ASSESSING THE RISK-RETURN 2009

TRADE-OFF IN LOANS PORTFOLIOS

Javier Mencía

Documentos de Trabajo
N.º 0911

Electronic
Electroniccopy
copyavailable
availableat:
at:[Link]
[Link]
ASSESSING THE RISK-RETURN TRADE-OFF IN LOANS PORTFOLIOS

Electronic
Electroniccopy
copyavailable
availableat:
at:[Link]
[Link]
ASSESSING THE RISK-RETURN TRADE-OFF IN LOANS
PORTFOLIOS (*)

Javier Mencía (**)

BANCO DE ESPAÑA

(*)
A previous version of this paper has been circulated under the title “Assessing the risk, return and efficiency of
banks’ loans portfolios”. This paper is the sole responsibility of its author. The views represented here do not
necessarily reflect those of the Bank of Spain. I am grateful to Max Bruche, Enrique Sentana and seminar audiences
at the Bank of Austria, Bank of Spain and CEMFI for their comments and suggestions. Special thanks are due to
Alfredo Martín, for his valuable suggestions as well as for his help with the interest rate database. Of course, the
usual caveat applies. Address for correspondence: Alcalá 48, E-28014 Madrid, Spain, tel: +34 91 338 5414,
fax: +34 91 338 6102.
(**)
[Link]@[Link].

Documentos de Trabajo. N.º 0911

2009

Electronic
Electroniccopy
copyavailable
availableat:
at:[Link]
[Link]
The Working Paper Series seeks to disseminate original research in economics and finance. All papers
have been anonymously refereed. By publishing these papers, the Banco de España aims to contribute
to economic analysis and, in particular, to knowledge of the Spanish economy and its international
environment.

The opinions and analyses in the Working Paper Series are the responsibility of the authors and, therefore,
do not necessarily coincide with those of the Banco de España or the Eurosystem.

The Banco de España disseminates its main reports and most of its publications via the INTERNET at the
following website: [Link]

Reproduction for educational and non-commercial purposes is permitted provided that the source is
acknowledged.

© BANCO DE ESPAÑA, Madrid, 2009

ISSN: 0213-2710 (print)

ISSN: 1579-8666 (on line)
Depósito legal:
Unidad de Publicaciones, Banco de España

Electronic copy available at: [Link]

Abstract

This paper analyses the risk and return of loans portfolios in a joint setting. I develop a model
to obtain the distribution of loans returns. I use this model to describe the investment op-
portunity set of lenders using mean-variance analysis with a Value at Risk constraint. I also
obtain closed form expressions for the interest rates that banks should set in compensation
for borrowers’ credit risk under absence of arbitrage opportunities and I use these rates as
a benchmark to interpret actual loans’ prices. Finally, I study the risk-return trade-off in an
empirical application to the Spanish banking system.

Keywords: Credit risk, Probability of default, Asset Pricing, Mean-Variance allocation, Sto-
chastic Discount Factor, Value at Risk.

JEL classification: G21, G12, G11, C32, D81, G28.

Electronic copy available at: [Link]

1 Introduction
Standard capital market theory states that there is a risk-return trade-off in equilib-
rium. The more risk one is willing to take, the higher the return one will be able to
get. This relationship has been extensively analysed in the context of liquid assets that
trade in organised markets (see e.g. Fama and MacBeth, 1973; Ghysels, Santa-Clara, and
Valkanov, 2005). However, much less is known about its implications on the behaviour
of banks as risk managers and profit maximisers. Banks aggregate profits have been ana-
lysed by Behr et al. (2007) and Hayden, Porath, and von Westernhagen (2007), among
others, who find that more specialisation tends to yield higher returns but also a higher
level of risk. However, the optimal degree of bank specialisation is not analysed by these
papers. In addition, they aggregate the different activities of banks, that is, they do not
separate market and credit activities. As already explained, the features of liquid assets
are well known. Thus, my goal is to focus on the less well understood features of the lend-
ing business, which is the main role of banks as transformers of short term investments
(deposits) into long term ones (loans).
The risk and return of loans portfolios has generally been analysed separately. On
the one hand, the Basel II framework has originated the development of many quantitat-
ive models to estimate the loss distribution of loans portfolios (see Embrechts, Frey, and
McNeil, 2005, for a textbook review of the literature). On the other hand, a parallel liter-
ature that studies the determinants of interest rates has simultaneously grown during the
previous years (see e.g. Martı́n, Salas, and Saurina, 2007; Mueller, 2008). Unfortunately,
to the best of my knowledge, an specific common framework to analyse the risk and return
characteristics of bank’s loans portfolios is still missing in the literature, probably due to
the technical difficulties that the characteristics of loans entail. In particular, banks set
the interest rates of the loans that they grant. Hence, the prices of these products are
not the result of trading in an open market, which implies that the lenders may be price
setters rather than price takers. In addition, the returns that banks are finally able to
obtain may be smaller than the required rates in the presence of default risk, but their
exact value is not observable. Therefore, it is necessary to develop a model in order to
infer returns from the available information.

BANCO DE ESPAÑA 9 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 In this paper, I propose a flexible although analytically tractable model to derive
loans returns from interest rates and historical loan default rates. Since this information
is generally available for banks and supervisors, my model can be readily applied either
as a risk management or as an off-site supervisory tool. I consider a general structure in
which defaults are driven by stochastic probabilities of default, which may be correlated
with recovery rates. I make my model operational by expressing probabilities of default
as a probit function of an underlying multivariate Gaussian vector of state variables. I am
able to obtain closed form expressions for the expected returns, variances and covariances
between different loans. The covariance matrix of returns does not only depend on the
distribution of the probabilities of default, but also on the granularity of the portfolios. I
analyse risk and return jointly using the mean-variance analysis of Markowitz (1952) to
obtain the set of efficient portfolios. In this sense, the properties of the return distribution
of loans, and in particular the absence of probability mass at the right tail, make variance
a suitable measure of risk in this context. Thus, I can assume that banks would like to
minimize the variance of their loans portfolio for any given target expected return, from
which I can obtain the investment opportunity set on the mean-variance space. This
set may be restricted by the minimum capital requirements imposed by the regulator or
possibly by an even more stringent rating target. Both conditions can be interpreted
as a constraint on the minimum return that the bank must obtain, which technically
corresponds to a bound on the admissible Value at Risk (VaR). Sentana (2003) and
Alexander and Baptista (2006) have previously considered mean-variance analysis with a
VaR constraint when returns are elliptical. In this sense, I extend their approach into the
non-elliptical statistical framework of this paper.
I also analyse the risk-return trade-off from a pricing point of view. In this regard, my
goal is to study whether banks set interest rates taking into account the potential default
risk of their borrowers. To do so, I first derive theoretical closed form expressions for the
spreads over the risk-free rate that banks should require to ensure absence of arbitrage.
Although the market power exercised by banks may yield significant deviations from this
setting, this result provides a useful benchmark against which I can interpret the evolution
of actual spreads. I obtain the model-based rates using the exponentially affine stochastic
discount factor (SDF) proposed by Gourieroux and Monfort (2007). In this setting, I can

BANCO DE ESPAÑA 10 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

fully characterise the risk-neutral measure that is equivalent to this SDF.
I consider an empirical application to Spanish loans. I use quarterly data from the
Spanish credit register, from 1984.Q4 to 2008.Q4, to estimate a dynamic probit for the
probabilities of default and obtain the granularity of empirical portfolios. I consider an
additional database in which banks inform about the average interest rates for several
classes of loans, which is available from 1990.Q1. With this information, I compute the
mean-variance frontiers at different periods and analyse the historical evolution of the
expected values and standard deviations of loans returns. Finally, I study the deviations
of the actual mean-variance frontiers from those that would result in an arbitrage free
setting.
The rest of the paper is organised as follows. I introduce a general framework to model
the loan return distribution in Section 2. I discuss a simple feasible parametrisation of my
model in Section 3 and introduce mean-variance analysis in Section 4. Then, I develop
an arbitrage free model to price loans in Section 5. Section 6 presents the results of the
empirical application. Finally, Section 7 concludes. Proofs and auxiliary results can be
found in appendices.

