5

Loan-cor(R,V)=0.05
Loan-cor(R,V)=0.45
Bond-cor(R,V)=0.15
Bond-cor(R,V)=0.05
-1.0

Bond-cor(R,V)=0.45

0.0 0.2 0.4 0.6 0.8 1.0

kappa
Stochastic Collateral & Drift Merton Model

266 BIS Papers No 58

5. Empirical analysis: calibration of models
In this section we describe our strategy for estimating parameters of the different models for
LGD by full-information maximum likelihood (FIML.) This involves a consideration of the LGD
implied in the market at time of default tiD for the ith instrument in recovery segment s,
denoted LGDi , s ,t D . This is the expected, discounted ultimate loss-given-default LGDi , s ,t E at
i i

i
E
time of emergence t as given by any of our models m, LGD P
s ,m θ 
s ,m over the resolution
E D
period t i ,s  t i ,s :

EtP  LGDi , s ,t E 
LGDi ,s ,t D   i 
 LGDsP,m  θ s ,m 
1  ri,s 
i t E

D i ,s i ,st D

(5.1)
where θ s ,m is the parameter vector for segment s under model m, expectation is taken with
respect to physical measure P, discounting is at risk adjusted rate appropriate to the
D
instrument ri , s and it is assumed that the time-to-resolution tiE,s  tiD,s is known.

In order to account for the fact that we cannot observe expected recovery prices ex ante, as
only by coincidence would they coincide with expectations, we invoke market rationality to
postulate that for a segment homogenous with respect to recovery risk the difference
between expected and average realized recoveries should be small. We formulate this by
defining the normalized forecast error as:

LGDsP, m  θ s ,m   LGDi , s ,t E
i , s  i

LGDi , s ,t D  t  t E
i,s
D
i,s
i
(5.2)
This is the forecast error as a proportion of the LGD implied by the market at default (a “unit-
free” measure of recovery uncertainty) and the square root of the time-to-resolution. This is a
mechanism to control for the likely increase in uncertainty with time-to-resolution, which
effectively puts more weight on longer resolutions, increasing the estimate of the loss-
severity. The idea behind this is that more information is revealed as the emergence point is
approached, hence a decrease in risk. Alternatively, we can analyze
LGDsP, m  θ s ,m   LGDi , s ,t E
 i,s  i
, the forecast error that is non-time adjusted, and argue that
LGDi , s ,t D
i

its standard error is proportional to tiE, s  tiD, s , which is consistent with an economy in which
information is revealed uniformly and independently through time (Miu and Ozdemir, 2005).
Assuming that the errors i , s in (5.2) are standard normal,8 we may use full-information
maximum likelihood (FIML), by maximizing the log-likelihood (LL) function:

8
If the errors are i.i.d and from symmetric distributions, then we can still obtain consistent estimates through
ML, which has the interpretations as the quasi-ML estimator.

BIS Papers No 58 267

 N sD
 sP,m  arg max LL  arg max log    θ  
LGD
θ s ,m θ s ,m

i 1
 i,s s ,m   
  LGD P  θ   LGD E
N sD  (5.3)
 arg max  log   
s ,m s ,m i , s ,ti

θ s ,m i 1
  LGDi , s ,t D  ti , s  ti , s
E D 
  i  
This turns out to be equivalent to minimizing the squared normalized forecast errors:
N  LGDsP,m  θ s ,m   LGDi , s ,t E 
2
  Ns
 s 1  2
 arg min  E   arg min  i , s , m 
P

LGD s ,m  i

θ s ,m
 i 1 ti , s  ti , s
D
 LGDi , s ,t D   θ s ,m  i 1 
  i   (5.4)
We may derive a measure of uncertainty of our estimate by the ML standard errors from the
Hessian matrix evaluated at the optimum:
1

  2 LL  2
ˆˆ
Σ   
 θ s ,m θ s ,m
θ T
s ,m

θ s ,m  θˆ s ,m
(5.5)

6. Data and estimation results

We summarize basic characteristics of our data-set in Tables 1 and 2, and the maximum
likelihood estimates are shown in Table 3. These are based upon our analysis of defaulted
bonds and loans in the Moody’s Ultimate Recovery (MURD™) database release as of
August, 2009. This contains the market values of defaulted instruments at or near the time of
default,9 as well as the values of such pre-petition instruments (or of instruments received in
settlement) at the time of default resolution. This database is largely representative of the
U.S. large-corporate loss experience, from the mid-1980’s to the present, including most of
the major corporate bankruptcies occurring in this period.
Table 1 shows summary statistics of various quantities of interest according to instrument
type (bank loan, bond, term loan or revolver) and default type (bankruptcy under Chapter 11
or out-of-court renegotiation). First, we take the annualized return or yield on defaulted debt
from the date of default (bankruptcy filing or distressed renegotiation date) to the date of
resolution (settlement of renegotiation or emergence from Chapter 11), henceforth
abbreviated as “RDD”. Second, the trading price at default implied LGD (“TLGD”), or par
minus the trading price of defaulted debt at the time of default (average 30-45 days after
default) as a percent of par. Third, our measure of ultimate loss severity, the dollar loss-
given-default on the debt instrument at emergence from bankruptcy or time of final
settlement (“ULGD”), computed as par minus either values of pre-petition or settlement
instruments at resolution. We also summarize two additional variables in Table 1, the total
instrument outstanding at default, and the time in years from the instrument default date to
the time of ultimate recovery.

9
This an average of trading prices from 30 to 45 days following the default event. A set of dealers is polled
every day and the minimum /maximum quote is thrown out. This is done by experts at Moody’s.

