Portfolio Selection with Probabilistic Utility and

Markov Chain Monte Carlo

Pietro Rossi, Massimo Tavoni, Robert Marschinski and
Flavio Cocco
Prometeia S.r.l., Via Marconi 43, 40100 Bologna, Italy
Fondazione Eni Enrico Mattei, C.so Magenta, 63 20100 Milano, Italy
Institute for Physics, University of Potsdam, Germany
E-mail: pietro.rossi@prometeia.it, massimo.tavoni@feem.it,
robert.marschinski@pik-potsdam.de, flavio.cocco@prometeia.it
To be published on Annals of Operational Reseach ( Volume on Financial Modelling)
Abstract.
In this paper, we present a novel portfolio selection technique which replaces the
traditional maximization of the utility function with a probabilistic approach inspired
by statistical physics. We no longer seek the single global extremum of some chosen
utility function, but instead reinterpret the latter as a probability distribution of
optimal portfolios and select the portfolio that is given by the mean value with respect
to this distribution. This approach has several attractive features, when comparing
it to the standard maximization of expected utility. First, it signicantly reduces
the over-pronounced sensitivity to external parameters that plague optimization
procedures. Second, it mitigates the commonly observed concentration on too few
assets; and, third, it provides a natural and self consistent way to account for the
incompleteness of information. Empirical results supporting the proposed approach
are presented by using articial data to simulate nite-sample behaviour and out-of-
sample performance. With regard to the numerics, we carry out all integrals by using
Markov Chain Monte Carlo, where the chains are generated by an adapted version of
Hybrid Monte Carlo.
Keywords: Portfolio Selection, Estimation Error, Probabilistic Utility, Markov
Chain Monte Carlo, Asset Allocation
Introduction
Classical portfolio selection [26] by Maximization of Expected Utility (MEU) suers
from well-documented drawbacks [27]: it often leads to extreme and hardly plausible
portfolio weights, which additionally are very sensitive to changes in the expected
returns. Moreover, it does not take into account informational uncertainty, since
its straightforward optimization procedure imposes complete faith on the estimated
parameters. Historical data provide some information on future returns, but it is well
known that simple-minded use of this information often leads to nonsense because
estimation uncertainty (noise) overwhelms the value of the information contained in
the data. In fact, the position of the extrema of a function tends to be highly sensitive
to irrelevant distribution details and it is quite simple to build examples where a minimal
parameter variation induces a very large shift in the location of the global extremum.
In other words, over-tting quickly becomes an issue when data - as in all cases - is
limited.
The issue of uncertainty in expected returns and its implications for portfolio
selection has been extensively analyzed in the relevant literature: starting with the
work of Bawa, Brown and Klein [5], many authors have since addressed the problem,
often resorting to a Bayesian framework [4, 6, 20, 23, 15, 27, 12]. More recently, with
a growing debate on asset return predictability (which will not be addressed here), the
issue has recaptured the attention of academia [2, 3, 7, 22, 18].
In our opinion, the choice of a particular utility function alone is not sucient
for removing the instability arising from external parameters and changing data sets,
while the Bayesian approach, although quite eective in doing so, leads to a whole new
discourse on how to supply justications for the choice of prior distributions.
Parameters (means, covariances, skewness, etc.) determined by observation of
historical data are not the only source of trouble for portfolios based on the optimization
with respect to an objective function: all of the procedures based on expected utility
maximization in addition suer from the presence of a scalar parameter related to the
investors risk aversion. The value of this parameter cannot be set by theory but the
resulting portfolio composition still tends to be quite sensitive to its value. Actually, due
to a complete lack of scale for this risk-aversion parameter, it is usually adjusted ex post
by hand, i.e. by merely observing where the dynamics happen and dening an ad hoc
scale according to the simple prescription increase the parameter if you want a more
aggressive - meaning riskier - portfolio. This turns out to produce highly unstable and
inconsistent portfolios, portfolios that might change (and usually do change) signicantly
for an apparently small shift in risk-aversion, and might even become less aggressive
than a neighbouring portfolio with a lower risk-aversion. This will be discussed in more
detail in Section 3.
In this paper we want to propose a dierent methodology, which explicitly aims
