Working Paper

Series
_______________________________________________________________________________________________________________________

National Centre of Competence in Research

Financial Valuation and Risk Management
Working Paper No. 253

Technical Analysis Compared to Mathematical Models

Based Under Misspeecification
Christophette Blanchet-Scallient
Awa Diop
Raina Gibson
Denis Talay
Etienne Tanr
Kathryn Kaminski

First version: August 2005

Current version: August 2005

This research has been carried out within the NCCR FINRISK project on
Conceptual Issues in Financial Risk Management
___________________________________________________________________________________________________________

Technical Analysis Compared to Mathematical Models Based

Methods Under Misspecification
Christophette Blanchet-Scalliet1 , Awa Diop2 ,
2 ,
Rajna Gibson3 , Denis Talay2 , Etienne Tanre
4
Kathryn Kaminski
Abstract
We aim to compare financial technical analysis techniques to strategies which depend on a mathematical model. In this paper, we consider the moving average indicator and an investor using a risky
asset whose instantaneous rate of return changes at an unknown random time. We construct mathematical strategies. We compare their performances to technical analysis techniques when the model is
misspecified. The comparisons are based on Monte Carlo simulations.

Keywords
Stochastic models, Model specification, Portfolio allocation, chartist.
JEL classification
G11, G14, C15, C65
AMS classification
60G35, 93E20, 91B28, 91B26, 91B70.

Acknowledgment
Financial support by the National Centre of Competence in Research Financial valuation and Risk
Management (NCCR FINRISK) is gratefully acknowledged.
NCCR-FINRISK is a research program supported by the Swiss National Science Foundation.

1 Laboratoire
Dieudonn
e, universit
e Nice Sophia-Antipolis Parc Valrose 06108 Nice Cedex 2, FRANCE,
blanchet@math.unice.fr
2 INRIA,
Projet
OMEGA,
2004
route
des
Lucioles,
BP93,
06902
Sophia-Antipolis,
France,
{Awa.Diop,Denis.Talay,Etienne.Tanre}@sophia.inria.fr.
Awa Diop gratefully acknowledges financial support from the
National Centre of Competence in Research Conceptual Issues in Financial Risk Management (NCCRFINRISK),
3 NCCR FINRISK, Swiss Banking Institute, University of Zurich, Plattenstrasse 14, Zurich 8032, Switzerland, Email: rgibson@isb.unizh.ch
4 MIT Operations Research Center, E40-149, Cambridge, MA 02142, (617) 2537412 (voice), katykam@mit.edu (email).

Introduction

In the financial industry, there are three main approaches to investment: the fundamental approach, where
strategies are based on fundamental economic principles, the technical analysis approach, where strategies
are based on past prices behavior, and the mathematical approach where strategies are based on mathematical models. The main advantage of technical analysis is that it avoids model specification, and thus
calibration problems, misspecification risks, etc. On the other hand, technical analysis techniques have
limited theoretical justification, and their efficiency is questionable (see [8]).
The purpose of this study is to address the following problem: Consider a nonstationary financial economy.
It is impossible to specify and calibrate models which can capture all the sources of instability during a long
time interval. In other words, one can only pretend to divide a long investment period into sub-periods such
that, in each one of these sub-periods, the market can reasonably be supposed to follow some particular model
(e.g., a stochastic differential system with a fixed volatility function). Because of the market instability, each
sub-period is short. Therefore, one can only use small amounts of data during each sub-period to calibrate
the model, and the calibration errors can be substantial. However, hedging strategies, portfolio management
strategies, etc., highly depend on the underlying model for the market evolution, and also on the values
of the parameters involved in the model. One can conclude that, in nonstationary economies, one can use
strategies which have been optimally designed under the assumption that the market is perfectly described
by a prescribed model, but these strategies are extremely misleading in practice because the prescribed
model does not fit the actual evolution of the market. In such a situation, one can prefer to use a technical
analysis technique whose aim is essentially to capture some basic trends of the market without assuming
model dependency.
Thus, it might be useful to compare the performance obtained by using erroneously calibrated mathematical models with the one associated with technical analysis techniques.
To our knowledge, this question has not been investigated in the literature. The purpose of this paper is
to present its mathematical complexity and preliminary results.
Here we consider the case of an asset whose instantaneous expected rate of return changes at an unknown
random time. We compare the performance of traders who respectively use:
a technical analysis technique,
a strategy which is optimal when the mathematical model is perfectly specified and calibrated,
mathematical strategies for misspecified situations.
We point out a large difference between the second strategy (optimal allocation) and the other ones: it
depends on the traders choice of a utility function.
Our study is divided into two parts: a mathematical part which, when possible, provides analytical
formulae for portfolios managed by means of mathematical and technical analysis strategies; a numerical
part which provides quantitative comparisons between the various strategies.
A short version of this paper is going to be published [3].

Description of the Setting and Organization of the Paper

Consider two assets which are traded continuously. The first one is an asset without systematic risk, typically
a bond (or bank account), whose price at time t evolves according to the following equation
(
dSt0 = St0 rdt,
S00
= 1.

The second asset is a stock subject to systematic risk. We model the evolution of its price at time t by the
linear stochastic differential equation
(


dSt = St 2 + (1 2 )1(t ) dt + St dBt ,
(2.1)
S0 = S 0 ,
where (Bt )0tT is a one-dimensional Brownian motion on a given probability space (, F, P). At the
random time , which is neither known, nor directly observable, the instantaneous return rate changes from
1 to 2 . A simple computation shows that


Z t
2
1( s) ds
)t + (2 1 )
St = S 0 exp Bt + (1
2
0
=: S 0 exp(Rt ),
where the process (Rt )t0 is defined as


Z t
2
Rt = Bt + 1
1( s) ds.
t + (2 1 )
2
0

(2.2)

95

St

90
85
80
75
70
65
60

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

=0.78
Figure 1: A Trajectory of S For 1 = 0.2, 2 = 0.2, = 0.15, = 2.
In section 3, we explicit the expectation of the terminal logarithmic wealth of a technical analyst who
uses the moving average indicator to detect the time at which the trend of the stock switches. We also
discuss numerical results.
In sections 4 and 5, we examine the performance of a trader whose strategy is based on a mathematical
model.
In section 4, we examine the optimal portfolio allocation strategy. We explicit the optimal wealth and
strategy of a trader who perfectly knows all the parameters 1 , 2 , and , and thus fully describe the best
financial performances that one can expect within our model.
In section 5, we consider a trader who aims to detect the change time as early and reliably as possible.
Notice that the trader only observes the price process (St )t0 . He thus aims to select a stopping time
adapted to the filtration generated by (St ), which serves as an alarm signal and satisfies the following two
requirements:
3

a) The detection of should occur with a delay which is as small as possible;

b) False alarms (i.e. alarms before ) should occur with a probability as small as possible.
In [10] and [11] Shiryayev considers minimization of the expected delay E( )+ subject to the constraint
P( < ) on the probability of a false alarm, for some given [0, 1]; he also considers the estimator
S which minimizes P( < ) + cE( )+ for some prescribed constant c R+ . A similar problem
has been considered by Karatzas [6] who studies the estimator K which minimizes E| |, that is, the
amount of time by which the alarm signal misses the change point . Consider the Model and Detect
strategy which consists in investing ones wealth in the bond up to S or K , and in the stock afterwards.
Unfortunately one cannot expect a tractable formula for the expectation of the terminal logarithmic wealth
resulting from this strategy. We thus study the Model and Detect strategy numerically.
In section 6 we consider the performances of the optimal portfolio allocation strategy and of the Model
and Detect strategy when the trader misspecifies the parameters of the model.
Finally, in section 6.6 we compare the misspecified mathematical strategies to the technical analysis
techniques. In particular, in the case of a large uncertainty on the drift coefficient, we discuss whether one
should use technical analysis methods or techniques based on mathematical models.

Assumptions and Notation

The -algebra generated by the observations at time t is denoted by
FtS := (Su , 0 u t) , t [0, T ].
Notice that the Brownian motion (Bt )0tT is not adapted to the filtration (FtS )t0 .
The Brownian motion (Bt )t0 and the random variable are assumed to be independent.
The law of the change point is the sum of a Dirac mass at 0 and an exponential law of parameter :
P ( > t) = (1 p)et ,

t 0,

(2.3)

for some known coefficients 0 p < 1, > 0.

