Beruflich Dokumente
Kultur Dokumente
Abstra
t
The purpose of this paper is to demonstrate how to evaluate sto
hasti
programming
models, and more spe
i
ally to
ompare two dierent approa
hes to asset liability
management. The rst uses multistage sto
hasti
programming, while the other is
a stati
approa
h based on so-
alled
onstant rebalan
ing or xed mix. Parti
ular
attention is paid to the methodology used for the
omparison. The two alternatives
are tested over a large number of realisti
s
enarios
reated by means of simulation.
We nd that due to the ability of the sto
hasti
programming model to adapt to
the information in the s
enario tree, it dominates the xed mix approa
h.
Key words:Simulation, sto
hasti
programming, portfolio sele
tion, asset liability
management, performan
e measurement, nonlinear programming.
1 Introdu tion
Lysaker, Norway
2 Present address: Molde University College, Servi
ebox 8, NO{6405 Molde, Nor-
way
In the literature, there are infrequent reports on empiri
al testing of the per-
forman
e of sto
hasti
programming models. Building sto
hasti
simulation
models to see how a sto
hasti
programming model performs in pra
ti
e is
a
omplex task, be
ause it involves solving a large number of sto
hasti
pro-
grams, whi
h in itself is diÆ
ult.
Kusy & Ziemba (1986) test a two-stage simple re
ourse model and
ompare
this to a sto
hasti
de
ision tree model. However, a new s
enario tree is not
generated ea
h time the horizon is rolled forward. Cari~no & Turner (1998)
ompare the xed mix-rule with a true dynami
sto
hasti
optimization-based
model. The model horizon is not rolled forward, and no out-of-sample data
are used in the test pro
edure. Cari~no, Myers & Ziemba (1998) show the
histori
al performan
e of an asset liability model used by Yasuda Kasai, a
Japanese insuran
e
ompany. Vassiadou-Zeniou & Zenios (1996) and Zenios,
Holmer, M
Kendall & Vassiadou-Zeniou (1998) also do validation ba
k-testing
based on histori
al data. Independently, Kouwenberg (2001) has developed and
tested a pension fund asset liability management model using rolling horizon
simulations.
In Se
tion 2 we des
ribe the sto
hasti
dynami
approa
h and the xed mix
approa
h. Se
tion 3 shows the simulation methodology that is applied to
om-
pare the two approa
hes. In Se
tion 4 we present the numeri
al results before
the
on
lusions are given in Se
tion 5.
2
2 The sto
hasti
dynami
and the xed mix approa
hes
We use a multi period, sto
hasti
asset liability management model developed
for the Norwegian mutual life insuran
e
ompany Gjensidige Spareforsikring.
A mathemati
al des
ription is provided in Hyland & Walla
e (2001a). The
portfolio sele
tion problem is modeled at the strategi
level, where
apital is
allo
ated among a few aggregated asset
lasses su
h as sto
ks and bonds. The
obje
tive is to maximize the expe
ted portfolio value at the end of the horizon
net of
osts, subje
t to rebalan
ing and legal
onstraints on balan
e gures.
The
osts are
omposed of transa
tion
osts 3 and imputed
osts asso
iated
with the violation of the legal
onstraints. These imputed
osts are used as
the measure of risk in the model.
Se
tion 2.1 des
ribes the two approa
hes to the model, while Se
tion 2.2 brie
y
des
ribes the s
enario generation pro
ess.
Both the xed mix and the dynami
versions of the model require de
isions to
be made at dis
rete points in time and dis
rete probability distributions for
the un
ertain variables. The dynami
model
an be explained by
onsidering
a s
enario tree, of whi
h a two period (three stage) example is shown in Figure
1. The top node represents the de
isions today and the nodes further down
t = 1 (today)
t=2
t=3
Fig. 1. The s
enario tree. The nodes represent de
isions, while the ar
s represent
realizations of the un
ertain variables.
represent
onditional de
isions at later stages. The ar
s linking the nodes
represent realizations of the un
ertain variables. This approa
h will
apture the
dynami
s of de
ision making, sin
e de
isions are adjusted to the realizations
of the un
ertainties.
