International Journal of Economics and Finance; Vol. 5, No.

3; 2013
ISSN 1916-971X E-ISSN 1916-9728
Published by Canadian Center of Science and Education
221

An Investment Strategy Based on Stochastic Unit Root Models
Mamadou A. Kont
1

1
Universit Gaston Berger, UFR Sciences Economiques et de Gestion, BP 234 Saint Louis, Sngal
Correspondence: Mamadou A. Kont,

Universit Gaston Berger, UFR Sciences Economiques et de Gestion,
BP 234 Saint Louis, Sngal. Tel: 221-77-622-1350. E-mail: mamadou-abdoulaye.konte@ugb.edu.sn or
kontedoudou@yahoo.fr

Received: September 29, 2012 Accepted: January 17, 2013 Online Published: February 26, 2013
doi:10.5539/ijef.v5n3p221 URL: http://dx.doi.org/10.5539/ijef.v5n3p221

Abstract
An algorithm is presented that locally approximates the nonlinearity of stochastic unit root (STUR) models by n
linear models. The previous integer n is chosen so that the Hadamard matrix of order n can be defined. The
strategy STUR(n), then consists in creating n linear models from this Hadamard matrix and taking their average
forecast. A purchase (sell) signal is made if the obtained average forecast is positive (negative). Subsequently, a
comparison is made with respect to competing models (Moving average strategies) to assess their ability to
forecast the variation of five international indexes. It is found, after taking account transaction costs, that STUR(n)
generates generally the highest profitability in the out-of-sample data.
Keywords: forecasting, trading rules, random coefficient autoregressive models, efficiency market hypothesis
1. Introduction
The question of Efficiency Market Hypothesis (EMH) has been studied for many years by both academics and
market participants. The aim is to see if the assumptions of market frictionless and traders rationality are a good
description of real markets where microstructure (transaction costs, information asymmetry, etc.) and noise
traders are present. This is an ongoing debate and there has been no consensus. That is why some authors have
tried to reconcile the EMH and Behavioral finance arguments through dynamic systems, see for example Lo
(2005) and Kont (2010). The empirical studies of this hypothesis are based generally on three classes. The first
is traditional regression models. Their aim is to test the validity of the EMH in its weak form through traditional
time series forecast such as Auto Regressive Moving average models ARMA(p,q). If the market is supposed to be
a nonlinear dynamic system, one may consider nonlinear models such as the Random Coefficient Autoregressive
RCA(p) or regime switching models among others. The traditional regression models also contain analysis tools
based on firms fundamental (dividend, Book-to-Market, etc.). In this case, the objective is to test the EMH in its
semi-strong form (fundamental analysis). We refer to Ou and Penman (1989) and references therein.
The second class uses Technical Analysis tools such as Moving Average, Support and Resistance methods. This
approach is widely applied by traders to detect trends or reversal effects by using information such that prices,
trading volume, etc. Here, the validity of EMH is tested through its weak form, see for example Sullivan et al.,
(1999).
The last class, based on Machine Learning (Genetic Algorithms, Neural Networks methods) investigates the
EMH in its weak and semi-strong form as for the class of traditional regression models. Their difference is that
Machine learning models are self-adaptive methods in that there are few a priori assumptions about the
relationship between inputs while the traditional regression models make strong assumptions (parametric
approach).
The paper belongs to the first class where the nonlinearity of financial asset prices is modeled by RCA(p). This
econometric model generates the main stylized facts of financial time series, see Yoon (2003). It may be also
related to an Agent Based Model with a switching phenomenon between fundamentalists and noise traders, see
for example Kont (2011). There are many methods proposed in the literature to estimate its parameters for
trading or forecast purpose. For example Nicholls and Quinn (1981) employed the traditional least squares and
the maximum likelihood methods, see also Granger and Swanson (1997). Wang and Ghosh (2002) use Bayesian
approach while Sollis et al. (2000) work with Kalman filter. We follow here another approach consisting to
approximate the RCA(1) model by n simple linear models where n is any integer such that the Hadamard matrix
www.ccsenet.org/ijef International Journal of Economics and Finance Vol. 5, No. 3; 2013
222