268 BIS Papers No 58

The preponderance of this sample is made up of bankruptcies as opposed to out-of-court
settlements, 1,322 out of a total of 1,398 instruments. We note that out-of-court settlements
have lower LGDs by either the trading or ultimate measures, 37.7% and 33.8%, as compared
to Chapter 11’s, 55.7% and 51.6%, respectively; and the heavy weight of bankruptcies are
reflected in how close the latter are to the overall averages, 54.7% and 50.6% for TLGD and
ULGD, respectively. Interestingly, not only do distressed renegotiations have lower loss
severities, but such debt performs better over the default period than bankruptcies, RDD of
37.3% as compared to 28.1%, as compared to an overall RDD of 28.6%. We also note that
the TLGD is higher than the ULGD by around 5% across default and instrument types,
55.7% (37.7%) as compared to 51.6% (33.8%) for bankruptcies (renegotiations). We also
see that loans have better recoveries by both measures as well higher returns on defaulted
debt, respective average TLGD, ULGD and RDD 52.5%, 49.3% and 32.2%. I
In Table 2 we summarize ULGD, TLGD and RDD by major collateral categories and seniority
classes. We observe for this sample that either LGD measure appears to weakly exhibit the
usual decreasing pattern observed in the literature with respect to higher seniority class, but
this relationship is not consistent with respect to collateral categories. On the other hand,
while also not monotonic, we a somewhat stronger relationship for RDD, as these tend to be
higher for either better secured or more highly ranked instruments. We have average TLGD
(ULGD) of 53.3% (49.3%), 51.6% (35.0%), 56.0% (38.0%), 58.5% (36.5%) and 65.8%
(33.5%) for Revolving Credit/Term Loan, Senior Secured Bonds, Senior Unsecured Bonds,
Senior Subordinated Bonds and Junior Subordinated Bonds, respectively. The
corresponding averages of RDD in descending order of seniority class are 32.2%, 36.6%,
23.8%, 33.2% and 15.6% - an overall decreasing albeit non-monotonic pattern. On the other
hand, for this particular sample and segmentation of collateral codes, we fail to see much of
a rank ordering as we might have expected. We have average TLGD (ULGD) of 66.5%
(65.0%), 41.6% (32.9%), 50.6% (47.6%), 61.6% (48.6%), 59.3% (59.4%) and 57.4%
(51.46%) for Cash, Accounts Receivables & Guarantees, Inventory/Most Assets &
Equipment, All Assets & Real Estate, Non-Current Assets & Capital Stock, PPE/Second Lien
and Unsecured/Other Illiquid Collateral, respectively. Even just focusing upon the split
between secured and unsecured, we fail to see much (any) of a difference in TLGD (ULGD),
57.58% vs. 53.40% (37.69% vs. 36.13%), respectively. The corresponding averages of RDD
in descending order of collateral quality are: 22.6%, 33.2%, 33.8%, 46.2%, 29.0% and 24.1%
- a humped shaped pattern. However, RDD is higher for secured as compared to unsecured,
34.5% vs. 3.6%, respectively.

BIS Papers No 58 269

 Table 1
Characteristics of loss-given-default and return on
defaulted debt observations by default and instrument type
(Moody's Ultimate Recovery Database 1987-2009)

Bankruptcy Out-of-Court Total

Standard Standard Standard

Count Average Error of Count Average Error of Count Average Error of
the Mean the Mean the Mean

Return on
1
Defaulted Debt 28.32% 3.47% 45.11% 19.57% 29.19% 3.44%
Bonds and Term Loans

2
LGD at Default 55.97% 0.96% 38.98% 3.29% 55.08% 0.93%

Discounted LGD
3 1072 51.43% 1.15% 59 33.89% 3.05% 1131 50.52% 1.10%
Time-to-
4
Resolution 1.7263 0.0433 0.0665 0.0333 1.6398 0.0425
Principal at
5
Default 207'581 9'043 416'751 65'675 218'493 9'323
Return on
1
Defaulted Debt 25.44% 3.75% 44.22% 21.90% 26.44% 3.74%

2
LGD at Default 57.03% 1.97% 37.02% 5.40% 55.96% 1.88%
Bonds

Discounted LGD
3 837 52.44% 1.30% 47 30.96% 3.00% 884 51.30% 1.25%
Time-to-
4
Resolution 1.8274 0.0486 0.0828 0.0415 1.7346 0.0424
Principal at
5
Default 214'893 11'148 432'061 72'727 226'439 11'347
Return on
1
Defaulted Debt 26.93% 7.74% 10.32% 4.61% 25.88% 7.26%

2
LGD at Default 54.37% 1.96% 33.35% 8.10% 53.03% 1.93%
Revolvers

Discounted LGD
3 250 52.03% 2.31% 17 33.33% 7.63% 267 50.84% 2.23%
Time-to-
4
Resolution 1.4089 0.0798 0.0027 0.0000 1.3194 0.0776
Principal at
5
Default 205'028 19'378 246'163 78'208 207'647 18'786
Return on 26.161 18.872
1
Defaulted Debt 32.57% 5.71% % % 32.21% 5.49%

2
LGD at Default 53.31% 9.90% 38.86% 7.22% 52.50% 3.21%
Loans

Discounted LGD
3 485 50.00% 1.68% 29 38.31% 5.79% 514 49.34% 2.25%
Time-to-
4
Resolution 1.3884 0.0605 0.0027 0.0000 1.3102 0.0816
Principal at
5
Default 193'647 11'336 291'939 78'628 199'192 16'088

270 BIS Papers No 58

 Table 1 (cont)
Characteristics of loss-given-default and return on
defaulted debt observations by default and instrument type
(Moody's Ultimate Recovery Database 1987-2009)

Bankruptcy Out-of-Court Total

Standard Standard Standard

Error of Error of Error of
Count Average the Mean Count Average the Mean Count Average the Mean

Return on
1
Defaulted Debt 28.05% 3.17% 37.33% 15.29% 28.56% 3.11%

2
LGD at Default 55.66% 0.86% 37.72% 3.12% 54.69% 0.84%
Total

Discounted LGD
3 1322 51.55% 1.03% 76 33.76% 2.89% 1398 50.58% 0.99%
Time-to-
4
Resolution 1.6663 0.0384 0.0522 0.0260 1.5786 0.0376
Principal at
5
Default 207'099 8'194 378'593 54'302 216'422 8'351
1
Annualized return or yield on defaulted debt from the date of default (bankruptcy filing or
distressed renegotiation date) to the date of resolution (settlement of renegotiation or emergence
from Chapter 11). 2 Par minus the price of defaulted debt at the time of default (average 30-45
days after default) as a percent of par. 3 The ultimate dollar loss-given-default on the defaulted
debt instrument = 1 - (total recovery at emergence from bankruptcy or time of final settlement)/
(outstanding at default). Alternatively, this can be expressed as (outstanding at default - total
ultimate loss)/(outstanding at default). 4 The total instrument outstanding at default. 5 The time
in years from the instrument default date to the time of ultimate recovery.