In the xed mix model we assume the following de
ision rule: The portfolio
is rebalan
ed to xed proportions (for instan
e 70% bonds and 30% sto
ks)
3 The presen
e of transa
tion
osts, and the tradeo between short term risk and
return and long term risk and return, makes a dynami
sto
hasti
approa
h ne
es-
sary.
3
at all future de
ision nodes. This means that at all de
ision nodes, assets are
bought and sold so that the xed mix is maintained. The optimization prob-
lem is to nd the proportions that maximize the obje
tive fun
tion. Note the
dieren
e between a xed mix model and a xed mix investment strategy. In
the xed mix investment strategy, the proportions are kept
onstant over a
long investment horizon, while applying a xed mix model means that the
proportions are
hanged every time the model is rerun. In this paper we
om-
pare the results of a xed mix model with the results of a sto
hasti
, dynami
model. See Perold & Sharpe (1988) for a des
ription of the xed mix and other
de
ision rules.
The xed mix approa
h does not require the tree stru
ture, but allows the
un
ertainties to be des
ribed in terms of streams of out
omes (or s
enarios),
as illustrated in Figure 2. However, we have applied the same s
enario tree
Fig. 2. Possible des
ription of the un
ertainties with the xed mix approa
h. In this
approa
h the model is unable to exploit the perfe
t foresight implied after the rst
stage in this tree.
for both approa
hes. Doing this ensures that we get
omparable results and
means that the only dieren
e between the two model formulations are the
onstraints in the xed mix model assuring that the portfolio is rebalan
ed to
the xed mix at every de
ision node. If the approa
hes were to be
ompared
with respe
t to both solution time and solution quality, this pro
edure might
disfavour the xed mix approa
h 4 . However, this problem does not arise sin
e
this paper is
on
erned with
omparing the quality of the solutions given the
quality on the input data, and not solution times.
4
tree so that these statisti
al properties are preserved. This is done by letting
(random) out
omes and probabilities in the s
enario tree be de
ision vari-
ables in a nonlinear optimization problem where the obje
tive is to minimize
the square distan
e between the statisti
al properties spe
ied by the de
i-
sion maker and the statisti
al properties of the
onstru
ted tree (Hyland &
Walla
e 2001b).
Generally, expe
tations for nan
ial markets will depend on the state of the
e
onomy. Some statisti
al properties are
learly state dependent, while others
might be spe
ied independently of the state. As an example of state depen-
den
y,
onsider the volatility of sto
k returns. Empiri
al studies have shown
the ee
t
alled volatility
lustering, meaning that the volatility in
reases af-
ter a period of extreme returns. We model this ee
t by letting the volatility
(standard deviation) in period t + 1 be a fun
tion of the out
omes' deviation
from the mean in period t. For more details of the s
enario generation system,
see Se
tion 4.1.
This se
tion illustrates how we test the quality of the solutions obtained from
ea
h approa
h. In pra
ti
e, only the rst stage solution will be used for a
-
tual de
ision making, whatever approa
h is used. The
onditional de
isions at
stages two and onward are only made in order to nd the right in
entives for
the rst stage de
isions. If the model system is run for example ea
h quarter, a
new instan
e of the model will be generated and solved at the end of a quarter.
We want to test how good ea
h method is on average, for a large number of
realisti
e
onomi
s
enarios, denoted simulation s
enarios. A single simulation
s
enario
onsists of realizations of the un
ertain variables in ea
h simulation
period. To test the solutions of the two approa
hes we pro
eed as follows: At
the beginning of the rst period, our de
ision model tells us what to do, and at
the end of the rst period we see the
onsequen
es of our de
isions. Given that
information and the new state of the e
onomy, we make a new model instan
e
and obtain the de
ision for the beginning of the se
ond period. Based on the
out
omes in the se
ond period, we
al
ulate the
onsequen
es of this de
ision.
