Professional Documents
Culture Documents
a r t i c l e
i n f o
Article history:
Received 10 August 2007
Accepted 1 December 2009
Available online 6 December 2009
Keywords:
OR in government
Multiple objective programming
Risk analysis
Stochastic programming
Public debt management
a b s t r a c t
While raising debt on behalf of the government, public debt managers need to consider several possibly
conicting objectives and have to nd an appropriate combination for government debt taking into
account the uncertainty with regard to the future state of the economy. In this paper, we explicitly consider the underlying uncertainties with a complex multi-period stochastic programming model that captures the trade-offs between the objectives. The model is designed to aid the decision makers in
formulating the debt issuance strategy. We apply an interactive procedure that guides the issuer to identify good strategies and demonstrate this approach for the public debt management problem of Turkey.
2009 Elsevier B.V. All rights reserved.
1. Introduction
Public debt management (PDM) is concerned with meeting the funding requirements of a country that arise from budgetary and other
nancial liabilities of the government. More specically, it can be dened as the process of establishing and executing a strategy for managing the governments debt to raise the required amount of funding, pursue its cost/risk objectives, and meet any other public debt management goals the government may have set, . . . (International Monetary Fund The World Bank, 2003, p. 5). Public debt managers have a
range of nancial instruments, securities, at their disposal and have to form a specic portfolio, in terms of maturity, currency and interest
types, that would suit the governments objectives. Given the exposure of public sector balances and the countrys nancial stability to
public debt, the overall structure of public debt portfolio is fundamentally important for a countrys macroeconomic stability. Once a government is in a nancial problem, i.e. facing difculties in fullling scal liabilities or having to pay excessive costs when issuing debt, this
has spill-over effects on the entire economy, as demonstrated by a number of recent macroeconomic crises in several emerging countries.
Therefore, the nancial liability portfolio of the government must be effectively managed.
Given its importance, the problem of designing the public debt management strategy, in terms of setting the composition of nancing,
draws attention of both practitioners and academicians from various perspectives. Alesina et al. (1990) elaborate on the choice of maturity
of public debt and argue that issuing debt at long maturities that is evenly concentrated in time will boost public condence and reduce
perceived likelihood of condence crisis about debt default. Missale and Blanchard (1994) claim that government can use the maturity of
debt to show her commitment to anti-inationary policies and thus should prefer short-maturity or indexed debt. The tax smoothing approach assumes that the main reason for the government to change taxes is to meet the long-term nancing constraint, and the objective is
to smooth taxes by choosing the optimal composition of debt with respect to maturity and contingencies. There is uncertainty about macroeconomic variables such as public expenditures, tax base, etc. and therefore, the composition of debt matters (Barro, 1995, 2003). The
argument is that debt can serve as a buffer against tax rate changes if the government can issue debt with costs that are lower when
net tax receipts are lower. A good review of theoretical and practical concepts regarding public debt management can be found in Dornbush and Draghi (1990) and Leong (1999).
Debt management authorities take a practical point of view and apply concepts and tools derived from those employed by private nancial institutions. The simulation models of PDM ofces are generally derived from the Value at Risk (VaR) concept widely used by banks
1
The ideas expressed in this study are only those of the authors and do not necessarily reect the views and policies of the Turkish Treasury.
* Corresponding author.
E-mail addresses: emre.balibek@hazine.gov.tr (E. Balibek), koksalan@ie.metu.edu.tr (M. Kksalan).
0377-2217/$ - see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2009.12.001
206
and other nancial rms. For sovereigns, this approach is modied into Cost-at-Risk or Cash-Flow-at-Risk simulation models by which
the cost and risk performances of alternative debt management strategies are tested under various macroeconomic simulation scenarios
(see Danmarks Nationalbank (2005) and Bergstrm et al. (2002) for two country examples). Bolder (2003) explains the simulation model
for debt strategy analysis in Canada. More recently, Bolder and Rubin (2007) try to combine simulation and optimization approaches in
debt strategy analysis. Countries also apply other methods like stress testing or scenario analysis to compare different PDM strategies
(see International Monetary Fund The World Bank (2003) and OECD (2005) for discussions on debt management practices of selected
countries).
The general aim in these applications is to quantify costs and risks associated with policy choices. On the other hand, there is a need for
providing support to decision makers (DM) to nd efcient solutions. To the best of our knowledge, this is the rst study that explicitly
considers multiple objectives and guides the DM towards desirable strategies. In this paper, we model the PDM problem as a multi-period
stochastic program with multiple objectives. We formulate the problem as a deterministic equivalent linear programming model, in which
the decision variables are the amounts of different types of bonds to be issued, accounting for the cash-ow constraints for the government.
We develop our approach as a decision aid tool in analyzing the trade-offs between alternative courses of action. In this context, we identify
efcient solutions based on different preferences and apply an interactive Multi-Criteria Decision Making (MCDM) approach to guide the
DMs in developing the debt strategy. We demonstrate how sovereign DMs can experiment with such a tool in a practical setting, drawing
on the case of Turkey.
The following section denes the public debt management problem and discusses its main features. We then present our multi-stage
stochastic programming (SP) model, developed to guide issuance decisions. Section 4 discusses how we employ the SP model to obtain
efcient solutions and use an interactive algorithm by which the DMs can experiment to explore alternatives. The proposed methods
are illustrated in Section 5.
207
Scenario Paths
Set Strategy
for Year 1
quarter
decision stage 0
10
11
12
billion USD payable in the rst 3 months, while the countrys foreign reserves stood at a level of 6.3 billion USD (Cassard and Folkerts-Landau, 1997).