at the procedure for selecting portfolios - independent of the details of a given utility
function or how the underlying distribution is modelled. We replace the traditional
maximization problem by suggesting a dierent interpretation of the utility function:
we consider it to be the logarithm of the probability density for the portfolio to assume
a given composition, and we dene as preferred the expected value of the portfolios
weights with respect to that distribution.
This approach was conceived through inspiration by statistical physics. The
dynamics of the so called canonical ensemble are strikingly similar to the problem of
portfolio selection: in fact, such a system is subject to two competing inuences: on
one hand it tends to minimize its internal energy, but on the other it also tries to
maximize its entropy. The relative weight between these two rivals is determined
by the systems (constant) temperature: at a high temperature entropy dominates,
since strong internal uctuations bar the particles from settling down in any particular
and energetically convenient state; conversely, at low temperatures the system will be
more likely to assume a state of low internal energy, even in a conguration that is
quite particular, e.g. highly asymmetric. Now, if we view internal energy as utility
with reversed sign, entropy as a measure of portfolio diversication, and temperature
as a measure of missing information on parameters (i.e. estimation error), the exact
formal analogy with portfolio selection becomes evident. Although noticed already by
other authors, such as Bouchaud et al[9] and Piotrowski & Sladkowski[28], this close
resemblance has not yet been fully utilized and put into practice as suggested here.
As it will be shown by means of numerical examples, our probabilistic utility
approach leads to an improved portfolio selection procedure, meaning that we are able
to overcome the excessive sensitivity to external parameters, and we can account for
incomplete information without losing track of the historical data time series.
The nal contribution of our paper concerns the numerical methodology employed
to perform all of the relevant integrals. Most of these cannot be computed explicitly
and therefore we will resort to a dynamical Monte Carlo integration or Markov Chain
Monte Carlo. To enhance performance we have used a variation of the Metropolis-
Hastings prescription known as Hybrid Monte Carlo, that rst appeared in the physics
literature in 1987 [14]. We refer the reader to the appendix for a more extensive
discussion.
The performance of our proposed method is tested by comparing our best
portfolio estimator (PU from now on) with the traditional utility maximization. To
isolate the eects of our probabilistic approach we assume a common utility function
and a deterministic parameter distribution (no Bayesian modeling).
We test the portfolio sensitivity to the estimation errors of market parameters
(means, covariances). As is well known, asset log-returns can be very noisy, so that
inferring the distribution with the required precision can turn out to be a dicult task;
what we explore in the numerical exercise is the portfolio composition dependency on
estimation errors. As will be shown in Section 3, our method outperforms the simplistic
optimization: volatility is reduced without impacting performance. Finally, we are also
able to report a signicant improvement with regard to the instability of the selection
Although in[9] a free-utility, F = U TS
q
is dened which is minimized by slightly modifying the
usual Markowitz approach.
process induced by the risk-aversion parameter.
The paper is organized as follows: in Section 1 we propose our method, in which
MEU optimization is replaced by an expectation with respect to (a transformation
of) the utility function. In Section 2 we deal with numerical integration, and report
empirical results in Section 3. Conclusions and nal remarks are presented in Section 5.
1. The Recommended Portfolio Approach
Let us denote with the set of parameters (weights) that identify a portfolio, and
with U the set of parameters that characterize our utility function, like risk aversion,
investment horizon etc. Let us assume furthermore that the expected utility is computed
with respect to a distribution characterized by parameters, for instance expected excess
returns, that we will denote collectively by . The expected utility may then be written
as a function
u = u(, U, ).
In classical asset allocation theory, the prescription would be to select portfolios that
maximize the expected utility, independent of the magnitude of the estimation error
associated with . In our framework, we consider the expected utility as proportional
to the logarithm of a probability measure in portfolio space, fully conditional on U and
:
P(|U, ) = Z
1
(, U, ) exp