The pro
ess
ontinues for the third period, and in prin
iple, indenitely. Keep
in mind that we do not use the information on what will happen when we
make the de
isions. Ea
h time we make a de
ision, the future is sto
hasti
,
and the information we do use is in the form of s
enario trees. These trees
are
reated solely based on past and
urrent information, and not on future
out
omes in the simulation s
enarios, i.e. presently unavailable information.
Denote the s
enarios
ontained in these trees optimization s
enarios.
5
3.1 The test pro
edure
(1) Based on the market expe
tations, generate optimization s
enarios and
use the two approa
hes to obtain the present de
ision.
(2) Based on the same market expe
tations, generate a high number of sim-
ulation s
enarios.
(3) For ea
h out
ome in ea
h simulation s
enario, generate an optimization
s
enario tree and solve the problem using both the dynami
and the xed
mix approa
hes.
(4) Based on the
onse
utive de
isions and out
omes,
al
ulate the e
onomi
performan
e of both approa
hes.
Modeling the problem with a nite horizon is another potential error sour
e.
With few periods, the potential gain for a dynami
approa
h over a stati
one
is small.
6
Market expe
tations:
? ?
xDynami
xFixed mix
Fig. 3. Steps 1 and 2 of the test pro
edure. In step 1 we
onstru
t optimization
s
enarios that are
onsistent with the market expe
tations, and optimize using both
approa
hes to obtain rst stage de
isions, denoted xDynami
and xFixed mix. In step
2 we generate a high number of simulation s
enarios based on the same market
expe
tations. Consult Se
tion 4 and Appendi
es A and B to see how the simulation
s
enarios are generated.
4 Numeri al example
This se
tion presents the numeri
al results. We study a mutual life insuran
e
ompany with a portfolio similar to that of Gjensidige Spareforsikring anno
1997, using a somewhat distorted data set. Assume that the
ompany invests
in the money markets (i.e. bonds with less than a year to maturity) and
sto
ks. We let the money markets be represented by the three month NIBOR
(Norwegian InterBank Oering Rate) and sto
ks be represented by the Oslo
Sto
k Ex
hange Total index. The starting values of assets and liabilities are
given in Table 1. The liabilities are the basis for the
al
ulation of the legal
requirements, and the risk is modeled by penalizing the violations of the legal
requirements. For details, see the model formulation in Hyland & Walla
e
(2001a).
We employ a three period (four stage) model with a total of 1250 s
enarios
to obtain the present solutions from the dynami
approa
h and the xed mix
approa
h,
orresponding to step 1 in Se
tion 3.1. Ea
h period is one year, so
that the total planning horizon is three years. To obtain the future solutions,
7
Step 3: Generate s
enario
tree and obtain
the solutions
for both
approa
hes
/ w
Se
ond ?stage
xDynami
xFixed mix
Third ?stage
xDynami
xFixed mix
Fig. 4. Step 3: Generating
onditional trees and new solutions. After the out
ome
represented by the dotted line, new information is re
eived, and the model horizon
is rolled forward. A new s
enario tree is generated
onditional on the information
in that out
ome. The model is solved to obtain a new present solution, in the gure
alled the se
ond stage solution. The next out
ome in the simulation s
enario is
represented by the thi
k line, whi
h in a similar way gives rise to the third stage
solution. Thus for ea
h s
enario, we obtain the solutions for
onse
utive stages after
generating s
enario trees
onditional on previous out
omes.
Fig. 5. Out
omes and de
isions for one s
enario. Sin
e de
isions and out
omes are
known for this s
enario, in step 4 we
an
al
ulate the obje
tive fun
tion value at
the end of the third period. So ea
h simulated s
enario gives rise to one perfor-
man
e number for the dynami
sto
hasti
approa
h, and another for the xed mix
approa
h.
