The performan

e of sto hasti dynami and

xed mix portfolio models ?

Stein-Erik Fleten , Kjetil Hyland 1, Stein W. Walla e 2

Department of Industrial E onomi s and Te hnology Management, Alfred Getz v.
1, Norwegian University of S ien e and Te hnology, NO{7491 Trondheim, Norway

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

The purpose of this work is to demonstrate how to quantitatively ompare

two models of the same underlying de ision problem, and more spe i ally
to ompare two dierent approa hes to portfolio management. The rst is a
? A epted for publi ation in European Journal of Operational Resear h.
 Corresponding author. Tel.: +47-73 59 12 96; fax: +47-73 59 36 03.
Email addresses: Stein-Erik.Fleteniot.ntnu.no (Stein-Erik Fleten),
Kjetil.Hoylandgjensidige.no (Kjetil Hyland), Stein.Walla ehiMolde.no
(Stein W. Walla e).
1 Present address: Gjensidige Asset Management ASA, P.o. box 276, NO{1324

Lysaker, Norway
2 Present address: Molde University College, Servi ebox 8, NO{6405 Molde, Nor-

way

Preprint submitted to Elsevier S ien e 8 June 2001

multistage sto hasti program, while the other is based on the simple de ision
rule of onstant rebalan ing, also alled xed mix. The omparison is done in
a fair and realisti way. Realisti here means a situation that is lose to the
pra ti al use of the models, where the models are rerun at regular intervals.
An out-of-sample test pro edure like the one applied in this paper is suitable
for this kind of omparison. The test methodology is standard, but this paper
is one of the few attempts to use it on a real problem.

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.

The test methodology is explained in the ontext of a spe i portfolio prob-

lem, but the on lusions are general enough to be used in other areas. We
nd that a dynami sto hasti approa h dominates a xed mix approa h, but
that the degree of domination is mu h smaller when the models are om-
pared out-of-sample than when they are ompared in-sample. The reason for
this is that in an out-of-sample ontext, where the random input data is (at
least) stru turally dierent from the in-sample s enarios, the sto hasti pro-
gramming model loses its advantage in optimally adapting to the information
available in the s enario tree. Furthermore, the performan e of the xed mix
approa h will improve, be ause the asset mix is updated at every stage.

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.

2.1 How the two approa hes dier

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.

2.2 S enario generation

Generation of s enarios an be based on simulation or onstru tion. The ap-

plied s enario generation method is a ombination of the two. We let the
de ision maker spe ify the market expe tations by any statisti al properties
that are onsidered relevant for the problem to be solved, and onstru t the
4 For the xed mix approa h we have hosen to in lude the transa tion osts in-
urred in every state at every stage due to rebalan ing to the xed mix. Thus we
have as many re ourse variables as for the dynami approa h and the solution time
for the xed mix approa h is at least as large as for the dynami approa h in our
ase.

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 ied 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 ied 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.

3 Comparing the performan e of the sto hasti dynami and the

xed mix models

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, indenitely. 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

Our testing methodology follows a sequen e of four major steps, assuming

that the diÆ ult task of spe ifying market expe tations is already done:

(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.

Be ause our s enario generation system is su h that optimization s enario trees

an be generated onditional on sampled out omes, we an generate suÆ ient
input to the two optimization approa hes for ea h out ome in ea h simulation
s enario (sin e we assume that our de isions do not in uen e the sto hasti ity,
we ould in prin iple have generated all s enario trees in advan e). Of ourse,
for the very rst stage we have no prior stage-information, so the s enario tree
will be generated based only on the urrent market expe tations.

The pro edure is illustrated in the Figures 3{5 below.

3.2 Potential error sour es

If the simulation s enarios dier from the optimization s enarios then a po-

tential bias toward one of the approa hes is introdu ed in the test pro edure.
What ould happen is that the dieren e between the information in the opti-
mization s enarios and the simulation s enarios makes the dynami approa h
a priori worse o. The dynami approa h uses the information in the opti-
mization s enarios better than the xed mix approa h, but it ould be using
misleading information so that it onsistently makes relatively bad de isions.