u(, U, )

. (1)
The symbol has to be interpreted as is distributed according to, and Z(, U, )
is a normalization constant dened by
Z(, U, ) =

D()
[d] exp

u(, U, )

,
where D() stands for the integration domain of . The recommended portfolio ,
given U and , is dened as the expected or mean value of :
(U, ) = Z
1
(, U, )

D()
[d] exp

u(, U, )

. (2)
This denition reects the afore mentioned close analogy between the problem of
portfolio selection and the statistical physics of canonical ensembles. In the later, a
systems equilibrium state will emerge at the optimal trade o between lowest possible
internal energy and maximal entropy. A states probabilistic weight hence depends on
an energetic contribution exp (E), where E is a function of the systems state, and an
entropic contribution, which - since every microstate has the same a priori probability
- is constant and given by the hypervolume any microstate occupies in the phase space.
If we replace E by u, then by , and nally introduce Z for a proper normalization,
we recover Eq.1.
Interestingly, in statistical physics Z plays a prominent role as the so called partition function.
Once the probability density function is known, the equilibrium value of any
observable is obtained by calculating its mean value over all possible states. In
particular, Eq.2 represents the formal analogue of computing such a systems position.
The meaning of the constant parameter , which for now is not set by theory,
can easily be derived by noting that its statistical physics analogue, , is inversely
proportional to the systems temperature, and hence determines the relative weighting
between entropy and energy. In our case, energy becomes (minus) utility, and in fact
works as an adjustable amplier of the utility contribution exp(u): for large it will
dominate and hence produce portfolios with an expected utility close to the possible
maximum, which will actually coincide with traditional MEU in the limit of
(since all the measure is concentrated where u(, U, ) has its maximum), while for
small values of the utility function becomes increasingly smoothed out and hence tends
to weigh the dierent portfolios more evenly. Utility considerations become irrelevant
in the limit case of 0, for which Eq.2 leads to the uniform portfolio with equal
allocations for all assets, and hence - as expected from the picture of statistical physics
- to the portfolio with the highest possible entropy.
For a rational investor, the uniform portfolio would in fact be preferred if no
information on future asset performance was available. Conversely, any portfolio
conguration that maximizes expected utility would be acceptable if we had perfect
knowledge of future asset performance. It follows that is in fact a measure of exactly
how much we have of this information.
With a set of stationary historical data, we can bootstrap from the data, build
scenarios, and compute unbiased estimates of the expected utility; or, if we have a
data model, and we believe that historical data are drawn from some distribution, we
can use the time series to estimate the distribution parameters. In both situations our
condence on the value of the expected utility will in some way be linked positively to
the length of the available time series data set. It seems a reasonable assumption for
to exhibit the following asymptotic behavior:
lim
N0
(N) = 0, (3)
lim
N
(N) = , (4)
where N is the size of the data set. The simplest such form is
= N