H of order n can be defined. The latter is an n n matrix with all its elements being either 1 or 1, and such that
HH
T
=n*I
n
where H
T
is the transpose of H, I
n
is the identity matrix of order n. Therefore, the Hadamard
matrix columns is an orthogonal binary basis of R
n
explaining why it is widely used in physics particularly in
the field of signal transmission. The integer n, in this study, must satisfy the constraint n , n/12 or n/20 is a power
of 2. The prediction is then made by taking the average forecast of these n linear models extracted from the
Hadamatrix since many researchers agree that combining multiple forecasts leads to increased accuracy, see
Granger and Ramanathan (1984).
The paper contributes in two ways to the literature. First, contrarily to other forecasting methods, the estimation
procedure is made locally to capture traders feedback or interaction since the variance and other higher
moments of STUR model do not exist. Only n data are used in the linear regression models where 8n50. This
constraint gives us exactly height (08) strategies STUR(n) with n{8, 12, 16, 20, 24, 32, 40, 48}. The second
contribution shows an application of the Hadamard basis to reduce the complexity of a problem (from
exponential to linear) for trading purposes.
The paper is divided into four additional sections. Section 2 presents our methodology and its competing
strategies to forecast the variation of asset prices of five international indexes (CAC 40, DAX 30, FTSE 100,
Nikkei 225, S&P 500). Section 3 describes the data and the methodology used in the empirical application.
Section 4 presents the empirical results and the last section concludes.
2. Some Forecasting Rules
2.1 Our Methodology
Consider the following stochastic unit root STUR(1) model defined by:
t t t t
y b y r - -
-1
) 1 ( (1)
0 ) , cov( , ) ( , ) ( , 0 ) ( ) (
2 2 2 2

t t t t t t
b E b E E b E r o r c r
where (
t
) is an i.i.d Gaussian process and y
t
=logS
t
(log of asset prices).
The properties of eq. (1), to replicate financial times series, have been studied by (Yoon, 2003). The econometric
model is also related to an agent based model with interaction between fundamentalist and noise traders, see
(Kont, 2011). It is a special case of the Random Coefficient Autoregressive RCA(1) model which is defined by:
y
t
= (+b
t
)y
t1
+
t

. ) , cov( , ) ( , ) ( , 0 ) ( ) (
2 2 2 2
c r o r c r
t t t t t t
b E b E E b E
(Nicholls and Quinn, 1982) have shown that the RCA(1) process (y
t
) is a finite second-order stationary moment
if the condition
2
+
2
<1 is satisfied. Since =1 in our case, the stationary condition is violated (Note 1). That
means conventional methods based on this assumption such as Maximum likelihood method cannot perform, see
Yoon (2006) (Note 2). Consequently, we propose a methodology that locally approximates the nonlinearity of
asset prices by n linear models. For this purpose, b
t
is supposed to take only two values and at any time.
Therefore, it may be rewritten as b
t
=X
t1
where for any t, X
t1
=1 or X
t1
=1. The equation (1) becomes

t t t t t t t t
y X c y y S S r r o - - - -
- - - - 1 1 1 1
log log (2)
where a constant c is added, as usual, in the regression model.
For the moment, the estimation cannot be proceed because the variable X
t1
is not known. To circumvent this
problem, regressions models are used conditional on the path of (X
t
). For example, in the equation (2), if it is
decided to use n data for the estimation process, we will have 2
n
paths for (X
t
) , t=1,,n since X
t
takes 1 or
1 at any time. Each trajectory generates a linear model with three input variables X
t1
, y
t1
and the constant
variable c. Therefore, the nonlinearity is approximated by 2
n
linear models (Note 3). This feature comes at a
cost, as we need to store a binary matrix of size n2
n1
to make all linear regressions (Note 4). Generally, we
need the parameter n to be big for estimation precisions but not too much to keep a local approximation. In the
application, it is taken 8n50. To solve the dimensionality problem, techniques similar to Component Principal
www.ccsenet.org/ijef International Journal of Economics and Finance Vol. 5, No. 3; 2013
223

Analysis (CPA) in exploratory data analysis are used. Namely, we extract "n orthogonal linear models". Here,
the orthogonality of two models i and j is defined by the orthogonality of their corresponding paths (X
i
tk
) and
(X
j
tk
), k=1,,n. If n is constrained to be an integer such that n/2, n/12 or n/20 is equal to 2
k
, kN, then
Hadamard matrices exist. For example for n=2, the Hadamard matrix is
(