BIS Papers No 58 271

 Table 2
Loss-given-default by seniority ranks and collateral types
(Moody's Ultimate Recovery Database 1987-2009)

Cash, Non-
Inventory, Unsecured
Accounts All Assets Current PPE &
Most & Other Total Total Total
Collateral Type Receivables & Real Assets & Second
Assets & Illiquid Unsecured Secured Collateral
& Estate Capital Lien
Equipment Collateral
Guarantees Stock

Count 39 8 367 38 29 33 32 482 514

Revolving Credit / Term Loan

Average 66.81% 46.60% 51.95% 59.94% 55.02% 45.63% 46.25% 53.79% 53.31%
LGD at
Default1 Standard
Error 4.44% 11.79% 1.70% 5.27% 6.08% 5.07% 5.20% 1.47% 1.42%

Average 64.38% 56.03% 48.58% 50.62% 56.53% 30.70% 31.78% 50.51% 49.34%
Ultimate
LGD2 Standard
Error 5.09% 13.85% 1.91% 6.10% 6.88% 6.17% 5.20% 1.47% 1.42%

Return on Average 22.57% -5.80% 33.49% 35.68% 46.07% 22.39% 19.77% 33.03% 32.21%
Defaulted
Standard
Debt3
Error 18.20% 30.27% 6.89% 15.01% 27.64% 8.12% 7.93% 5.83% 5.49%

Count 2 38 41 35 7 142 3 139 142

Average 61.50% 40.19% 36.02% 62.99% 61.24% 51.67% 50.73% 51.59% 51.57%
LGD at
Senior Secured Bonds

Default1 Standard
Error 36.50% 5.50% 5.03% 4.71% 11.63% 2.48% 23.79% 2.76% 2.74%

Average 76.81% 23.87% 36.67% 46.70% 60.32% 49.68% 50.15% 34.88% 35.04%
Ultimate
LGD2 Standard
Error 19.39% 3.90% 5.61% 5.71% 12.68% 3.19% 28.95% 2.96% 2.94%

Return on Average 23.86% 47.53% 35.03% 55.99% 14.33% 17.44% -27.66% 38.02% 36.63%
Defaulted
Standard
Debt3
Error 40.63% 7.18% 22.04% 20.10% 27.41% 6.34% 36.65% 9.05% 8.92%

Count 0 0 1 0 1 459 452 9 461

Average 0.00% 0.00% 85.00% N/A 80.00% 55.83% 55.94% 56.63% 55.96%
Senior Unsecured Bonds

LGD at
Default1 Standard
Error N/A N/A N/A N/A N/A 1.42% 1.43% 10.36% 1.42%

Average 0.00% 0.00% 78.76% N/A 74.25% 48.33% 38.14% 32.03% 38.00%
Ultimate
LGD2 Standard
Error N/A N/A N/A N/A N/A 1.78% 1.79% 10.68% 1.77%

Return on Average 0.00% 0.00% 86.47% n 119.64% 23.40% 23.71% 25.62% 23.75%
Defaulted Standard
Debt3 Error N/A N/A N/A N/A N/A 4.80% 4.86% 22.61% 4.78%

Count 0 0 1 0 1 159 158 3 161

Average 0.00% N/A 85.00% N/A 90.50% 58.09% 57.98% 83.46% 58.48%
Senior Subordinated Bonds

LGD at
Default1 Standard
Error N/A N/A N/A N/A N/A 2.48% 2.50% 4.58% 2.47%

Average N/A N/A 74.72% N/A 97.74% 54.51% 36.50% 40.47% 36.46%
Ultimate
LGD2 Standard
Error N/A N/A N/A N/A N/A 2.89% 2.90% 23.36% 2.87%

Return on Average 0.00% N/A 57.45% N/A -45.98% 33.57% 31.01% 150.30% 33.23%
Defaulted Standard
Debt3 Error N/A N/A N/A N/A N/A 10.44% 10.18% 147.62% 10.32%

272 BIS Papers No 58

 Table 2 (cont)
Loss-given-default by seniority ranks and collateral types
(Moody's Ultimate Recovery Database 1987-2009)

Cash, Non-
Inventory, Unsecured
Accounts All Assets Current PPE &
Most & Other Total Total Total
Collateral Type Receivables & Real Assets & Second
Assets & Illiquid Unsecured Secured Collateral
& Estate Capital Lien
Equipment Collateral
Guarantees Stock

Count 0 1 0 0 0 119 117 3 120

Average N/A 27.33% 0.00% N/A N/A 66.15% 66.58% 37.42% 65.81%
Junior Subordinated Bonds

LGD at
Default1 Standard
Error N/A N/A N/A N/A N/A 2.50% 2.48% 22.25% 2.50%

Average N/A 20.15% 0.00% N/A N/A 65.36% 33.62% 32.77% 33.54%
Ultimate
LGD2 Standard
Error N/A N/A N/A N/A N/A 3.06% 3.11% 18.92% 3.06%

Return on Average N/A 72.13% 0.00% N/A N/A 15.11% 15.74% 9.49% 15.59%
Defaulted
Standard
Debt3 Error N/A N/A N/A N/A N/A 10.93% 11.11% 31.36% 10.85%

Count 41 28 407 79 66 777 762 636 1398

Average 66.53% 41.57% 50.55% 61.56% 59.31% 57.41% 57.58% 53.40% 55.66%
LGD at
Default1 Standard
Total Instruments

Error 4.41% 6.18% 1.63% 3.39% 3.86% 1.09% 1.10% 1.28% 0.84%

Average 64.98% 32.93% 47.60% 48.58% 59.43% 51.46% 37.69% 36.13% 36.99%
Ultimate
LGD2 Standard
Error 4.90% 5.99% 1.82% 3.99% 4.28% 1.35% 1.37% 1.43% 0.99%

Return on Average 22.63% 33.17% 33.82% 46.22% 28.96% 24.12% 34.46% 23.63% 28.56%
Defaulted
Standard
Debt3 Error 17.36% 11.74% 6.56% 12.02% 13.90% 3.94% 3.97% 4.89% 3.11%

1 2
Par minus the price of defaulted debt at the time of default (average 30-45 days after default) as a percent of par. The
ultimate dollar loss-given-default on the defaulted debt instrument = 1 - (total recovery at emergence from bankruptcy or time
of final settlement)/(outstanding at default). Alternatively, this can be expressed as (outstanding at default - total ultimate
3
loss)/(outstanding at default). Annualized return or yield on defaulted debt from the date of default (bankruptcy filing or
distressed renegotiation date) to the date of resolution (settlement of renegotiation or emergence from Chapter 11).

In Table 3 we present the full-information maximum likelihood estimation (FIML) results of

the leading model for ultimate LGD derived in this paper, the two-factor structural model of
ultimate loss-given-default, with systematic recovery risk and random drift (2FSM-SR&RD)
on the recovery process.10 The model is estimated along with the optimal foreclosure
boundary constraint.
We first discuss the MLE point estimates of the parameters governing the firm value process
and default risk, or the “PD-side”. Regarding the parameter  , which is the volatility of the
firm-value process governing default, we observe that estimates are decreasing in seniority
class, ranging from 9.1% to 4.3% from subordinated bonds to senior loans. As standard
errors range in 1% to 2%, increasing in seniority rank, these differences across seniority
classes and models are generally statistically significant. Regarding the MLE point estimates
of the parameter  , which is the drift of the firm-value process governing default, we
observe estimates are increasing in seniority class, ranging from 9.6% to 18.6% from
subordinated bonds to loans, respectively.