Liquidity risk or re-funding/re-nancing risk as it is sometimes called is also a critical concern for public debt managers, especially in
developing countries. This type of risk may arise from the level of cash reserves of the government, for example due to a decline in tax
revenues or from the lenders reluctance in renewing their loans. However, in many cases, the level of funds from net revenues available
to governments to service debt is very low when compared to public debt redemptions. On the other hand, market conditions that affect
investor demand are not controllable by debt managers. Therefore, the guiding principle for countries is to spread maturities across the
maturity spectrum to avoid bunching of maturity payments (OECD, 2005, p. 34). In this regard, many countries target total redemptions
over the near term and try to extend maturities to control liquidity risk which is crucial for a governments reputation. Lenders, other government institutions, public employees and in the end all the parties in the economy will be affected from the governments liquidity crises.
In many cases, there can also be a trade-off between the cost and liquidity risk objectives, since reducing the liquidity risk may require
long-term borrowing at high costs and/or keeping a certain level of excess cash reserve which also induce a cost for the government. Aiming to minimize market risk may dictate to borrow xed-rate long-term bonds whose repayments accumulate at a certain point in time
which in turn induces a certain level of liquidity risk.
While formulating the issuance strategies the authorities also need to consider several constraints. To begin with, the amount of funds
raised should not be less than those required by the budget. Governments generally hold a cash account which serves as a buffer to cover
unexpected cash needs and this allows borrowing more or less than needed for a certain period of time. However, there are also limitations
to the levels of this account, i.e. governments cannot over or under borrow continuously. Thus, the inter-temporal budgetary and cash account constraints must be satised.
The size and efciency of a countrys nancial markets, the governments ability to access international markets and other macroeconomic environmental conditions impose several limitations on the type of securities the debt managers can issue. For under-developed or
developing countries, where the level of domestic savings and efciency of internal nancial markets are limited, the main option is to opt
for funds from international markets. Some developing countries have a functioning domestic nancial market, but they also need to issue
debt in foreign currency to lengthen maturity when domestic lenders prefer shorter maturities. Given a certain instrument set, public debt
managers should also consider market constraints with regard to the availability of funds and the demand for different types of securities.
In a volatile environment, creditors may not be willing to extend long-term loans, and the governments insistence on lengthening maturities may result in a funding-crisis. Institutional investors, individuals, banks might have different preferences which may impose several
constraints on the size of bonds to be offered to different market segments.
The debt strategy is not a one-off decision. The PDM problem embodies a sequence of decisions that would allow the governments debt
portfolio to adjust to changing environmental conditions. A decision made now for the portfolio structure is subject to revision in the future
depending on changing outlooks for the macro-economy. Therefore, debt managers should incorporate this exibility in their decision
making processes. They need to consider the effects of the potential for adjusting decisions in the future, since future decisions will be contingent on the previous actions and prevailing market conditions.
3. A multi-objective stochastic programming approach
In recent years, (SP) models have been increasingly used to address real life multi-period asset and liability management problems. A
seminal contribution was made by Bradley and Crane (1972) who proposed a multi-stage model for bond portfolio management. More
recently, Carino et al. (1994) applied SP to the assetliability management problem of the insurance industry, and Zenios et al. (1998)
and Topaloglou et al. (2008) formulated models for a portfolio of xed income securities. Nielsen and Poulsen (2004) proposed a multistage SP model for managing mortgage backed loans. Volosov et al. (2004) developed a two-stage decision model for foreign exchange
exposure management. Grill and stberg (2003) have applied an optimization approach for debt management. Yu et al. (2003) provide
a bibliography of SP models in nancial optimization. Extensive collections of SP models for nancial problems can also be found in Ziemba
and Mulvey (1998) and Dupacova et al. (2002).
The general approach of multi-stage SP models in representing uncertainty is forming a scenario tree that reects the evolution of random variables in each stage of the decision horizon, by discretizing their joint probability distributions. The simplest approach is to use
208
historical data regarding random variables without any modelling and claim that future will replicate the past, i.e. sampling from different
points in time for generating scenarios. Another method is trying to create scenarios that replicate the empirical characteristics of random
variables such as the moment matching method of Hoyland and Wallace (2001). A more sophisticated method requires econometric modelling such as Boender (1997) and Villaverde (2003) who use vector autoregressive (VAR) time series models for scenario generation.
We formulate the PDM problem as a linear deterministic equivalent model based on a scenario tree representation of macroeconomic
factors that affect the cost of debt. The model has a multi-stage structure that takes into account sequential decisions concerned with debt
issuance policies. We assume that at the beginning of each year the government sets a borrowing strategy, which embodies the amount of
bonds to be offered in each month (or quarter) of the following year and revises this strategy annually. We start with a given liability cashow scheme (arising from the current debt portfolio) and a set of anticipated scenarios about future states of relevant macroeconomic variables such as the interest and exchange rates. The objective is to specify a sequence of bond issuance decisions at discrete points in time.
Since the funding requirements might be different in each quarter, total funds raised through issuance of bonds might also be different.
We present a general n-stage model in which each period is divided into several sub-periods, t. (If the periods correspond to years, sub-periods can be months or quarters.) Decisions are made at the start of each period for the following sub-periods, i.e. issuance decisions are not revised in each sub-period, but only at decision stages. The scenarios between decision stages combine to form a sequence of joint realizations for
a certain period. These sequences of scenarios are linked at the decision nodes and we have scenario paths covering the entire planning horizon.
