See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/284794164

A genetic network programming model for portfolio optimization by generating

risk-adjusted trading rules

Conference Paper · January 2014

CITATIONS READS

2 98

2 authors, including:

Akbar Esfahanipour
Amirkabir University of Technology
47 PUBLICATIONS   543 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

MGP-INTACTSKY: Multitree Genetic Programming-based learning of INTerpretable and ACcurate TSK sYstems for dynamic portfolio trading View project

Volatility Modeling and Forecasting View project

All content following this page was uploaded by Akbar Esfahanipour on 03 January 2016.

The user has requested enhancement of the downloaded file.

A genetic network programming model for portfolio
optimization by generating risk-adjusted trading rules

Akbar Esfahanipour Maryam Tayari

Assistant Professor, Department of Industrial Engineering Master of Science in Industrial Engineering, Garmsar
and Management Systems, Amirkabir University of Branch, Amirkabir University of Technology,
Technology, Tehran, Iran
Tehran, Iran [email protected]
[email protected]

Abstract— Genetic network programming (GNP) as an in generating trading rules instead of forcing model to use
evolutionary computation method has been used for stock only two signals.
trading recently. Former researches confirm the efficiency The foundation of portfolio optimization was laid by
of trading rules which are created by GNP. In this paper, Markowitz in 1959, where he proposed a mean-variance
GNP has been applied for stock portfolio optimization by optimization model. He states that investors should
generating risk-adjusted trading rules. There are two main
novelties in this paper: 1) we use conditional Sharp ratio as
decide the allocation of their investment on the basis of a
a risk-adjusted measure for generating trading rules, 2) in trade-off between risk and expected returns [2].Although
our GNP model, binary trading rules have been extended to there are many nonlinear constraints in the real-world
more realistic rules which are called trinary rules using problems, evolutionary computation methods were
three signals of buy, sell and no trade. We applied our GNP developed to solve these kinds of problems.
model on ten stocks from Tehran Stock Exchange (TSE). Genetic algorithm (GA) is a heuristic search method
The numerical results show that our proposed model with inspired by nature which has been applied to financial
three signals outperformed the previous model with two problems. GA as the most popular heuristic optimization
signals of buy and sell in terms of excess return and excess techniques was developed by Holland in 1975 [3]. Lin et
risk adjusted return.
al. solved the model of portfolio selection problem
through taking into consideration the multi-objective
Keywords: genetic network programming; portfolio genetic algorithm [4]. Oh et al. proposed a new portfolio
optimization; technical trading rules; risk-adjusted measures; selection algorithm based on portfolio beta using GA [5].
Tehran Stock Exchange (TSE) Xia Lau Yang applied GA method with a dynamic
portfolio optimization system to improve the
I. INTRODUCTION efficiency of the stock portfolio [6]. Chang et al.
Evolutionary Computation is well-known for producing introduced a heuristic approach to portfolio optimization
the solutions in optimization problems. Genetic Network problems in different risk measures using GA [7].
Programming (GNP) is an extended method of Genetic Genetic programming (GP) has been developed by Koza
Algorithm (GA) and Genetic Programming (GP). It has in 1992 for the first time as an extension of GA [8]. The
been clarified that GNP is an effective method mainly for main difference between GP and GA is the representation
dynamic problems since GNP represents its solutions of the solutions. In GP, the population members are not
using graph structures. Compact structures and implicit fixed length character strings that encode possible
memory functions are some of inherent characteristics of solutions to the problem at hand, they are programs that,
GNP's graph structure that contribute to creating effective when executed, are the candidate solutions to the
action rules [1]. problem. These programs are expressed in genetic
In this paper, we intend to describe a model of stock programming as parse trees rather than as lines of code
portfolio optimization which has two important novelties [9]. Although GP is widely used in the financial field, it
relative to the previous studies. First, using risk adjusted occasionally causes some bloating problems for its tree
measure as a fitness function, second, using three signals structure. GP has been applied for stock trading model by