10
Estimates for the baseline Merton structural model (BMSM) and for the Merton structural model with
stochastic drift (MSM-SD) are available upon request.

BIS Papers No 58 273

274

Table 3
Full information maximum likelihood estimation of option theoretic two-factor structural model of ultimate loss-given-default
with optimal foreclosure boundary, systematic recovery risk and random drift in the recovery process
(Moody's Ultimate Recovery Database 1987-2009)

Recovery Segment Parameter σ1 μ2 β3 ν4 σR5 πRβ6 πRν7 (βσ)0.5 κα8 α9 ηα10 ς11

Est.
Seniority Class

4.32% 18.63% 18.16% 36.83% 41.06% 19.55% 80.45% 12.82% 3.96% 37.08% 48.85% 20.88%
Revolving Credit / Term Loan
Std. Err. 0.5474% 0.9177% 0.7310% 1.3719% 0.4190% 0.0755% 4.2546% 3.2125% 0.9215%

Est. 5.47% 16.99% 16.54% 30.41% 34.62% 22.83% 77.17% 11.64% 4.40% 33.66% 44.43% 18.99%
Senior Secured Bonds
Std. Err. 0.5314% 0.8613% 0.6008% 1.3104% 0.7448% 0.0602% 3.5085% 2.6903% 0.8297%

Est. 6.82% 14.16% 13.82% 24.38% 28.02% 24.30% 75.70% 9.71% 5.50% 28.07% 37.04% 15.83%
Senior Unsecured Bonds
Std. Err. 0.5993% 1.0813% 1.3913% 1.9947% 0.6165% 0.0281% 2.8868% 2.2441% 0.6504%

Est. 8.19% 11.33% 12.02% 17.35% 21.11% 32.43% 67.57% 7.76% 4.42% 22.45% 29.68% 12.69%
Senior Subordinated Bonds
Std. Err. 0.6216% 1.0087% 1.0482% 1.0389% 0.9775% 0.0181% 2.0056% 2.0132% 1.0016%

Est. 9.05% 9.60% 10.24% 12.37% 16.06% 40.66% 59.34% 5.97% 3.34% 18.80% 18.69% 9.43%
Subordinated Bonds
Std. Err. 0.6192% 1.0721% 1.0128% 1.0771% 0.9142% 0.0106% 2.0488% 2.0014% 1.0142%

Value Log-Likelihood Function -371.09

Degrees of Freedon 1391
P-Value of Likelihood Ratio Statistic 4.69E-03
Time Diagnostic

Area Under ROC Curve 93.14%

In-Sample /

Statistics

Komogorov-Smirnov Stat. (P-Val.) 2.14E-08

McFadden Pseudo R-Squared 72.11%

BIS Papers No 58

Hoshmer-Lemeshow Chi-Squared (P-Values) 0.63

1 2 3
The volatility of the firm-value process governing default. The drift of the firm-value process governing default. The sensitivity of the recovery-rate process to the systematic governing
4
default in (or the component of volatility in the recovery process due to PD-side systematic risk). The sensitivity of the recovery-rate process to the systematic governing collateral value (or the
5 2 2 6
component of volatility in the recovery process due to LGD-side systematic risk). The total volatility of the recovery rate process: sqrt(β +ν ). Component of total recovery variance
2 2 2 7 2 2 2 8
attributable to PD-side (asset value) uncertainty: β /(β +ν ). Component of total recovery variance attributable to LGD-side (collateral value) uncertainty: ν /(β +ν ). The speed of the mean-
9 10
reversion in the random drift in the recovery rate process. The long-run mean of the random drift in the recovery arte process. The volatility of the random drift in the recovery rate
11
process. The correlation of the random processes in drift of and the level of the recovery rate process.
These too are statistically significant across seniorities. The fact that we are observing
different estimates of a single firm value process across seniorities is evidence that models
which attribute identical default risk across different instrument types are misspecified – in
fact, we are measuring lower default risk (i.e., lower asset value volatility and greater drift in
firm-value) in loans and senior secured bonds as compared to unsecured and subordinated
bonds.
A key result regards the magnitudes and composition of the components of recovery volatility
across maturities inferred from the model calibration. The MLE point estimates of the
parameter  , the sensitivity of the recovery-rate process to the systematic factor governing
default (or due to PD-side systematic risk), increases in seniority class, from 10.2% for
subordinated bonds to 18.2% senior bank loans. On the other hand, estimates of the
parameter , the sensitivity of the recovery-rate process to the systematic factor governing
collateral value (or due to LGD-side systematic risk), are greater than  across seniorities,
and similarly increases in from 12.4% for subordinated bonds to 36.8% for bank loans. This
monotonic increase in both  and  as we move up in the hierarchy of the capital structure
from lower to higher ranked instruments has the interpretation of a greater sensitivity in the
recovery rate process attributable to both systematic risks, implying that total recovery
volatility  R     increases from higher to lower ELGD instruments, from 16.1% for
2 2

subordinated bonds to 41.1% for senior loans. However, we see that the proportion of the
total recovery volatility attributable to systematic risk in collateral (firm) value, or the LGD
(PD) side, is increasing (decreasing) in seniority from 59.3% to 80.5% (40.7% to 19.6%) from
subordinated bonds to senior bank loans. Therefore, more senior instruments not only exhibit
greater recovery volatility than less senior instruments, but a larger component of this
volatility is driven by the collateral rather than the asset value process.
The next set of results concern the random drift in the recovery rate process. The MLE point
estimates of the parameter  , the speed of the mean-reversion in, is hump-shaped in
seniority class, ranging from 3.3% subordinated bonds, to 5.5% for senior unsecured bonds,
to 4.0% for loans, respectively. Estimates of the parameter  , the long-run mean of the
random drift in the recovery rate process, increase in seniority class from 18.8% for
subordinated bonds to 37.1% for senior bank loans. This monotonic increase in  as we
move from lower to higher ranked instruments has the interpretation of greater expected
return of the recovery rate process inferred from lower ELGD (or greater expected recovery)
instruments as we move up in the hierarchy of the capital structure. We see that the volatility
of the random drift in the recovery rate process  , increases in seniority class, ranging from
18.7% to 48.9% from subordinated bonds to senior loans, respectively. The monotonic
increase in  as we move from lower to higher ranked instruments has the interpretation of
greater volatility in expected return of the recovery rate process inferred from lower ELGD (or
greater expected recovery) instruments as we move up in the hierarchy of the capital
structure. Finally, estimates of the parameter  , the correlation of the random processes in
drift of and the level of the recovery rate process, increases in seniority class from 9.4% for
subordinated bonds to 20.9% for senior bank loans.
Finally with respect to parameter estimates, regarding the MLE point estimates of the
correlation between the default and recovery rate processes  in the 2FSM-SR&RD, we
observe estimates are increasing in seniority class, ranging from 6.0% to 12.8% from
subordinated bonds to loans, respectively.
We conclude this section by discussing the quality of the estimates and model performance
measures. Across seniority classes, parameter estimates are all statistically significant, and
the magnitudes of such estimates are in general distinguishable across segments at
conventional significance levels. The likelihood ratio statistic indicates that we can reject the