The issuance strategy for the sub-periods of the rst year is not scenario specic and is set to be implemented across all scenarios. At the
end of the rst year we have a new liability portfolio and we now have to make a new set of decisions incorporating this new portfolio
structure, thus the updated cash-ow scheme is contingent on the scenario realization in the interim (rst year) and the current scenario
tree branches. Thus decisions, other than the rst stage decision, are path-dependent and we have a SP problem with recourse. Fig. 1 illustrates the structure of a problem with three periods each divided into four quarters.
One main assumption we make is that the macroeconomic environment is independent of the governments policy actions with regard
to public borrowing. That is, the amount and the type of bonds the government decides to issue do not affect the level of interest rates in
the market. This is a realistic assumption, especially for countries that have deep and liquid bond markets with many issuers and lenders.
Even though the constraints imposed by the characteristics of the economy and the nancial markets may differ from one country to
another, the objectives in debt management are similar and the general structures of liabilities are comparable. For example, Turkey and
Canada issue the same type of ination linked bonds. The oating rate notes of the Turkish Treasury are similar to those of Italy. Therefore,
the generic model can serve for different country characteristics with small modications. There can be as many periods as relevant, and
the number of sub-periods in each period can vary. The model can also consist of two dissimilar periods, the rst period corresponding to
year one and the second covering all remaining years in the model horizon.
3.1. Notation
The model is based on cash-ow equations that guarantee that the total out-ows of the government match the inows. We include a
cash account in our model that would absorb any excess or short borrowing that might occur when certain scenarios are realized since cash
out-ows are based on some parameters that are scenario specic. Thus, the debt managers set the amount of each bond to be issued in all
sub-periods of the following period considering the possibilities for the level of the nancing requirement. If, in some cases, the total debt
raised is more (less) than needed, the excess (short) amount is injected into (withdrawn from) the cash account of the government.
Parameters and index sets
T
N
t
Ti
S
s
ps
J1
J2
J3
J
J 0t
J 0g;t
mj
cj
ut;j
PSt
Y s;t;j
set of type g bonds with maturities shorter than t, J 0g;t fj 2 J i : t mj > 0g; g 1; 2; 3
CB0
ns; t
s pays
The decisions are made at the nodes of the scenario tree, thus nodes are where scenario paths branch. The parameter ns; t denotes the
node in which the issuance decision is made for time t under scenario s. For all the scenarios in period 1, the decision is made in node 0:
ns; t 0;
for 8s 2 S and t 6 T 1 :
209
r st;c
est;j
interest earned on the cash balance of the government at time t under scenario s
exchange rate prevailing at time t for instrument j under scenario s (est;j 1 for local currency instruments)
Lst
liability payments xed before the decision horizon for time t under scenario s
The decision variables are dened for each node of the scenario tree:
Decision variables:
amount of instrument j to be issued in period t under scenario s, decided at decision point ns; t
ns;t
X t;j
Auxiliary variables:
Ist
Dst
F st
TC s
C st
CBst
Rst
VR
cv s
PR
cps
; Ist ; Dst ; F st ; TC s ; C st ; CBst ; Rst ; VR; cv s ; PR; cps are in units of numeraire currency.
3.2. Constraints
The constraints of the model include inter-temporal cash-ow equations that provide for balance between the governments payments,
amortization and interest payments, and revenues. There is also a balance equation for the cash account. We also include constraints
regarding the marketability of the bonds, as there might be bond specic limitations for the amount of issuance due to the structure of
market demand. Below are the constraints of our model:
Total principal paid back at time t, scenario s:
Dst
ns;tmj
X tmj ;j
j2J 0t
8t; s:
3:1
This equation sums the principal values of all the bonds that mature at time t for a specic scenario s.
Total net interest paid at time t, scenario s:
Ist
X
j2J 01;t
est;j
ns;tm
X tmj ;j j r stmj ;j
estmj ;j
0
t1
X
X
j2J 02;t
ns;s
stmj
est;j
ess;j
t1
X X
j2J 03;t
stmj
ns;s
est;j
ess;j
Rst
8t; s:
3:2
The interest cash-ow equation for scenario s, consists of the interest paid on maturing zero coupon bonds and the coupons paid for live
xed and oating rate bonds at time t, all adjusted for changes in the underlying exchange rate. The interest paid is computed by multiplying the principal value of a bond by the applicable interest rate, which is xed at time of issuance for zero and xed coupon bonds and
at the start of coupon period for variable rate notes. The interest earned on the cash account is deducted from the interest payments.
Appreciation in the value of foreign currency debt at maturity:
F st
X
j2J 0t
ns;tm
X tmj ;j j
est;j
estmj ;j
1
8t; s:
3:3
Countries that issue foreign currency denominated debt also have to pay the difference in the market value of debt due to exchange rate
uctuations.
The cash-ow balance:
ns;t
X t;j
8t; s:
3:4
j2J
The cash-ow balance equation indicates that the total amount of bonds issued at time t (for scenario s) and the amount used from the
governments cash account (C st can take both positive and negative values) should equal the sum of debt repayments, including principal
and interest, and the non-debt liabilities of the government, accounting for the primary surplus available for time t.
210
CBst CB0 C st
CBst CBst1 C st
8s and if t 1;
8s; 8t : t > 2:
3:5
3:6
The cash account balance must be adjusted after each time step, taking into account in and out-ows.
Interest earned from the cash account balance:
8s and if t 1;
3:7
8s; 8t : t > 2:
3:8
The balance at the governments cash account generates revenue for the government and the amount is based on the outstanding balance
at the end of previous time step
Non-negativity:
CBst P 0 8t; s;
ns;t
X t;j
3:9
P 0 8t; s; j:
3:10
We assume that the government does not allow its cash account to deplete. The amount of bonds issued cannot as well be negative (nobuybacks are allowed).