, (5)
with and constants strictly greater than 0. All of the simulations carried out in
this paper will have = 1 and = 1. The limit will recover the standard
maximization approach. It is obviously interesting to ask whether a more sophisticated
relation between and N, e.g. =
N
J
with J representing the number of considered
assets, could lead to a better algorithm, in particular to one that makes a more eective
use of the available information. However, since already the simple relation = N leads
to great improvements, these questions will be addressed in future research.
Finally, since we choose to insist on a distribution to describe the portfolio, it is
natural to identify the error associated with the estimate of with its standard deviation
in terms of the same distribution. An unbiased estimate of the standard deviation can
be computed, at no extra cost, while computing the integral in Eq.(2).
2. Numerical Integration
2.1. Markov Chain Monte Carlo Integration
The integral in Eq.(2) is easily carried out by Markov Chain Monte Carlo (MCMC)
integration [16, 21].
We do not know how to sample directly from that p.d.f., and we are forced to devise
a Markov chain that relaxes to the desired distribution. After several experiments with
variations of the Metropolis-Hastings, we resorted to an implementation of the Hybrid
Monte Carlo method. Details of the algorithm can be found in the Appendix B.1.
Once relaxation has been achieved we can run the Markov chain for few more steps in
order to perform measurements. Relaxation or thermalization is not a trivial issue but
a thorough discussion of the problems involved would bring us too far from the subject
of this paper. We choose to defer this discussion to a forthcoming paper focusing on the
implementation of the numerical integration scheme.
2.2. The Utility Function
The selection of a good utility function is not the subject of this paper, nor is it
particularly relevant for our results. In this paper we will consistently show results
for two two of them. One is the well established mean-variance Markovitz utility, while
for the other, that we shall call CZC, we refer the interested reader to Appendix C.
3. Empirical Results
In this section the performance of our proposed PU method will be analyzed in various
contexts.
First we will compare the performance of our PU approach with MEU in terms of
out of sample performance. The out of sample will be evaluated looking both at the
utility function and the portfolio returns. Finally, we assess the sensitivity towards the
risk aversion parameter: what is the behavior of the portfolio composition as varies?
How is it aected by a dierent data sample? Historical data used to infer distribution
parameters consists of 8 monthly indexes covering the period from January 1988 to
January 2002. In Tab.1 we show the list of titles employed; this set of data, in the
following, will be referred to as full sample.
3.1. Out of sample performace: I. The utility function
We have carried out the following test: for a given length L, we have generated a time
series of length L from the distribution guessed from the full set of input data. This series
has been used to guess distribution parameters and build the portfolio accordingly. With
Table 1. List of assets employed. The full set of the data goes from Jan 1988, to
Jan 2002. The used acronyms have the following meaning: MSCI = Morgan Stanley
Capital Index, JPM = JPMorgan Index, ML = Merrill Lynch Index . Data source:
Datastream. Data types : Price Index for equities, Total Return Index for bonds. All
the samples are in local currency, unadjusted for ination. The index titles refer to
the Datastream mnemonics.
Assets Description
MSNAMR MSCI North America Equity
MSPACF MSCI Pacic Equity
MSEROP MSCI Europe Equity
JPMUSU JPM US Government Bond
JPMJPU JPM Japan Government Bond
JPMEIL JPM Europe Government Bond
MLHMAU ML US Corporate High Yield
JPEC3M JPM Euro Cash
the portfolio obtained in this way we have computed the utility function corresponding
to the original true distribution, and compared it to the maximum achievable value
of the utility as obtained by seeking the extremum of the original distribution. For each
length we have carried out 1000 such experiments and computed the quantity:
W =
W
T
W
|W
T
,
and what we plot on the y-axis is the expectation E [W] and its associated standard
deviation, where W
T
is the true utility as computed with the portfolio that maximizes
the original distribution, while W is the utility obtained with the portfolio given by the
guessed distribution.
In Fig.1 we show the results of this experiment using the mean-variance utility
function, while in Fig.2 the same experiment is carried out with the CZC utility.
In both cases we would expect:
lim
L
E [W] = 0,
toghether with the variance. For reasonable L, the relevant parameters are both the
mean value as well as the standard deviation. In this respect we can observe that the
PU approach is systematically superior with the mean-variance utility, while, with the
CZC utility results are more undecided.
3.2. Out of sample performace: II. The portfolio return
In this test we took the full set (jan 1988-dec 2002) and broke it into two parts. The
rst part comprised the 156 months from january 1988 through december 2000, and
the second with the remaining months. With each data set we determined the vectors
of mean
1
,
2
and correlation matrices
1
,
2
respectively. With the parameter set

1
,
1
we generated a a time series of length 156 and used this time series to recompute
Figure 1. Average distance form the maximally achievable utility in out of sample
experiments for the mean-variace portfolio.
distribution parameters and the associate optimum portfolio p. With the parameter set

2
,
2
we generated a 2 year long scenario, and we tracked the evolution of the portfolio
along the twenty for months. This process has been repeated about 1000 times and we
measured the mean and the standard deviation of the portfolio resturn. Furthermore
the experimnt was carried out bith with the mean-variance utility function and the CZC
function. Results of this numerical experimnt are shown in gures 3 and 4, where we
plot respectively for the mean-variance utility function and the CZC function.
-2
-1
0
1
2
3
4
5
0 50 100 150 200 250 300 350 400
O
u
t

o
f

s
a
m
p
l
e

d
i
s
t
a
n
c
e

f
r
o
m

m
a
x
.

u
t
i
l
i
t
y
Length of time series
CZC Utility Function
MEU:
PU:
Figure 2. Average distance form the maximally achievable utility in out of sample
experiments for the CZC portfolio.
20
30
40
50
60
70
80
90
100
0 10 20 30 40 50 60 70 80 90 100
S
q
u
a
r
e

r
o
o
t

d
i
s
t
a
n
c
e

f
r
o
m

T
R
U
E

p
o
r
t
f
o
l
i
o
Length of time series
CZC Utility Function
0
10
20
30
40
50
60
70
80
90
100
0 10 20 30 40 50 60 70 80 90 100
S
q
u
a
r
e