2 General framework
Consider an economy with two periods: t = 0, 1. There is a risk-free asset, whose
return is r, and K different types of loans. These loans, which are risky because of the
presence of credit risk, may be interpreted as belonging to different economic sectors, or
just as a means of classifying borrowers with different characteristics (e.g.: corporates vs.
households). In each of these groups, there are Nk loans, for k = 1, · · · , K. I denote the
volume of loan i from group k as Lki , while rk will be the net interest rate required by
the lender at t = 1 for each loan type. Interest rates are set at t = 0.
I now turn to modelling borrowers’ credit risk. Borrowers may default at t = 1,
where default is driven by a binary variable Dki that takes a value of 1 if i defaults
and 0 otherwise. The probability of default of any loan from group k will be given by
the stochastic variable πk . In fact, πk is not even observable at t = 1 unless Nk grows to
infinity.1 In case of default, the lender will obtain a recovery rate δik , which is a proportion
1


Specifically, it can be shown that πk = limNk →∞ i Dik /Nk under standard regularity conditions. I

BANCO DE ESPAÑA 11 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 of the total amount that is owed. I will assume that δik , which is also a stochastic variable,
can be decomposed as δik = δk ψik . δk is the systematic component of the recovery rate,
since it is common to to all the loans of the k-th category and it may also be correlated
with other categories. In contrast, ψik will be an idiosyncratic term, modelled as an
iid positive variable with mean 1. Hence, this effect may alter the volatility of recovery
rates, but it does not change their expected values nor their correlations with other types
of loans. I will assume that defaults are conditionally independent given πk and δik for
k = 1, · · · , K.
Consider the portfolio of loans held by a particular bank at time 0. Its value will be
the cost of the initial investment, which is the sum of the outstanding debt at this period,
i.e.

K
p= pk ,
k=1

where
Nkt−1

pk = Ljk , (1)
j=1

denotes the outstanding debt of the k-th category of loans, for k = 1, · · · , K. One period
later, the pay-offs generated by each class of loans can be expressed as
Nk
 Nk

Zk = (1 + rk ) Ljk (1 − Djk ) + δjk Ljk Djk . (2)
j=1 j=1

Intuitively, each borrower will either repay the principal plus interests or default, in which
case the lender will only receive the recovery rate times the outstanding amount.
From (1) and (2), it is straightforward to write the return or yield generated by loans
from group k as:

Zk − p k
yk = ,
pk

Nk
j=1 (1 + rk − δjk )Ljk Djk
= rk − , (3)
Nk L̄k

where
Nk
1 
L̄k = Ljk .
Nk j=1

will exploit this feature in the empirical application.

BANCO DE ESPAÑA 12 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

Using the conditional independence property, it can be shown that the expected value of
(3) is given by
E(yk ) = rk − (1 + rk − E(δk ))E(πk ) + cov(πk , δk ). (4)

Hence, the expected return obtained by the lender may in practice differ from the required
interest rate rk because of the presence of credit risk. Specifically, this value will be smaller
than rk under a positive expected loss given default, which is realised with expected
probability E(πk ). This is a well known feature (see e.g. Feder, 1980). In addition, the
correlation between probabilities of default and recovery rates also affects the expected
return. For instance, a negative correlation would yield a further reduction in expected
returns, since recoveries would then be smaller in bad times, which is precisely when
probabilities of default are higher.

3 A simple implementation of the model

To make the framework of Section 2 operational, I need to make specific statistical
assumptions. In particular, I model probabilities of default in terms of the probits

πk = Φ(xk ), (5)

for k = 1, · · · , K, where Φ(·) is the standard normal cdf, and x = (x1 , x2 , · · · , xK )
is

a vector of Gaussian state variables. For the sake of concreteness, I will express this
distribution as
x ∼ N [µ(θ), Σ(θ)], (6)

where θ is a vector of p free parameters, and µ(θ) and Σ(θ) are, respectively, the mean
and covariance matrix of x. Interestingly, despite the flexibility of this framework, it is
still possible to derive general closed form expressions for the moments of (5):

Proposition 1 Consider a loan with probability of default (5). Then,

 
µk
E(πk ) = Φ √ , (7)
1 + σkk

E(πk πj ) = ωkj
 
µ µj σkj
= Φ2 √ k , , (8)
1 + σkk 1 + σjj (1 + σkk )(1 + σjj )

BANCO DE ESPAÑA 13 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

where µk is the k-th component of µ(θ), σkj is the element of Σ(θ) on row k and column
j, Φ(·) is the cdf of the standard normal distribution, and Φ2 (·, ·, ρ) denotes the cdf of a
bivariate normal distribution with zero means, unit variances and correlation ρ.

From Proposition 1 it is straightforward to show that the expected probability of

default at t = 0 is E(πk ). With this result in mind, I can express the conditional default
correlation as follows.

Proposition 2 Consider a loan with probability of default (5). Then

ωkj − E(πk )E(πj )

cor(Dki , Dji ) =  ,
[E(πk ) − E 2 (πk )][E(πj ) − E 2 (πj )]

where E(πk ), E(πj ) and ωkj are defined in (7) and (8).

4 Mean-Variance frontier with a VaR constraint

In this section, I develop a framework in which banks can decide the optimal combin-
ation of different types of loans in their lending policies, such that they maximise their
expected return for a given risk appetite. Hence, I need to consider an appropriate meas-
ure of risk to address this issue. In this sense, notice that, since δjk will always be smaller
than the gross interest rate, the distribution of (3) is bounded on its right tail by the
maximum yield that can be obtained, which is rk . Hence, risk is always undesirable in
this setting since it implies receiving a smaller return than the require rate. This is an
interesting difference with equity return distributions, in which positive extreme returns
may also occur. Nevertheless, the variance is still a popular measure of risk in applications
to stock returns, even though it is well known that a high variance may not necessarily be
undesirable if it is due to high unexpected returns (see e.g. Peñaranda, 2007). Such a con-
cern does not apply in this paper, because there is no right tail. Hence, I can adequately
assess the risk of the loans portfolios by means of their covariance matrix.
In consequence, mean-variance analysis is a natural framework to analyse efficiency
in this setting, since I can assess profitability with expected returns, and use variances
as a measure of risk. Let ω denote a vector containing the proportion of credit that a
particular bank allocates to each of the K different types of loans. This vector is chosen
at time 0. Hence, ω
y, where y = (y1 , y2 , · · · , yK )
, constitutes the return of the credit

BANCO DE ESPAÑA 14 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

portfolio of this bank at t = 1. The expected return and variance are given by ω
E(y)
and ω
V (y)ω, respectively.
I assume that banks would like to minimise the variances of their portfolios for every
expected return µ0 , subject to the restriction

ω
y > τ0 . (9)

This constraint ensures that banks will obtain a return higher than τ0 , which may be
necessary to comply with the minimum capital required by the regulator. For instance, if
(9) is violated, the bank may have to raise additional funds. Alternatively, if it is targeting
a high rating, it may need to satisfy an even more restrictive minimum capital requirement.
In general, though, it cannot be ensured that (9) will hold with probability 1. Hence, I
will set the minimum probability 1 − α with which (9) must hold, which is equivalent to
specifying a maximum admissible VaR. Notice that the higher this confidence level is the
more I will restrict the set of admissible portfolios.
In sum, this maximisation problem can be expressed in the following terms:

min ω
V (y)ω, (10)

ω

such that

ω
E(y) = µ0 , (11)

Pr(ω
y > τ0 ) ≥ 1 − α, (12)

ω
ι = 1, (13)

and
0 ≤ ωk ≤ 1, k = 1, · · · , K, (14)

where ι is a vector of K ones. I have introduced (13) and (14) to ensure that banks can
neither short sell nor leverage their investments. If I allow for the presence of a risk-free
asset, I do not need to impose (13). In this case, the mean-standard deviation frontier
would be the straight line that is tangent to the frontier without a risk-free asset and
intercepts the zero standard deviation axis at the risk-free rate return.
Following Sentana (2003), this problem can be decomposed in two simpler ones. First,
the region described by the VaR constraint (12) can be analysed separately by Monte

BANCO DE ESPAÑA 15 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Carlo simulation techniques (see Appendix C). And secondly, the mean-variance frontier
can be computed regardless of restriction (12), that is, by minimising (10) subject to
(11), (13) and (14). This frontier can be easily obtained by numerical optimisation, using
Proposition 1 to express the expected return in closed form, and the following Proposition
to obtain the analytical formula of the covariance matrix.