Marketability:
ns;t
X t;j
6 ut;j
8t; s; j:
3:11
The marketability constraint accounts for the demand for government bonds.
Min z1
ps TC s :
3:12
s2S
Here, the cost denition TC s is to be determined taking into account the relevance of possible alternative measures. A possible formulation is provided in Section 5.1.
3.3.2. Market risk
With regard to market risk, we consider measures that preserve LP solvability (see Mansini et al. (2007) for linear risk measures used in
portfolio optimization models) and use the Conditional Cost-at-Risk (CCaR) concept. This is based on the Conditional Value-at-Risk
(CVaR) measure, also referred as the mean excess loss or the expected shortfall, that has emerged as an alternative risk measure as
a response to the limitations of VaR.
Despite their popularity, Value-at-Risk models, used to obtain a measure of the maximum potential change in value of a portfolio of
nancial instruments with a given probability over a pre-set horizon (RiskMetrics, 1996, p. 6), have undesirable mathematical characteristics such as lack of subadditivity (VaR of a portfolio can be larger than the total of that of individual assets). It is difcult to optimize when it
is calculated from scenarios (see Pug, 2000) and does not provide any information about the level of risk if the condence level is exceeded.
The portfolios CVaR is the expected loss given that the loss is greater than (or equal to) its VaR. In other words, it is the expected value of
100a% worst costs over the entire scenario set at a given level of a. Pug (2000) has shown that CVaR possesses the required properties of
coherent risk measures in the sense identied by Artzner et al. (1999). Rockafellar and Uryasev (2000) show that CVaR can be efciently
minimized using linear programming in a scenario-based framework. For a government that is concerned with the level of interest costs
rather than the value of the debt portfolio, the CVaR measure can be turned into a Conditional Cost-at-Risk (CCaR) metric.
To compute the CCaR value, we dene an auxiliary variable, cv s , which takes positive values when a certain level, VR, is exceeded. In the
optimal solution, the VR value equals the associated VaR level for a given a:
Min z2 VR
1X
ps cv s
s2S
3:13
s:t: cv s P TC s VR 8s;
cv s P 0 8s:
211
3:14
3:15
Depending on the DMs concerns or preferences with regard to market risk, it is also possible to use other measures such as the worstcase cost. The government might also have a target level for the debt service expenditures, and any deviation above this level due to market
conditions can be a measure of market risk. If high deviations are more important than lower ones, one can assign different weights to
different levels of excess cost and invent a piece-wise linear objective function.
3.3.3. Liquidity risk
While the cost and the market risk can be measured in accounting terms, liquidity or re-nancing risk, is associated with the actual debt
service or total cash ows of the government For this risk objective, we again adopt a Conditional at-Risk measure:
Min z3 PR
1X
ps cps
3:16
s2S
cp P 0 8s:
3:17
3:18
This Conditional Payment-at-Risk (CPaR) formulation considers the highest debt service level in a single time step for each scenario and
aims to minimize the expected value of 100a% of highest payments over the entire scenario set for a given level of a. In other words, the
CPaR metric accounts for the expected highest possible payment level under worst-case scenarios. The governments can compare this
amount with their available resources or with the level of funds they would be able to generate in extreme market conditions to manage
liquidity risk.
Alternative formulations are also possible depending on the preferences of governments such as a metric that would quantify the variability of cash ows through time to serve as a means for smoothing out debt repayments.
4. MCDM approaches to the PDM problem
MCDM literature contains many examples which combine MCDM tools with the nancial decision making process. Zopounidis (1999)
and Steuer and Na (2003) present extensive bibliographies on the subject showing that methods like multi-objective/goal programming,
outranking relations approaches, Analytical Hierarchical Process (AHP) have been applied to the elds of portfolio analysis, nancial planning, budgeting, risk analysis, corporate management, etc.
In our framework, the SP model forms the basis on which the DMs explore the trade-offs between different objectives. We experiment
with our model, which has the following form, for possible achievements of the objectives and discuss the results to demonstrate how sovereign DMs can utilize these models in making their decisions:
P\Min" z fz1 x; z2 x; z3 xg
s:t:
x 2 X;
where zi ; i 1; 2; 3 denote the objective functions, x and X represent the decision variables and the feasible set, respectively. We use quotation marks since the minimization of a vector is not a well dened operation. When multiple criteria are considered, it is unusual to have a
single solution that is best for all criteria, and typically, one needs to sacrice in some criteria in order to improve in others. In general, if we
have p criteria, a solution x 2 X is said to be efcient if there does not exist x0 2 X such that zi x0 6 zi x for all i, and zi x0 < zi x for at least
one i. If x 2 X is efcient then its image in the criterion space fz1 x; z2 x; . . . ; zp xg is said to be non-dominated.
4.1. Identifying efcient solutions using the PDM model
To help the DMs explore the outcomes of alternative issuance strategies, we obtain a set of non-dominated solutions, i.e. a portion of
the efcient frontier (E) by utilizing an achievement scalarizing program (see Steuer (1986, pp. 400405) for a discussion on achievement scalarizing functions). We rst identify an ideal point z in the criterion space where each objective attains its respective minimum by solving variants of the problem with each individual objective separately. These values then become the targets towards
which feasible solutions are aimed. To accomplish this, we project the ideal reference point onto the non-dominated surface. We employ a weighted Tchebycheff metric to discover the projected point on the surface, which is dened by the criterion vector that has the
shortest weighted Tchebycheff distance to the ideal point. This projection is obtained by solving the following achievement scalarizing
program:
Min b e
p
X
zi x
4:1
i1
s:t: b P ki zi x zi
i 1; . . . ; p;
x 2 X;
4:2
4:3
where e is a very small positive constant, which guarantees that the solution obtained is non-dominated and the CCaR and CPaR metrics are
properly computed, i.e. in line with Rockafellar and Uryasev (2000). The approach is illustrated in Fig. 2. By changing the values of ki , i.e. the
weights assigned to the Tchebycheff distance and solving the above program iteratively, we end up with a set of different points on the efcient surface.