We denote the value of the portfolio at time t by Wt .
We denote by Ft the conditional a posteriori probability (constructed through the observation of the
process (St )) that the time change occurs within the interval [0, t]:


Ft := P t/FtS .
(2.4)
We denote by (Lt )t0 the following exponential likelihood-ratio process:

1
Lt = exp
(2 1 )Rt
2

 
1
2
) t .
2 (2 1 )2 + 2(2 1 )(1
2
2
The parameters 1 , 2 , > 0 and r 0 are such that
1

2
2
< r < 2
.
2
2

(2.5)

Technical Analysis

3.1

Introduction

Technical analysis is a trading approach based on the prediction of the future evolution of a financial instrument price using only its price (or/and volume) history. Thus, technical analysts compute indicators
which result from the past history of transaction prices and volumes. These indicators are used as signals
to anticipate future changes in prices (see, e.g., the book by Steve Achelis [1]) assuming that:
The price of a stock is governed by the law of supply and demand.
These stocks evolve in trends for one discernible period.
These discernible tendencies repeat themselves in a regular fashion.
By detecting a pattern, technical analysts attempt to trade in order to benefit from the trend if the stock
behaves as in the past. A very large number of technical analysis indicators are available. Here, we limit
ourselves to the moving average indicator because it is simple and often used to detect changes in rates
of returns. To obtain its value, one averages the closing prices of the stock during the most recent time
periods. When prices are trending, this indicator reacts quickly to recent price changes.

3.2

Moving Average Indicator for the Stock Prices

Consider a trader who takes decisions at discrete times of a regular partition of the interval [0, T ] with step
T
t = N
:
0 = t0 < t1 < . . . < tN = T ; tk = kt.
We denote by t {0, 1} the proportion of the agents wealth invested in the risky asset at time t, and by
Mt the moving average indicator of the prices defined as
Mt

1
=

Su du.

(3.1)

At time 0, the agent knows the prices before time 0 of the risky asset. We suppose that he has enough
data to compute M0 . At each tn , n [1 N ], the agent invests all his wealth into the risky asset if the
price Stn is larger than the moving average Mtn . Otherwise, he invests all the wealth into the riskless asset.
Consequently,
(3.2)
tn = 1(St M ) .
n

tn

Denote by x the initial wealth of the trader. The wealth at time tn+1 is
Wtn+1 = Wtn

!
St0n+1
Stn+1
tn + 0 (1 tn ) ,
Stn
Stn

and therefore, since St0n+1 /St0n = exp(rt),

WT = x

N
1
Y





tn exp(Rtn+1 Rtn ) exp(rt) + exp(rt) .

n=0

In Fig. 2(a), we present a typical trajectory of the stock price with parameters:
1
2

= 0.2
= 0.2

= 2
r
= 0.15 T

= 0.0
= 2.0

(3.3)

price S_t
Moving average
of order 0.8

90
85

Prices

Proportion in risky asset

95

80
75
70
65
60

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

1.4

Proportion
in the stock

1.2
1
0.8
0.6
0.4
0.2
0

Time

0.2

0.4

0.6

(a)

0.8

Time

1.2

1.4

1.6

1.8

(b)

4.9

log(Wt )

log of wealth

4.85
4.8
4.75
4.7
4.65
4.6
4.55
4.5

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time
(c)

Figure 2: A Nominal Trajectory

We also present the moving average of order = 0.8 in Fig. 2(a). In Fig. 2(b), we present the proportion of
the wealth of the trader invested in the risky asset. In Fig. 2(c), we present the logarithm of the wealth of
the trader for this particular trajectory.
In Fig. 3, we present t 7 E(log(Wt )), where Wt is the wealth of the technical analyst with a moving
average of order 0.8.

3.3

The Particular Case of the Logarithmic Utility Function

In the case of a logarithmic utility function the expected utility of wealth can be explicited as follows.
Proposition 3.1. Consider a technical analyst whose strategy is defined as in (3.2). Then the expected
logarithmic utility of his wealth is
 
 

WT erT
2
(1)
E log
= 2
r T p
x
2



2
1 eT  (2)
(1)
(3)
+ t 2
r
(p

p
)e
+
p

2
1 et
t(2 1 )(et t)

1 eT (3)
p ,
1 et

4.8

Technical
Analyst

E(log(Wt ))

4.78
4.76
4.74
4.72
4.7
4.68
4.66
4.64
4.62
4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time t
Figure 3: Technical Analyst with a Moving Average of Order = 0.8
where we have set
2

Z Z

(1)

p =

(2)

p =

2y

1(

y2

R4

(2 / /2) (1 + z 2 )



z
2y
2
2
e
i2 /2
dzdy
2 y

z 2 3/2

z1
+z2
y1

(2 / /2) ( v) (1 + z22 )



z2
2
2 2 y2 i2 (v)/2
e
2 y2

3/2

z2 2
2y2

(1 / /2) v (1 + z12 )



z1
2y
2
2
1 i 2 v/2
e
2 y1

3/2

z1 1
2y1

ev dy1 dz1 dy2 dz2 dv,

and
2

(3)

p =

Z
0

z 1 3/2
e
2y

y
2 /4y

ze
iy (z) =
y

(1 / /2) (1 + z 2 )



z
2
2 2 y i2 /2
dzdy
2 y

ez cosh uu

/4y

sinh u sin(u/2y)du.

Remark 3.2. From proposition 3.1, one can use numerical optimization procedures to optimize the
choice of . As we will see in the next subsection, inadequate choices of may weaken the performance of
the technical analyst strategy.
Proof. See Appendix.

3.4

Empirical Determination of a Good Windowing

One can optimize the choice of by using Proposition 3.1 and deterministic numerical optimization procedures, or by means of Monte Carlo simulations. In this subsection we present results obtained from Monte
7

Carlo simulations which show that inadequate choices of may weaken the performance of the technical analyst strategy. For each value of we have simulated 500,000 trajectories of the asset price and computed the
expectation E log(WT ) by a Monte Carlo method. In all our simulations the empirical variance of log(WT )
is around 0.04. Thus the Monte Carlo error on E log(WT ) is of order 5.104 with probability 0.99. The
number of trajectories used for these simulations seems to be too large; however, considered as a function of
, the quantity E log(WT ) varies very slowly, so that we really need a large number of simulations to obtain
the smooth curves (cf. Fig. 4).
The parameters used to obtain Fig. 4(a) and Fig. 4(b) are all equal but the volatility. It is clear from
the figures that the optimal choice of varies. When the volatility is 5 percent, the optimal choice of is
around 0.3 whereas, when the volatility is 15 percent, the optimal choice of is around 0.8.
The parameters used to obtain Fig. 4(b) and Fig. 4(c) are all equal but the maturity. The optimal choice
of is around 0.3 when the maturity is 2 years, and is around 0.4 when the maturity is 3 years.
4.8

4.88
4.87

4.79

E(log(W_T))

E(log(W_T))

4.86
4.78
4.77
4.76
4.75
4.74

4.85
4.84
4.83
4.82
4.81
4.8
4.79

4.73
4.72

4.78
0

0.2

0.4

0.6

0.8

1.2

1.4

4.77

1.6

order of moving average = delta

0.1

0.2

0.3

0.4

0.5

0.6

0.7

order of moving average = delta

(a) 1 = 0.2; = 2; r = 0.0;

2 = 0.2; = 0.15; T = 2.0

(b) 1 = 0.2; = 2; r = 0.0;

2 = 0.2; = 0.05; T = 2.0

5.07

E(log(W_T))

5.06

5.05

5.04

5.03

5.02

5.01

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

order of moving average = delta

(c) 1 = 0.2; = 2; r = 0.0;

2 = 0.2; = 0.05; T = 3.0

Figure 4: E(log(WT )) in Terms of .

We conclude:
- When the volatility decreases, the choice of becomes more important: the curve is flat for large
volatilities only.
When the volatility is high, the losses decrease when choosing a large window size.
In all cases, a too small windowing is sub-optimal,
As |1 | or 2 decreases (respectively, increases), the choice of becomes less (respectively, more)
important.
- The choice of depends on the arrival rate.

2 - This parameter has a strong effect on the importance of the windowing size. In Figs. 20, 21, and 22
the curves are flatter when 2 increases. This observation confirms the intuition; if the future drift is
not large enough, the detection of will be difficult.
Fig. 5 shows that, when = 2, the time horizon has a significant effect on the optimal choice of . When
the time horizon is small, fo example when T = 1, one has better to underestimate than overestimate it.
When T is large, one has better to overestimate . Of course, the critical values for T highly depend on .

T=1

T=2
4.81

4.66
4.8
4.65
4.79
4.64

4.63

4.78

0.5

1.5

4.77

0.5

T=3

1.5

1.5

T=4

5.2
5.18

4.98

5.16
5.14

4.96
5.12
5.1

4.94

5.08
4.92

0.5

1.5

5.06

0.5

Figure 5: Time Horizon and Optimal Windowing: E(log(WT )) as a function of .

To conclude this analysis, we examine the effects of the utility functions on the optimal choice of . We
have chosen HARA utility functions:
1
1

> 1: U (W ) = (W )

= 1: U1 (W ) = log(W ).
In Fig. 6, we represent E(U (WT )) as a function of the windowing size for different values of the risk
aversion parameter . Fig. 6 shows that, for HARA utility functions, the choice of the optimal does
not depend strongly on the risk tolerance. This suggests that our results obtained for a logarithmic utility
remain robust even when the traders utility function displays another risk aversion coefficient .