1 1
1 1
2
H

that is a basis of R
2
. Recursively, we can define the matrix H
4
, H
8
, , by using the following formula
(

n n
n n
n
H H
H H
H
2

This approach allows to pass from exponential (2
n
) to linear (n) complexity since now for n data used in the
equation (2), n linear models are also employed where their paths correspond to the columns of the Hadamard
matrix of order n. For each model, determined by the path of (X
t
), the parameters and c are estimated by the
Ordinary Least Square method. Then a forecast is made at time t+1 through the equation
t t t t t
y X c y y r o
1 1
- -
- -
(3)
A recursive regression is applied. At any time, the previous n data are used in the regression model to
determine the new estimated parameters. We denote by ) (n SUR the strategy that consists to take the average
forecasts of all n "orthogonal linear models". The procedure to create the buy and sell signals is then simple: a
buy (sell) signal is produced if the average forecast, denoted by
1

- t
a r , is positive (negative). To reduce the
number of transaction costs, we enhance the strategy by allowing static positions in the case where the forecast
signal is not significant. In other words, the following strategy is applied for ) (n SUR .

-
- >

- -
. ) , 1 (
30 / )) min( ) (max( ; | ) | ) (
) , (
1 1
otherwise n t position
R R c c a r if a r sign
n t position
t t t t

where R
t
=logS
t
logS
t1
represents the index return at time t (Note 6). The sign function is defined by
sign(x)=1 if x>0, sign(x)=1 if x<0 and sign(0)=0.
2.2 Competing Trading Rules
If the market is supposed to be efficient, an optimal strategy is to buy and hold an index. The strategy B&H
consists therefore to be long on the index at any time and consequently there are no transaction costs. We
consider also simple and exponential moving average strategies that have been widely used by traders to capt
momentums or reversal effects. The idea is to consider two moving average series M(n,t) and M(m,t) with
different lengths n and m. If we denote by (S
t
) the asset price process, the simple and exponential moving
average are defined, for a given length k>0, by respectively the equations (4) and (5).

i t
k
i
S
k
t k M
-
-

1
0
1
) , (
(4)
1
) 0 , ( , ) 1 , ( ) 1 ( ) , ( S k M with S t k M t k M
t
- - - / / (5)
If m<n then M(m,t) (resp. M(n,t)) is called the short-term moving average (resp. the long-term moving average).
The decision rule for taking positions is specified as follows. If the short-term moving average M(m,t) intersects
the long-term moving average M(n,t) from below, a long position is taken. Conversely, if the M(n,t) is
intersected from above, a short position is taken. The moving average strategies are implemented by using the
Matlab function movavg. Note that in these strategies, transaction costs appear only when an intersection appears
between M(m,t) and M(n,t). In the decision making process of traditional regression models, if a threshold is not
used, the number of transactions may be very high.
www.ccsenet.org/ijef International Journal of Economics and Finance Vol. 5, No. 3; 2013
224

3. Data and Methodology
In this paper, we consider the daily closing prices of five international indexes CAC 40, Dax 30, FTSE 100,
Nikkei 225 and S&P 500 obtained from Yahoo Finance website. All time series have the same length of data as
shown by Table 1.

Table 1. Interval of study of the five international indexes
Index CAC 40 DAX 30 FTSE 100 NIKKEI 225 SP 500
in-sample 30 Jun 2000 09 Aug 2000 09 May 2000 04 Jan 2000 20 Apr 2000
19 Feb 2009 13 Feb 2009 30 Jan 2009 25 Dec 2008 30 Jan 2009
out-of-sample 20 Feb 2009 16 Feb 2009 02 Feb 2009 26 Dec 2008 02 Feb 2009
30 Dec 2011 30 Dec 2011 30 Dec 2011 30 Dec 2011 30 Dec 2011
Total data 2943 2943 2943 2943 2943