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:

Step 1: Generate s enario

tree for rst
stage and Step 2: Create
solve the R simulation
problem using both approa hes s enarios

? ?
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.

First stage xDynami xFixed mix

Se ond stage xDynami xFixed mix

Third stage xDynami xFixed mix

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

orresponding to step 3 in Se tion 3.1, two-period (three stage) models of 72

s enarios are used. Although this way the third stage de isions are made based
on a model whose horizon is beyond the horizon of the simulation s enarios,
the xed mix and dynami approa hes are treated alike, thus this is not a
signi ant sour e of bias.
The model is programmed in AMPL (Fourer, Gay & Kernighan 1993) and
solved using MINOS 5.4 (Murtagh & Saunders 1983).
Se tion 4.1 illustrates the pro ess of generating s enarios, both for simulation
and optimization. The two versions of the model are solved for dierent levels
of risk aversion, and the results from omparing the quality of the solutions
are given in Se tion 4.2.

4.1 S enario generation

This se tion explains how the market expe tations are spe ied, in luding
state dependen ies, and how optimization s enarios and simulation s enarios
are generated.

4.1.1 Spe i ation of market expe tations

The basis for both the simulation s enarios and optimization s enario trees
are user supplied per entiles for the rst period marginal distributions of sto k
returns and interest rates, the orrelation between the asset lasses and the
denition of the state dependent statisti al properties.
The per entiles supplied by the user are given in Table 2. Marginal distribution
fun tions are tted to the per entiles, as explained in Appendix A.
The orrelation between sto ks and interest rates is assumed to be 0.2 in all

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

to a large degree of volatility lustering),  is the average standard deviation

it

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

state dependent standard deviation of return on sto ks in period t and  is a t

risk premium onstant in period t.

4.1.2 Generating optimization s enarios

For generating optimization s enario trees we have assumed that the relevant
statisti al properties are the rst four moments of the marginal distributions,
the orrelation and the des ription of the state dependent properties. From the
marginal distribution fun tions, whi h are derived as explained in Appendix
A, we an al ulate all marginal moments. Table 3 ontains spe i ations of
the marginal distribution properties for period 1 and the state independent
marginal distribution properties for periods 2 and 3. The volatility lustering
parameter,  , in Equation (1) is set to 0.3 for both assets, the mean reversion
i

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
dened 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 ied 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 dened
in Se tion 4.1.1.

4.1.3 Generating simulation s enarios

For generating the simulation s enarios we do not use the al ulated moments
in 3, but sample from the tted umulative distribution fun tions dire tly. To
apture the orrelation, interest rates are sampled onditional on ea h sto k
return. For ea h sto k return, the onditional interest rate distribution from
whi h to sample is found, and an interest rate sample is drawn from this
distribution. For details, see Appendix B.
For the subsequent periods, the means and the standard deviations are up-
dated for mean reversion and volatility lustering a ording to Equations (1),
(2) and (3). In the end we have a number of simulation s enarios ontaining
subsequent sto k return and interest rate pairs for three periods, satisfying the
same statisti al properties as the rst-stage optimization s enarios, in luding
the state dependen ies.

4.2 Numeri al results

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

0.18 (37.5, 49.4)

(27.5, 34.6) Dynamic
0.16
Fixed mix
(13.6, 20.2)
0.14
(0.4, 5.9)
0.12

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

probability that s enario s o urs, is a weight parameter allowing the de i- j

sion maker to weigh the relative importan e of dierent shortfall types,  is

a path dependent dis ount fa tor 8 , depending on the s enario s 2 S and the
st

time period t 2 T , x is the shortfall of type j 2 J , in time period t 2 T

under s enario s 2 S . The obje tive is minimized for dierent required levels
sjt

of minimum expe ted portfolio returns.