BIS Papers No 58 275

null hypothesis that all parameter estimates are equal to zero across all ELGD segments, a
p-value of 4.7e-3. We also show various diagnostics that assess in-sample fit, which show
that the model performs well-in sample. The area under receiver operating characteristic
curve (AUROC) of 93.1% is high by commonly accepted standards, indicating a good ability
of the model to discriminate between high and low LGD defaulted instruments. Another test
of discriminatory ability of the models is the Kolmogorov-Smirnov (KS) statistic, the very
small p-value 2.1e-8 indicating adequate separation in the distributions of the low and high
LGD instruments in the model.11 We also show two tests of predictive accuracy, which is the
ability of the model to accurately quantify a level of LGD. The McFadden psuedo r-squared
(MPR2) is high by commonly accepted standards, 72.1%, indicating a high rank-order
correlation between model and realized LGDs of defaulted instruments. Another test of
predictive accuracy of the models is the Hoshmer-Lemeshow (HL) statistic, high p-values of
0.63 indicating high accuracy of the model to forecast cardinal LGD.

7. Downturn LGD
In this section we explore the implications of our model with respect to downturn LGD in the
2FSM-SR&RD. This is a critical component of the quantification process in the Basel II
advanced IRB framework for regulatory capital. The Final Rule (FR) in the U.S. (OCC et al,
2007) requires banks that either wish, or are required, to qualify for treatment under the
advanced approach to estimate a downturn LGD. We paraphrase the FR, this is an LGD
estimated during an historical reference period during which default rates are elevated within
an institution’s loan portfolio.
In Figures 6 through 8 we plot the ratios of the downturn LGD to the expected LGD. This is
derived by conditioning on the 99.9th quantile of the PD side systematic factor in the 2FSM-
SR&RD. We show this for loans and bonds, as well as for different settings of key
parameters (  , or  ) in each plot, with other parameters set to the MLE estimates.
We observe that the LGD mark-up for downturn is monotonically declining in ELGD, which is
indicative of lower tail risk in recovery for lower ELGD instruments. It is also greater than
unity in all cases, and approaches 1 as ELGD approaches 1. This multiple is higher for
bonds than for loans, as well as for either higher PD-LGD correlation  , collateral

specific volatility  or volatility in the drift of the recovery rate drift process  ; although these
differences narrow for higher ELGD. For example, in Figure 6, we see that for loans having

ELGD of 15% and


= 10% (=20%), the ratio of downturn to ELGD is about 2 (2.5); but
for ELGD of 50%, this is about 1.5 (1.6); and for ELGD of 80%, this about 1.2 (1.3). And for

bonds having ELGD of 15% and


= 10% (=20%), the ratio of downturn to ELGD is about
2.5 (23); but for ELGD of 50%, this is about 2 (2.2) ; and for ELGD of 80%, this is about 1.6
(1.7).

11
In these tests we take the median LGD to be the cut-off that distinguishes between a high and low realized
LGD.

276 BIS Papers No 58

8. Model validation
In this final section we validate our preferred model, the 2FSM-SR&RD. In particular, we
implement an out-of-sample and out-of-time analysis, on a rolling annual cohort basis for the
final nine years of our sample. Furthermore, we augment this by resampling on both the
training and prediction samples, a non-parametric bootstrap (Efron [1979], Efron and
Tibshirani [1986], Davison and Hinkley [1997]). The procedure is as follows: the first training
(or estimation) sample is established as the cohorts defaulting in the 10 years 1987-1996,
and the first prediction (or validation) sample is established as the 1997 cohort. Then we
resample 100,000 times with replacement from the training sample the 1987-1996 cohorts
and for the prediction sample 1997 cohort, and then based upon the fitted model in the
former we evaluate the model based upon the latter. Then we augment the training sample
with the 1997 cohort, and establish the 1998 cohort as the prediction sample, and repeat
this. This is continued until we have left the 2008 cohort as the holdout. Finally, to form our
final holdout sample, we pool all of our out-of-sample resampled prediction cohorts, the
12 years running from 1997 to 2008. We then analyze the distributional properties (such as
median, dispersion and shape) of the two key diagnostic statistics: the Spearman rank-order
correlation for discriminatory (or classification) accuracy, and the Hoshmer-Lemeshow
Chi-Squared (HLCQ) P-values for predictive accuracy, or calibration.

Figure 6: Ratio of Ultimate Downturn to Expected LGD vs. ELGD at 99.9th Percentile of PD-Side Systematic Factor Z
5

Loans/corr(R,V)=0.1
Loans/corr(R,V)=0.2
Loans/cor(R,V)=0.05
Bonds/corr(R,V)=0.1
Bonds/corr(R,V)=0.2
4

Bonds/cor(R,V)=0.05
DLGD/ELGD

3
2
1

0.0 0.2 0.4 0.6 0.8 1.0

ELGD
Stochastic Collateral & Drift Merton Model (Parameters Set to MLE Estimates for Loans & Bonds)

BIS Papers No 58 277

 Figure 7: Ratio of Ultimate Downturn to Expected LGD vs. ELGD at 99.9th Percentile of PD-Side Systematic Factor Z

5
Loans/stdev(R|V)=0.1
Loans/stdev(R|V)=0.2
Loans/stdev(R|V)=0.05
Bonds/stdev(R|V)=0.1
Bonds/stdev(R|V)=0.2
4

Bonds/stdev(R|V)=0.05
DLGD/ELGD

3
2
1

0.0 0.2 0.4 0.6 0.8 1.0

ELGD
Stochastic Collateral & Drift Merton Model (Parameters Set to MLE Estimates for Loans & Bonds)