Proposition 3 Consider K portfolios of loans whose returns are given by (3) for k =
1, · · · , K. Then,
 
V (yk ) = (1/gk ) E (1 + rk − δk ψik )2 E(πk ) − (1 + rk − E(δk ))2 ωkk

+(1 + rk − E(δk ))2 [ωkk − E 2 (πk )], (15)

and

cov(yk , yj ) = (1 + rk − E(δk ))(1 + rj − E(δj ))

×[ωkj − E(πk )E(πj )], (16)

for k = j, where

Nk 2
i=1 Lik
gk =
Nk 2 . (17)
i=1 Lik

Therefore, both the variances and covariances depend on the moments of recovery
rates and probabilities of default. In the case of the variances, (15) is the sum of two
components. The first one is proportional to the reciprocal of (17), which can be inter-
preted as a granularity parameter. For any finite number of loans, (17) will always be
finite. However, if I let Nk grow to infinity while the size of the loans remains bounded,
i.e.
sup{Lik , 0 ≤ i ≤ Nk } ≤ Lk+ < ∞,

then it can be shown that gk will tend to infinity and the first term in (15) will disappear.
Therefore, gk increases as granularity grows to infinity. Thus, the first component in
(15) comprises a diversifiable risk that can be reduced by just increasing the number of
borrowers in the portfolio. This risk essentially captures what is usually known as “jump
to default” risk (see e.g. Collin-Dufresne et al., 2003).
In practice, banks might not be able to place their loans portfolios on the mean-
variance frontier, since their lending business is constrained by several restrictions. To

BANCO DE ESPAÑA 16 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

begin with, loans are illiquid assets that cannot be easily traded in the market. Hence,
it may be costly to rebalance these portfolios. In this sense, many banks have securitised
their credit portfolios until 2008 to be able to expand in the areas that they perceived as
more profitable. However, rebalancing has become much more costly after the subsequent
collapse of this market. In addition, the granularity of loans portfolios may also have a
strong impact on the degree of efficiency that banks can obtain. From this perspective,
there will be a “feasible” frontier, which takes the granularity of banks’ portfolios as
given, and an “infeasible” one that assumes that banks can actually get infinitely granular
portfolios.
For illustrative purposes, I now consider a simple example with two categories of loans,
such that (6) can be expressed as:
   √ √ 
x1 Φ−1 (0.05) √1.01 0.01
√ ρ 0.001
x= ∼N ; , (18)
x2 Φ−1 (0.025) 1.1 ρ 0.001 0.1

For the sake of brevity, I will refer to these portfolios as portfolios 1 and 2. It is not clear
a priory which of them is safer. On the one hand, the first type of assets have a expected
probability of default of 5%, while this value is only 2.5% for the second type of assets.
On the other hand, the behaviour of category 2 is more volatile, since the variance of x2
is ten times larger than the variance of x1 . Figure 1 shows the effect of granularity on the
standard deviations of portfolios 1 and 2. Both portfolios have high standard deviations
for small granularity, because the smaller the granularity, the higher the idiosyncratic
risk. However, portfolio 1 is riskier for small granularity due to its higher unconditional
probability of default. Introducing more loans in each portfolio can reduce their risk. This
reduction is smaller for portfolio 2, since the high standard deviation of x2 dominates over
the larger expected probability of default of 1 as the portfolios become more granular.
For infinite granularity, only the undiversifiable risk remains, which is again higher for x2 .
Figure 2 shows the mapping between the underlying correlation parameter ρ and
the correlation between the returns of portfolios 1 and 2. For very small granularity,
the correlation is approximately zero regardless of ρ, because most of the risk is purely
idiosyncratic in this case. In more granular portfolios, though, ρ becomes increasingly
informative about the return correlation. Eventually, it approximately corresponds to the
actual return correlation on infinitely granular portfolios (g → ∞).

BANCO DE ESPAÑA 17 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Following with this example, I show in Figure 3 the effect of different correlations
between portfolios 1 and 2 on the mean-standard deviation frontier. I assume infinite
granularity and interest rates r1 = 4.5% and r2 = 5%. Since there are only two types of
loans in this example, this figure simply shows the combination line between portfolios
1 and 2 for different values of ρ. This line is more convex when the correlation is small.
Thus, it is easier to obtain diversification gains by combining several portfolios when the
loans have small correlations. However, these benefits become increasingly smaller as the
correlation increases. For instance, if the correlation is close to 1, then any combination
of portfolios 1 and 2 will approximately lie on a straight line between these two assets.

5 Asset pricing implications

So far, I have treated the interest rates charged by lenders as exogenously given val-
ues. Nevertheless, in a competitive market banks should demand interest rates that
compensate for the risk of borrowers’ defaults to ensure absence of arbitrage opportunit-
ies. Of course, actual prices may substantially differ from those of this competitive pricing
scheme since this is an illiquid market in which some counterparties may exercise market
power. However, arbitrage free prices can provide a useful benchmark against which I can
compare actual data. To formalise this issue, I introduce an asset pricing framework by
modelling the stochastic discount factor (SDF). The SDF is a very useful instrument to
obtain prices under absence of arbitrage. For instance, consider an asset with payoff Z at
t = 1. Then, its price at t = 0 can be obtained from the pricing formula:

p = E(M Z). (19)

where M is the SDF. The SDF can be related to the marginal rate of substitution between
t = 0 and t = 1 in equilibrium models (see e.g. Gourieroux and Monfort, 2007). Pricing
assets with an SDF is equivalent to working under a different statistical measure Q in
which agents are risk-neutral. Under this new measure, (19) becomes

1
p= E Q (Z). (20)
1+r
Notice that, in the particular case in which the SDF is constant, the actual and the
risk-neutral measures will coincide.

BANCO DE ESPAÑA 18 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 If I use (20) to price the portfolio with payoffs (2), then I can express the spread that
banks should charge over the risk-free rate as

[1 + r − E Q (δk )]E Q (πk ) − cov Q (πk , δk )

rk − r = . (21)
1 − E Q (πk )

Hence, (21) is an increasing function of the risk-adjusted probability of default and a

decreasing function of the expected recovery rate under Q. In addition, all else equal,
spreads should be higher if the correlation between probabilities of default and recovery
rates becomes smaller. In this sense, Altman et al. (2005) find negative correlations
between recoveries and default rates in corporate bonds.
Since the market is incomplete in a discrete time model such as this one, there exists
a multiplicity of SDF’s that are compatible with the valuation formula (19). Hence, to
operationalise this result I assume a priory an exponentially affine form for the SDF,

M = ν0 exp(ν
1 x), (22)

where ν0 and ν 1 are, respectively, a positive scalar and a K × 1 vector. The exponential
structure ensures that the SDF is always positive, which in turn guarantees the absence of
arbitrage opportunities (see e.g. Cochrane, 2001, chapter 4). In addition, this specification
corresponds to the Esscher transform used in insurance (see Esscher, 1932), as well as in
derivative pricing models (Bertholon, Monfort, and Pegoraro, 2003; León, Mencı́a, and
Sentana, 2007). As Gourieroux and Monfort (2007) argue, the SDF of several important
equilibrium models, such as the consumption based CAPM, can be expressed with this
structure. I can also obtain the distribution of x under the risk-neutral measure from the
following result.

Proposition 4 Let the SDF be given by (22). Then, the distribution of x under the
risk-neutral measure Q is multivariate Gaussian with mean vector

µQ (θ) = µ(θ) + Σ(θ)ν 1 (23)

and covariance matrix Σ(θ).