212
z2
An Efficient Solution
Tchebycheff contours
z1
Fig. 2. Illustration of the Tchebycheff program.
Min b e
p
X
zi x
4:4
i1
s:t: b P ki zi x hi hdi i 1; . . . ; p;
x 2 X;
4:5
4:6
where e is a very small positive constant and h is the step size along direction d. The DM is assisted with a graphical display where the
changes in the objective function values are depicted based on different d and h values. Korhonen and Laakso solve the achievement scalarizing program for h going from 0 to 1. The kinks of the objective function value trajectories occur at h values that correspond to basis changes
in the solution of the linear program.
5. An application: Examples from the case of Turkey
We illustrate the ideas discussed in the previous sections for the case of sovereign debt management in Turkey using our generic SP
model. We assume that the government prepares an annual borrowing program at the beginning of each year deciding on the issuance
strategy for the following four quarters. This decision is then to be revised at the beginning of the next year.
Our model covers a period of 3 years, in line with the central governments Medium Term Fiscal Plan (MTFP). The short-horizon of the
Turkish case limits the models capabilities in this application, especially in consideration of the liquidity risk objective. The model tends to
choose longer-term bonds that mature beyond the decision horizon when this objective is minimized. However, it was not possible to obtain longer-term scal assumptions.
As a simplication, we assume that the existing forward liabilities forecast as of that date are scenario-independent Lst Lt 8s as well
as the non-debt cash ows of the government PSst PSt 8s. The parameters Lt are based on the available data at the Turkish Treasury website (www.hazine.gov.tr) as of December 2005 and PSt values are in line with the scal policy assumptions in the MTFP 20062008 details
of which can be found at the Ofcial Gazette (no. 25863) dated 02.07.2005 (in Turkish). The cash account of the Treasury, held at the Central Bank of Turkey (CBT), is not remunerated due to the current legal framework. Therefore, the revenues on the cash account, Rst , are all set
to zero. The starting cash balance is assumed as zero for simplication purposes. The numeraire currency is the Turkish Lira (TRY).
Our model presents a selection of seven different kinds of bonds: four of which are TRY denominated zero-coupon bonds with maturities of 3, 6, 12 and 18 months. We also include a 3 year TRY xed-rate coupon bond (3YTF), a 3 year TRY variable rate coupon bond indexed to 6 month Treasury Bill yields (3YTV), and a 3 year USD denominated xed-rate coupon bond (3YUF). All coupons are semi-annually
redeemed. The variable rate bond is assumed to be issued with a xed spread over the prevailing 6-month interest rate. We include two
marketability constraints: we assume that in one quarter the amount of 3-month bills the Treasury can issue is capped at 10 billion TRY
and the market availability for the USD denominated bonds is 3 billion USD per quarter, considering the size and preferences of the lenders
in those segments of the debt market.
5.1. The SP model
In our application, we evaluate interest costs in accrual terms so that the interest payments can be attributed to the periods they are
generated. However, for Turkey, the interest charge is not the sole source of cost for the government. There are bonds issued in other
213
currencies and any increase in debt repayments, including principal and interest, due to changes in the exchange rates, adds up to the cost
of debt. Thus, our cost denition covers not only the interest charges to accrue during the planning period but also the change in the market
value of foreign currency denominated debt (see Turkish Treasury, 2004, p. 59). Foreign currency linked debt that mature beyond the decision horizon are marked to market value at the end of the model period.
As a result, we include the following accrual cost measure in our model:
Total accrued cost at T in scenario s:
As
T1
X
j2J
sTmj 1
X
j2J 3
ns;s
X s;j
esT;j
ess;j
!
1
T1
X
j2J 1
sTmj 1
T 1
X
esT;j 1
ns;s
X s;j 1 Y s;T;j rss;j s
es;j cj
sTm 1
ns;s
X s;j
X X
T1
esT;j T s s
esT;j 1
ns;s
r
X s;j 1 Y s;T;j rsTcj 1;j s
s;j
s
es;j
mj
es;j cj
j2J 2 sTmj 1
8t; s;
5:1
where cj is 2, since the regular coupon period for Treasury bonds is assumed as 6 months.
The accrued cost is calculated for bonds and bills that have not yet matured at time T. It consists of the changes in the value of the bonds
due to movements in the exchange rate and the interest that has accumulated on a bond since its issuance (for zero-coupon bonds) or its
last coupon payment (for coupon bonds) up to time T. We compute the accrued interest rate by multiplying the effective interest rate by
the fraction of days that have passed since the last the coupon payment to the coupon period. We then adopt the following cost denition,
which is the sum of actual interest payments made in cash and interest costs accrued:
TC s
T
X
Ist F st As :
5:2
t1
As measures of market and liquidity risks, we use the CCaR and CPaR metrics, respectively. In this application we calculate these values
at the 90th percentile of their respective distributions (i.e. a 10%). Thus the model becomes:
P\Min"
z fz1 ; z2 ; z3 g
Yt C
l
X
Ai Y ti et ;
5:3
i1
et is the vector of
5:4
5:5
F 0 F X:
5:6
2
The parameters of the model cannot be disclosed due to the condentiality reasons. However, estimation of the model parameters is a straightforward process which can be
carried out by using commercially available econometrics packages.