Figure 8: Ratio of Ultimate Downturn to Expected LGD vs. ELGD at 99.9th Percentile of PD-Side Systematic Factor Z
5

Loans/stdev(alpha)=0.1
Loans/stdev(alpha)=0.2
Loans/stdev(alpha)=0.05
Bonds/stdev(alpha)=0.1
Bonds/stdev(alpha)=0.2
4

Bonds/stdev(alpha)=0.05
DLGD/ELGD

3
2
1

0.0 0.2 0.4 0.6 0.8 1.0

ELGD
Stochastic Collateral & Drift Merton Model (Parameters Set to MLE Estimates for Loans & Bonds)

Before discussing the results, we briefly describe the two alternative frameworks for
predicting ultimate LGD that are to be compared to the 2FSM-SR&RD developed in this
paper. First, we implement a full-information maximum likelihood simultaneous equation
regression model (FIMLE-SEM) for ultimate LGD, which is an econometric model built upon
observations in URD at both the instrument and obligor level. FIMLE is used to model the
endogeneity of the relationship between LGD at the firm and instrument levels in an internally
consistent manner. This technique enables us to build a model that can help us understand
some of the structural determinants of LGD, and potentially improve our forecasts of LGD.
This model contains 199 observations from the URD™ with variables: long-term debt to
market value of equity, book value of assets quantile, intangibles to book value of assets,
interest coverage ratio, free cash flow to book value of assets, net income to net sales,
number of major creditor classes, percent secured debt, Altman Z-Score, debt vintage (time
since issued), Moody’s 12-month trailing speculative grade default rate, industry dummy,
filing district dummy and a pre-packaged bankruptcy dummy. Detailed discussion of the
results can be found in Jacobs and Karagozoglu (2011). The second alternative model we
consider addresses the problem of non-parametrically estimating a regression relationship, in
which there are several independent variables and in which the dependent variable is

278 BIS Papers No 58

bounded, as an application to the distribution of LGD. Standard non-parametric estimators of
unknown probability distribution functions, whether or not conditional or not, utilize the
Gaussian kernel (Silverman (1982), Hardle and Linton (1994) and Pagan and Ullah (1999)).
It is well known that there exists a boundary bias with a Gaussian kernel, which assigns non-
zero density outside the support on the dependent variable, when smoothing near the
boundary. Chen (1999) has proposed a beta kernel density estimator (BKDE) defined on the
unit interval [0,1], having the appealing properties of flexible functional form, a bounded
support, simplicity of estimation, non-negativity and an optimal rate of convergence n 4/5 in
finite samples. Furthermore, even if the true density is unbounded at the boundaries, the
BKDE remains consistent (Bouezmarni and Rolin, 2001), which is important in the context of
LGD, as there are point masses (observation clustered at 0% and 100%) in empirical
applications. We extend the BKDE (Renault and Scalliet, 2004) to a generalized beta kernel
conditional density estimator (GBKDE), in which the density is a function of several
independent variables, which affect the smoothing through the dependency of the beta
distribution parameters upon these variables. Detailed derivation of this model can be found
in Jacobs and Karagozoglu (2007), who also present a “horse-race” as herein between
GBKDE the FIMLE-SEM.
Results of the model validation are shown in Table 4 and Figures 9-10. We see that while all
models perform decently out-of-sample in terms of rank ordering capability, FIMLE-SEM
performs the best (median = 83.2%), the GBKDE the worst (median = 72.0%), and our
2FSM-SR&RD in the middle (median = 79.1%). It is also evident from the table and figures
that the better performing models are also less dispersed and exhibit less multi-modality.
However, the structural model is closer in performance to the regression model by the
distribution of the Pearson correlation, and indeed there is a lot of overlap in these.
Unfortunately, the out-of-sample predictive accuracy is not as encouraging for any of the
models, as in a sizable proportion of the runs we can reject adequacy of fit (ie p-values
indicating rejection of the null of that the model fits the data it at conventional levels). The
rank ordering of model performance is the same as for the Pearson statistics: FIMLE-SEM
performs the best (median = 24.8%), the GBKDE the worst (median = 13.2%), and our
2FSM-SR&RD in the middle (median = 23.9%); and the structural model developed herein is
comparable in out-of-sample predictive accuracy to the high-dimensional regression model.
We conclude that while all models are challenged in predicting cardinal levels of ultimate
LGD out-of-sample, it is remarkable that a relatively parsimonious structural model of
ultimate LGD can perform so closely to a highly parameterized econometric model.

BIS Papers No 58 279

 Table 4
1
Bootstrapped out-of-sample and out-of-time classification and predictive accuracy
model comparison analysis of alternative models for ultimate loss-given-default
(Moody's Ultimate Recovery Database 1987-2009)

Test
Model GBKDE4 2FSM-SR&RD5 FIMLE-SEM6
Statistic
Time 1 Year Ahead Prediction

Median 0.7198 0.7910 0.8316

Spearman Standard Deviation 0.1995 0.1170 0.1054
Rank-Order
Out-of-Sample /

th
Correlation2 5 Percentile 0.4206 0.5136 0.5803
th
95 Percentile 0.9095 0.9563 0.9987

Hoshmer- Median 0.1318 0.2385 0.2482

Lemeshow Standard Deviation 0.0720 0.0428 0.0338
Chi- th
Squared 5 Percentile 0.0159 0.0386 0.0408
(P-Values)3 th
95 Percentile 0.2941 0.5547 0.5784
1
In each run, observations are sampled randomly with replacement from the training and prediction samples, the
model is estimated in the training sample and observations are classified in the prediction period, and this is
repeated 100,000 times. 2 The correlation between the ranks of the predicted and realizations, a measure of the
3
discriminatory accuracy of the model. A normalized average deviation between empirical frequencies and
average modelled probabilities across deciles of risk, ranked according to modelled probabilities, a measure of
4
model fit or predictive accuracy of the model. Generalized beta kernel conditional density estimator
5
model. Two-factor structural Merton systematic recovery and random drift model. 6 Full-information
maximum likelihood simultaneous equation regression model. 199 observations with variables: long-term debt to
market value of equity, book value of assets quantile, intangibles to book value of assets, interest coverage ratio,
free cash flow to book value of assets, net income to net sales, number of major creditor classes, percent secured
debt, Altman Z-Score, debt vintage (time since issued), Moody’s 12-month trailing speculative grade default rate,
industry dummy, filing district dummy and prepackaged bankruptcy dummy.