Therefore, as in other financial applications (see e.g. Black and Scholes, 1973), I only
have to modify the mean of x to change the measure, whereas the covariance matrix

BANCO DE ESPAÑA 19 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 remains unaltered. Once I have done this change, I can price any asset by just computing
its discounted expected price. This device can be used to price any combination of the
original assets, as well as derivatives or collateralised debt obligations (CDO’s). To obtain
the required spread of a loan from the k-th category I only need to introduce the expected
probability of default under Q in (21). From Proposition 1,

µk + ν
1 σ •k
E Q (πk ) = Φ √ , (24)
1 + σkk
where σ •k is the k-th column of Σ(θ). In consequence, arbitrage-free spreads are the
result of a combination of data-based information and preference-based factors. To begin
with, they depend on variables from the actual probability measure, such as the recovery
rates or the parameters of the actual probability of default. All these parameters are
specific of each category of loans, except for σ •k , which captures the covariance of the
credit risk of loans from a given type k with the remaining groups. However, the impact
of σ •k on E Q (πk ) depends on ν 1 , which reflects the consensus of the market about how
to map systematic credit risk into prices. As already mentioned, these coefficients are
closely related to the utility preferences of the agents, and in particular to their marginal
rate of intertemporal substitution.

6 Empirical application
I consider an empirical application to the Spanish banking system. I distribute loans
in three categories: corporate, consumption loans and mortgages. Although the empir-
ical evidence suggests that loans from different economic sectors may not have the same
exposures to common shocks (see Jiménez and Mencı́a, 2009), I aggregate all corporate
loans into just one group due to the absence of data about discrepancies between interest
rates across economic sectors.
I use data from the Spanish credit register to estimate the distribution of the probab-
ilities of default. This database has information about every loan with a volume above
e6,000. Since this threshold is very small, I can safely assume that the data is represent-
ative of the whole banking system. There are two variables in the database that are of
particular interest for this paper.2 In particular, I will obtain the volumes of the existing
2
See Jiménez and Saurina (2004) and Jiménez, Salas, and Saurina (2006) for a thorough description

BANCO DE ESPAÑA 20 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 loans from the outstanding amounts that are reported in the credit register. In addition,
the database also has extremely useful information about the default situation of each
loan. It indicates which loans are overdue, and how long they have been in this situation.
Based on this information, I have computed historical quarterly series of the default fre-
quencies for each type of loans from 1984.Q4 to 2008.Q4. Default frequencies are defined
as the ratio of the number of loans that defaulted during a particular quarter over the
total number of loans. Following the definition given by the Basel II framework, I assume
that a loan is in default if it has been overdue for more than 90 days. I will use default
frequencies as proxies of πkt , which I model with a Gaussian distribution as in (6).3 In
what follows, I introduce a t subindex to take into account the time dimension. In order
to estimate the dynamics of πkt , I transform them with the inverse of the standard normal
cdf. Then, I consider the following vector autoregressive model for the transformed series:

xt = [Ik − diag(α)]−1 µ + diag(α)xt−1 + βft + diag(γ)1/2 εt , (25)

where µ, α, β and γ are k × 1 vectors, diag(α) yields a diagonal matrix whose diagonal
terms are given by α, ft is a latent factor such that

ft = ϕft−1 + 1 − ϕνt ,

while εt and νt are independent standard Gaussian vectors. Hence, correlations in this
model are driven by a common latent factor that may have time series autocorrelation.
In this way, I can account for two important stylised facts: persistence in default rates
and correlation between the rates of default from different categories. Therefore, this
model provides an equilibrium between flexibility in the time series and cross-sectional
correlation structures and parsimony in terms of the number of parameters employed.
Figure 4a shows the historical evolution of default frequencies. The three series display
a highly cyclical pattern, although corporate loans seem to be the category that was more
sensitive to the 1993 recession. Corporate loans are also the category with higher default
rates at the 2008 crisis. Table 1 shows the maximum likelihood parameter estimates of
of the database.
3
This estimation approach relies on the property that πkt can be recovered from default rates if Nkt is
sufficiently large (see footnote 1). This is a reasonable approximation in this case because I am considering
the whole loans portfolio of Spanish banks. Hence, I have about one million loans on average in each
group.

BANCO DE ESPAÑA 21 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 the model specified in (25). The high estimates of α show that there is strong time
series autocorrelation.4 In addition, the autocorrelation of ft , combined with the positive
and significant factor loadings in β, indicate a strong cross-sectional correlation in which
common shocks tend to have lasting effects on the evolution of risks.
Once I have estimated the model for the probabilities of default, I need to consider
the interest rates of loans in order to compute the moments of their return distribution.
Although the Spanish credit register does not include data about interest rates, I can
use the values reported by banks for corporate loans, consumption loans and mortgages.
Banks declared the average interest rate for each type of loans from 2003.Q1 to the
end of the sample. Before 2003.Q1, though, they only informed about average marginal
interest rates, that is, the interest rates of the new operations. Thus, to obtain a consistent
database for the whole sample period, I have estimated the age of each loan in the database
at each quarter prior to 2003.Q1. With this information, I have constructed series of
average interest rates by considering the appropriate marginal rates for each cohort of
loans, and weighting them by the proportion of the total credit in the economy that they
represent. This procedure yields a consistent time series of interest rates from 1990.Q1 to
2008.Q4, as shown in Figure 4b. This Figure also shows the evolution of treasury rates
throughout this period, which will be used as a proxy for the risk-free rate. Notice the
presence of important jumps in the treasury rates due to devaluations in the Spanish
peseta (see e.g. Gimeno and Marqués, 2008, for more detailed information).
As for recovery rates, I assume that they are constant. Specifically, I set them at 0.65,
0.75 and 0.85 for corporate loans, consumption loans and mortgages, respectively, which
are consistent with the results reported by Spanish banks for the QIS5.5 The assumption
of constant recovery rates can of course be relaxed, but it will be rather difficult calibrate
their distribution in practice due to the absence of time series data about this variable in
the case of loans.
Based on these estimates I consider the mean-variance frontier in Figure 5. Since
the granularity parameter g of the whole portfolio of Spanish loans is higher than 104
4
It could be argued that these coefficient might actually be 1. In this sense, although it would be
possible to estimate the model in first differences, I prefer using levels because the stationarity of the
series ensures the existence of unconditional probabilities of default.
5
Fifth Quantitative Impact Survey.

BANCO DE ESPAÑA 22 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 in all cases, the diversifiable risk due to lack of granularity will be negligible in this
application.6 I consider four panels for the frontiers at 2003.Q4, 2005.Q4, 2008.Q1 and
2008.Q4. Each panel also includes the line above which a non-negative excess return
over the treasury rate will be obtained with 99.9% probability. In general, consumption
loans yield the highest expected excess return, although at the cost of a slightly higher
standard deviation than mortgages. In contrast, corporate loans yield an even higher
standard deviation and a smaller excess return. It can also be observed that the mean-
standard deviation frontier has moved down (lower expected values) and to the right
(higher standard deviations) since 2003, although the most important change has occurred
within 2008. In this sense, the aggregate position of the loans portfolios was well above
the border of the VaR constraint in all periods except in last one, in which it is slightly
below. Hence, the probability that the Spanish loans portfolio would yield a non-negative
excess return in 2008.Q4 was smaller than 99.9%, albeit higher than 99.5% (also shown
in Figure 5d).
In Figure 6a, I consider the evolution over the sample period of the expected returns
and standard deviations of the aggregate portfolio of Spanish loans. Both moments have
decreased from the beginning of the 1990’s until 2005, but the value for 2008.Q4 shows
that the standard deviation has sharply risen at the end of the sample. Thus, high risk is
positively correlated with high gross expected returns. However, the risk-return trade-off,
if it exists, should affect excess returns rather than pure returns, because the risk-free
rate is mainly driven by monetary policy and not by the underlying risk of borrowers.
In this sense, I consider returns in excess of the treasury rate in Figure 6b. It can be
clearly observed that banks obtained high spreads on periods with very high standard
deviations, such as those close to the 1993 recession. Similarly, there were small expected
excess returns and standard deviations on the last decade of the sample. However, despite
this broad picture there is not such a clear pattern when I compare specific periods that
are not at the opposite extremes of the credit cycle.
In order to understand the deviations of actual interest rates from an arbitrage free
asset pricing model, I compare the actual mean-standard deviation frontiers with those
that would result from the pricing model of Section 5. To do so, I estimate the coefficients
6
Nevertheless, the granularity of the loans portfolios of individual banks may be much smaller, de-
pending on the size of the portfolio.