4
4.1

The Optimal Portfolio Allocation Strategy

Introduction

In this section our aim is to explicit the optimal wealth and strategy of a trader who perfectly knows all the
parameters 1 , 2 , and . Of course, this situation is unrealistic. However it is worth computing the best
9

2.232

4.81

2.23
4.8

2.228
2.226

4.79
2.224
2.222

4.78

2.22
4.77

0.5

1.5

2.218

0.5

1.5

1.5

=1.7

=1.4
3.97

7.3
7.28

3.96

7.26
3.95

7.24

3.94

7.22
7.2

3.93
3.92

7.18
0

0.5

1.5

7.16

0.5

Figure 6: HARA Utility Functions and Optimal Windowing

1 = 0.2, 2 = 0.2, = 0.15, = 2
on the abscissa: ; on the ordinate: E(U (WT ))
performance that one can expect within our model. This performance can be seen as an optimal benchmark
for misspecified allocation strategies using a mathematical model and for technical analysis strategies.
To be able to compare this optimal strategy to a technical analyst strategy, we impose constraints on
the portfolio. Indeed, a technical analyst is only allowed to invest all his/her wealth in the stock or the
bond. Therefore, the proportions of the traders wealth invested in the stock are constrained to lie within
the interval [0, 1].
To compute the constrained optimal wealth we apply the martingale approach to stochastic control
problems as developed by Karatzas, Shreve, Cvitanic, etc. More precisely, we follow and carefully adapt
the martingale approach to the well known optimal asset allocation problem studied by Merton [9]. We
emphasize that our situation differs from the Merton problem by two aspects:
The drift coefficient of the dynamics of the risky asset is not constant over time (since it changes at
the random time ).
Here we must face some subtle measurability issues since the traders strategy needs to be adapted
with respect to the filtration generated by (St ): As already noticed, the drift change at the random
time makes this filtration different from the filtration generated by the Brownian motion (Bt ).
Let t be the proportion of the traders wealth invested in the stock at time t; the remaining proportion
1 t is invested in the bond. For a given, non random, initial capital x > 0, let Wx, denote the wealth
process corresponding to the portfolio ( ).





dWtx, = Wtx, rdt + t 1 r + (2 1 )1( t) dt + dBt .
(4.1)
Now, let U () denote a utility function. We suppose that U is strictly increasing, concave, of class C 2 (0, )
10

and satisfies
U 0 (0+) = lim U 0 (x) = ,

U 0 () = lim U 0 (x) = 0.

x0

Let A(x) denote the set of admissible portfolios, that is,

A(x) := { (FtS )t progressively measurable proc. s.t. W0x, = x, Wtx, > 0 for all t > 0, [0, 1]}.
The investors objective is to maximize his/her expected utility of wealth at the terminal time T . He/she
solves the following optimization problem P:
sup EU (WTx, ).

V (x) :=

A(x)

4.2

A Representation of the Price Process S in Terms of the Innovation Process

F

We aim to characterize the optimal wealth and a corresponding optimal strategy. To this end, we have to
develop technical calculations. We first give an (FtS )t representation of the a posteriori probability process
(Ft , t 0)
Lemma 4.1. The conditional
a posteriori probability Ft that the change point has appeared before time


t, that is, Ft := P t|FtS , is
t
Ft =
,
(4.2)
1 + t
where
t =

p t
e Lt +
1p

Lt (ts)
e
ds.
Ls

(4.3)

In addition, the exponential likelihood-ratio process Lt is


 

1
1
2
2
Lt = exp
(2 1 )Rt 2 (2 1 ) + 2(2 1 )(1
) t .
2
2
2
Proof. In order to compute Ft , we start with a probability space (, F, Q) that can support both a standard
and independent random variable : [0, ) with distribution Q[ > t] = (1p)et
Brownian motion B
which means that Gt := (, B
s ; 0 s
for t 0; we denote by (Gt )t0 the filtration generated by and B,
t). In view of Girsanovs theorem, the process
t
Bt = B

(2 1 )

1( s) ds

is a (Gt )t0 -Brownian motion under the measure of probability P such that
dP 
 = exp
dQ Gt

Z
0

1
s 1
(2 1 )1( s) dB

2 2

t
2

(2 1 )

1( s) ds


1
1
2
+

(2 1 )(Bt B )1( t) 2 (2 1 ) (t )
= exp

2


1
1
2

= exp
(2 1 )(Bt B ) 2 (2 1 ) (t ) 1( t) + 1( >t) .

2


Using (2.2), we observe that

t = 1
B


Rt (1

11


2
)t .
2

(4.4)

 and
Observe that under Q, the random variable is independent of the (Gt ; t 0)-Brownian motion B
thus of R. Thus, the above expression of dP/dQ can be written as follows
Lt
dP 
1( t) + 1( >t) := Zt
(4.5)
 =
dQ Gt
L
where Lt is defined as in (2.5). We now check that, on the probability space (, F, P), we have the same
model as the one presented in the introduction. Indeed, the random variable is G0 -measurable; then,
under P, is independent of the (Gt )t0 -Brownian motion B and we have P[ > t] = EQ [Z0 1 >t ] = Q[ > t].
Using the Bayes rule, one gets


EQ Zt 1( t) |FtS
S

 .
(4.6)
Ft = P[ t|Ft ] =
EQ Zt |FtS
and under Q, one gets
Using now the independence of B




Lt
1(0< t) + 1( >t) | FtS
EQ Zt |FtS = EQ Lt 1( =0) +
L
Z t
Lt s
e
ds + (1 p)et .
= pLt + (1 p)
L
s
0
On the other hand, we get


EQ Zt 1( t) |FtS = pLt + (1 p)

Z
0

Lt s
e
ds.
Ls

Then, going back to (4.6), we get

pLt + (1 p)
Ft =
pLt + (1 p)

R t Lt s
e
ds
0 L
s

.
R t Lt
s ds + (1 p)et
e
0 L
s

Moreover, we can show that the process (Ft )t0 satisfies the following stochastic differential equation
dFt = (1 Ft )dt +

(2 1 )
Ft (1 Ft )dB t ,

(4.7)

where


Z t
2
Rt (1
)t (2 1 )
Fs ds , t 0,
(4.8)
2
0
Ft
is the innovation process. Indeed, set t :=
; an easy computation leads to t = Ut + Vt , where
1 Ft
p t
e Lt ,
Ut =
1p
Z t
Lt
Vt =
e(ts) ds.
Ls
0
1
Bt =

From


dLt = Lt

1
1
2
(2 1 )dRt 2 (2 1 )(1
)dt
2

we deduce

1
1
2
= Ut 2 (2 1 )(1
) dt + 2 (2 1 )Ut dRt ,

U0 =
,
1p



1
2
1
= + Vt 2 (2 1 )(1
)
dt + 2 (2 1 )Vt dRt ,

= 0.

dUt

dV
t

V0

12

It follows that
dt =




2
1
1
+ t 2 (2 1 )(1
)
dt + 2 (2 1 )t dRt .

(4.9)

t
, and get the stochastic differential equation (4.7).
1 + t
Notice that B t defined as in (4.8) is a (FtS )t0 - Brownian motion.
We finally apply It
os formula to the process Ft =

We conclude this section with a result usefull to apply the martingale representation theorem in the
following section
Lemma 4.2. The filtration generated by the observations (FS ) is equal to the filtration generated by the
t ; t 0). In particular, each (F S ) martingale M admits a representation as
innovation process (B

Z
Mt = M0 +

s ,
s dB

where is an (FS ) adapted process.

Proof. Thanks to (4.8), B t is (FS ) adapted. Conversely, we write (4.8) as:


2
dRt = (1
) + (2 1 )Ft dt + dB t ,
2

(4.10)

Thanks to (4.7), F is (FB ) adapted. and we conclude that the process S (= exp(R )) is also (FB ) adapted.

And so (FB ) = (FS ).

4.3

The Case of General Utility Functions

Thus the equation (4.1) can be rewritten as follows

dWtx,
= {(1 t )r + t (1 + (2 1 )Ft )} dt + dB t .
Wtx,
As in Karatzas-Shreve [7], we introduce an auxiliary unconstrained market M defined as follows: Let D be
the subset of {FtS } progressively measurable proc. : [0, T ] R such that
T

(t) dt < , where (t) := inf(0, (t))

E
0

The bond price process S 0 () and the stock price S() satisfy
St0 () = 1 +

Su0 ()(r + (u))du,

Z
St () = S0 +

Su () (1 + (2 1 )Fu + (u) + (u))du + dB u .