All the series end to 30 December 2011, totaling N=2943 trading days. Their difference appears only on the
beginning period where the latter is chosen so that to have the same length of data than the Nikkei index.
Each data is after divided into two periods: the first period (in-sample data) contains 2207 (0.75*N) trading days.
The remaining data (736 or approximatively 0.25*N) is retained for the second period (out-of-sample data). The
use of many geographic zones (Asia, Europa, United States) is to test the robustness of the different algorithms.
The methodology is the following. For each class of trading rules, here STUR(n), Simple Moving Average
SMA(m,n) and Exponential Moving Average EMA(m,n), a training period (in-sample data) is used to find its best
model in terms of the Sharpe Ratio which is an economic gain adjusted for risk. If we let R
t
=logS
t
logS
t1
, the
log return of the index at time t, then the Sharpe ratio (SR) is defined for any strategy, say k, by
) (
) (
) (
k
k RM
k SR
o

(6)
where

1
1
) , ( ,
1
) (
-

t t t
T
t
R k t position z z
T
k RM
(7)
5 , 0 2
1
] ) ( [
) (
1
) ( k RM z
T
k
t
T
t
-

o

Here position(t,k) takes 1 (1) if the strategy k is long (short) at time t and T represents the number of
predictions.
The second part consists to compare the performance of the best in-sample models with respect to the
out-of-sample data. The comparison is based on many criteria such as the Sharpe Ratio, the winning up periods
(W.U.P), the winning down periods (W.D.P), the correct directional changes (C.D.C) and the Maximum
Drawdown (M.D). Let
t
R
~
and
t
R be respectively the daily trading profit and actual return at time t ,
} 0 {
1
>

t
R t
Q
and
} 0 {
1
<

t
R t
F
the indicator functions for rise and fall at time t , and finally
} 0
~
, 0 {
1
> >

t t
R R
t
U and
} 0
~
, 0 {
1
> <

t t
R R
t
D the indicator functions for winning up and winning down periods, then the above expressions can
be defined by
T
C D C R R D M R R
t
R
T
t
c
i
t i
c
t
T t
i
t
i
c
t
} 0
~
{
1
, , 1 , , 1
1
1
. . , ),
~
( max
~
min . ,
~ ~
) (
>

(8)
t
T
t
t
T
t
t
T
t
t
T
t
F
D
D U W
Q
U
P U W

1
1
1
1
100 . . , 100 . .
(9)
where
i
t
i
c
t
R R
~ ~
1


is the cumulative trading returns up to time t (Note 5).
www.ccsenet.org/ijef International Journal of Economics and Finance Vol. 5, No. 3; 2013
225

Finally, we integrate the transaction costs in the analysis. Namely, it is supposed that any transaction implies a
constant cost of 20 basis points.
4. Results
We recall that the in-sample data contains approximatively 9 years of data for each index. The STUR(n) class,
with the constraint 8n50 and n, n/12 or n/20 is a power of 2, contains height (08) admissible strategies
characterized by the integer n valued in {8, 12, 16, 20, 24, 32, 40, 48}. The simple and exponential moving
average classes are parametrized by two integers m and n, representing respectively the lead and lag parameter.
In this study, sixteen (16) strategies are proposed for each Moving Average class with their parameters given by
m{1, 5, 10, 15,} and n{50, 100, 150, 200}. All these algorithms need some initial data to start the
forecasting procedure. For example, the STUR(n) strategy needs n+1 data to make the first forecast. For these
initial data, the agent decision is supposed to be always 1. The cost of one transaction is taken to be 20 basis
point i.e 0.2%.
Table 2 shows the performance of the best strategies in each class through the different indexes and through their
respective in-sample data given in Table 1.
For the STUR class, the best strategy is given by the parameter n=16 for the CAC, NIKKEI and S&P indexes
and by n=20 and 24 for the DAX and FTSE indexes, respectively. Overall, it is seen for the STUR class, the
approximation needs to be local or to have less data (n24) to generate good results.
For the Exponential Moving Average class, the lag parameter of the best strategy is always equal to n=150 for
the different indexes and the lead parameter lies to the set {10, 15}. For the Simple Moving Average class, the
lag parameter varies through indexes where the parameter n=150 is more frequent. The same remark applies also
for the lead parameter m where the mode is given by m=15. We also remark that for both moving average classes,
a small lead (m=1 or 5) does not give satisfactory in-sample results. All best competing models (STUR, EMA,
SMA), in the in-sample evaluation, generate economic gains or a positive Sharpe Ratio. Furthermore, except in
the FTSE index, the optimal strategy of the EMA class outperforms the other best models.