214
Table 1
Stability results (based on 50 independent replications).
Scenario tree
10 10 10
30 10 10
40 10 10
80 10 10
Average
St. Dev.
Average
St. Dev.
50.9
52.3
52.9
52.7
4.9
3.7
3.0
2.5
83.5
88.7
89.6
90.7
10.9
7.1
5.3
4.2
We then transform the 5 1 vector that contains the standard normal random variables by multiplying it by F and impose the resulting
vector to the VAR model as a random shock. The monthly paths created by the model are converted to quarterly gures by taking averages
over 3-month periods. Once we obtain simulated values for the short and 1 year interest rate, we compute the yields for maturities in-between by linear interpolation. For maturities longer than a year, a at yield curve is assumed. That is, interest rates are not allowed to vary
with maturity, but taken constant for maturities over a year. The scenario generation process is implemented on MATLAB 6.5.
5.3. Assessment of scenario tree generation
Since our scenario generation method is stochastic, it generates different scenarios on different runs. We need to ensure that solving the
SP model on different trees generated by the same method yields similar optimal values. Kaut and Wallace (2007) focus on this issue and
discuss the evaluation of the quality of scenario generation methods, dening some minimal requirements. Specically, they propose two
measures to test the suitability of a certain generation method for a given SP model: one related with the robustness of the tree generator
(stability) and the other regarding the bias it contains.
By what they dene as in-sample stability, the authors propose to check that the optimal objective values obtained from different
scenario tree instances are approximately identical. While in-sample stability is concerned with the variability of the optimal objective
function value, out-of-sample stability is related with the performance of the optimal solutions in the decision space. In this regard, we
need to test whether solutions obtained on different scenario trees yield similar results when plugged in the real problem. However, this is
not always possible since we may not have full information about the actual distributions that drive our stochastic variables. To ensure that
the scenario generation method contains no bias, we need to compare the optimal values in the scenario-based problem to that of the true
problem and see whether or not they are close to each other. This is again impossible in most cases since it requires solving the true problem optimally. As a proxy, Kaut and Wallace (2007) recommend the employment of a larger reference tree that is believed to have a better representation of the true stochastic process and use the results as a benchmark to test for a possible bias.
To test the in-sample stability of our method, we solved our SP model based on different instances obtained from the same scenario
tree generator, trying to optimize our objective functions separately. We have generated 50 independent and identically distributed scenario trees and solved our model separately on these to see how the optimal value of each objective function varies due to the stochasticity
included in our modelling framework. The model is implemented on GAMS 2.0 using CPLEX as the linear programming solver.
Table 1 depicts the averages and standard deviations of optimal cost and market risk values when models are solved on trees of different
sizes. The results are all based on a three-stage model, only the numbers of branches differ. The notation of 10 10 10 corresponds to a
three-stage tree with 10 branches from each node in each stage. Thus, in the nal stage there are 1000 branches. The results in Table 1 show
that we need to increase the dimensionality to achieve a signicant reduction in the variations of the optimal objective function values.
However, that adds to the complexity of the model, i.e. causes an increase in solution times.
The 3000-branch model is solved in the order of 510 minutes on a Pentium 4, 728 MB RAM PC when a single objective is minimized.
Solution times can be longer when we implement multi-objective approaches (the problem size reaches around 500,000 rows and columns). For our illustrative model, we are content with a 3000-branch scenario tree. As regards to measuring the bias in our model,
the optimal solutions of the 3000 and 8000 branch trees are comparable.
5.4. Experiments
The gures in Table 1 for the cost measure depict the level of interest expenditures due to issuances within the 3-year planning period,
while the values for the CCaR measure indicate the interest expenditures expected under the worst-case scenarios. The table points out
that while the expected cost is estimated to be around 50 billion TRY for a 3 year period, the government can expect to incur up to 90 billion
if the market conditions deteriorate. This potential increase is very signicant considering that the total level of expenditures in Turkish
Central Government Budget in 2006 was around 178 billion TRY.
We now provide the optimal borrowing strategies generated by our model when each decision criterion is optimized separately. Table 2
includes the issuance policy generated for four quarters of year one with respect to each objective. Since decisions for years 2 and 3 are
scenario dependent, we only include the bonds to be issued in the rst year. As expected, since our VAR model generally generated scenarios with declining interest rates in line with the macroeconomic environment in Turkey during the period of analysis, the model chooses
short-term or variable rate securities for the rst year when expected cost is minimized. Short-term rates are also lower on average. However, as far as the market risk is concerned, funding is raised through a combination of short-term and long-term xed-rate bonds. The
model aims to extend maturities to minimize liquidity risk. We acknowledge that this is mainly caused by the short decision horizon in
this application. The similarity in the issuance amounts of USD denominated bonds is due to the marketability constraint we include in
this model for this type of securities.
The model is then used to obtain a set of points on the efcient frontier as depicted in Fig. 3. The graph depicts the ranges of objective
function values and shows how the governments preferences affect these levels. The steep nature of the efcient surface in z3 , liquidity
risk, dimension suggests that we can obtain a sizeable improvement in the liquidity risk objective with relatively small sacrices in the
levels of cost and market risk.
215
Table 2
Optimal borrowing strategy for each objective (initial stage decisions).
90
Market Risk
59
58
57
56
55
54
53
52
80
Liq. Risk
70
60
Cost
50
40
Cost
theta
Fig. 4. Criterion value trajectories in reducing cost (billion TRY).