8
Table 1
Starting balan
e.
Assets Liabilities
Money markets (3 month NIBOR) 70 Insuran
e fund 87
Sto
ks 30 Supplemental reserves 4
Equity 6
Primary
apital 2
Other liabilities 1
Sum assets 100 Sum liabilities 100
This se
tion explains how the market expe
tations are spe
ied, in
luding
state dependen
ies, and how optimization s
enarios and simulation s
enarios
are generated.
9
Table 2
Per
entiles of the marginal distributions. For the money market, the market views
are expressed in terms of expe
tations for the interest rates, while for sto
ks the
expe
tations are given in terms of total returns.
Per
entile 0 0.05 0.25 0.5 0.75 0.95 1
Short term interest rates 1.5 2.4 3.6 3.9 4 5.8 7
Sto
ks -29 -23 3 15 17 19 59
periods.
As explained in Se
tion 2, some statisti
al properties are modeled as state
dependent, while others are assumed independent of the state of the e
onomy.
The state dependen
ies generally depend upon the
hara
teristi
s of the asset
lass. In this example we have modeled state dependent expe
ted returns and
volatilities for both asset
lasses. The other statisti
al properties are assumed
state independent, meaning that they are the same in all states of the e
onomy
at a
ertain point in time.
In order to
apture the volatility
lustering ee
t, the state dependent stan-
dard deviation (or volatility) in period t > 1 is given by
= (1 ) + jx
it i it i i;t 1 i;t 1 j; (1)
where i 2 fs; bg is the asset
lass index, either sto
ks (s) or interest rates (b),
2 [0; 1℄ is the volatility
lustering parameter for asset
lass i (a large leads
i i
of the out
ome of asset
lass i in period t, x is the out
ome of asset
lass i
it
in period t, and is the expe
ted out
ome of asset
lass i in period t. This
it
way the volatility will in
rease after extreme returns, and de
rease after more
normal returns, in line with empiri
al observations.
For interest rates we model a mean reversion ee
t 5 , and let the expe
ted
value of the interest rate at the end of period t > 1 be given by
= + (1 )x
bt i;t 1 ; (2)
where is the mean reversion fa
tor (a high leads to a large degree of mean
5 Mean reversion means that interest rates tend to revert to an average level. When
interest rates are high, the e
onomy slows down and interest rates tend to fall, and
when interest rates are low, the e
onomy booms and interest rates tend to rise. The
mean reversion ee
t usually needs more than two or three years to manifest itself
to a signi
ant degree, and a low value for the mean reversion fa
tor is
hosen in
the numeri
al example.
10
reversion), is the mean reversion level and x is the interest rate at the end
bt
of period t.
For sto
ks we assume that the expe
ted total return in ea
h period is given
by
=x
st b;t 1 + ;
t st (3)
where x is the state dependent short term interest rate in period t, is the
bt st
level and fa
tor ( and ) in Equation (2) are set to 4% and 0.2 respe
tively,
while the risk premium
onstant, , in Equation (3) is set to 0.3 in all periods.
t
Table 3
Market expe
tations used for generation of optimization s
enarios. The expe
tations
for the rst period are derived from the marginal distributions, see Appendix A
for details. While the skewness and kurtosis 6 are assumed state independent, the
expe
ted value and the standard deviation are assumed state dependent.