Table 2. Sharpe ratio of the best trading rules in each class (In-sample)
Class STUR(n ) EMA (m,n) SMA(m,n) Buy and Hold
CAC n=16 m=10, n=150 m=15, n=150
Sharpe Ratio 0.45 0.741 0.730 -0,37
Dax n=20 m=15, n=150 m=15, n=150
Sharpe Ratio 0.627 0.7825 0.5974 -0.218
FTSE n=24 m=15, n=150 m=15, n=150
Sharpe Ratio 0.4938 0.451 0.492 -0.210
Nikkei n=16 m=10, n=150 m=10, n=50
Ratio 0.222 0.611 0.528 -0.355
S&P n=16 m=15, n=150 m=15, n=200
Sharpe Ratio 0.274 0.4728 0.4436 -0.2916

On the other hand, the Buy and Hold Strategy has a negative mean in the in-sample data of all geographical
zones showing consequently a negative Sharpe ratio. This may be explained by the fact that all five indexes are
highly correlated and therefore the probability to have the same sign performance in the five indexes is very
high.
After getting the best strategy in each class, we make a comparison between them. Namely, three trading rules
are investigated for each index in their out-of-sample data given in Table 1. The aim is to see if it is possible to
do better than the benchmark strategy after taking into account transaction costs. To reduce the chance feature, a
long time series of out-of-sample is considered as containing around three years of data. The results are shown in
the Table 3 and Table 4 .

www.ccsenet.org/ijef International Journal of Economics and Finance Vol. 5, No. 3; 2013
226

Table 3. Out-of-sample performance of the best trading rules in each class (Part I)
CAC STUR(16) EMA (10,150) SMA(15,150) Buy and Hold
Sharpe Ratio 0.135 -0.72 -0.16 0.125
Transactions 4 20 6 0
M.D -0.35 -0.83 -0.52 -0.40
DAX 30 STUR(20) EMA (15,150) SMA(15,150) Buy and Hold
Sharpe Ratio 0.53 0.13 0.32 0.40
Transactions 4 4 4 0
M.D -0.33 -0.42 -0.30 -0.39
FTSE 100 STUR(24) EMA (15,150) SMA(15,150) Buy and Hold
Sharpe Ratio -0.11 -0.29 0.30 0.49
Transactions 2 10 4 0
M.D -0.37 -0.34 -0.29 -0.20
Nikkei 225 STUR(16) EMA (10,150) SMA(10,150) Buy and Hold
Sharpe Ratio 0.03 -0.46 -0.77 -0.01
Transactions 3 8 21 0
M.D -0.32 -0.59 -0.61 -0.33
S&P 500 STUR(16) EMA (15,150) SMA(15,200) Buy and Hold
Sharpe Ratio 0.09 -0.73 -0.20 0.63
Transactions 3 13 4 0
M.D -0.41 -0.66 -0.40 -0.21
Description: This table presents the out-of-sample values of the Sharpe ratio, the number of transactions and the Maximum Drawdown (M.D)
for each best strategy.

Table 4. Out-of-sample performance of the best trading rules in each class (Part II)
CAC 40 STUR(16) EMA (10,150) SMA(15,150)
C.D.C 50.41% 47.83% 50.82%
W.U.P 34.32% 52.82% 57.91%
W.D.P 66.94% 42.70% 43.53%
DAX 30 STUR(20) EMA (15,150) SMA(15,150)
C.D.C 52.58% 51.90% 51.77%
W.U.P 73.26% 77.12% 76.61%
W.D.P 29.39% 23.63 % 23.92%
FTSE 100 STUR(24) EMA (15,150) SMA(15,150)
C.D.C 51.90% 50.27% 50.82%
W.U.P 38.60% 63.73% 60.10 %
W.D.P 66.57% 35.43% 40.57%
Nikkei 225 STUR(16) EMA (10,150) SMA(10,150)
C.D.C 51.90% 52.58% 51.36 %
W.U.P 43.16% 48.16 % 51.32%
W.D.P 61.24% 57.30% 51.40%
S&P 500 STUR(16) EMA (15,150) SMA(15,200)
C.D.C 50.82% 51.63% 53.53%
W.U.P 46.96 % 66.67% 65.21%
W.D.P 55.69% 32.62% 38.77%
Description: This table presents the out-of-sample values of correct directional change (C.D.C), the winning up periods (W.U.P) and the
Winning down periods (W.D.P) for each best strategy.