r
o
o
t

d
i
s
t
a
n
c
e

f
r
o
m

T
R
U
E

p
o
r
t
f
o
l
i
o
Length of time series
CZC Utility Function
0
10
20
30
40
50
60
70
80
90
100
0 10 20 30 40 50 60 70 80 90 100
S
q
u
a
r
e

r
o
o
t

d
i
s
t
a
n
c
e

f
r
o
m

T
R
U
E

p
o
r
t
f
o
l
i
o
Length of time series
CZC Utility Function
Figure 3. Average portfoliio return for the mean-variace portfolio.
3.3. Out of sample performace: nal comments
The central values of both the out of sample experiments show remarkable agreement
between teh MEU and the PU approach, but it is quite clear that the PU approach ,
in term of volatility, is always at least as good as teh MEU approach, more often, it is
better. The interpretation of this result, is actually at the base of our approach, that is a
caution attitute towards distribution parameters, guessed form nite length distribution,
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0 5 10 15 20 25
P
o
r
t
f
o
l
i
o

s

r
e
t
u
r
n
Holding period
CZC Utility Function
MEU:
PU:
Figure 4. Average portfoliio return for the CZC portfolio.
turns out to be, statistically, a winning approach when used to guide behaviour for
3.4. Sensitiveness Towards Risk Aversion Parameter
As a third empirical investigation, we examine the algorithms stability for a given
portfolio prole (L=5% per yr, T= 1 yr). As previously stressed, all expected-utility
based procedures suer from the presence of a risk aversion parameter, dimensionless and
un-settable from theory. As a measure of instability, it seems then natural to compare
the sensitivity to for both the MEU optimization prescription and the PU method
we propose in this paper. Specically, we examine the behavior of a diversication
indicator, i.e. an indicator that measures the degree of concentration within a portfolio,
and consequently allows to identify the range of the parameter that mostly aects the
portfolio composition.
The simplest of such a - as Bouchaud et al [9] put it - entropy-like measure is the
quantity:
Y =
J

j=1

2
j
. (6)
which ranges from
1
J
(J=number of assets= 8 here), when the portfolio is totally
diversied (evenly spread), to 1 in case of complete concentration on one asset.
In Fig.5 and Fig.6 we report the behavior of Y with respect to for two dierent
data samples, the full sample and a slightly restricted 1988-2000 one. Looking at the
MEU graph in Fig.5, the behavior appears very erratic and the signicant range of
restricted to a relatively small interval, meaning that small changes in can produce
large modications in the portfolio composition. This is certainly not reassuring, given
that is only loosely tied to investors risk aversion, and its setting is not without
uncertainties. Back to Fig.5, what strikes even more is what happens when we look at
the results for a dierent data sample, in this case shortened by the last two years of
observations: the curve decidedly shifts to the right, and consequently the relevant range
of does the same, leading to dangerous risk prole mis-identications, and forcing to
re-calibrate (with all the associated uncertainties) the values of basically each time
new historical observations are added to the sample.
Coming now to the PU model, for which results are shown in Fig.6, the
diversication indicator displays a very dierent pattern: it indicates a more conservative
overall behavior, with values closer to the lower bound of
1
8
. It never concentrates all the
weights on a single asset, not even for the risk-neutral ( = 1) case. Most importantly,
Y exhibits a smooth pattern. This reduces the danger of mis-settings of and its
sensitiveness on the chosen sample; for dierent data sample, in fact, the curve shifts
but remains rather similar, leaving unaected the signicant range of .
1 2 3 4 5 6 7 8 9 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
D
i
v
e
r
s
i
f
i
c
a
t
i
o
n

I
n
d
e
x

Y

19882002
19882000
Figure 5. Portfolio diversication w.r.t. risk aversion and data sample: MEU
Procedure
4. A few words on alternate methods
A well acceptet procedure to account for parmeter uncertainty relays on the adoption
of bayesian statistics. In this case rather than the utility itself the function to maximize
is the expectation of the utility w.r.t the posterior distribution of the parameters.
E [U] =

dmP(m|x)U(, m, )
1 2 3 4 5 6 7 8 9 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1