BANCO DE ESPAÑA 23 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 of the SDF (22) by minimising the sum of the squared differences between the actual
interest rates and those implied by (21). Hence, these estimates yield the arbitrage free
pricing model that is closest to the data. Figure 7 compares the actual and model-based
interest rates that I have obtained. These rates do reflect some of the main characteristics
of the data, such as higher interest rates and spreads in the first half of the 1990’s and
a rise in spreads from 2006 to the end of the sample. However, there are still important
differences. In particular, from around 1996 until 2006 actual rates have been higher that
the model-based ones. Finally, the model can obviously not reproduce the effect of the
currency devaluations that I have already mentioned.
Figure 8a shows the implied arbitrage-free mean-standard deviation frontiers at several
periods. I can clearly observe a risk-return trade-off in this case. The more risky the
investment opportunity set is on one particular period, the higher the expected returns
that are demanded. In fact, if I compute the envelope of the investment opportunity set
for all the quarters in the sample, I obtain a line with positive slope that is also shown on
the graph. The efficient part of all the frontiers is tangent to this line at some point, as
can be checked for the 6 frontiers that I have represented on the graph.7 I compare the
actual mean-standard deviation frontiers with this envelope on Figure 8b. Many of these
frontiers are to the left of the envelope. This implies that lenders have been able to obtain
higher expected returns than in the competitive setting, probably because of their market
power. However, when the risks of the loss distribution increased as quickly as in 2008,
they were not able to increase the spreads accordingly at the same speed, as illustrated
by the fact that the 2008.Q4 frontier is to the right of the envelope. In consequence, the
movements of the mean-variance frontier for small changes in borrowers’ risks do not seem
to follow risk-return trade-off considerations. Nevertheless, the position of the frontier on
1993.Q4 shows that banks are eventually able to raise spreads to compensate for increases
in credit risk when the situation becomes extremely uncertain.
7
I have only plotted 6 periods to avoid the cluttering of the pictures. The envelop summarises the
results for all periods but the frontiers are also available on request.

BANCO DE ESPAÑA 24 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

7 Conclusions
This paper provides an analytical model to compare the risk and return of loans port-
folios in a joint framework. I propose the use of mean-variance asset allocation techniques
to carry out this analysis, allowing for a Value at Risk constraint in order to take into
account regulatory requirements. The behaviour of the return yielded by loans portfolios
is described by means of a flexible albeit easily implementable model, which is driven by
a vector of Gaussian state variables. In this sense, I obtain closed form expressions for
the first two moments of the distribution of returns. I interpret these formulas and show
that the variance of returns can be decomposed in diversifiable and non-diversifiable risk,
where the diversifiable component can be eliminated by increasing portfolio granularity.
In addition, I illustrate how this model can capture default correlations between different
loans. These correlations are introduced by means of the state variables. In this sense,
the correlations of the underlying Gaussian distribution can be approximately interpreted
as the default correlations under infinite granularity.
I develop an asset pricing model as a benchmark against which I can compare the
actual data. Specifically, I obtain closed form expressions under absence of arbitrage,
in which spreads are increasing functions of the risk-adjusted probabilities of default and
decreasing functions of expected recovery rates. In addition, they depend negatively on the
covariance between recovery rates and default probabilities. If I assume an exponentially
affine stochastic discount factor, I can show that the distribution of the state variables
under the risk-neutral measure is another Gaussian distribution with the same covariance
matrix but different means.
Finally, I consider an empirical application in which I study the loans portfolio of the
Spanish banking system. Mortgages are the safest category of loans, although they yield
smaller returns than consumption loans. In contrast, corporate loans are generally riskier
and more sensitive to recessions. My model also reflects how the investment opportunity
set on the mean-variance space has quickly moved to riskier values on the 2008 crisis.
When I compare the actual mean-standard deviation frontiers with those based on the
closest arbitrage-free setting, I find that banks are able to move the mean-standard de-
viation frontier to regions of higher expected returns in good times. This may probably

BANCO DE ESPAÑA 25 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 be due to their market power. However, I can still observe a risk-return trade-off when I
compare periods with very different standard deviations in the loans portfolios, such as
the peak and the trough of the last credit cycle.
An interesting avenue for future research would be to explore the effect of macroeco-
nomic variables on the mean-variance frontier, extending the idea of Jiménez and Mencı́a
(2009) into an analysis of risk and return. It would also be interesting to consider the eco-
nomic sectors of corporate loans to take into account any potential heterogeneity within
this group. Lastly, it would be helpful to study the risk-return trade-off at multiperiod
horizons.

BANCO DE ESPAÑA 26 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

References
Alexander, G. J. and A. M. Baptista (2006). Does the Basle Capital Accord reduce
bank fragility? An assessment of the value-at-risk approach. Journal of Monetary
Economics 53, 1631–1660.

Altman, E., B. Brady, A. Resti, and A. Sironi (2005). The link between default and
recovery rates: theory, empirical evidence and implications. Journal of Business 78,
2203–2227.

Behr, A., A. Kamp, C. Memmel, and A. Pfingsten (2007). Diversification and the banks’
risk-return-characteristics - Evidence from loan portfolios of German banks. Deutsche
Bundesbank Discussion Paper No. 05/2007.

Bertholon, H., A. Monfort, and F. Pegoraro (2003). Pricing and inference with mixtures
of conditionally normal processes. Mimeo CREST.

Black, F. and M. Scholes (1973). The pricing of options and corporate liabilities. Journal
of Political Economy 81, 637–655.

Cochrane, J. H. (2001). Asset pricing. Princeton: Princeton University Press.

Collin-Dufresne, P., R. S. Goldstein, and J. Helwege (2003). Is credit event risk priced?
Modeling contagion via the updating of beliefs. mimeo, Carnegie Mellon University.

Embrechts, P., R. Frey, and A. J. McNeil (2005). Quantitative risk management: concepts,
techniques and tools. Princeton: Princeton University Press.

Esscher, F. (1932). On the probability function in the collective theory of risk. Skand-
inavisk Aktuariedskrift 15, 165–195.

Fama, E. F. and J. D. MacBeth (1973). Risk, return and equilibrium: empirical tests.
Journal of Political Economy 81, 607–636.

Feder, G. (1980). A note on debt, assets and lending. Journal of Financial and Quantit-
ative Analysis 15, 191–200.

BANCO DE ESPAÑA 27 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Ghysels, E., P. Santa-Clara, and R. Valkanov (2005). There is a risk-return trade-off after
all. Journal of Financial Economics 76, 509–548.

Gimeno, R. and J. M. Marqués (2008). Uncertainty and the price of risk in a nominal
convergence process. Banco de España Working Paper 0802.

Gourieroux, C. and A. Monfort (2007). Econometric specification of stochastic discount

factor models. Journal of Econometrics 136, 509–530.

Hayden, E., D. Porath, and N. von Westernhagen (2007). Does diversification improve the
performance of German banks? Evidence from individual bank loan portfolios. Journal
of Financial Services Research 32, 123–140.

Jiménez, G. and J. Mencı́a (2009). Modeling the distribution of credit losses with observ-
able and latent factors. Journal of Empirical Finance 16, 235–253.

Jiménez, G., V. Salas, and J. Saurina (2006). Determinants of collateral. Journal of

Financial Economics 81, 255–281.

Jiménez, G. and J. Saurina (2004). Collateral, type of lender and relationship banking as
determinants of credit risk. Journal of Banking and Finance 28, 2191–2212.

León, A., J. Mencı́a, and E. Sentana (2007). Parametric properties of semi-nonparametric

distributions, with applications to option valuation. Forthcoming Journal of Business
and Economic Statistics.

Markowitz, H. M. (1952). Portfolio selection. Journal of Finance 7, 77–91.