Table 3 shows that for the Sharpe Ratio criterion, the STUR strategy gives overall the best results, namely three
over the five indexes. Then it is followed by the B&H strategy which performs two times over the five cases.
The results of SMA and EMA trading rules are not satisfactory in the out-of-sample data.
For the Maximum Drawdown (M.D) measure, it is found over all that the two best strategies are also given by
the STUR class and the Buy and Hold Strategy ( 2 over 5 indexes for each). Precisely, the STUR trading rule
obtains the good results from the CAC and Nikkei indexes while B&H does better in the FTSE and S&P indexes.
www.ccsen

The more
Maximum
It is noted
indexes. F
of the othe
index whe
Now, we a
directional
For the C
trading rul
FTSE. The
the W.U.P
However,
measure w
results and


We can re
since it be
Correct dir
Winning U
5. Conclus
In this pa
contributio
models. F
simplificat
The first i
data used
is to solve
of the prin
of R
n
and
the averag
positive (n
net.org/ijef
M.D, the little
m downside and
d that the riskie
For the profitab
er strategies fo
ere SMA gives
are interested
l changes in th
.D.C criterion
le gives better
e EMA, with a
P criterion, it is
in financial m
will play a maj
d sometimes th
Figure 1
sume the analy
elongs to the b
rectional chan
Up period crite
sion
aper, the stoch
on consists in
or this purpos
tions are also
s to constraint
in the regressi
e the dimension
ncipal compone
d then associat
ge forecast of t
negative). To
Inte
e is the downs
d profitability (
est strategies (
bility measure,
or the CAC, D
some interesti
in the percent
he rise and fall
n, the two bes
r results for the
a C.D.C of 52.
s EMA the bes
markets, invest
jor role for the
he difference is
1. The cumulat
ysis, by saying
best strategies
nge and the Wi
erion.
hastic unit roo
n approximat
se, two simpli
relevant to re
t the stochastic
ion, the nonlin
nality problem
ent analysis (P
te to each vect
these n models
diminish the
ernational Journa
ide risk. The c
(Sharpe Ratio)
(downside risk
, it is also seen
DAX, Nikkei a
ng results.
tage of correct
periods (W.U.
st strategies ar
e CAC and S&
58% only outp
st trading rules
tors are more
em. It is noted
s very significa
tive return path
g overall, the S
for many crite
inning down p
ot model is u
ting locally th
ifications are
duce the exec
c parameter to
nearities are ap
m by extracting
PCA) by taking
tor, a simple li
s and the decis
transaction co
al of Economics
227
cumulative retu
) over the out-o
k) are usually
n that the STU
and S&P index
t directional ch
.P and W.D.P)
re given by S
&P indexes wh
performs the o
in three over
concern to de
d that, for all f
ant as in the ca
h of the differe
STUR strategy
eria such as th
periods. Its poo
used to foreca
he nonlineariti
made to facil
ution time and
o take, at any t
pproximated by
g only the mos
g n orthogonal
inear model. T
sion making is
osts for profita
s and Finance
turns of Figure
of-sample data
reached by SM
UR and B&H c
xes. The excep
hanges (C.D.C
), see the eqs. (
SMA and STU
hile the STUR
ther trading ru
the five indice
etect falling pe
five indexes, t
ase of CAC an
ent trading rule
y outperforms
he Sharpe ratio
or results in th
ast the directi
ies of financi
litate the estim
d the memory
time, two poss
y 2
n
linear m
st significant li
l binary vector
Then, the strate
s simply to buy
ability reasons
e 1 illustrates b
a for the differ
MA and EMA
cumulative ret
ption appears
C), and the pe
(8) and (9) and
UR. The simpl
R strategy perfo
ules for the Nik
es and then it i
eriods. Conseq
the STUR stra
nd S&P indexe
es for each ind
the moving av
o, the Maximu
he out-of-samp
ion change of
ial time series
mation of the
y cost involved
sible values. C
models. The sec
inear models. W
rs (columns of
egy STUR(n) is
y (sell) if this
s, but also to
Vol. 5, No. 3;
both the conce
rent indexes.
in four among
turns end at th
only for the F
rcentage of co
d Table 4.
le moving ave
forms for DAX
kkei index. Bu
s followed by
quently, the W
ategy gives the
s.
dex
verage trading
um Drawdown
ple only come
f asset prices.
s by simple l
parameters. T
d by the algor
Consequently,
cond simplific
We follow the
f Hadamard ma
s defined by ta
average forec
capture signif
2013
ept of
g five
e top
FTSE
orrect
erage
X and
ut, for
SMA.
W.D.P
e best
rules
n, the
from
The
inear
These
ithm.
for n
ation
e idea
atrix)
aking
ast is
ficant
www.ccsenet.org/ijef International Journal of Economics and Finance Vol. 5, No. 3; 2013
228