Figure 6 shows the results of solving the rst stage models (referring to step
1 in Se tion 3.1) for both approa hes for dierent levels of minimum expe ted
portfolio return. We see that in this in-sample omparison the dynami ap-
proa h learly dominates the xed mix approa h sin e the target expe ted
p instan e, for an expe ted
returns are a hieved for lower levels of risk. For
return of 15.0%, the (square root of the) risk, F , is 0:42 for the dynami
approa h and 1:84 for the xed mix approa h. Observe that the xed mix
approa h must hoose a more aggressive portfolio (i.e. more sto ks) than the
8 It is path dependent be ause it depends on short term interest rates, whi h are
sto hasti .

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.

dynami approa h to a hieve the target expe ted in-sample return.

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

(DS3,FM3) ! = 0:47974 z = 0:01450 ^ = 1:976z p=0.5251

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 ied 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 ied 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 fulll 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.

5 Con lusion and future work

In this paper we have ompared the performan e of two alternative versions

of a portfolio model. The omparison is severely ompli ated by the fa t that
the portfolio sele tion pro ess involves dynami de ision making under un er-
tainty, so parti ular attention has been paid to the design of the out-of-sample
simulation test. The results show that the sto hasti dynami approa h weakly
dominates the xed mix approa h. We expe t that the degree of dominan e
would in rease if the number of stages in the de ision model is in reased, sin e
three stages probably is too low to fully take advantage of dynami de ision
making.
The applied simulation pro edure takes into a ount that a new instan e of
the de ision model will be rerun when new information is available. With
the new information, a new des ription of the un ertainties is generated and
the de ision models are resolved. The pro edure involves solving thousands of
sto hasti programs.
This work has not fo used on numeri al eÆ ien y. However, the stru ture of
the simulation program is ideal for parallel implementation. In reased numeri-
al eÆ ien y also enables testing more realisti models with more asset lasses
and de ision stages, and allows the sample size to be in reased. We leave this
for future work.
Another future resear h area is to improve the xed mix strategy by reating
more dynami de ision rules, so that the model behaves more like the true
dynami model.
Comparing the robustness of the approa hes with respe t to errors in the spe -
i ation of market expe tations as well as errors in the dis rete approximation
of the distributions, would be another interesting extension of this work.

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 ied 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.

A Spe ifying market expe tations

Expressing market expe tations an be done in many ways. We have hosen

to let the de ision maker supply per entiles for the marginal umulative prob-
ability distribution fun tions for all un ertain variables (this is how the asset

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

(i) Interest rate percentiles

1

0.8
Cumulative probability

0.6

0.4

0.2

0
0 0.02 0.04 0.06 0.08
Interest rates

(ii) Fitted marginal cdf (iii) Implied marginal pdf

1 120

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 dene 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 ied 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 ied 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).)

The following notation is used:

 orrelation between sto k returns and interest rates

x0
sb

s a given return on sto ks

s expe ted return on sto ks
b expe ted interest rate
j
s b expe ted interest rate given a return on sto ks
s standard deviation of sto ks returns
b standard deviation of interest rates
j
b s standard deviation given a return on sto ks
xqb;i per entile of interest rates
x j
qb s;i per entile of interest rates given a return on sto ks

The new mean, given the return on sto ks, is


 j = + b
(x 0  ) : (B.1)
s b b sb
s
s s

20
The new standard deviation is
 
 j =  1 2 :
b s b sb
(B.2)

So the linear transformation of the per entiles is given by


x j =j + (x b
) (B.3)
qb s;i
j s b
b s
qb;i b

for per entile i for interest rates, x . qb;i

Applying this linear transformation to ea h per entile, and then reating a

new distribution fun tion yields a distribution that has the new mean and
standard deviation, and un hanged skewness and kurtosis. The interest rate
is then sampled from this distribution.
For the se ond period, the means and the standard deviations are updated
for mean reversion and volatility lustering a ording to Equations (1), (2)
and (3). The orrelation between sto k returns and interest rates is assumed
onstant for all periods. The above s heme is reapplied for period three.

21