没有合适的资源?快使用搜索试试~ 我知道了~
首页credit risk optimization with conditional VaR criterion.pdf
credit risk optimization with conditional VaR criterion.pdf
需积分: 31 6 下载量 87 浏览量
更新于2023-03-03
收藏 99KB PDF 举报
This paper examines a new approach for credit risk optimization. The model is based on the Conditional Value-at-Risk (CVaR) risk measure, the expected loss exceeding Value-at-Risk. CVaR is also known as Mean Excess, Mean Shortfall, or Tail VaR.
资源详情
资源推荐
Digital Object Identifier (DOI) 10.1007/s101070000201
Math. Program., Ser. B 89: 273–291 (2001)
Fredrik Andersson · Helmut Mausser · Dan Rosen · Stanislav Uryasev
Credit risk optimization with Conditional Value-at-Risk
criterion
Received: November 1, 1999 / Accepted: October 1, 2000
Published online December 15, 2000 – Springer-Verlag 2000
Abstract. This paper examines a new approach for credit risk optimization. The model is based on the
Conditional Value-at-Risk (CVaR) risk measure, the expected loss exceeding Value-at-Risk. CVaR is also
known as Mean Excess, Mean Shortfall, or Tail VaR. This model can simultaneously adjust all positions
in a portfolio of financial instruments in order to minimize CVaR subject to trading and return constraints.
The credit risk distribution is generated by Monte Carlo simulations and the optimization problem is solved
effectively by linear programming. The algorithm is very efficient; it can handle hundreds of instruments
and thousands of scenarios in reasonable computer time. The approach is demonstrated with a portfolio of
emerging market bonds.
1. Introduction
Risk management is a core activity in asset allocation conducted by banks, insurance
and investment companies, or any financial institution that evaluates risks. This paper
examines a new approach for minimizing portfolio credit risk. Credit risk is the risk
of a trading partner not fulfilling their obligations in full on the due date or at any
time thereafter. Losses can result both from counterparty default, and from a decline
in market value stemming from the credit quality migration of an issuer or counter-
party. Traditionally used tools for assessing and optimizing market risk assume that the
portfolio return-loss is normally distributed. With this assumption, the two statistical
measures, mean and standard deviation, can be used to balance return and risk. The
optimal portfolio is selected on the “efficient frontier”, the set of portfolios that have the
best mean-variance profile [10]. In other words, this is the set of Pareto optimal points
with two conflicting criteria: mean and variance.
Although this traditional approach has proven to be quite useful in various applica-
tions, it is inadequate for credit risk evaluations because credit losses are characterized
by a large likelihood of small earnings, coupled with a small chance of losing a large
amount of the investment. Thus the loss distributions are, in general, heavily skewed
and standard optimization tools developedfor market risk are inadequate. This, together
F. Andersson: Ementor, Stortorget 1, 111 29 Stockholm, Sweden,
e-mail: fredrik.andersson@ementor.se, web: http://www.ementor.se
H. Mausser, D. Rosen: Algorithmics, Inc., 185 Spadina Avenue, Toronto, Ontario M5T 2C6, Canada,
web: http://www.algorithmics.com
S. Uryasev: University of Florida, Dept. of Industrial and Systems Engineering, PO Box 116595, 303 Weil
Hall, Gainesville, FL 32611-6595, e-mail: uryasev@ise.ufl.edu,
web: http://www.ise.ufl.edu/uryasev
Mathematics Subject Classification (1991): 20E28, 20G40, 20C20
274 Fredrik Andersson et al.
with the lack of historical data to estimate credit correlations, poses significant modeling
challenges, compared to market risk modeling and optimization.
To cope with skewed return-loss distributions, we consider Conditional Value-at-
Risk (CVaR) as the risk measure. This measure is also called Mean Excess Loss, Mean
Shortfall, or Tail VaR. By definition, β-CVaR is the expected loss exceeding β-Value-at-
Risk (VaR), i.e., it is the mean value of the worst (1 − β) ∗ 100% losses. For instance, at
β = 0.95, CVaR is the average of the 5% worst losses. CVaR is a currency-denominated
measure of significant undesirable changes in the value of the portfolio.