Asset
lass Distribution (End of) (End of) (End of)
Property Period 1 Period 2 Period 3
Money Exp. spot rate 0.04 State dep State dep
market|3 Standard dev. 0.01 State dep State dep
months Skewness 0.5 0.5 0.5
NIBOR Kurtosis 3.0 3.0 3.0
Domesti
Exp. return 0.085 State dep State dep
sto
ks Standard dev. 0.15 State dep State dep
Skewness -0.5 -0.5 -0.5
Kurtosis 4.0 4.0 4.0
11
For the present optimization problem (
orresponding to step 1 in Se
tion 3.1),
a three period (four stage) s
enario tree is generated. This has 50 out
omes
in the rst period, and 5 in the se
ond and the third, yielding a total of
1250 s
enarios. The s
enario tree is
onsistent with the market expe
tations
given in Table 3 and the
orrelation and state dependent statisti
al properties
dened in Se
tion 4.1.1. For the future optimization problems, we generate
two-period (three stage) s
enario trees with 12 out
omes in period 1 and 6
out
omes in period 2, leading to a total of 72 s
enarios. The rst period
spe
i
ations in these future optimization problems will be dependent on the
sampled out
ome. The state dependen
ies are spe
ied in the same way as
before. The generated trees are
onsistent with these spe
i
ations, in addition
to the state independent spe
i
ations in Table 3 and the
orrelation dened
in Se
tion 4.1.1.
Note that sin
e the xed mix formulation leads to a non-
onvex optimization
model, we would usually need global optimization routines to be sure that the
optimal solution is found. However, testing shows that for our data sets, the
problem is
onvex 7 .
6 The normal distribution has a kurtosis of 3.0. A kurtosis larger than 3.0 means
that the distribution is more peaked around the mean and have fatter tails than the
normal distribution.
7 Many model instan
es were generated, varying the starting balan
e, the market
expe
tations and the degree of risk aversion. Using many dierent starting values
for ea
h instan
e, the optimization always
onverged to the same solution.
12
0.22
0.2
Expected return over three periods
0.1 F
0 2 4 6 8 10 12
Fig. 6. Optimization results of step 1 in Se
tion 3.1. The eÆ
ient frontiers show the
tradeo between the expe
ted return and the risk, and is obtained by solving the
models for dierent required rates of expe
ted returns. The numbers in parentheses
show the initial investment in sto
ks for the dynami
and the xed mix approa
hes,
respe
tively.
The goal is to minimize risk subje
t to a minimum target expe
ted portfolio
return. Risk is measured in terms of shortfalls relative to legal requirements,
and is given by the expe
ted a
umulated quadrati
shortfalls:
2 3
X X X
F = min P s
4
j st x2
sjt
5; (4)
s 2S j 2J t2T
where S is the set of all s enarios, J is the set of shortfall types, P is the s
13
0.22
0.2
Expected return over three periods
0.18
DS3 FM3
0.16 Dynamic
DS2 FM2 Fixed mix
0.14
0.12
0.1 F
0 0.5 1 1.5 2 2.5
Fig. 7. Results from the simulation pro
edure. Ea
h point on the eÆ
ient fron-
tiers are found by solving 600 optimization programs, one for ea
h period in ea
h
simulation.
Figure 6 does not provide a fair
omparison sin
e it does not take into a
ount
that the models will be rerun in the future. The xed mix approa
h suers
under su
h assumptions due to the la
k of dynami
de
ision making. To make
a fair
omparison the testing pro
edure in Se
tion 3 is applied and 200 simula-
tion s
enarios of three periods are generated as des
ribed in Se
tion 4.1.3. The
reason for not in
reasing the sample size is that a single run (of whi
h there
are 8 in Figure 7) involves solving 600 sto
hasti
programs and requires more
than 10 hours solution time on a Sun Ultra 2 (with 200 simulation s
enarios
of three periods).
Figure 7 shows the results of this out-of-sample
omparison and we see that
the dynami
approa
h still dominates the xed mix approa
h, but to a smaller
degree.
Due to the apparent small dieren
e, statisti
al tests are performed. Note that
the expe
ted returns from the xed mix approa
h are generally higher than
for the dynami
approa
h, but the dynami
approa
h yields lower risk. Thus
it is not possible to test whether the dynami
approa
h has a higher expe
ted
return at a given level of risk, or that it gives lower expe
ted shortfall
osts
ompared to the xed mix approa
h for a given level of expe
ted return. The
tests must
ompare risk adjusted expe
ted values, su
h as a
onvex
ombina-
tion of expe
ted return and risk.