Let us now assume that the DMs in our problem would like to explore the areas where the level of liquidity risk is at its minimum to
achieve a cost reduction. To this end, we employ the visual interactive approach of Korhonen and Laakso (1986). Fig. 4 displays the effect of
altering the step size h in moving from the point where liquidity risk is at its lowest level in a direction towards the minimum of the cost
objective, i.e. d 3:6; 8:8; 94:0, with k 1=3; 1=3; 1=3. The trajectories indicate that in order to achieve a reduction in cost from its
current level, the DM has to admit signicant increases in the level of liquidity risk, which means an important degree of concentration
in the governments repayments, i.e. the piling up of bond redemptions within a single period. The apparent insensitivity of market risk
is due to representing both risk types on the same scale. The market risk does, in fact, vary sizeably within its own range.
Let us assume that despite the increase in the level of liquidity risk, the DM likes the solution at h 0:7 z1 52:9; z2 86:1; z3 83:7
among the solutions in Fig. 4. Table 3 contains the corresponding issuance strategy, which contains a mixture of short-term and long-term
216
Table 3
Optimal borrowing strategy for the selected solution (initial stage decisions).
1
2
3
4
Bond
3m bill (Bil. TRY)
10.0
19.8
40.9
10.0
43.8
78.0
Cost
54
90
Cost
86
Market Risk
53
82
Liq. Risk
52
78
0
0.1
0.2
0.3
0.4
Quarter
0.5
theta
Fig. 5. Criterion value trajectories in reducing market risk (billion TRY).
bonds for the rst year. The short-term securities provide for the chance to renew debt at expectedly lower levels of interest, while 3 year
xed-rate bonds help the debt management keep liquidity risk at acceptable levels.
Changing the direction (d) and h interactively with the DMs will produce different trajectories on which the DMs can analyze and
experiment with their decisions. Let us now suppose the DMs would like to explore solutions on the direction of reducing the market risk.
Fig. 5 plots the trajectories assuming that direction change occurs at h 0:7 in Fig. 4 (h 0 corresponds to h 0:7 in the previous gure) in
an attempt to reduce the market risk z2 , for example with d 2:7; 4:7; 17:3. An analysis of the gure reveals that in return for a one
billion TRY increase in the expected cost of debt, we can obtain savings in the order of three billion TRY in the levels of both risks. A debt
manager who has a lower risk appetite, especially in terms of market risk, may choose to accept this increase in cost to contain the charge
of debt under volatile conditions.
These experiments can be repeated with many different starting points and directions to be explored. Based on the DM preferences, we
can identify various combinations of short- and long-term, xed and oating rate bonds that will guide the government in meeting its overall cost and risk objectives.
6. Conclusions
The existing methods employed for public debt strategy analysis rely on enumeration of costs and risks associated with given nancing
strategies under various different macroeconomic scenarios. Since, these methods are limited with user-supplied alternatives; they do not
guarantee efcient solutions. In this paper, we formulate the PDM strategy problem as a multi-objective stochastic programming model. In
our analysis, we rst identify several efcient solutions to understand the ranges of objective functions and to obtain an overall view of the
trade-offs using the developed SP model. We then experiment in different regions of the efcient surface based on DM preferences until
obtaining a satisfying solution using an interactive procedure. The procedure relies on DM involvement and provides a graphical representation of the trajectories of the objective function values. This provides for assessing preferences, exploring distinct alternative solutions
and guiding the DMs in making debt strategy solutions.
Our experiments with the case of Turkey show that this framework can be of practical use in a real setting. In our example, the model
suggests issuing short-term bonds to minimize expected cost and longer-term xed-rate securities to decrease the level of market risk as
expected. The DMs can solve the model attaching different weights to the objectives and gain insights about the resulting debt strategy
compositions. With the help of such a quantitative tool, the sovereign debt issuers will have the means to see the effects of different risk
and cost preferences on the debt issuance policy. Our generic model can serve for different country characteristics with small modications, i.e. by changing the number of periods, including different types of instruments, incorporating the relevant cost and risk measures.
Experimentation on the model can help the assessment of the DMs preferences with regard to associated criteria, which are not only
crucial in debt management policies, but also in other nancial decisions of the government. Formulation of robust debt management strategies is critical for preparing a sound government budget which has to account for expected expenditures and potential increases in their
level. Governments must meet their payment commitments at all times, and therefore many countries include buffers in their budgets to
cope with the effects of uctuations in the economy. Our PDM model can also be used to assess the level of contingencies in the national
budget with regard to debt service and interest costs.
Disclosing mainline results from the modelling work to general public opinion can also help the debt management ofces convey
their strategies to market participants and inform stakeholders, such as tax payers, about the objectives of the public debt management
policy. Similar models can also be employed by independent organizations that conduct research on the macroeconomic policies of the
government. This will provide them the means to comment on the actions and plans of the government in raising debt on behalf of the
citizens.
217
References
Alesina, A., Pratti, A., Tabellini, G., 1990. Public condence and debt management: A model and a case study of Italy. In: Dornbush, R., Draghi, M. (Eds.), Public Debt
Management: Theory and History. Cambridge University Press, pp. 94118.
Artzner, P., Delbaen, J.M., Eber, J.M., Heath, D., 1999. Coherent measures of risk. Mathematical Finance 9 (3), 203228.
Barro, R.J., 1995. Optimal Debt Management. National Bureau of Economic Research, Working Paper 5327.
Barro, R.J., 2003. Optimal management of indexed and nominal debt. Annals of Economics and Finance 4, 115.