CVaR maybe compared with thewidely accepted VaR riskperformancemeasure for
which various estimation techniques have been proposed, see, e.g., [7,15]. VaR answers
the question: what is the maximum loss with the confidence level β ∗ 100% over a given
time horizon? Thus, its calculation reveals that the loss will exceed VaR with likelihood
(1 − β) ∗ 100%, but no information is provided on the amount of the excess loss, which
may be significantly greater. Mathematically, VaR has serious limitations. In the case of
a finite numberof scenarios, it is a nonsmooth, nonconvex,and multiextremum function
[11] (with respect to positions), making it difficult to control and optimize. Also, VaR
has some other undesirable properties, such as the lack of sub-additivity [1,2].
By contrast, CVaR is considered a more consistent measure of risk than VaR. CVaR
supplements the information provided by VaR and calculates the quantity of the excess
loss. Since CVaR is greater than or equal to VaR, portfolios with a low CVaR also
have a low VaR. Under quite general conditions, CVaR is a convex function with
respect to positions [17], allowing the construction of efficient optimization algorithms.
In particular, it has been shown in [17], that CVaR can be minimized using linear
programming (LP) techniques. The minimum CVaR approach [17] is based on a new
representation of the performance function that allows the simultaneous calculation of
VaR and minimization of CVaR. A simple description of the approach for minimization
of CVaR and optimization problems with CVaR constraints can be found in [21]. Since
CVaR can be minimized by LP algorithms, a large number of instruments and scenarios
can be handled. LP techniques are routinely used in financial planning applications, see,
for instance, paper [4] which applies a penalty approach to controlling risks. However,
comparing to the paper [4], we directly handle the quantile-based constraints with
a specified confidence level.
Similar measures as CVaR have been earlier introduced in the stochastic program-
ming literature, although not in financial mathematics context. The conditional expecta-
tion constraints and integrated chance constraints described in [14] may serve the same
purpose as CVaR. The reader interested in other applications of optimization techniques
in finance area can find relevant papers in [22].
In credit risk evaluations, we are interested in the losses experienced in the event of
counterparty default or credit quality migration in the course of a day, a year or other
standardized period. Several approaches are available for estimating credit risk [5,6,13,
18,19]. Probably, the most influential contribution in this field has been J.P. Morgan’s
CreditMetrics methodology [5].
Bucay and Rosen [3] conducted a case study and applied the CreditMetrics method-
ology to a portfolio of bonds issued in emerging markets. The portfolio consists of 197
bonds, issued by 86 obligors in 29 countries. Bond maturities range from a few months
to 98 years and the portfolio duration is approximately five years. The mark-to-market
Credit risk optimization with Conditional Value-at-Risk criterion 275
value of the portfolio is $8.8 billion. Mausser and Rosen [12] applied the regret op-
timization framework to minimize the credit risk of this portfolio. In this paper, we
analyzed the same portfolio and minimized the credit risk using the Minimum CVaR
approach. We have used the dataset of Monte Carlo scenarios generated at Algorithmics
Inc. [3,12]. Using the CreditMetrics methodology, a large number of scenarios is cal-
culated based on credit events such as defaults and credit migrations. By evaluating the
portfolio for each scenario, the loss distribution is generated. Monte Carlo simulation
tools are widely used for evaluating credit risk and other risks of portfolios containing
non-linear instruments, such as options (see, for instance, [11,17,15]).
The optimization analysis conducted for the portfolio of bonds may be briefly sum-
marized as follows. First, each single position (a position corresponds to an obligor)
is optimized. This provides the best hedge, i.e. the position that gives the minimum
CVaR when holding the other positions fixed. Then, all portfolio positions are simultan-
eously adjusted to minimize the portfolio CVaR under trading and budget constraints.