Martı́n, V. Salas, and J. Saurina (2007). A test of the law of one price in retail banking.
Journal of Money, Credit and Banking 39, 2021–2040.

Mueller, P. (2008). Credit spreads and real activity. mimeo, Columbia Business School.

Peñaranda, F. (2007). Portfolio choice beyond the traditional approach. Forthcoming

Revista de Economı́a Financiera.

Sentana, E. (2003). Mean-Variance portfolio allocation with a value at risk constraint.

Revista de Economı́a Financiera 1, 4–14.

BANCO DE ESPAÑA 28 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

A Auxiliary results
Proposition 5 Let Φ(·) be the cdf of the standard normal distribution. Then, if z ∼
N (0, 1),
a
E [Φ (a + bz)] = Φ √ . (A1)
1 + b2

Proposition 6 Let Φ(·) be the cdf of the standard normal distribution. Then, if f, z1 , z2
are independent standard normal variables,
 2   
 a1 a2 2
bi
E Φ (ai + bi f + ci zi ) = Φ2  2 2
, 2 2
; 
2 2
(A2)
i=1 1 + b 1 + c 1 1 + b 2 + c 2 i=1 1 + b i + c i

where Φ2 (p1 , p2 | ρ) is the cdf of a bivariate normal distribution with zero means, unit
variances and correlation ρ.

Proposition 7 Let Φ(·) be the cdf of the standard normal distribution. Then, if x ∼
N (µ, σ 2 ),
1 2 2 µ + aσ 2
E[exp(ax)Φ(x)] = exp a σ + µa Φ √ (A3)
2 1 + σ2

B Proofs of propositions
Proposition 1

Let us first consider (7). Since xt satisfies (6), I can always express πkt as

√
πkt = Φ(µkt + σkkt εkt ), (B4)

where εkt is a standard normal variable. Then, I can easily obtain (7) from Proposition
5.
As for (8), I can use the properties of the Gaussian distribution to express πjt as
⎡  ⎤
σkjt σkkt σjjt − σkjt
2
πjt = Φ ⎣µjt + √ εkt + εjt ⎦ , (B5)
σkkt σkkt

where εjt is independent of εkt . Hence, from Proposition 6, I can show that the expected
value of the product of (B4) and (B5) satisfies the required result.

BANCO DE ESPAÑA 29 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Proposition 2

As usual, I can express cor(Dkit , Djit |It−1 ) as

E(Dkit Djit |It−1 ) − E(Dkit |It−1 )E(Djit |It−1 )
cor(Dkit , Djit |It−1 ) = 
2
[E(Dkit |It−1 ) − E 2 (Dkit |It−1 )][E(Djit
2
|It−1 ) − E 2 (Djit |It−1 )]
E(Dkit Djit |It−1 ) − πkt|t−1 πjt|t−1
= 
(πkt|t−1 − πkt|t−1
2
)(πjt|t−1 − πjt|t−1
2
)

I can easily compute the remaining cross moment by exploiting the conditional independ-
ence Dkit and Djit given πkt and πjt :

E(Dkit Djit |It−1 ) = E[E(Dkit |πkt , πjt , It−1 )E(Djit |πkt , πjt , It−1 )|It−1 ],

= E[πkt πjt |It−1 ] = ωkjt|t−1 .

Proposition 3

By the law of iterated expectations, I have that

V [ykt |It−1 ; θ] = E [V [ykt |πkt , It−1 ; θ]|It−1 ; θ] + V [E[ykt |πkt , It−1 ; θ]|It−1 ; θ] , (B6)

where

E[ykt |πkt , It−1 ; θ] = rkt − (1 + rkt − δk0 )πkt ,

 
V [ykt |πkt , It−1 ; θ] = (1/gkt ) E (1 + rkt − δikt )2 πkt − (1 + rkt − δk0 )2 πkt
2
.

Hence, (15) follows directly from introducing the results of Proposition 1 in (B6). Sim-
ilarly, I can also exploit the law of iterated expectations to express cov[ykt , yjt |It−1 ; θ]
as

cov[ykt , yjt |It−1 ; θ] = E [cov[ykt , yjt |πkt , πjt , It−1 ; θ]|It−1 ; θ]

+cov [E[ykt |πkt , πjt , It−1 ; θ], E[yjt |πkt , πjt , It−1 ; θ]|It−1 ; θ] ,

where
cov[ykt , yjt |πkt , πjt , It−1 ; θ] = 0

due to the conditional independence property of the model, and

cov [E[ykt |πkt , πjt , It−1 ; θ], E[yjt |πkt , πjt , It−1 ; θ]|It−1 ; θ] = (1 + rkt − δk0 )(1 + rjt − δj0 )

×cov(πkt , πjt |It−1 ; θ),

which yields (16).

BANCO DE ESPAÑA 30 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

Proposition 4

The risk-neutral probability density function of xt can be expressed as

f Q (xt ) = (1 + rt )Mt−1,t f (xt ), (B7)

where f (xt ) is the density of xt under the actual measure:


f (xt ) = exp[−(1/2)(xt − µt (θ))
Σ−1
t (θ)(xt − µt (θ))]/ 2π|Σt (θ)|. (B8)

If I introduce (22) and (B8) in (B7), it is straightforward to show the required result when
I substitute ν0t for

exp [−ν
1t µt (θ) − (1/2)ν
1t Σt (θ)ν 1t ]

ν0t = ,
1 + rt
which ensures that (19) is satisfied for the risk-free asset.

Proposition 5

I can rewrite (A1) as

 ∞  ∞  a+bz 
Φ (a + bz) φ(z)dz = φ (s) ds φ(z)dz,
−∞ −∞ −∞

where φ(·) is the pdf of the standard normal distribution. By means of the change of
variable t = s − bz, I can obtain

 ∞  a+bz  ∞  a
φ (s) φ(x)dsdx = φ (t + bz) φ(z)dtdz.
−∞ −∞ −∞ −∞

It is straightforward to show that φ (t + bz) φ(z) is the pdf of the bivariate normal distri-
bution:    
z 0 1 −b
∼N ; .
t 0 −b 1 + b2
This proves the result.

Proposition 6

By the law of iterated expectations, I can write

 2   2 
 
E Φ (ai + bi f + ci zi ) = E E [ Φ (ai + bi f + ci zi )| f ] . (B9)
i=1 i=1

BANCO DE ESPAÑA 31 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Using Proposition 5, I can express (B9) as
 2   2  
  ai + bi f
E E [ Φ (ai + bi f + ci zi )| f ] = E Φ 
i=1 i=1 1 + c2i
 ∞ 2
  √
ai +bi f 
1+c2
i
= φ (si ) dsi φ(f )df (B10)
−∞ i=1 −∞


By means of the changes of variables ti = si 1 + c2i − bi f , for i = 1, 2, I rewrite (B10) as
 ∞  a1  a2    
t1 + b1 f t2 + b2 f
φ  φ  φ(f )dt1 dt2 df
−∞ −∞ −∞ 1 + c21 1 + c22

It can be shown that this is the integral of the pdf of the following trivariate distribution:
⎡ ⎤ ⎡⎛ ⎞ ⎛ ⎞⎤
f 0 1 −b1 −b2
⎣ t1 ⎦ ∼ N ⎣⎝ 0 ⎠ ; ⎝ −b1 1 + b21 + c21 b1 b2 ⎠⎦ .
t2 0 −b2 b1 b2 2 2
1 + b2 + c2

Proposition 7

I can rewrite (A3) as


exp[−(x − µ)2 /(2σ 2 )]
E[exp(ax)Φ(x)] = exp(ax)Φ(x) √ dx
2πσ 2

1 2 2 exp[−(x − (µ + aσ 2 ))2 /(2σ 2 )]
= exp a σ + µa Φ(x) √ dx.
2 2πσ 2
If I consider the change of variable y = (x − (µ + aσ 2 ))/σ, I can obtain

1 2 2
E[exp(ax)Φ(x)] = exp a σ + µa Φ(µ + aσ 2 + σy)φ(y)dy
2
1 2 2
= exp a σ + µa E[Φ(µ + aσ 2 + σy)].
2

Then, if I use Proposition 5 I can obtain the required result.