D
i
v
e
r
s
i
f
i
c
a
t
i
o
n

I
n
d
e
x

Y
19882002
19882000
Figure 6. Portfolio diversication w.r.t. risk aversion and data sample: PU
Procedure
where x represents the observed data. In our view this approach would not solve the
problem given that the result would stil come out of a single procedure of computing
the extremum and the problem concerning paramter sensitivity would present itlsef as
an excess of sensitivity towards the priors. In order to understand the issue one must
bear in mind that even an exact knowledge of the parameters generating the underlying
distributions would not suce. In fact whenven we address an asset allocation problem
we deal with a nite realisation of a given distribution. The observed parameters, in the
next few years or months might shift considerably from the exact ones. We believe that
even a Bayesian approach , to account for parameters, uncertainty would have to be
supplemented with a resample of the forthcoming nite realisation. The correct thing to
do would then be to compute the average of the portfolio, with respect the distribution
of nite size relisation of the underlying distribution. In this respect we consider our
approach more economical in the sense that we do not have to tackle the delicate issue
of choosing priors and we can do without resampling. The interested reader can convice
hersef that, in the case of a Markowitz portfolio, our approach will give the same results
as the Bayes-resampled, provided priors for mean values are chosen suitably.
5. Conclusion And Final Remarks
The purpose of this work was to address and improve some of the well known
weaknesses of portfolio selection by maximizing expected utility. In particular, we
wanted to overcome the excessive sensitivity to external parameters and to account
for informational uncertainty. We achieved this by employing a dierent interpretation
of the utility function. We have tested the proposed method against traditional expected
utility maximization, using articial data to simulate nite-sample behavior, and have
shown superior performance of our method as compared to the simplistic optimization.
We also managed to signicantly improve the intrinsic instability with respect to the
risk-aversion parameter (lack of continuity) that plagues all maximization approaches.
As for future lines of research, we might be interested in relaxing the normality
assumption, for instance by modeling the data with a mixture of Gaussian distributions.
Additionally, our theory led to a parameter that could easily be determined in their
asymptotic behavior, but whose value on the intermediate range was not clear ( in
Eq.5). , the smoothing parameter within our probabilistic utility certainly has an
important inuence on the overall performance of our method; it would therefore be
interesting to ask whether a more sophisticated prescription than the by us employed
= N could lead to an enhanced overall performance, i.e., in particular, to a faster
convergence toward any true optimal portfolio.
Acknowledgments
The authors thankfully acknowledge the exchange of opinions with T. Kennedy and A.
Pellisetto for useful insight on the MCMC integration.
Appendix A. Random Portfolios
Let us denote with p a particular portfolio comprising J assets. It is instructive to
ask what would be the average square distance from it if we were to draw random
portfolios. In this context random means portfolios drawn uniformly from the
hyperplane characterized by the equality constraint:
J

j=1
p
j
= 1
and the J inequality constraints
p
i
0, (A.1)
. . . , (A.2)
p
j
0. (A.3)
The expected value of a function with respect to this measure is dened as:
E[O] Z

1
0
dp
1

1p
1
0
dp
2
. . .

J1
j=1
p
j
0
dp
J
O(p
1
, . . . , p
J1
, 1
J1

j=1
p
j
), (A.4)
where Z is a normalization constant dened by
1 = Z

1
0
dp
1

1p
1
0
dp
2
. . .

J1
j=1
p
j
0
dp
J
.
Recalling the result

1
0
dp
1

1p
1
0
dp
2
. . .

J1
j=1
p
j
0
dp
J1
p
a
1
1
. . . p
a
J
J1
=

J
j=1
(1 + a
j
)
(

J
j=1
(1 + a
j
))
, (A.5)
we have:
Z = (J) (A.6)
E[p
i
] =
1
J
(A.7)
E[p
i
p
j
] =
1
J(J + 1)
for i = j (A.8)
E[p
2
i
] =
2
J(J + 1)
(A.9)
and the average square distance is given by:
E[
J

j=1
(p
j
p
j
)
2
] =
J

j=1
p
2
j

2
J(J + 1)
. (A.10)
Appendix B. MCMC For Portfolio Optimization
Appendix B.1. Hamiltonian MCMC
The algorithm we used in this paper is well known in the physics literature with the
name of hybrid Monte-Carlo ([14]). In this appendix we limit ourselves to a short
introduction.
Let us denote with the portfolios variables, then expectations of functions of
will not be aected if we replace exp [U()] with the distribution
G(, ) exp

U(p)