Fig. 9 - Densities of Pearson Correlations for LGD Prediction

100,000 Repetitions Out-of-Sample and Out-of-Time 1997-2008
Simulataneous Equation Regression Model
2-Factor Merton Structural Model
Non-parametric Beta Kernel Density Model
2.5

2.0
Probability Density

1.5

1.0

0.5

0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2
Pearson Correlation

280 BIS Papers No 58

 Fig.10 - Densities of Hoshmer-Lemeshow P-Values for LGD Prediction
100,000 Repetitions Out-of-Sample and Out-of-Time 1997-2008
Simulataneous Equation Regression Model
2-Factor Merton Structural Model
Non-parametric Beta Kernel Density Model
8
Probability Density

0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
P-Value of HL Statistic

9. Conclusions and directions for future research

In this study, we have developed a theoretical model for ultimate loss-given-default, having
many intuitive and realistic features, in the structural credit risk modeling framework. Our
extension admits differential seniority within the capital structure, an independent process
representing a source of undiversifiable recovery risk with a stochastic drift, and an optimal
foreclosure threshold. We have analyzed the comparative statics of this model and
compared these to a baseline structural model. In the empirical analysis we calibrated
alternative models for ultimate LGD on bonds and loans, having both trading prices at default
and at resolution of default, utilizing an extensive sample of agency-rated defaulted firms in
the Moody’s URD™. These 800 defaults are largely representative of the US large corporate
loss experience, for which we have the complete capital structures, and can track the
recoveries on all instruments to the time of default to the time of resolution.
We demonstrated that parameter estimates vary significantly across models and recovery
segments, finding that the estimated volatilities of the recovery rate processes and their
random drifts are increasing in seniority; in particular, for first-lien bank loans as compared to
senior secured or unsecured bonds. We argued that this as reflects the inherently greater
risk in the ultimate recovery for higher ranked instruments having lower expected loss
severities. In an exercise highly relevant to requirements for the quantification of a downturn
LGD for advanced IRB under Basel II, we analyzed the implications of our model for this
purpose, finding the later to be declining for higher expected LGD, higher for lower ranked
instruments, and increasing in the correlation between the process driving firm default and
recovery on collateral. Finally, we validated our leading model derived herein in an out-of-
sample bootstrapping exercise, comparing it to two alternatives, a high-dimensional
regression model and a non-parametric benchmark, both based upon the same URD data.
We found our model to compare favorably in this exercise.
We conclude that our model is worthy of consideration to risk managers, as well as
supervisors concerned with advanced IRB under the Basel II capital accord. It can be a

BIS Papers No 58 281

valuable benchmark for internally developed models for ultimate LGD, as this model can be
calibrated to LGD observed at default (either market prices or model forecasts, if defaulted
instruments non-marketable) and to ultimate LGD measured from workout recoveries. Risk
managers can use our model as an input into internal credit capital models.

References
Acharya, V. V., Bharath, S. T, Srinivasan A., 2007, “Does industry-wide distress affect
defaulted firms? – evidence from creditor recoveries”. Journal of Political Economy, 85(3),
787-821.
Altman, E.I., 1968. “Financial ratios, discriminant analysis and the prediction of corporate
bankruptcy”, Journal of Finance, 23, 589-609.
Altman, E.I., R. Haldeman, Narayanan, P., 1977. “Zeta analysis: a new model to predict
corporate bankruptcies”, Journal of Banking and Finance, 1:1, July, 29-54.
Altman, E.I., Eberhart, A.,1994. “Do seniority provision protect bondholders’ investments?,
Journal of Portfolio Management”, Summer, 67-75.
Altman, E.I., Kishore, V.M, 1996. “Almost everything you wanted to know about recoveries
on defaulted bonds”, Financial Analysts Journal, Nov/Dec.
Altman, E.I., Resti, A, Sironi, A., 2001. “Analyzing and explaining default recovery rates”,
ISDA Research Report, London, December.
______, 2003. “Default recovery rates in credit risk modeling: a review of the literature and
empirical evidence”, Working paper, New York University Salomon Center, February.
Altman, E.I., Brady, B, Resti, A, Sironi, A 2005. “The link between default and recovery rates:
theory, empirical evidence and implications”, Journal of Business 78, 2203-2228.
Altman, E.I., 2006. “Default recovery rates and LGD in credit risk modeling and practice: an
updated review of the literature and empirical evidence”, Working paper, NYU Salomon
Center, November.
Araten, M., Jacobs, Jr., M. and Varshney, P, 2004. “Measuring loss-given-default on
commercial loans for the JP Morgan Chase wholesale bank: an 18 year internal study”, RMA
Journal, 86, 96-103.
Asarnow, Elliot and David Edwards, 1995, “Measuring loss on defaulted bank loans: a
24-year study”, Journal of Commercial Lending, Vol. 77, No.7, pp. 11-23.
Bakshi, G., Madan, D., Zhang, F., 2001. “Understanding the role of recovery in default risk
models: empirical comparisons and implied recovery rates”, Federal Reserve Board, Finance
and Economics Discussion Series #37.
Basel Committee on Banking Supervision, 2003. The new Basel accord, Consultative
Document. Bank for International Settlements (BIS), April.
______, 2005. Guidance on paragraph 468 of the framework document. BIS, July.
______, 2006. International convergence on capital measurement and capital standards: a
revised framework. BIS, June.
Bjerksund, P., 1991. “Contingent claims evaluation when the convenience yield is stochastic:
analytical results”, Working paper, Norwegian School of Economics and Business
Administration.
Black, F. Scholes, M.,1973, “The pricing of options and corporate liabilities”, Journal of
Political Economy 81 (3): 637–654.