Bergstrm, P.B., Holmlund, A., Lindberg, S., 2002. The SNDOs Simulation Model for Government Debt Analysis. Swedish National Debt Ofce (SNDO). <http://
www.riksgalden.se> (retrieved on 29.01.2008).
Boender, G.C.E., 1997. A hybrid simulation/optimization scenario model for asset/liability management. European Journal of Operational Research 99, 126135.
Bolder, D.J., 2003. A Stochastic Simulation Framework for the Government of Canadas Debt Strategy. Bank of Canada, Working Paper 2003-10.
Bolder, D.J., Rubin, T., 2007. Optimization in a Simulation Setting: Use of Function Approximation in Debt Strategy Analysis. Bank of Canada, Working Paper 2007-13.
Bradley, S.P., Crane, D.B., 1972. A dynamic model for bond portfolio management. Management Science 19 (2 Application Series), 139151.
Carino, D.R., Kent, T., Myers, D.H., Stacy, C., Sylvanus, M., Turner, A.L., Watanabe, K., Ziemba, W.T., 1994. The RussellYasuda Kasai model: An asset/liability model for a
Japanese Insurance Company using multistage stochastic programming. Interfaces 24 (1), 2949.
Cassard, M., Folkerts-Landau, D., 1997. Risk Management of Sovereign Assets and Liabilities. IMF, Working Paper WP/97/166.
Danmarks Nationalbank, 2005. Danish Government Borrowing and Debt. <http://www.nationalbanken.dk> (retrieved on 29.01.2008).
Dornbush, R., Draghi, M. (Eds.), 1990. Public Debt Management: Theory and History. Cambridge University Press, Cambridge.
Dupacova, J., Hurt, J., Stepan, J., 2002. Stochastic Modelling in Economics and Finance. Kluwer Academic Publishers, Dordrecht.
Grill, T., stberg, H., 2003. A Financial Optimization Approach to Quantitative Analysis of Long Term Government Debt Management in Sweden. Masters Thesis, Dept. of
Mathematics, Linkping Institute of Technology.
HM Treasury, 2007. Debt and Reserves Management Report. <http://www.dmo.gov.uk> (retrieved on 04.07.2007).
Hoyland, K., Wallace, K., 2001. Generating scenario trees for multistage decision problem. Management Science 47 (2), 295307.
International Monetary Fund The World Bank, 2003. Guidelines for Public Debt Management-Accompanying Document and Selected Case Studies. International Monetary
Fund Publication Services, Washington, D.C.
Kaut, M., Wallace, S.W., 2007. Evaluation of scenario-generation methods for stochastic programming. Pacic Journal of Optimization 3 (2), 257271.
Korhonen, P., Laakso, J., 1986. A visual interactive method for solving the multiple criteria problem. European Journal of Operational Research 24, 277287.
Leong, D., 1999. Debt Management Theory and Practice. HM Treasury, Treasury Occasional Paper No10.
Mansini, R., Ogryczak, W., Speranza, M.G., 2007. Conditional value at risk and related linear programming models for portfolio optimization. Annals of Operations Research
152 (1), 227256.
Missale, A., Blanchard, O.J., 1994. The debt burden and debt maturity. American Economic Review 84 (1), 309319.
Nielsen, S.S., Poulsen, R., 2004. A two-factor stochastic programming model of Danish mortgage-backed securities. Journal of Economic Dynamics and Control 28, 12671289.
Organisation for Economic Cooperation and Development (OECD), 2005. Advances in Risk Management of Public Debt. OECD Publishing No. 6, Paris.
Pug, G.C., 2000. Some Remarks on the value-at-risk and the conditional value-at-risk. In: Uryasev, S. (Ed.), Probabilistic and Constrained Optimization: Methodology and
Applications. Kluwer Academic Publishers, Boston, pp. 272281.
RiskMetrics, 1996. Technical Document, fourth ed., JP Morgan, New York.
Rockafellar, R.T., Uryasev, S., 2000. Optimization of conditional value-at-risk. The Journal of Risk 2 (3), 2141.
Steuer, R.E., 1986. Multiple Criteria Optimization: Theory, Computation and Application. John Wiley, New York.
Steuer, R.E., Na, P., 2003. Multicriteria decision making combined with nance: A categorized bibliography. European Journal of Operational Research 150 (3), 496515.
Topaloglou, N., Vladimirou, H., Zenios, S.A., 2008. A dynamic stochastic programming model for international portfolio management. European Journal of Operational
Research 185 (3), 15011524.
Turkish Treasury, 2004. Public Debt Management Report. <http://www.hazine.gov.tr> (retrieved on 30.08.2009).
Villaverde, M., 2003. Global Fund Management Using Stochastic Optimization. Judge Institute of Management, University of Cambridge, Research Papers in Management
Studies WP12/2003.
Volosov, K., Mitra, Spagnolo, F., Lucas, C., 2004. Treasury Management Model with Foreign Exchange Exposure. Centre for Analysis of Risk and Optimization Modelling
Applications, Brunel University, Technical Report CTR/28/04.
Yu, L.Y., Ji, X.D., Wang, S.Y., 2003. Stochastic programming models in nancial optimization: A survey. Advanced Modelling and Optimization 5 (1), 75201.
Zenios, S.A., Holmer, M.R., McKendall, R., Vassiadou-Zeniou, C., 1998. Dynamic models for xed income portfolio management under uncertainty. Journal of Economic
Dynamics and Control 22, 15171541.
Ziemba, W.T., Mulvey, J.M. (Eds.), 1998. Worldwide Asset and Liability Modelling. Cambridge University Press, Cambridge.
Zopounidis, C., 1999. Multicriteria decision aid in nancial management. European Journal of Operations Research 119, 404415.