2

, (B.1)
then starting from a pair (
n
,
n
), the updating rule is dened as follows:
Step 1 Sample as a normal variable with mean zero and variance 1.
Step 2 For a time interval T, integrate Hamiltons equations
d
i
dt
=
U

i
(B.2)
d
i
dt
=
i
, (B.3)
toghether with the boundary conditions
(0) = , (B.4)
p(0) = p
n
; (B.5)
Step 3 With probability
= min

1, exp(G(p(T), (T)) G(p

n
, )

, (B.6)
set p
n+1
= p(T), and with probability 1 set p
n+1
= p
n
.
The clever idea behind this algorithm rests on the observation that, if step 2 is
carried out exactly, Hamiltons equation enforce G(p(T), (T)) = G(p
n
, ) and every
proposed conguration is accepted. In general the acceptance rate will be controlled
by the numerical accuracy of our integration scheme. A good scheme is the interleaved
leap frog that, for nite integration step t and xed trajectory length (that is, scaling
the number of steps in the integration scheme with 1/t), is guaranteed to have errors
O(t
2
).
Appendix C. The Utility Function Employed
The selection of a good utility function is not the subject of this paper, nor is it
particularly relevant for our results. Whenever the return distribution is assumed to
be normal, as in our framework, the explicit solutions of all the utility functions are but
a combination of rst and second distribution moments. Still, a non-trivial dierence
arises when standard deviation terms are included, since they are able to generate a time
horizon eect, i.e. an eect that favors less risky assets on the short range and turns on
risky ones, with higher returns, on the long range. We are aware of the academic debate
on this topic, testied by a considerable amount of related literature [10, 11, 25, 29, 30],
and we believe this to be a desirable feature for a utility function. The probability for the
riskier assets to outperform the less risky ones, in fact, approaches one asymptotically
with time, being it the error function of the ratio between mean and standard deviation,
which grows with the square root of time.
However, utility functions of standard use in nancial economics (such as those who
exhibit Constant Relative Risk Aversion) do not fall in this category. The standard
deviation terms are directly related to non-regular utility functions that measure risk
with the concepts of, say, Value at Risk, Loss Probability etc., i.e. with the so-called
downside risk measures [8]. In this way risk is measured by the expected amount by
which a specied target is not met: this might better describe how the investor perceives
risk, as documented by results from behavioral nance [17], and is more in line with
some recent ALM practice.
For these reasons we employed the following expected utility function, drawn from
the article of Consiglio et al [13]:
E

u(, L, , T)

=
N
T

n=1
t[E(U(nt)) E(D(nt))], (C.1)
where U(nt) and D(nt) are the upside and downside, respectively, of the
portfolio return at time nt against a xed target return L, and is a weight indicating
the investor risk aversion. The time horizon T is built out of N
T
intermediate time
intervals t such that T = N
T
t, is a sequence of N
T
values for (nt), n =
1, . . . , N
T
. The model takes a target-all time view, and the allocation is such that
staying as close to the target return trajectory at all times is the primary concern. A risk
averse investor will want to keep as far as possible from target return under-performance
situations, and will favour paths close to the target line.
Modeling the distributions for the single period log-return with the normal
N(m, ),
exp

( m)
T

1
( m)
2

(2)
J
||
, (C.2)
we obtain by straightforward (tedious) Gaussian integration an explicit expression
for the utility function:
E

u(, L, , T)

=
N
T

n=1
t

ntMf
2
(nt)

ntSf
1
(nt)