We compute the optimal allocation strategy for each auxiliary unconstrained market driven by a process
(see Proposition 4.3). We conclude with Proposition 4.4 which links the optimal strategy for the constrained
problem with the set of optimal strategies for auxiliary unconstrained markets.
For each auxiliary unconstrained market, let A(x, ) denote the set of admissible strategies, that is,
A(x, ) := { FtS progressively measurable process s.t.W0, = x, Wt, > 0 for all t > 0}.

13

We have to solve the following problem P :

V (, x) :=

sup
A(,x)

where

Ex U (WT, ),

dWt,
= (r + (t))dt + t ((1 r) + (2 1 )Ft + (t)) dt + t dB t .
Wt,

Proposition 4.3. For each D, the optimal wealth is



RT
1
WT, = (U 0 )
yHT erT 0 (t)dt
Rt

Wt,

ert+

(s)ds

Ht

(4.11)





RT
RT
1
yHT erT 0 (s)ds |FtS .
E HT erT 0 (s)ds (U 0 )

Moreover, the optimal portfolio satisfies

t,

t
1 r + (2 1 )Ft + (t)
Rt
+

Ht Wt, ert 0 (s)ds

!
,

where Ft is defined as in (2.4), y stands for the Lagrange multiplier, that is, is such that
E(HT exp(rT

(t)dt) (U 0 )

(yHT exp(rT

(t)dt))) = x.

Ht is the exponential process defined by

Z t


1 r + (s) (2 1 )Fs
+
dB s

0
2 !
Z 
1 t 1 r + (s) (2 1 )Fs
+
ds ,

2 0

Ht = exp

and is a FtS adapted process which satisfies

E

RT
HT e(rT 0 (t)dt)

0 1

(U )


RT

yHT e(rT 0 (t)dt) FtS

s dB s .

=x+
0

Proof. We follow Karatzass method (see for example Karatzas [5]). For A(x, ), remember that
dWt,
= (r + (t))dt + t ((1 r) + (2 1 )Ft + (t)) dt + t dB t .
Wt,
Define
t := (1 r) + ( u2 1 )Ft + (t)


Z t
,
,

Wt := Wt exp rt
(s)ds .
0

We thus have

t,
W

t,
dW

t, = t t dt + t dBt ,
W
Z t 


Z t
1 2 2

= x exp
s s s ds +
s dBs .
2
0
0
14

(4.12)

ft, = W
t, Mt is an exponential
We now search an exponential martingale Mt (independant of ) such that W
martingale. Set
Z t

Z t
s 1
s dB
2s ds .
Mt = exp
2 0
0
Then needs to satisfy
1
1
1
2
(s + s ) = s s 2s 2 s2 ,
2
2
2
from which s = s and Mt = Ht . Thus, for all ,

ft,
dW
t 
dBt .
=

t
ft,

(4.13)

ft, , 0 t T ) is a non-negative (FtS , P)- local martingale and so a super martingale.

Therefore the process (W
Consequently,
"
!#
Z T
,
E HT WT exp rT
(t)dt
x.
0

We now introduce the convex dual of U () :

(y) := max [U (x) xy] ,
U

y > 0.

0<x<

Using a duality method, we obtain: for all > 0, A(x, ), y 0,

!#
"
Z T
,

(t)dt)
E[U (WT )] E U yHT exp(rT
0

+ yE

HT WT,

Z
exp(rT

"
(yHT exp(rT
E U

(t)dt)

(t)dt)) + yx.
0

This inequality is an equality if and only if



R
W , = (U 0 )1 yH exp(rT T (t)dt) ,
T
T
0i
h

R
E H W , exp rT T (t)dt = x.
T
T
0
The coefficient y is introduced to satisfy the constraint.
Now, we need to verify that there exists a portfolio such that the process
 #
"


R
R
T
T

E HT exp(rT 0 (s)ds)(U 0 )1 yHT exp(rT 0 (s)ds) FtS

Xt :=
R
t
Ht exp(rt 0 (s)ds)
is its wealth process. We use the martingale representation property of the Brownian filtration in order to
find the optimal strategy . Indeed, there exists a predictable process such that
Z t


R
R
rT 0T (t)dt
0 1
rT 0T (t)dt
S
s dB s .
E HT e
(U ) (yHT e
) / Ft = x +
0

15



R
et = Xt Ht exp rt t (s)ds , we obtain
In particular, with the notation X
0
et = t dB
t .
dX
Consider the strategy
t

1 r + (2 1 )Ft + (t)
t
Rt
+
rt
(u)du

0
X t Ht e

!
.

In view of (4.13), we have

ft,
et
t
dW
dX
=
dBt =
.

,
e
e
ft
Xt
Xt
W

By uniqueness arguments, we obtain Xt = Wt, .

Proposition 4.4. If there exists e such that
V (e
, x) = inf V (, x)

(4.14)

then an optimal portfolio , for the unconstrained problem P is also an optimal portfolio for the constrained
original problem P and
,

Wt = Wt ,e .
(4.15)
An optimal porfolio allocation strategy is
t

:=

1 r + (2 1 )Ft + e(t)
t
R
+
rt 0t
e (s)ds

Ht W t e

!
,

(4.16)

where Ft defined in (2.4) satisfies


R t s 1
p
e Lt
+ 0 e
Ls ds
1p
.

Ft =
Rt
p
+ 0 es L1
1 + et Lt
ds
s
1p
t

Proof. See the proof in Karatzas-Shreve [7] p. 275.

4.4

The Particular Case of the Logarithmic Utility Function

Proposition 4.5. If U () = log() and the initial endowment is x, then the optimal wealth process and
strategy are


Rt
x exp rt + 0 e (s)ds
Wt,x =
,
(4.17)
Hte
1 r + (2 1 )Ft + e(t)
t =
,
(4.18)
2
where
e(t) =

(1 r + (2 1 )Ft )

1 r + (2 1 )Ft
< 0,
2
1 r + (2 1 )Ft
if
[0, 1],
2
otherwise,

if

2
(1 r + (2 1 )Ft )

and, as above,
e (t) = inf (0, e(t)) .
16

(4.19)

Proof. If U (x) = log(x) for each D, the solution of the unconstrained problem is
t,

Wt,

V (, x)

1 r + (2 1 )Ft + (t)
,
2


Rt
x exp rt + 0 (s)ds
,
Ht
"Z
#
T

(t) dt

log x + rT + E
0

"Z

+E
0

Set (t) :=
satisfies

1
2

1 r + (2 1 )Ft + (t)

2

#
dt .

(4.20)

1 r + (2 1 )Ft
2
1
+
. Then the process e defined by (4.19)
and u(, , ) := +


2
1 1 r + (2 1 )Ft + (t)

u(e
(t), (t), ) = min (t) +
.
D
2

Moreover
"Z
"Z

2 #

2 #
T
T
1 1 r + (2 1 )Ft + e(t)
3 1 r + (2 1 )Ft

E
e(t) +
dt E
dt < ,
2

0
0 2
and thus V e(x) < for all x.
Remark 4.6. The optimal strategies for the constrained problem are the projections on [0, 1] of the
optimal strategies for the unconstrained problem. In addition, notice that these strategies depend on the
traders choice of a utility function, whereas the technical analysis technique and the Model and Detect
strategies of section 5 have portfolio weights that are pre-specified without assuming any particular utility
function.
Remark 4.7. In the case of the logarithmic utility function, when t is small and thus smaller than the
change time with high probability, one has Ft close to 0; since, by hypothesis, one also has 1r
0,
2
the optimal strategy is close to 0 ; after the change time , one has Ft close to 1, and the optimal strategy
is close to min(1, 2r
2 ). In both cases, we approximately recover the optimal strategies of the constrained
Merton problem with drift parameters equal to 1 or 2 respectively.
Using the explicit value (4.19) of e(t), one can obtain an explicit formula for the value function V e(x)
corresponding to the optimal strategy.
Thanks to Lemma 4.1, we are able to compute t .
To conclude this section, we explicit the expected logarithm of the optimal wealth. We only consider
here the simple case p = 0 (the expression in the general case is not difficult to obtain but is untractable).
In view of (4.20) and (4.19) one has
V (x) =V (e
, x)

= log(x) + rT + EP


1 r + (2 1 )Ft

Z
1
+ EP
2

1 r + (2 1 )Ft

2

17

1
1 r

2 (1 r) dt
Ft >

2 1

2 (1 r) dt .
<Ft <

2 1
2
1

We recall the notation introduced in Lemma 4.1:

t
,
Ft =
1 + t
dP 
 = Zt ,
dQ Gt


EQ Zt |FtS = et (1 + t ).
We have
V (x) = log(x) + rT
"


Z T
t
2
t
EQ e (1 + t ) 1 r + (2 1 )
+

1 + t
2
0
#

1
t
2 (1 r) dt
>
1 +
2 1
t
"

2
Z
t
1 T
1 + t
EQ et

r
+
(

)
+
1
2
1
2 0
2
1 + t
#

1
1 r
t
2 (1 r) dt.
<
<

2 1
2
1 1 + t

Using (4.9) and (4.4), we have

dt = (1 + t )dt +

2 1
t .
t d B

(4.21)

Therefore

t = exp



  Z t

 
(2 1 )2
(2 1 )2
2 1
2 1
Bt +
B

exp

u du.
u

2
0

t
is known: See Yors result (8.3) quoted below. We denote by g(a, t) the density of
The law of

finally obtain:

t
.