Given that the dynami
approa
h gives lower expe
ted return and risk, a
onservative test, i.e. one that is not biased in favor of the dynami
approa
h,
an be
onstru
ted using a
onvex
ombination where there is a relatively
14
large weight on expe
ted return. The following utility fun
tion is employed:
max U = ! Expe
ted portfolio value (1 ! )Risk (5)
where ! 2 [0; 1℄ is the weight parameter governing the degree of risk aversion,
and where Risk = F is our measure of risk in this problem.
To test the hypothesis that the dynami
sto
hasti
approa
h dominates the
xed mix approa
h, the following statisti
is used:
z = ! (VDynami
VFixed mix) (1 ! ) (FDynami
FFixed mix) (6)
where V is sample return, and F is sample risk. The probability that the mean
of z is positive
orresponds to the probability of dominan
e and is found via
bootstrapping (Efron & Tibshirani 1993), sin
e the distribution of z is not
normal. A bootstrap of z is a ve
tor of the same length as z with elements
pi
ked at random (with repla
ement) from the elements of z . The probability of
the mean of z being positive is estimated as
ounting the number of bootstrap
z having positive mean and dividing by the number of bootstrap runs.
The test
ompares the performan
e of the two approa
hes for the pairs (DS2,
FM2) and (DS3,FM3) displayed in Figure 7. There will be a weight ! for ea
h
pair. Weights are
hosen su
h that utility maximization yields the points FM2
and FM3 respe
tively, given that one must
hoose a point on the xed mix
eÆ
ient frontier 9 .
The result of the test is shown in Table 4. The probability that the dynami
approa
h dominates the xed mix approa
h is higher than 50%, however the
dieren
e between the performan
es is not statisti
ally signi
ant. If we
hoose
a weight
orresponding to the average of the slopes of the eÆ
ient frontiers at
FM2 and DS2, we get a probability of dominan
e of 0.8650. For (FM3, DS3),
we get p = 0:8531.
9 Utility maximization means that the marginal rate of substitution of risk for
return equals the negative of the slope of the xed mix frontier:
dEV U=F 1 !
= =
dF U=EV !
where EV is the expe
ted portfolio value and F is the risk along the frontier in
Figure 7. Thus ! = 1=(1 + dd d ), where d
EV
F
ddEV
F
is an estimate of the derivative of
the xed mix eÆ
ient frontier, e.g. at FM2. This means that we
an nd a weight
that
orresponds to a utility fun
tion whose maximum is attained at FM2. Sin
e
we
hoose the xed mix frontier, the test results will not be biased in favor of the
dynami
approa
h.
15
Table 4
Test of whether the dynami
approa
h dominates the xed mix approa
h. A hun-
dred thousand bootstrap runs were used. The p-value reported is the (bootstrap)
probability of dominan
e. The number of observations is 200.
(DS2,FM2) ! = 0:17645 z = 0:04143 ^ = 0:8418 p=0.7456
z
The reason that the dynami
approa
h dominates to a smaller degree, is that
it exploits information in the optimization s
enarios that is not present in the
simulation s
enarios. When
omparing the approa
hes in-sample, the dynami
approa
h has the advantage of the ability to perfe
tly adapt to the informa-
tion given. In the out-of-sample
ase, the information given is no longer per-
fe
t, so an advantage of sto
hasti
programming over the alternative has been
removed. The performan
e of the xed mix approa
h will be
loser to the per-
forman
e of the dynami
approa
h be
ause in the simulations the asset mix
is allowed to be adjusted after ea
h stage. In other words, sin
e we are using
rolling horizon simulations and allow the \xed" mix to
hange at ea
h stage
and in every state, the xed mix model a
tually be
omes relatively dynami
.