, (C.3)
where:
f
1
(t) = ( 1)
e
t
2

2
(C.4)
f
2
(t) =
1 +
2

1
2
erfc(

t) (C.5)
=
M
2S
(C.6)
M =
T
mL (C.7)
S
2
=
T
. (C.8)
As expected, the explicit solution for this specic form of utility is a function of
the portfolio mean, variance and -most importantly- standard deviation. It incorporates
a competing eect between average return and standard deviation with dierent time
scaling properties: the standard deviation contribution is proportional to 1, and can
be traced back to the imperfect cancellation between positive and negative deviations
from ideal line. We thus obtain the desired dependency of the optimal portfolio on the
chosen time horizon: the longer the horizon (ceteris paribus), the more aggressive the
optimal allocation.
References
[1] Ait-Sahalia, Y. and M. Brandt. (2001). Variable Selection for Portfolio Choice. Journal of Finance
56, 1297-1351.
[2] Balduzzi, P. and L. Liu. (2000). Parameter Uncertainty and International Investment. Working
Paper, Boston College
[3] Barberis, N. (2000). Investing for the Long Run when Returns are Predictable. Journal of Finance
55, 225-264.
[4] Barry, C.B. (1974). Portfolio Analysis under Uncertain Means, Variances and Covariances.
Journal of Finance 29 515-522
[5] Bawa, V., S. Brown, and R. Klein. (1979). Estimation Risk and Optimal Portfolio Choice. North
Holland, Amsterdam
[6] Black, F. and R. Litterman. (1992). Global Portfolio Optimization. Financial Analyst Journal 09
28-43
[7] Brennan, M. (1998). The Role of Learning in Dynamic Portfolio Decisions. European Finance
Review 1 295-306
[8] Bouchaud, J.P. and M. Potters. (2000). Theory of Financial Risks. Cambridge University Press
[9] Bouchaud, J.P., M. Potters, and J.P. Aguilar. (1997). Missing Information and Asset Allocation.
Unpublished paper: available for download at www.science-nance.fr
[10] Campbell, J. and L.M. Viceira. (2001). Who Should Buy Long-Term Bonds?. American
Economic Review 91 99-127
[11] Campbell, J. and L.M. Viceira. (2002). Strategic Asset Allocation: Portfolio Choice for Long-
Term Investors. Clarendon Lectures in Economics, Oxford University Press
[12] Chopra, V. and W.T. Zembia. (1993). The Eect of Errors in Means, Variances and Covariances
on Optimal Portfolio Choice. Journal of Portfolio Management Winter 1993 6-11
[13] Consiglio, A., F. Cocco, and S.A. Zenios. (2001). The value of integrative risk management for
insurance products with guarantees. Journal of Risk Finance 2 6-16
[14] Duane, S., A.D. Kennedy, B.J. Pendleton, and D. Roweth. (1987). Hybrid Monte Carlo.
Phys.Lett. B195 216-222
[15] Frost, P. and J. Savarino. (1986). An Empirical Bayes Approach to Ecient Portfolio Selection.
Journal of Financial and Quantitative Analysis 21 293-305
[16] Gilks, W.R., S. Richardson, and D.J. Spiegelhalter. (1996). Markov Chain Monte Carlo in
Practice. Chapman & Hall
[17] Hirshleifer, D. (2001). Investor Psychology and Asset Pricing. Journal of Finance 56 1533-1597
[18] ter Horst, J.R., F.A. de Roon, and B.J.M. Weker. (2002). Incorporating Estimation Risk in
Portfolio Choice. Working Paper, Tilburg University
[19] Ingersoll, J. (1987). Theory of Financial Decision Making. Rowman & Littleeld
[20] Jobson, J.D. and R. Korkie. (1980). Estimation for Markowitz Ecient Portfolios. Journal of
the American Statistical Association 75 544-554
[21] Johannes, M. and N. Polson. (2002). MCMC Methods in Financial Econometrics. Handbook of
Financial Econometrics Forthcoming
[22] Johannes, M., N. Polson, and J. Stroud. (2002). Sequential Optimal Portfolio Performance:
Market and Volatility Timing. Working Paper, Columbia University
[23] Jorion, P. (1985). International Portfolio Diversication with Estimation Risk. Journal of
Business 58 259-278
[24] Kirkpatrick, S., C.D. Gelatt, and M.P. Vecchi. (1983). Optimization by Simulated Annealing.
Science 220 (4598), 671-680
[25] Kritzman, M. and D. Rich. (1998). Beware of the Dogma: The Truth about Time Diversication.
The Journal of Portfolio Management, 24 66-77
[26] Markowitz, H. (1952). Portfolio Selection. Journal of Finance 7 77-91
[27] Michaud, R.O. (1989). The Markowitz Optimization Enigma: Is Optimized Optimal?. Financial
Analyst Journal 45 31-42
[28] Piotrowski, E.W. and J. Sladkowski. (2001). The Thermodynamics of Portfolios. Acta Physica
Polonica B, 32 597-604
[29] Samuelson, P.A. (1994). The Long-TermCase for Equities. The Journal of Portfolio Management
21
[30] Thorley, S. (1995). The Time Diversication Controversy. Financial Analyst Journal 06 68-76