We

V (x) = log(x) + rT

Z T Z "
a
2

+
1 r + (2 1 )

2 1 + r
1+a
2
0
0
a>

2 2 + r
#

2
1
a

1 r + (2 1 )
1
1 r
2 1 + r
2
1+a

<a<
r
2 2 + r
2
et (1 + a)g(a, t)dadt.

4.5

Numerical Examples

Fig. 7(a) shows a nominal trajectory of the price St . Fig. 7(b) shows the corresponding optimal allocation
strategy. Fig. 7(c) shows the logarithm of the wealth. We keep the parameters used to obtain Fig. 2, that is,
1
2

= 0.2
= 0.2

= 2
r
= 0.15 T

= 0.0
= 2.0

Fig. 8 shows the function t 7 E log(Wt ) in the case where the trader uses the optimal allocation strategy
with initial wealth x = 100.
18

95

Proportion in risky asset

St

90
85
80
75
70
65
60

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

1.4

Proportion
in the stock

1.2
1
0.8
0.6
0.4
0.2
0

=0.78

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time

(a)

(b)

4.95

log of wealth

4.9

log(Wt )

4.85
4.8
4.75
4.7
4.65
4.6
4.55
4.5

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time
(c)

Figure 7: A Nominal Trajectory

Two Model and Detect Strategies

The optimal portfolio allocation strategy in the previous section supposes that the trader is able to perfectly
detect the time , which is both the time change of the drift term of the model, and the time at which the
trader reinvests his portfolio. In this section we describe the wealth of a trader who does not perfectly detect
but, at least, uses an optimal detection procedure to decide when he reinvests his portfolio. In order to
facilitate the comparison with the technical analyst, we here assume that the portfolio weights do not satisfy
a maximisation procedure. Rather, they will be arbitrarily set to = 0 or 1. We continue to suppose that
the trader perfectly knows all the parameters of the model.
We consider two detection methods: the first one has been proposed by Karatzas [6], and the other one
has been proposed by Shiryayev [11]. These two methods aim to find a stopping rule that detects the
instant . The time is interpreted as the time of alarm, it can occur before (in this case, it corresponds
to the false alarm), or after . So, the amount of time by which the stopping rule misses the change
point is given by | |.

5.1

The Detection Procedures

We first adapt Karatzass method to compute the optimal stopping rule K that minimizes the expected
miss
R() := E| |
(5.1)

19

4.85

Log of wealth

Optimal trader
4.8

4.75

4.7

4.65

4.6

0.2

0.4

0.6

0.8

Time

1.2

1.4

1.6

1.8

Figure 8: Optimal Allocation Strategy

over all stopping rules , when is assumed to have the a priori exponential distribution (2.3). Supposing
2
2
< r < 2 , we compute the wealth of the trader who uses the strategy which consists in putting
1
2
2
all of his money in the bond until K , and in the stock after K :
WT =

xS0K
ST 1(K T ) + xST0 1(K >T ) .
S K

The detection method proposed by Shiryayev (namely the Variant B in [11]) consists in computing
B(c) := inf {P( < ) + cE( )+ }

and the corresponding optimal stopping time for a given c > 0.

Remark 5.1. We finally mention a recent work by Beibel [2] who considered the model (2.1) with


2
2
1 2 r 2 2 . He studied the problem of maximizing E er S 1(<) over all stopping times
adapted to the filtration generated by (St ). We do not study the resulting detection procedure here.

5.2

The Particular Case of the Logarithmic Utility Function

In this section, we restrict ourselves to the detection procedure introduced by Karatzas [6].
The key point is the computation of the optimal stopping rule K . The proof of the following proposition
is postponed to the appendix.
Proposition 5.2. Consider the process R defined by (2.2), being a random time which cannot be
detected exactly and has the a priori distribution (2.3). The stopping rule K which minimizes the expected
miss E| | over all the stopping rules with E() < is




Z t

p
p
 t
s 1
K
= inf t 0 e Lt
+
e
Ls ds
,
1p
1 p
0
where Lt is defined as in (2.5) and p is the unique solution in ( 21 , 1) of the equation
Z
0

1/2

(1 2s)e/s 2
s
ds =
(1 s)2+

with = 2 2 /(2 1 )2 .
20

1/2

(2s 1)e/s 2
s
ds
(1 s)2+

Proof. We adapt Karatzass method in [6] to our specific case. Denote by S the collection of stopping rules
: [0, ) such that E() < . Rewrite (5.1) as follows:


R() = E ( )+ + ( )+ = E( )+ + E( ) E( ).
Then, using (2.3) and the notation (2.4), we get
1p
R() =
+E

1p
+E

Z

1( s) ds

1( >s) ds

1p
=
+ 2E

21( s) 1 1(s) ds

1
Fs
2


ds.

We thus obtain
1p
+ 2 inf E
R(p) := inf R() =
S
S

1
Fs
2


ds.

It now remains to follow Karatzas arguments (see [6]).

The terminal wealth of a trader who uses this detection procedure is
K

WT = xer exp (BT BK ) + (1

2
)(T K )
2
!

1(K T )

+ (2 1 )[(T )+ (K )+ ]
+ x exp(rT )1(K >T ) .

If one uses Shiryayevs detection method instead of Karatzasone, the optimal stopping time which
minimizes


B(c) = inf P( < ) + cE( )+

is
S (B ) := inf{t 0; Ft B }
for all c > 0, where F is the conditional a posteriori probability solution of (4.7) and the parameter B is
defined as the root in (0, 1) of the equation
Z
0

2 2
1
exp
2
(2 1 ) y


1
y(1 y)2

y
1y

2 2
(2 1 )2

(2 1 )
2
1
exp
2 2 c
(2 1 )2 B

dy



B
1 B

2 2
(2 1 )2

For Karatzas and Shiryayevs detection procedures, we are able to write an explicit formula for E(log(WT ))
(similar to the formula in Proposition 3.1). Given that the formula is complex, we do not include it here but
it is available upon request.

21

95

Proportion in risky asset

St

90
85
80
75
70
65
60

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

1.2
1
0.8
0.6
0.4
0.2
0

=0.78

Karatzas trader
Shiryaev trader

1.4

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time

(a)

(b)

4.95

Karatzas trader
Shiryaev trader

4.9

log of wealth

4.85
4.8
4.75
4.7
4.65
4.6
4.55
4.5
4.45

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time
(c)

Figure 9: A Nominal Trajectory

5.3

Numerical Examples

Figs. 9(b) and 9(c) show the allocation strategies of traders using Karatzas and, respectively, Shiryayev
detection procedures. The underlying trajectory of the stock price is shown in Fig. 9(a). Here, c = 3.0 and
the other parameters are described below:
1
2

= 0.2
= 0.2

= 2
r
= 0.15 T

= 0.0
= 2.0

Fig. 10 shows the time evolution of the respective expectations of wealths: the curves are merged. The
two detection strategies have the same performances for these values of parameters.

5.4

Comparison with Technical Analysis Techniques

We consider the ideal case where the parameters are perfectly known and numerically compare the performances of two change time detection strategies (one based on technical analysis, the other one based on
Karatzas or Shiryayevs detection rule), and the optimal portfolio allocation strategy. Fig. 11 show the time
evolution of the performances of these various strategies.
We present in the Appendix 8.3 a serie of numerical studies. We take different values of the parameters
and we compare the performances of these four strategies.
We observe:
When the volatility increases, the technical analysts performance seems to decrease even if he uses a
22

E[log(Wt )]

4.85

K. Model and detect

4.8

S. Model and detect

4.75

4.7

4.65

4.6

0.2

0.4

0.6

0.8

Time

1.2

1.4

1.6

1.8

Figure 10: Karatzas and Shiryayev Model and Detect Strategies

4.85

Optimal allocation
K. Model and detect
Technical Analyst

E[log(Wt )]

4.8
4.75
4.7

4.65
4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time

Figure 11: Comparison

good windowing size . When the volatility is small, a technical analyst using an optimized windowing
size can succeed to have a performance close to the optimal one.
When 2 is small, the strategy using Shiryayev detection rule is sensitive to the choice of the parameter
c. However the performances of Shiryayev Model and Detect strategy remain close to the optimal
allocation strategy performances.

6
6.1

Misspecified Models
Introduction

In practice, it is extremely difficult to know parameters exactly. If one may hope to calibrate 1 and
relatively well owing to historical data, the value of 2 cannot be determined a priori (i.e. before the
occurrence of the drift change), and the law of cannot be calibrated accurately because of the lack of data
concerning .