The simulation s
enarios and the optimization s
enarios are ne
essarily dif-
ferent, for example regarding the stru
ture of the evolution of information. In
the optimization s
enarios the number of out
omes per stage in
reases expo-
nentially, but in the simulation s
enarios the number is
onstant at 200 for
stages 2 and 3. In parti
ular, the simulation s
enarios do not
ontain so-
alled
worst
ase s
enarios where asset
lass returns are negative in all subsequent
periods. Both the simulation s
enarios and the optimization s
enarios satisfy
the spe
ied statisti
al properties, but for the optimization s
enarios, after
the worst
ase out
ome in the rst and se
ond period, one of the subsequent
out
omes also has a low return. In the simulation s
enarios there is no reason
why a low return out
ome should follow a low return out
ome in the previous
period. These worst
ase s
enarios have a signi
ant ee
t on the per
eived
risk of the portfolio approa
hes as seen in Figure 6, due to the quadrati
na-
ture of risk. Although the probability of these worst
ase s
enarios is very
small, the shortfall
osts a
umulated are so large that they ae
t the overall
expe
ted shortfall
ost. This
an be seen from the dieren
e in the s
ale of risk
in Figures 6 and 7, where the risk in the out-of-sample
ase is mu
h smaller 10 .
Further
omparing Figures 6 and 7 we see that the out-of-sample expe
ted
return of the dynami
approa
h is roughly the same as in-sample, while for
the xed mix approa
h the expe
ted return is higher. This is be
ause the
10 Itis hard to judge whether it is the optimization s
enarios or the simulation
s
enarios that are more realisti
; both sets mat
h the spe
ied statisti
al properties
in
luding state dependen
ies. If the number of simulation s
enarios were in
reased,
a worst
ase s
enario might o
ur,
ausing the risk in Figure 7 to in
rease.
16
xed mix approa
h generally has a higher share of sto
ks than the dynami
approa
h. At the se
ond and third stages in the simulation, the xed mix
approa
h will
hoose more aggressive portfolios in order to fulll the return
requirement. This explains why the eÆ
ient frontier for the xed mix approa
h
is shifted not only to the left, but also upwards from Figure 6 to Figure 7.
Referen es
Cari~no, D. R. & Turner, A. L. (1998), Multiperiod asset allo
ation with derivative
assets, in W. T. Ziemba & J. M. Mulvey, eds, `Worldwide Asset and Liability
17
Modeling', Cambridge University Press, Cambridge, U.K., pp. 182{204.
Cari~no, D. R., Myers, D. H. & Ziemba, W. T. (1998), `Con
epts, te
hni
al issues
and uses of the Russell-Yasuda Kasai nan
ial planning model', Operations
Resear
h 46(4), 450{462.
Efron, B. & Tibshirani, R. J. (1993), An Introdu
tion to the Bootstrap, Chapman
& Hall, New York.
Fourer, R., Gay, D. M. & Kernighan, B. W. (1993), AMPL: A Modeling Language
for Mathemati
al Programming, The S
ienti
Press, San Fran
is
o.
Hyland, K. & Walla
e, S. W. (2001a), `Analyzing legal restri
tions in the
Norwegian life insuran
e business using a multistage asset liability management
model', European Journal of Operations Resear
h 134(2), 65{80.
Hyland, K. & Walla
e, S. W. (2001b), `Generating s
enario trees for multistage
de
ision problems', Management S
ien
e 47(2), 295{307.
Kouwenberg, R. R. P. (2001), `S
enario generation and sto
hasti
programming
models for asset liability management', European Journal of Operational
Resear
h 134(2), 51{64.
Kusy, M. I. & Ziemba, W. T. (1986), `A bank asset and liability management model',
Operations Resear
h 34(3), 356{376.