In this framework, the CVaR-return efficient frontier of the portfolios is also calculated.
As mentioned above, the minimization of CVaR automatically leads to a significant
improvement of VaR.
The remainder of this paper is organized as follows. First, we present the minimum
CVaR approach [17]. Then, following [3,12], we describe the bond portfolio. In this
section, we give a brief description of the CreditMetrics methodology used to calcu-
late the portfolio loss distribution. Then, we describe the optimization model and its
parameters. Finally, we present the analysis and concluding remarks.
2. Minimum CVaR approach
This section describes the approach to minimization of CVaR. Let f : IR
n
× IR
m
→ R
be the loss function which depends upon the control vector x ∈ IR
n
and the random
vector y ∈ IR
m
. We use bold face for the vectors to distinguish them from scalars. We
considerthat therandomvectory hasthe probabilitydistributionfunction p : IR
m
→ IR .
However, the existence of the density is not critical for the considered approach, this
assumption can be relaxed. Denote by (x, a) the probability function
(x,α) =
f(x,y)≤ α
p(y) dy , (1)
which, by definition, is the probability that the loss function f(x, y) does not exceed
some threshold value α. The VaR (or percentile) function α(x,β) is defined as follows
α(x,β) = min{α ∈ IR : (x,α) ≥ β} . (2)
Let us consider the following CVaR performancefunction (x) which is the conditional
expectedvalueof the loss f(x, y) underthe conditionthat it exceeds the quantile α(x,β),
i.e.,
(x) = (1 − β)
−1
f(x,y)≥ α(x,β)
f(x, y) p(y) dy . (3)
276 Fredrik Andersson et al.
The decision vector x belongs to the feasible set X ⊂ IR
n
. It was shown in [17] that
the minimization of the excess loss function (x) on the feasible set X ⊂ IR
n
can be
reduced to the minimization of the function
F
β
(x,α) = α + (1 − β)
−1
y∈IR
m
( f(x, y) − α)
+
p(y) dy , (4)
on the set X × IR , where b
+
is the positive part of the number b, i.e., b
+
= max{0, b}.
The following equality is valid
min
α∈IR
F
β
(x,α) = (x),
and the optimal solution α of this problem is VaR. This follows from the fact that the
derivative of the function F
β
(x,α)with respect to α equals 1+ (1− β)
−1
((x,α)− 1).
Equating this derivative to zero gives (see details in [17])
(x,α) = β.
Consequently,
min
x∈X,α∈IR
F
β
(x,α) = min
x∈X
min
α∈IR
F
β
(x,α) = min
x∈X
F
β
(x,α(x,β)) = min
x∈X
(x).
Thus, by minimizing the function F
β
(x,α) we can simultaneously find the VaR and
optimal CVaR. Under general conditions, the function F
β
(x,α) is smooth [20] (key
conditions: the density p(y) and the loss function f(x, y) are smooth and the gradient
of the function f(x, y) with respect to y is not equal zero).
Thefunction F
β
(x,α), givenbyequation (4),is convexinα (discussions of properties
of convex functions can be found, for instance, in [16,9]). F
β
(x,α)is convex in x, if the
function f(x, y) is convex in x. We can use various approaches to calculate the integral
function F
β
(x,α). If the integral in (4) can be calculated or approximated analytically,
then we can use nonlinear programming techniques to optimize the function F
β
(x,α).
In this paper, we approximate(4) using scenarios y
j
, j = 1,... ,J, which are sampled
from the density function p(y), i.e.,
y∈IR
m
(f(x, y) − α)
+
p(y) dy ≈ J
−1
J
j=1
(f(x, y
j
) − α )
+
.
If the loss function f(x, y
j
) is convex, and the feasible set X is convex, we can solve the
convex optimization problem
min
x∈X,α∈IR
˜
F
β
(x,α) , (5)
where
˜
F
β
(x,α)
def
= α + ν
J
j=1
(f(x, y
j
) − α )
+
剩余18页未读,继续阅读
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功