C Value at Risk constraint

For each possible variance σ02 , I estimate the points on the mean-variance space for
which (12) holds with equality by solving the following problem:

min ω
t E[yt |It−1 ; θ]

ωt

BANCO DE ESPAÑA 32 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

such that

ω
t V [yt |It−1 ; θ]ω t = σ02 ,

Pr[ω
t yt > τ0 |It−1 ; θ] ≥ 1 − α,

where Pr[ω
t yt > τ0 |It−1 ; θ] is estimated for each ω t by computing the proportion of times
that M replications of the DGP of ω
t yt given It−1 is above τ0 . I use M = 100000 in the
empirical application.

BANCO DE ESPAÑA 33 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Table 1
Model of the probabilities of default
Sample 1984.Q4-2008.Q4

Parameter Estimate S.E.

µ
Corporate -2.309 0.046∗∗
Consumption -2.070 0.035∗∗
Mortgages -3.050 0.051∗∗
α
Corporate 0.993 0.018∗∗
Consumption 0.997 0.007∗∗
Mortgages 0.990 0.012∗∗
β
Corporate 0.040 0.005∗∗
Consumption 0.029 0.004∗∗
Mortgages 0.027 0.006∗∗
γ
Corporate 0.001 0.000∗∗
Consumption 0.001 0.000∗∗
Mortgages 0.002 0.000∗∗
.
ϕ 0.464 0.116∗∗

Model:

xt = [I3 − diag(α)]µ + diag(α)xt−1 + βft + diag(γ)1/2 εt


ft = ϕft−1 + 1 − ϕ2 νt

Notes: Two asterisks indicate significance at the 5% level. xt are probit transformed default frequencies
for the whole Spanish banking system, where εt ∼ N (0, I3 ) and νt ∼ N (0, 1) are iid independent variables.

BANCO DE ESPAÑA 34 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 1: Standard deviations of two illustrative portfolios

1.2
Loans type 1
Loans type 2
1

0.8

0.6

0.4

0.2

0
0 1 2 3 4
10 10 10 10 10
Granularity Parameter

Notes: Example based on two types of loans whose defaults are conditionally independent given the evolution of
their respective probabilities of default π1t = Φ(x1t ) and π2t = Φ(x2t ), where
√ √
x1t Φ−1 (0.05) √ 1.01 √0.01 ρ 0.001
xt = ∼ N ; .
x2t Φ−1 (0.025) 1.1 ρ 0.001 0.1

BANCO DE ESPAÑA 35 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 2: Correlation between two illustrative portfolios

0.8 Granularity=1
Granularity=10
0.6 Granularity=100
Granularity=1000
0.4
Granularity=∞
0.2

−0.2

−0.4

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ρ

Notes: Example based on two types of loans whose defaults are conditionally independent given the evolution of
their respective probabilities of default π1t = Φ(x1t ) and π2t = Φ(x2t ), where
√ √
x1t Φ−1 (0.05) √ 1.01 √0.01 ρ 0.001
xt = ∼ N ; .
x2t Φ−1 (0.025) 1.1 ρ 0.001 0.1

BANCO DE ESPAÑA 36 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 3: Mean-standard deviation frontier for two illustrative portfolios

4.9
Portfolio 2
ρ=0
4.8 ρ=.4
ρ=.8
4.7

4.6
Mean

4.5

4.4

4.3

4.2
Portfolio 1

4.1
0.05 0.1 0.15
Standard Deviation

Notes: Infinite granularity. Two types of loans whose defaults are conditionally independent given the evolution of
their respective probabilities of default π1t = Φ(x1t ) and π2t = Φ(x2t ), where
√ √
x1t Φ−1 (0.05) √ 1.01 √0.01 ρ 0.001
xt = ∼ N ; .
x2t Φ−1 (0.025) 1.1 ρ 0.001 0.1
Interest rates are r1 = 4.5% and r2 = 5% for the first and second types of loans, respectively.

BANCO DE ESPAÑA 37 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 4a: Historical evolution of default frequencies (%)
3
Corporate
Consumption
2.5 Mortgages

1.5

0.5

0
Dec84 Sep87 Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06

Figure 4b: Historical evolution of interest rates (%)

18
Corporate
16 Consumption
Mortgages
Treasury
14

12

10

2
Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06

BANCO DE ESPAÑA 38 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 5: Mean-standard deviation frontier with a Value at Risk constraint

(a) 2003.Q4 (b) 2005.Q4

cons
2.5 Frontier 2.5 Frontier

BANCO DE ESPAÑA
′

39
Pr(w yt−r>0)=99.9% Pr(w′ yt−r>0)=99.9%
2 2
cons

1.5 1.5
mort
1 1

µ0−r
µ0−r
0.5 0.5 mort
corp

DOCUMENTO DE TRABAJO N.º 0911

corp
0 0

−0.5 −0.5

−1 −1
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3
σ0 σ0

(c) 2008.Q1 (d) 2008.Q4

2.5 Frontier 2.5 Frontier

′
Pr(w yt−r>0)=99.9% Pr(w′ yt−r>0)=99.9%
2 2
cons Pr(w′ yt−r>0)=99.5%
1.5 1.5 cons

mort
1 mort 1

µ0−r
µ0−r

0.5 0.5
corp

Electronic copy available at: [Link]

0 0

−0.5 −0.5

−1 −1 corp
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3
σ0 σ0
Notes: Excess returns with respect to the Spanish treasury rates. A star denotes the situation of the Spanish aggregate loans portfolio at the four periods.
“corp”, “cons” and “mort” denote corporate loans, consumption loans and mortgages, respectively. 100,000 simulations have been employed to estimate
the the border of the area that yields positive returns with 99.9% probability.
 Figure 6a: Risk vs. expected returns (%)
16
199012
199112
14
199212
199312

12
199412
199512

10
199612
µ0

8 199712

199812
199912
200012
6 200112
200212
200312 200712 200812
4 200612
200412
200512
0 0.05 0.1 0.15 0.2 0.25 0.3
σ0
Figure 6b: Risk vs. expected excess returns (%)
3

199112
199812
2.5 199712 199312

199912 199412

2
199212
µ0−r

1.5 199612
200112
200212

200312
1 200012

200512 199012
200712
0.5 200412
200812
200612
199512
0
0 0.05 0.1 0.15 0.2 0.25 0.3
σ0
Note: for the sake of clarity, only the fourth quarter of each year is plotted in the figures.

BANCO DE ESPAÑA 40 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 7: Historical interest rates and fit provided by the arbitrage-free pricing model
(a) Corporate. Interest rates (b) Corporate. Spread over treasury rates
20 8
Data Data
18 Arbitrage−free closest fit 7 Arbitrage−free closest fit
16
6
14
5
12
4
10
3
8

6 2

4 1

2 0
Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06 Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06

(c) Consumption. Interest rates (d) Consumption. Spread over treasury rates
20 8
Data Data
18 Arbitrage−free closest fit 7 Arbitrage−free closest fit
16
6
14
5
12
4
10
3
8

6 2

4 1

2 0
Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06 Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06

(e) Mortgages. Interest rates (f) Mortgages. Spread over treasury rates
20 8
Data Data
18 Arbitrage−free closest fit 7 Arbitrage−free closest fit
16
6
14
5
12
4
10
3
8

6 2

4 1

2 0
Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06 Jun90 Mar93 Dec95 Sep98 May01 Feb04 Nov06

Notes: The arbitrage-free rates have been obtained from an exponentially affine stochastic discount factor whose
parameters minimise the sum of square errors between the model based rates and the actual ones.

BANCO DE ESPAÑA 41 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

 Figure 8a: Mean-standard deviation frontiers implied by the pricing model
5
1993.Q4
4.5
1997.Q4
4 2003.Q4
2005.Q4
3.5 2008.Q1
2008.Q4
3
µ0−r

2.5

1.5

0.5

0
0 0.05 0.1 0.15
σ0

Figure 8b: Actual mean-standard deviation frontiers

5
1993.Q4
4.5 1997.Q4
2003.Q4
4
2005.Q4
3.5 2008.Q1
2008.Q4
3
µ0−r

2.5

1.5

0.5

0
0 0.05 0.1 0.15
σ0

Notes: The solid line indicates the envelope of all the quarterly investment opportunity sets in mean-standard
deviation space implied by the pricing model.