informations, an endogenous threshold c is used to activate a decision. Namely, the trader will transact if the
average forecast return is superior in absolute value to the threshold, elsewhere the previous position is
conserved. It is found that the strategies from STUR class dominate overall the moving average trading rules
(simple and exponential) and also the Buy and Hold strategy for the Sharpe criterion.
These interesting results may be explained by two facts. First, it is known that random coefficient autoregressive
models are able to fit well financial asset prices. So it is expected to have satisfactory results when this
econometric model is used for forecasting. The second reason is due to our estimation procedure which is local
and allows to capture feedback or interaction of traders rather using methods based on stationarity assumptions
in variance or higher moments that are violated in our case of stochastic unit root model.
References
Dunis, C. L., Williams, M. (2003). Applications of Advanced Regression Analysis for Trading and Investment. C.
In L. Dunis, J. Laws & P. Nam (Eds.), Applied Quantitative Methods for Trading and Investment. John
Wiley & Sons.
Granger, C. W. J., & Ramanathan, R. (1984). Improved Methods of Combining Forecasts. Journal of
Forecasting, 3, 197-204. http://dx.doi.org/10.1002/for.3980030207
Granger, C. W., & Swanson, N. R. (1997). An introduction to stochastic unit-root processes. Journal of
Econometrics, 80, 35-62. http://dx.doi.org/10.1016/S0304-4076(96)00016-4
Kont, M. A. (2011). A link between Random Coefficient AutoRegressive Models and some kind of Agent Based
Models. Journal of Economic Interaction and Coordination, 6(1), 83-92.
http://dx.doi.org/10.1007/s11403-010-0077-3
Kont, M. A. (2010). Behavioural Finance and Efficent Markets: Is the joint hypothesis really the problem? IUP
Journal of Behavioral Finance, 7, 9-19.
Lo, A. (2005). Reconciling Efficient Markets with Behavioral Finance: The Adaptive Markets Hypothesis.
Journal of Investment Consulting, 7(2), 21-44.
Nicholls, D. F., & Quinn, B. G. (1982). Random Coefficient Autoregressive Models: An Introduction. Lecture
Notes in Statistics, 11. New York: Springer-Verlag. http://dx.doi.org/10.1007/978-1-4684-6273-9
Nicholls, D., & Quinn, B. (1981). The Estimation of Random Coefficient Autoregressive Models. II. Journal of
Time Series Analysis, 2, 185-203. http://dx.doi.org/10.1111/j.1467-9892.1981.tb00321.x
Ou, J., & Penman, S. (1989). Accounting measurement, Price-Earnings Ratio, and the information content of
security prices. Journal of Accounting Research, 27, 111-152. http://dx.doi.org/10.2307/2491068
Sollis, R., Leybourne, S., & Newbold, P. (2000). Stochastic unit roots modelling of stock price indices. Applied
financial economics, 10, 311-315. http://dx.doi.org/10.1080/096031000331716
Sullivan, R., Timmermann, A., & White, H. (1999). Data-snopping, technical trading rule performance, and the
boostrap. The journal of finance, 54, 1647-1691. http://dx.doi.org/10.1111/0022-1082.00163
Wang, D., & Ghosh, S. K. (2002). Bayesian Analysis of Random Coefficient Autoregressive Models. Model
Assisted Statistics and Applications, 3(2), 281-295.
Yoon, G. (2003). A simple model that generates stylized facts of returns. Department of Economics, UCSD.
Paper 2003-04.
Yoon, G. (2006). A note on some properties of STUR processes. Oxford Bulletin of Economics and Statistics, 68,
253-261. http://dx.doi.org/10.1111/j.1468-0084.2006.00161.x