6.2

Misspecification of the Parameters

We consider the case where each parameter is estimated with lack of precision. The trader believes that the
stock price is


dSt = St 2 + (1 2 )1(t ) dt + St dBt ,
(6.1)
where the law of is exponential with parameter . We suppose that the true stock price is given by (2.1).
Our aim is to study the misspecified optimal allocation strategy and the misspecified model and detect
strategy.
23

Notation. As above we set Rt = log(St ), where St is the actual price. One need to compute Lt and Ft to
apply the optimal allocation strategy and the model and detect strategies. Here, we define the approximated
quantities computed when the model is misspecified Lt and F t as follows:


 
1
1
2
2
Lt = exp
( 1 )Rt 2 (2 1 ) + 2(2 1 )(1
) t ,
2
2 2
2
Rt
1
et Lt 0 es Ls ds
.
Ft =
Rt
1
1 + et Lt 0 es Ls ds

6.3

On the Misspecified Optimal Allocation Strategy

Observing the stock price St , the trader computes a pseudo optimal allocation by using the erroneous
parameters 1 , 2 , and . Thus, the value of his misspecified optimal allocation strategy is
t = proj[0,1]

(1 r + (2 1 )F t )
,
2

and the corresponding wealth is

Wt

rt

Z

= e exp

u d(eru Su )

6.4

On Misspecified Model and Detect Strategies

The erroneous stopping rule is

K


Z t
1
t
= inf t 0, e Lt
es Ls ds
0

p
1 p

1
where p is the unique solution in ( , 1) of the equation
2
Z 1/2
Z p
(1 2s)e/s 2
(2s 1)e/s 2
ds =
ds
s
s
(1 s)2+
(1 s)2+
0
1/2
with = 2 2 /(2 1 )2 .
The value of the corresponding portfolio is
0
W T = xS
K

6.5

ST
1 K
+ xST0 1(K >T ) .
SK ( T )

Numerical Example

We now numerically compare the performance of two traders who respectively use a misspecified model and
the true model.
We first compare the performance when parameters 1 , and are well estimated and the error is only
on 2 . 1 = 0.2, = 0.15, r = 0.0 and = 2.0, and we assume that these values are perfectly known by
the trader. A contrario 2 is misspecified. Its true value is 2 = 0.2. Figure 12 (optimal allocation strategy)
and Figure 13 (model and detect strategy using Karatzas procedure) show the functions t E(log(Wt ))
for three values of 2 . It suggests that it is better to overestimate 2 (2 > 2 ) than to underestimate it
(2 < 2 ).
We now compare numerically the performance when only parameter is misspecified.

24

2 = 0.2
2 = 0.3
2 = 0.1

4.85

E[log(Wt )]

4.8
4.75
4.7
4.65
4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time

Figure 12: Misspecified Optimal Allocation Strategy - Erroneous 2

2 = 0.2
= 0.3
2
2 = 0.1

4.85

E[log(Wt )]

4.8

4.75

4.7

4.65

4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time
Figure 13: Misspecified Karatzas Model and Detect Strategy - Erroreneous 2
We choose exact values as above. Fig. 14 and 15 show the functions t E(log(Wt )) for three values
of . With the optimal strategy, the trader have better performance if he/she overestimate than he/she
underestimate it (Fig. 14). With the Model and Detect strategy, we have the converse conclusion but the
three performances are very close in this last case.

6.6

A Comparison with Technical Analysis Techniques

We can now address our main question: Is it better to invest according to a mathematical strategy based on a
misspecified model, or according to a strategy which does not depend on any mathematical model? Because
of the analytical complexity of all the explicit formulae that we have obtained for the various expected
utilities of wealth at maturity, we have not yet succeeded to find a mathematical answer to this question,
1
even in asymptotic cases (when 2
is large, e.g.). Since this part of our work is still in progress, we
2
present here a few numerical results obtained from Monte Carlo simulations. Consider the following case
study.
Parameters of the model
1
2

r
True values
0.2 0.2 2 0.15 0.0
Parameters used by the trader
1
2

r
Misspecified values (case I)
0.3 0.1 1.0 0.25 0.0
Misspecified values (case II)
0.3 0.1 3.0 0.25 0.0
Figure 16 shows that the technical analyst overperforms misspecified optimal allocation strategies when
25

= 2.0

= 3.0

= 1.0

4.85

E[log(Wt )]

4.8

4.75

4.7

4.65

4.6

0.2

0.4

0.6

0.8

Time

1.2

1.4

1.6

1.8

Figure 14: Misspecified Optimal Allocation Strategy - Erroneous

= 2.0

= 1.0

= 3.0

E[log(Wt)]

4.85

4.8

4.75

4.7

4.65

4.6

0.2

0.4

0.6

0.8

1.2

Time

1.4

1.6

1.8

Figure 15: Misspecified Karatzas Model and Detect Strategy - Erroreneous

the parameter is underestimated.
We have looked for other cases where the technical analyst is able to overperform the misspecified optimal
allocation strategies. Consider the case where the true values of the parameters are given in Table 1. Table 2
summarizes our results. It must be read as follows. For the misspecified values 2 = 0.1, = 0.25, = 1,
if the trader chooses 1 in the interval (0.5, 0.05) then the misspecified optimal strategy is worse than
the technical analysts one. In fact, other numerical studies show that a single misspecified parameter is not
sufficient to allow the technical analyst to overperform the Model and Detect traders.
Surprisingly, other simulations show that the technical analyst may overperform the misspecified optimal
allocation strategy but not the misspecified model and detect strategy. One can also observe that, when
2 /1 decreases, the performances of well specified and misspecified model and detect strategies decrease:
see Figs. 17-19.
Remark 6.1. Fig. 17 shows that, despite a large misspecification, model and detect strategies seem
to be able to keep good performances and the optimal trading strategy with misspecified parameters is
out-performed by the technical analysis strategy. This result suggests that statistical techniques or technical
anslysis approaches could be more attractive when the parameters are misspecified.
However, when the market first shows a strong downward drift and a small positive trend after the time
change, the misspecified model and detect strategies have low performances when is over-estimated. This
suggests that, in some circumstances, one has better to underestimate 1 .
1 Results

available upon request

26

4.85

MSP Case II
Technical
Analyst

E[log(Wt )]

4.8
4.75

MSP Case I
4.7

4.65
4.6

0.2

0.4

0.6

0.8

Time

1.2

1.4

1.6

1.8

Figure 16: Technical Analysis May Overperform Misspecified Optimal Allocation Strategies
Remark 6.2. We compare here the performance of the technical analyst using the best choice of
windowing size with misspecified mathematical strategies. Obviously, the technical analyst does not know
this best windowing but one can observe from Fig. 20, 21 and 22 that the performance of the technical
analyst remains close to the perfect knowledge case.
In conclusion, the figures 17, 18, and 19 suggest that there is no universal solution to the problem of
misspecification. It seems that when the drifts are significantly strong on both side and in particular the
upward drift, the performance of the Model and Detect strategies can be quite robust, yet their performance
seems to deteriorate quickly when lambda is strongly miss-specified and the upward drift is not very strong.
Since the second drift is in fact the hardest to estimate because we will not have any a priori information,
we take figures 18 and 19 as a caution for the potential demise of a Model and Detect or optimal strategy in
favor of the technical strategy. This discussion allows us to understand the potential benefits of a technical
strategy by better understanding the limitations of mathematical approaches which are parameterized.

27

4.85

4.85

technical analysis
exact parameters
misspecified parameters

4.8

4.8
4.75

E[log(W )]

E[log(WT)]

4.75

technical analysis
exact parameters
misspecified parameters

4.7

4.7

4.65

4.65

4.6

4.6

4.55
0

0.5

1
Time (years)

1.5

4.55
0

0.5

(a) Karatzas

4.85

exact
parameters

E[log(WT)]

4.75

= 0.2
12 = 0.2
= 0.15
=2

4.7

4.65
4.6
4.55
0

0.5

1.5

(b) Shiryayev

technical analysis
exact parameters
misspecified parameters

4.8

1
Time (years)

1
Time (years)

1.5

misspecified
parameters
_
_1 = 0.3
_2 = 0.1

_ = 0.25
=1

(c) Optimal

Figure 17: Strategies With Exact and Misspecified Parameters

Table 1: True Values of the Parameters

Parameter
1
2

True Value
-0.2
0.2
0.15
2

Table 2: Misspecified Values and Range of the Parameters

1
2

(-0.5,-0.05)
0.1
0.25
1

1
2

-0.3
(0,0.13)
0.25
1

28

1
2

-0.3
0.1
(0.2,)
1

1
2

-0.3
0.1
0.25
(0,1.5)

E[log(WT)]

4.7

4.75

technical analysis
exact parameters
misspecified parameters

4.7
E[log(W )]

4.75

4.6

0.5

1
Time (years)

1.5

(a) Karatzas

4.75

E[log(WT)]