282 BIS Papers No 58

Black, F., Cox, J.C., 1976. “Valuing corporate securities: some effects of bond indenture
provisions”, Journal of Finance 31, 637-659.
Bouezmarni, T., Rolin, J.M., 2001. “Consistency of beta kernel density function estimator”,
Canadian Journal of Statistics 31, 89–98.
Chen, S.X., 1999. “Beta kernel estimators for density functions”, Computational Statistics &
Data Analysis 31, 131-145.
Carey, M., Gordy, M. 2007. “The bank as grim reaper: debt decomposition and recoveries on
defaulted debt”, Working paper, Federal Reserve Board.
Credit Suisse Financial Products, 1997, “Creditrisk+ - A Credit Risk Management
Framework”, Technical Document.
Davison, A.C. and D.V. Hinkley. 1997. “Bootstrap Methods and their Application”, Cambridge
University Press, Cambridge.
Dev, A, Pykhtin, M. 2002, “Analytical approach to credit risk modeling”, Risk, March,
S26–S32.
Duffie, D., 1998, “Defaultable term structure models with fractional recovery of par”, Working
paper, Graduate School of Business, Stanford University.
Duffie, D., Singleton, K.J., 1999. “Modeling the term structures of defaultable bonds”, Review
of Financial Studies 12, 687-720.
Duffie, D., Lando, D., 2001. “Term structure of credit spreads with incomplete accounting
information”, Econometrica 69, 633-664.
Efron, B., 1979. “Bootstrap methods: another look at the jackknife”, American Statistician 7,
1-23.
Efron, B., Tibshirani R, 1986. “Bootstrap methods for standard errors, confidence intervals,
and other Measures of statistical accuracy”, Statistical Science 1, 54-75.
Finger, C., 1999, “Conditional approaches for CreditMetrics™ portfolio distributions.
CreditMetrics™”, Monitor, April.
Franks, J., Torous, W., 1989. “An empirical investigation of U.S. firms in reorganization”,
Journal of Finance 44, 747-769.
Friedman, C. and Sandow, S. 2003a. “Ultimate recoveries”, Risk, 69-73.
Frye, J., 2000a. “Collateral damage”, Risk, April, 91-94.
______, 2000b. “Collateral damage detected”, Federal Reserve Bank of Chicago, Emerging
Issues Series, October, 1-14.
______, 2000c. “Depressing recoveries”, Risk, November, 108-111.
______, 2003. “A false sense of security”, Risk, August, 63-67.
Geske, R., 1977. “The valuation of corporate liabilities as compound options”, Journal of
Financial and Quantitative Analysis, 12 541-552.
Gordy, M., 2000. “A comparative anatomy of credit models”, Journal of Banking and Finance,
January, 119-149.
Gupton, G.M., Finger, C., Bhatia, M, 1997. Creditmetrics™ technical document. J.P. Morgan
& Co., New York.
Gupton, G.M., Gates, D. Carty, L.V., 2000. “Bank loan loss-given-default”, Moody’s Investor
Services, Global Credit Research, November.
Gupton, G.M., Stein, R.M., 2005. LossCalcTM V2: dynamic prediction of LGD. Moody’s KMV.

BIS Papers No 58 283

Härdle, W., Linton O.B. (1994). Applied nonparametric methods. in The Handbook of
Econometrics, vol. IV, eds. D.F. McFadden and R.F. Engle III. North Holland.
Haug, E.G., 2006. “Complete Guide to Option Pricing Formulas”, McGraw Hill, New York,
N.Y.
Hull, J.,White, A., 2002. “The impact of default risk on options and other derivative
securities”, Journal of Banking and Finance 19, 299-322.
Jacobs, Jr., M., Karagozoglu, A, 2007. “Understanding and predicting ultimate loss-given-
default on bonds and loans,” Working paper, U.S. Office of the Comptroller of the Currency
and Hofstra University.
Jacobs, Jr., M., Karagozoglu, A, 2011. “Modeling ultimate loss given default on corporate
debt,” The Journal of Fixed Income, 21:1 (Summer), 6-20.
Jarrow, R.A., 2001. “Default parameter estimation using market prices”, Financial Analysts
Journal 55:5, 75-92.
Jarrow, R.A., Turnbull, S.R., 1995. “Pricing derivatives on securities subject to credit risk”,
Journal of Finance 50, 53-86.
Jarrow, R.A., Lando, D., Turnbull, S.R.,1997. “A Markov model for the term structure of credit
spreads”, Review of Financial Studies 10, 481-523.
Jarrow, R.A., Jokivuolle, E., Peura, S., 2003. “Incorporating collateral value uncertainty in
loss-given-default estimates and loan-to-value ratios”, European Financial Management,
Vol. 9. 299-314.
Jones, E.S., Mason, E., Rosenfeld, E. , 1984. “Contingent claims analysis of corporate
structures: an empirical investigation”, Journal of Finance 39, 611-627.
Keisman, D., van de Castle, K., 2000. “Suddenly structure mattered: insights into recoveries
of defaulted debt”, Corporate Ratings, Commentary, Standard and Poors.
Kim, I.J., Ramaswamy, S., Sundaresan, S, 1993. “Does default risk in coupons affect
valuation of corporate bonds?: A contingent claims model”, Financial Management 22:3,
117-131.
Lando, D., 1998. “Cox processes and credit-risky securities”, Review of Derivative Research
2, 99–120.
Leland, H.E., 1994. “Corporate debt value, bond covenants, and optimal capital structure”,
Journal of Finance 49, 1213-1252.
Leland, H.E., Toft, K.B., 1996. “Optimal capital structure, endogenous bankruptcy and the
term structure of credit spreads”, Journal of Finance 51, 987-1019.
Litterman, R., T. Iben, 1991. “Corporate bond valuation and the term structure of credit
spreads”, The Journal of Portfolio Management, Spring, 52-64.
Longstaff, F.A., Schwartz, E.S., 1995. “A simple approach to valuing risk fixed and floating
rate debt”, Journal of Finance 89, 789-819.
Madan, D., Unal H., 1995. “Pricing the risks of default”, Review of Derivatives Research 2,
121-160.
Merton, Robert C., 1974. “On the pricing of corporate debt: The risk structure of interest
rates”, Journal of Finance, 29, 449–470.
Office of the Comptroller of the Currency / Department of the Treasury, Federal Reserve
System, Office of Thrift Supervision / Department of the Treasury; Risk-Based Capital
Standards: Advanced Capital Adequacy Framework— Basel II, Federal Register Vol. 72
No. 235, December 7, 2007.

284 BIS Papers No 58

Pagan A., Ullah, A.(1999), “Nonparametric Econometrics”, CambridgeUniversity Press,
Cambridge.
Pykhtin, M., 2003. “Unexpected recovery risk”, Risk, August, 74-78.
Renault, O., Scalliet, O., 2004. “On the way to recovery: a nonparametric bias free estimation
of recovery rate densities”, Journal of Banking & Finance 28, 2915–2931
Silverman, B. W., 1982. “Kernel density estimation using the fast Fourier transform”, Applied
Statistics, Royal Statistical Society, 33.
Vasicek, O.A., 1984. Credit valuation. KMV Corporation, March.
Wilson, T.C., 1998. “Portfolio Credit Risk”, Federal Reserve Bank of New York, Economic
Policy Review, October, 71-82.
Zhou, C., 2001. “The term structure of credit spreads with default risk,” Journal of Banking
and Finance 25, 2015-2040.

BIS Papers No 58 285