Lurie, P. M. & Goldberg, M. S. (1998), `An approximate method for sampling
orrelated random variables from partially-spe
ied distributions', Management
S
ien
e 44(2), 203{218.
Murtagh, B. A. & Saunders, M. A. (1983), MINOS 5.4 user's guide, Te
hni
al
Report SOL 83{20R. Revised Feb 1995, Department of Operations Resear
h,
Stanford University.
Perold, A. F. & Sharpe, W. F. (1988), `Dynami
strategies for asset allo
ation',
Finan
ial Analysts Journal 44(1), 16{27.
Vassiadou-Zeniou, C. & Zenios, S. A. (1996), `Robust optimization models for
managing
allable bond portfolios', European Journal of Operational Resear
h
91(2), 264{273.
Zenios, S. A., Holmer, M. R., M
Kendall, R. & Vassiadou-Zeniou, C. (1998),
`Dynami
models for xed-in
ome portfolio management under un
ertainty',
Journal of E
onomi
Dynami
s & Control 22(10), 1517{1541.
18
allo
ation managers in the life insuran
e
ompany prefer to express the mar-
ket expe
tations), see Figure A.1. An approximating
umulative distribution
0.8
Cumulative probability
0.6
0.4
0.2
0
0 0.02 0.04 0.06 0.08
Interest rates
100
0.8
80
0.6
Density
60
0.4
40
0.2 cdf
percentiles 20
0 0
0 0.02 0.04 0.06 0.08 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
Interest rates Interest rates
Fig. A.1. (i) User-supplied per
entiles for interest rates at the end of the rst pe-
riod. (ii) A
umulative distribution fun
tion is tted to the per
entiles. (iii) The
probability distribution fun
tion for interest rates.
fun
tion is tted to these per
entiles. The properties that are listed in Table
3 are
al
ulated from the fun
tion that is tted to the per
entiles.
In addition, we let the user spe
ify the
orrelation between all sto
hasti
vari-
ables, and dene the state dependent properties.
The approximating
umulative distribution fun
tions are found using a NAG
(Numeri
al Algorithms Group) C Library routine for interpolating data. This
method does not guarantee that the se
ond derivative
hanges sign only on
e,
in the
ase of Figure A.1
ausing a somewhat pe
uliar form near the top
of the distribution. However, the resulting fun
tion is monotoni
, so we are
guaranteed that the
urve will have the properties of a
umulative distribution
fun
tion, and that the user spe
ied per
entiles are t exa
tly (in
luding the
0% and 100% points).
19
B Generating the s
enarios that are input to the simulation
This se
tion explains how simulation s
enarios for interest rates and return
on sto
ks are generated. The method works fast, so the time spent on this
pro
edure is negligible
ompared to the a
tual simulation.
The simulation s
enarios are the ones we use to test the performan
e of the
two approa
hes, and they are the basis for generating optimization s
enarios
for later stages (but not for the rst stage, as explained in the paper). They are
sampled from the probability distributions that are tted to the user supplied
per
entiles.
The sto
k returns are sampled rst, using so-
alled inversion sampling. Uni-
form random numbers are sampled and input to the inverse of the
umulative
distribution fun
tion to yield the random returns. Sin
e we have spe
ied a
er-
tain
orrelation between interest rates and sto
k returns, the interest rates are
sampled
onditional on the sto
k returns. The distribution fun
tion for inter-
est rates
onditional on a sto
k return represents a distribution with dierent
mean and standard deviation, and it is found by means of a linear transforma-
tion of the per
entiles. This transformation is found using the formulas for the
onditional expe
tation and varian
e of a bivariate normal distribution. The
skewness or kurtosis of the distribution is not
hanged signi
antly when the
per
entiles are adjusted, so this represents a feasible sampling pro
ess (An al-
ternative way of sampling from marginal distributions with
orrelated random
variables is given by Lurie & Goldberg (1998).)
20
The new standard deviation is
j = 1 2 :
b s b sb
(B.2)
21