BANCO DE ESPAÑA 42 DOCUMENTO DE TRABAJO N.º 0911

Electronic copy available at: [Link]

BANCO DE ESPAÑA PUBLICATIONS

WORKING PAPERS1

0801 ENRIQUE BENITO: Size, growth and bank dynamics.

0802 RICARDO GIMENO AND JOSÉ MANUEL MARQUÉS: Uncertainty and the price of risk in a nominal convergence
process.
0803 ISABEL ARGIMÓN AND PABLO HERNÁNDEZ DE COS: Los determinantes de los saldos presupuestarios de las
Comunidades Autónomas.
0804 OLYMPIA BOVER: Wealth inequality and household structure: US vs. Spain.
0805 JAVIER ANDRÉS, J. DAVID LÓPEZ-SALIDO AND EDWARD NELSON: Money and the natural rate of interest:
structural estimates for the United States and the euro area.
0806 CARLOS THOMAS: Search frictions, real rigidities and inflation dynamics.
0807 MAXIMO CAMACHO AND GABRIEL PEREZ-QUIROS: Introducing the EURO-STING: Short Term INdicator of Euro
Area Growth.
0808 RUBÉN SEGURA-CAYUELA AND JOSEP M. VILARRUBIA: The effect of foreign service on trade volumes and
trade partners.
0809 AITOR ERCE: A structural model of sovereign debt issuance: assessing the role of financial factors.
0810 ALICIA GARCÍA-HERRERO AND JUAN M. RUIZ: Do trade and financial linkages foster business cycle
synchronization in a small economy?
0811 RUBÉN SEGURA-CAYUELA AND JOSEP M. VILARRUBIA: Uncertainty and entry into export markets.
0812 CARMEN BROTO AND ESTHER RUIZ: Testing for conditional heteroscedasticity in the components of inflation.
0813 JUAN J. DOLADO, MARCEL JANSEN AND JUAN F. JIMENO: On the job search in a model with heterogeneous
jobs and workers.
0814 SAMUEL BENTOLILA, JUAN J. DOLADO AND JUAN F. JIMENO: Does immigration affect the Phillips curve? Some
evidence for Spain.
0815 ÓSCAR J. ARCE AND J. DAVID LÓPEZ-SALIDO: Housing bubbles.
0816 GABRIEL JIMÉNEZ, VICENTE SALAS-FUMÁS AND JESÚS SAURINA: Organizational distance and use of collateral
for business loans.
0817 CARMEN BROTO, JAVIER DÍAZ-CASSOU AND AITOR ERCE-DOMÍNGUEZ: Measuring and explaining the volatility
of capital flows towards emerging countries.
0818 CARLOS THOMAS AND FRANCESCO ZANETTI: Labor market reform and price stability: an application to the Euro
Area.
0819 DAVID G. MAYES, MARÍA J. NIETO AND LARRY WALL: Multiple safety net regulators and agency problems in the
EU: Is Prompt Corrective Action partly the solution?
0820 CARMEN MARTÍNEZ-CARRASCAL AND ANNALISA FERRANDO: The impact of financial position on investment:
an analysis for non-financial corporations in the euro area.
0821 GABRIEL JIMÉNEZ, JOSÉ A. LÓPEZ AND JESÚS SAURINA: Empirical analysis of corporate credit lines.
0822 RAMÓN MARÍA-DOLORES: Exchange rate pass through in New Member States and candidate countries of the EU.
0822 RAMÓN MARÍA-DOLORES: Exchange rate pass-through in new Member States and candidate countries of the EU.
0823 IGNACIO HERNANDO, MARÍA J. NIETO AND LARRY D. WALL: Determinants of domestic and cross-border bank
acquisitions in the European Union.
0824 JAMES COSTAIN AND ANTÓN NÁKOV: Price adjustments in a general model of state-dependent pricing.
0825 ALFREDO MARTÍN-OLIVER, VICENTE SALAS-FUMÁS AND JESÚS SAURINA: Search cost and price dispersion in
vertically related markets: the case of bank loans and deposits.
0826 CARMEN BROTO: Inflation targeting in Latin America: Empirical analysis using GARCH models.
0827 RAMÓN MARÍA-DOLORES AND JESÚS VAZQUEZ: Term structure and the estimated monetary policy rule in the
eurozone.
0828 MICHIEL VAN LEUVENSTEIJN, CHRISTOFFER KOK SØRENSEN, JACOB A. BIKKER AND ADRIAN VAN RIXTEL:
Impact of bank competition on the interest rate pass-through in the euro area.
0829 CRISTINA BARCELÓ: The impact of alternative imputation methods on the measurement of income and wealth:
Evidence from the Spanish survey of household finances.
0830 JAVIER ANDRÉS AND ÓSCAR ARCE: Banking competition, housing prices and macroeconomic stability.
0831 JAMES COSTAIN AND ANTÓN NÁKOV: Dynamics of the price distribution in a general model of state-dependent
pricing.
0832 JUAN A. ROJAS: Social Security reform with imperfect substitution between less and more experienced workers.
0833 GABRIEL JIMÉNEZ, STEVEN ONGENA, JOSÉ LUIS PEYDRÓ AND JESÚS SAURINA: Hazardous times for monetary
policy: What do twenty-three million bank loans say about the effects of monetary policy on credit risk-taking?
0834 ENRIQUE ALBEROLA AND JOSÉ MARÍA SERENA: Sovereign external assets and the resilience of global
imbalances.
0835 AITOR LACUESTA, SERGIO PUENTE AND PILAR CUADRADO: Omitted variables in the measure of a labour
quality index: the case of Spain.
0836 CHIARA COLUZZI, ANNALISA FERRANDO AND CARMEN MARTÍNEZ-CARRASCAL: Financing obstacles and
growth: An analysis for euro area non-financial corporations.

1. Previously published Working Papers are listed in the Banco de España publications catalogue.

Electronic copy available at: [Link]

0837 ÓSCAR ARCE, JOSÉ MANUEL CAMPA AND ÁNGEL GAVILÁN: Asymmetric collateral requirements and output composition.
0838 ÁNGEL GAVILÁN AND JUAN A. ROJAS: Solving Portfolio Problems with the Smolyak-Parameterized Expectations Algorithm.
0901 PRAVEEN KUJAL AND JUAN RUIZ: International trade policy towards monopoly and oligopoly.
0902 CATIA BATISTA, AITOR LACUESTA AND PEDRO VICENTE: Micro evidence of the brain gain hypothesis: The case of Cape Verde.
0903 MARGARITA RUBIO: Fixed and variable-rate mortgages, business cycles and monetary policy.
0904 MARIO IZQUIERDO, AITOR LACUESTA AND RAQUEL VEGAS: Assimilation of immigrants in Spain: A longitudinal analysis.
0905 ÁNGEL ESTRADA: The mark-ups in the Spanish economy: international comparison and recent evolution.
0906 RICARDO GIMENO AND JOSÉ MANUEL MARQUÉS: Extraction of financial market expectations about inflation and interest rates
from a liquid market.
0907 LAURA HOSPIDO: Job changes and individual-job specific wage dynamics.
0908 M.a DE LOS LLANOS MATEA AND JUAN S. MORA: La evolución de la regulación del comercio minorista en España y sus
implicaciones macroeconómicas.
0909 JAVIER MENCÍA AND ENRIQUE SENTANA: Multivariate location-scale mixtures of normals and mean-variance-skewness portfolio
allocation.
0910 ALICIA GARCÍA-HERRERO, SERGIO GAVILÁ AND DANIEL SANTABÁRBARA: What explains the low profitability of Chinese banks?
0911 JAVIER MENCÍA: Assessing the risk-return trade-off in loans portfolios.

Unidad de Publicaciones
Alcalá, 522; 28027 Madrid
Telephone +34 91 338 6363. Fax +34 91 338 6488
e-mail: publicaciones@[Link]
[Link]

Electronic copy available at: [Link]