4.7

4.5
0

0.5

technical analysis
exact parameters
misspecified parameters

exact
parameters
= 0.3
12 = 0.1
= 0.15
=2

4.6

4.55

0.5

1
Time (years)

1
Time (years)

1.5

(b) Shiryayev

4.65

4.5
0

4.6
4.55

4.55
4.5
0

4.65

4.65

technical analysis
exact parameters
misspecified parameters

1.5

misspecified
parameters
_
_1 = 0.1
_2 = 0.3

_ = 0.25
=1

(c) Optimal

Figure 18: Strategies With Exact and Misspecified Parameters

29

4.7

technical analysis
exact parameters
misspecified parameters

E[log(WT)]

4.7

4.75

technical analysis
exact parameters
misspecified parameters
E[log(W )]

4.75

4.65

4.6

4.6

4.55
0

0.5

1
Time (years)

1.5

(a) Karatzas

4.75

technical analysis
exact parameters
misspecified parameters

0.5

exact
parameters
= 0.3
12 = 0.1
= 0.15
=2

4.65

4.6

4.55
0

4.55
0

0.5

1
Time (years)

1
Time (years)

1.5

(b) Shiryayev

E[log(WT)]

4.7

4.65

1.5

misspecified
parameters
_
_1 = 0.1
_2 = 0.3

_ = 0.25
=4

(c) Optimal

Figure 19: Strategies With Exact and Misspecified Parameters

30

Conclusion and Perspectives

We have compared trading strategies designed from possibly misspecified mathematical models with trading
strategies based on technical analysis techniques. We have made explicit the traders expected logarithmic
utility of wealth in all cases under study. Unfortunately, the explicit formulae are not propitious to mathematical comparisons. Therefore, we have used Monte Carlo numerical experiments, and observed from these
experiments that technical analysis techniques may overperform mathematical techniques in the case of severe parameter misspecifications. Our study also brings some information on the range of misspecifications
for which this observation holds true.
We are now considering the case where the instantaneous expected rate of return of the stock changes
at the jump times of a Poisson process, and the value of this rate after each time change is unknown. We
follow two new directions: stochastic control techniques for switching models and, jointly with M. Martinez
(INRIA) and S. Rubenthaler (University of Nice Sophia Antipolis), filtering techniques.
We also plan to consider other technical analysis indicators than the moving average indicator.

31

8
8.1

Appendix
Proof of Proposition 3.1

We recall that


2
+
1
t + (2 1 ) (tj+1 max(, tj ))
2

+ (Btj+1 Btj ) .

Stj+1
= exp
Stj

(8.1)

Now, using (3.3) and (8.1), we obtain


E log

WT
x


=

N
1
X






E log tj Stj+1 /Stj exp(rt) + exp(rt)

j=0

N
1
X

rtE

1(tj =0)

j=0

N
1
X

(
E

1(tj =1) (1

j=0

2
)t
2
+

+ (2 1 ) (tj+1 max(, tj )) + (Btj+1


Btj )

In view of the definition of tj in (3.2) we have


E log

WT
x


=


N
1 

X
2
r t
P Stj Mtj
1
2
j=0

+ (2 1 )

N
1
X

1(Stj Mt ) (tj+1 max(, tj ))+

j=0

+ rT +

N
1
X

n
E

1(Stj Mt ) (Btj+1 Btj )

j

j=0

Notice that the last term above is equal to zero since Btj+1 Btj is independent of Gtj , and thus of 1(St Mt ) .
j
j
In addition, as the Brownian motion B and the change point are independent, one has
Z
n
o
+
+
E 1(St Mt ) (tj+1 max(, tj ))
=
(tj+1 max(u, tj ))
j

R+



P Stuj Mtu,
eu du,
j
where Stuj and Mtu,
respectively are
j


2
Stuj = exp (1
)tj + (2 1 )(tj u)+ + Btj ,
2
Z tj
1
Mtu,
=
S u ds.
j
tj s

32

(8.2)

Thus, (8.2) becomes

n
E

1(Stj Mt ) (tj+1 max(, tj ))+ = t

tj

0
tj

+ t
tj

tj+1

+
tj



P Stuj Mtu,
eu du
j



P Stuj Mtu,
eu du
j



(tj+1 u)P Stuj Mtu,
eu du
j

:= I1 (j) + I2 (j) + I3 (j).

On I1 (j) : when u is less than tj we have
)
(
Z
n
o
1 tj Ssu
u,
u
ds
Stj Mtj
= 1
tj Stuj
(


 )
Z


1 tj
2
= 1
exp
2
(s tj ) + Bs Btj
ds
tj
2
Set
Bs Btj = (Bs Btj ) (Btj Btj ).
Use the change of variable s s (tj ) and the identity in law
L

s .
Btj +s Btj = B
It comes:
i
h
(1)
p = P Stuj Mtu,
j
"




 #

Z
2
1

2

s ds .
+ B
s + B
exp
2
= P exp
2
2
0
2
The right handside does not depend neither on tj nor on u). Therefore
Z tj
(1)
I1 (j) = t p
eu du
0


(1)
= t 1 e(tj ) p
We now compute the sum over j:
N
1
X
j=0




1 eN t
(1)
.
I1 (j) = p t N e
1 et
(1)

In order to compute the probability p , we use the following Lemma (see Yor [12] or BorodinSalminen [4,
formula 1.20.8 p.618]).
Lemma 8.1. Let B be a real Brownian Motion. Let > 0 and be in R. Let V be a geometric Brownian
Motion:
2
Vs = e s+Bs .
It holds that
Z


Vs ds dy ; Vt dz

P
0

2 2 t (1 + z 2 )



z
2
2 2 y i 2
=
e
t 2 y dydz,
2y
2
z 1

33

(8.3)

where

ze /4y

iy (z) :=

ez cosh uu

/4y

sinh u sin(u/2y)du.

We deduce:
" Z
# Z Z
2 2 (1 + z 2 )


1

z
1
z
2
2 2 y i2 /2
Vs ds V =
P
e
dzdy.
0
2y
2 y
y
0
We now use (8.4) with =
N
1
X
j=0

2
2

(8.4)

12 . It comes:




1 eN t
I1 (j) = t N e
1 et
Z

z 2 3/2
(2 / /2)
(1 + z 2 )
exp

2y
2
2 2 y


i2 /2

z
2 y


dzdy.

On I2 (j): For u [tj , tj ] one has






2
2
u
u + 2
(tj u) + Btj
Stj = exp
1
2
2
"Z





 #
Z tj
u
1
2
2
u,
Mtj =
exp
1
s + Bs ds +
s + Bs ds .
exp
2
tj
2
2
u

(8.5)
(8.6)

Use the change of variable s s (tj ) and the identity in law

L
Btj +s Btj = B
s.

Set u
= u (tj ). It comes
h

"

P Stuj Mtu,
= P exp
j



1 2 /2 u
+ 2 2 /2 ( u
) + B

exp




s ds
1 2 /2 s + B

exp




s ds
2 2 /2 s + B

!#

"
= P exp





B
u
2 2 /2 ( u
) + B



Z u
2
1



exp
(

)s
+

B
1
s ds
2
2
0

exp
1
u
+ Bu
2
!#
Z




1
2
s B
u ) ds .
+
exp 2 /2 (s u
) + (B
u


s+u B
u is a Brownian Motion independent of (B
v ; 0 v u
We again use that B
). As in Lemma 8.1 set
"
#
Z u
Z
h
i
1
1 u (2)
(2)
u,
u
(1)
P Stj Mtj = P Vu
Vs ds +
Vs ds ,
(1)
0
0
Vu
34

1
1
where V (1) and V (2) are two independent geometric Brownian Motions with parameters 1 = 2 and

2
1
2
2 = 2 . Then a straightforward computation leads to

2


N
1
X
1 eN t

I2 (j) = te
1 et
j=0
2

R4

1(

y2

z1
+z2
y1

(2 / /2) ( v) (1 + z22 )



z2
2y
2
2
2
i2 (v)/2
e
2 y2

3/2
z 2
) 2

2y2

(1 / /2) v (1 + z12 )



z1
2
2 2 y1 i2 v/2
e
2 y1

3/2
z1 1

2y1
v

dy1 dz1 dy2 dz2 dv



1 eN t
(2)

p .
= te
1 et
On I3 (j): We proceed as above (computation of I1 (j). We observe remark that P(Stuj Mtu,
) does not
j
depend on u [tj , tj+1 ], so that
N
1
X

I3 (j) =

j=0

1 eN t
t e t + 1
t
1e
2

(1 / /2) (1 + z 2 )



z
2
2 2 y i2 /2
e
dzdy
2y
2 y
0
y

(3)
1 eN t
t et + 1 p .
=
1 et
Z Z

z 1 3/2

Finally, we observe that

P(Stj

Mtj )

tj

P(Stuj

=
0

Mtu,
)eu du
j

tj

+
tj

P(Stuj Mtu,
)eu du
j

+
tj

P(Stuj Mtu,
)eu du
j

(1)

(2)

(3)

= p (1 e(tj ) ) + p e(tj ) + p etj



(1)
(2)
(1)
(3)
= p + etj (p p )e + p .
We conclude:
 



 

WT erT
2
2
1 eT  (2)
(1)
(1)
(3)
E log
= 2
r T p + t 2
r
(p

p
)e
+
p

x
2
2
1 et
t(2 1 )(et t)

1 eT (3)
p .
1 et

References
[1] S. Achelis. Technical Analysis from A to Z. McGraw Hill, 2000.
35

[2] M. Beibel and H. R. Lerche. A new look at optimal stopping problems related to mathematical finance.
Statist. Sinica, 7(1):93108, 1997.
Tanre. Technical analysis techniques versus
[3] C. Blanchet-Scalliet, A. Diop, R. Gibson, D. Talay, and E.
mathematical models: Boundaries of their validity domains. In H. Niederreiter and D. Talay, editors,
Monte Carlo and Quasi-Monte Carlo Methods 2004. Springer, 2005 (to appear).
[4] A. N. Borodin and P. Salminen. Handbook of Brownian Motion Facts and Formulae. Probability
and its Applications. Birkh
auser Verlag, Basel, second edition, 2002.
[5] I. Karatzas. Lectures on the Mathematics of Finance, volume 8 of CRM Monograph Series. American
Mathematical Society, Providence, RI, 1997.
[6] I. Karatzas. A note on Bayesian detection of change-points with an expected miss criterion. Statist.
Decisions, 21(1):313, 2003.
[7] I. Karatzas and S. E. Shreve. Methods of Mathematical Finance, volume 39 of Applications of Mathematics (New York). Springer-Verlag, New York, 1998.
[8] A. W. Lo, H. Mamaysky, and J. Wang. Foundations of technical analysis: Computational algorythms,
statistical inference, and empirival implementation. Journal of Finance, LV(4):17051770, 2000.
[9] R. C. Merton. Optimum consumption and portfolio rules in a continuous-time model. J. Econom.
Theory, 3(4):373413, 1971.
[10] A. N. Shiryaev. On optimum methods in quickest detection problems. Theory Probab. Applications,
8:2246, 1963.
[11] A. N. Shiryaev. Quickest detection problems in the technical analysis of the financial data. In Mathematical Finance Bachelier Congress, 2000 (Paris), Springer Finance, pages 487521. Springer, Berlin,
2002.
[12] M. Yor. On some exponential functionals of Brownian motion. Adv. in Appl. Probab., 24(3):509531,
1992.

36

Figures for optimal windowing

1=0.22=0.2
4.9

1=0.12=0.2
4.9
E[log(W )]

E[log(WT)]

=0.1
=0.2

4.7

4.6
0

4.8

0.5

1.5

4.7

4.6
0

1=0.22=0.1

4.7

4.75

=0.1
=0.2

4.65

1.5

4.7

=0.1
=0.2

4.65

4.6

4.55
0

0.5

1=0.12=0.1

E[log(WT)]

4.75

=0.1
=0.2

4.8

E[log(WT)]

8.2

0.5
1
1.5
delta (years)

4.6

4.55
0

0.5
1
1.5
delta (years)

Figure 20: Volatility and Optimal Windowing: E(log(WT )) T = 2, = 1

37

1=0.22=0.2

1=0.12=0.2
4.9

=0.1
=0.2

4.8

E[log(WT)]

E[log(WT)]

4.9

4.7

4.6
0

4.8

4.7

4.6
0.5

1.5

1=0.22=0.1
4.75

0.5

1.5

1=0.12=0.1
4.75

=0.1
=0.2

4.7

E[log(W )]

=0.1
=0.2

4.7

E[log(WT)]

=0.1
=0.2

4.65
4.6
4.55
0

0.5

1.5

4.65
4.6
4.55
0

0.5

delta (years)

1.5

delta (years)

Figure 21: Volatility and Optimal Windowing: E(log(WT )) T = 2, = 2

1=0.22=0.2

4.9

4.8

0.5

1.5

4.7
0

1=0.22=0.1
=0.1
=0.2

1.5

=0.1
=0.2

4.8
4.75

E[log(WT)]

4.75
4.7

4.65
4.6
0

0.5

1=0.12=0.1

E[log(W )]

4.8

=0.1
=0.2

4.9

4.8

4.7
0

1=0.12=0.2

E[log(WT)]

=0.1
=0.2

E[log(WT)]

4.7

4.65
0.5
1
1.5
delta (years)

4.6
0

0.5

1
1.5
delta (years)

Figure 22: Volatility and Optimal Windowing: E(log(WT )) T = 2, = 4,

38

8.3

Figures of Wealthes for Strategies With Exact Parameters

As announced in Section 5.4, we set here the figures corresponding to the performances of
1. Technical analysis strategy
2. Optimal allocation strategy
3. Shiryayev Model and Detect strategy
4. Karatzas Model and Detect strategy
for various values of the parameters (supposed to be perfectly known by all the traders).
1 = 0.2 ; 2 = 0.2 ; =0.1 ; = 1

1=0.1 ; 2=0.2 ; =0.1 ; =1

4.8

technical analysis
Optimal
Shiryaev
Karatzas

4.78

technical analysis
Optimal
Shiryaev
Karatzas

4.76

4.72

4.7

E[log(W )]

E[log(W )]

4.74

4.7

4.68

4.66

4.6
4.64

4.62

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

4.6

Time (years)

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

1.6

1.8

Time (years)

1=0.1 ; 2=0.1 ; =0.1 ; =1

1=0.2 ; 2=0.1 ; =0.1 ; =1

4.7

technical analysis
Optimal
Shiryaev
Karatzas

technical analysis
Optimal
Shiryaev
Karatzas

E[log(W )]

E[log(W )]

4.65

4.6

4.6

4.5

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

0.2

0.4

0.6

0.8

1.2

Time (years)

Time (years)

39

1.4

 =0.2 ; =0.2 ; =0.2 ; =1

1

1=0.1 ; 2=0.2 ; =0.2 ; =1

4.8

technical analysis
Optimal
Shiryaev
Karatzas

technical analysis
Optimal
Shiryaev
Karatzas

4.7

E[log(WT)]

E[log(W )]

4.7

4.6
4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

0.2

0.4

0.6

Time (years)

0.8

1.2

1.4

1.6

1.8

1.6

1.8

1.6

1.8

Time (years)

1=0.2 ; 2=0.1 ; =0.2 ; =1

1=0.1 ; 2=0.1 ; =0.2 ; =1

technical analysis
Optimal
Shiryaev
Karatzas

technical analysis
Optimal
Shiryaev
Karatzas

E[log(W )]

E[log(WT)]

4.65

4.6
4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

0.2

0.4

0.6

0.8

Time (years)

1=0.2 ; 2=0.2 ; =0.1 ; =2

4.9

1.2

1.4

1=0.1 ; 2=0.2 ; =0.1 ; =2

4.9

technical analysis
Optimal
Shiryaev
Karatzas

technical analysis
Optimal
Shiryaev
Karatzas
4.8

E[log(WT)]

4.8

E[log(W )]

Time (years)

4.7

4.7

4.6

4.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time (years)

0.2

0.4

0.6

0.8

1.2

Time (years)

40

1.4

 =0.2 ; =0.1 ; =0.1 ; =2

1

1=0.1 ; 2=0.1 ; =0.1 ; =2

technical analysis
Optimal
Shiryaev
Karatzas

4.72

technical analysis
Optimal
Shiryaev
Karatzas

4.7

4.7

E[log(W )]

E[log(W )]

4.68

4.66

4.64

4.6

4.62

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

4.6

Time (years)

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

1.6

1.8

1.6

1.8

Time (years)

1=0.2 ; 2=0.2 ; =0.2 ; =2

1=0.1 ; 2=0.2 ; =0.2 ; =2

4.9

4.85

technical analysis
Optimal
Shiryaev
Karatzas

technical analysis
Optimal
Shiryaev
Karatzas

4.8

E[log(W )]

4.75

E[log(W )]

4.8

4.7

4.7

4.6
4.65

4.5

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

4.6

0.2

0.4

0.6

Time (years)

0.8

1.2

1.4

Time (years)

1=0.2 ; 2=0.1 ; =0.2 ; =2

1=0.1 ; 2=0.1 ; =0.2 ; =2

4.7

technical analysis
Optimal
Shiryaev
Karatzas

4.7

technical analysis
Optimal
Shiryaev
Karatzas

4.68

4.66

4.64

E[log(W )]

E[log(WT)]

4.65

4.62

4.6

4.6

4.58

4.56
0

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Time (years)

0.2

0.4

0.6

0.8

1.2

Time (years)

41

1.4