www.iiec2014.com
Potvin et al. in 2004 [10]. Esfahanipour and mousavi Iranian companies are reported, and finally the paper
proposed a GP model to generate risk-adjusted technical closes with conclusion.
trading rules in stock market [11].
Genetic Network Programming (GNP) is an extended II. GENETIC NETWORK PROGRAMMING
method of GA and GP [1] that is applied in financial In this section, the structure of GNP method is explained.
fields recently. Since GNP represents its solutions using The traditional GNP is composed of a start node,
graph structures, which contributes to creating quite judgment nodes and processing nodes, which are
compact programs and has implicit memory function, it connected to each other. Fig. 1 shows a basic structure of
has been clarified that GNP is an effective method mainly GNP. Judgment nodes have if-then type branch decision
for complicated problems such as portfolio optimization functions, which return judgment results for assigned
systems [12]. GNP was used for stock trading by Mabu et inputs and determine the next node. Processing nodes
al., Chen et al, and Hirasawa et al. In 2008, Chen et al. take buying and selling actions for the stock trading.
extended a model of GNP with reinforcement learning While judgment nodes have conditional branches,
for generating trading rules on stock market using the processing nodes have no conditional branches. Start
stock prices of sixteen stocks of Tokyo stock market [13]. node is used to determine the first node to be executed,
They also extended a portfolio optimization model using which has no functions and no conditional branches.
GNP with control nodes (GNPcn) in 2009. Since GNP Once GNP is booted up, the execution starts from the
has a directed graph structure, the aim of GNPcn is to start node, then the next node is determined according to
improve the performance of GNP by extending the the connection from the current activated node [1].
evolutionary method in terms of breadth and depth of
searching procedure [14]. In 2010, Chen et al. extended a A. Basic Structure of Genetic Network Programming
model of portfolio optimization using time updating GNP with control nodes (GNPcn)
which can adapt to the change of stock prices and market In conventional GNP, the current node is not
conditions [12]. These approaches are essentially referred compulsorily transferred to the start node and there is no
to technical analysis since it determines the timing of terminal node in GNP. Therefore, once GNP is booted
buying and selling stocks based on the technical indices up, the successive activation of GNP system is carried
such as Relative Strength Index (RSI) and Rate of change out according to the network flow until the time limit.
(ROC). These studies use sum of the profit of trading in Since the current node is not compulsorily transferred to
terms of the difference between selling price and buying the start node, there is a possibility that some of the nodes
price as fitness for evolution [12, 13, 14]. Despite the are not used. GNPcn starts from a control node, and the
attractiveness of these works, they did not consider current node is transferred back to one of the control
transaction cost for evaluating fitness of generated rules. nodes after executing a certain number of processing
Transaction cost is brokerage fees that are payable for nodes. Consequently, the performance of GNPcn
each trade carried out. Since many trades take place in improves because the increase of the number of control
the buy and sell strategy, transaction cost would affect nodes contributes to search a solution space widely and
the profitability of trading rules. Some other parameters then to find many distinguished trading rules [14].
that have effect on return such as dividend and splits are Especially, the number of the control nodes can be
also important to consider in the fitness function. In our considered as breadth of the search of GNP, while the
proposed model, we consider these issues in the fitness number of processing nodes activated per control node
function by using a risk adjusted return measure named can be considered as depth of the search of GNP [15].
conditional sharp ratio which uses conditional value at GNPcn is also useful to solve the portfolio optimization
risk (CVaR) as an appropriate coherent risk measure. In model because of its ability to solve the multi-brands
our proposed GNP model, binary trading rules have been problem. In this paper, the number of control nodes is
also extended to more realistic rules which are called fixed and the number of processing nodes activated per
trinary rules using three signals of buy, sell and no trade. control node is evolved.
Additionally, we have included transaction costs,
dividend and splits in our GNP model for calculating
more accurate returns. We use Importance Index (IMX)
[13] as a pattern for evaluating of technical indexes in our
GNP model for efficient stock trading model. Here, GNP
is used to select appropriate actions as buying, selling,
and no trade.
This paper is organized as follows: In section 2, the
structure of GNP is described; in section 3, the proposed
GNP model is explained; in section 4, the computational
Fig. 1. Basic structure of GNP.
results of applying our model on the portfolio of 10
Fig. 2 shows the basic structure of GNPcn. GNPcn has
several control nodes, judgment nodes and processing

www.iiec2014.com
nodes. GNPcn uses one of the groups of control nodes for B. Genetic operators of GNPcn
one stock, so that GNPcn could deal with multi-stocks. In this subsection, the genetic operators of GNPcn are
When GNPcn deals with brand (or stock) ‘‘A”, GNPcn described as follows.
starts its node transition from control node ‘‘ ”, and the Crossover: Crossover is executed between two parents
current activated node returns to one of the control nodes and generates two offsprings. The procedure of crossover
( , ,…, successively after transiting m is as follows.
processing nodes from the last control node. GNP has (1) Select two individuals using tournament selection
two kinds of time delays: 1) time delays on judgment or twice and produce them as parents.
processing nodes, 2) time delays on nodes transition. The (2) Each node is selected as a crossover node with the
genotype expression of GNP node is also shown in Fig. probability of .
1. This figure describes the gene of node i, then the set of (3) Two parents exchange the genes of the corresponding
these genes represents the genotype of GNP individuals. crossover nodes.
Concretely speaking, represents the node type, 0 (4) Generated new individuals become the new ones of
means start node, 1 means judgment node, and the next generation.
2 means processing node. represents an If GNP exchanges the genes of judgment nodes, it must
identification number of the node function. In judgment exchange all the genes with suffix A, B, C ...
node, the gene shows a technical index that GNP simultaneously [1].
judges and in processing node, the gene shows Mutation: Mutation is executed in one individual, and a
actions that an agent can take such as buying, selling and new one is generated. The procedure of mutation is as
no trade. follows.
is a parameter which represents the threshold for (1) Select one individual as a parent using tournament
determining buying or selling stocks in a processing node. selection.
The role of judgment nodes, processing nodes and their (2) Mutation: there are four types of mutations which are
functions are explained in section 4. , , … show the selected in random with equal probabilities as follows.
node number of the next node. Judgment node determines o Change connection: Each node branch
the upper suffix of the connection genes to refer to ( , , …) is selected with the probability of
depending on the judgment result. is the time delay , and the selected branch is reconnected to
spent on the judgment or processing at node i, while another node.
, , … are time delays spent on the node transition o Change node function: Each node function ( )
from node i to the next node. If the judgment result is is selected with the probability of , and the
‘‘B”, GNP refers to and . Processing nodes always selected function is changed to another one.
refer to and since they have no conditional branches. o Change node number: The number of processing
nodes activated per control node (m) is changed
Similar to other evolutionary algorithms, crossover and
to the other value with the probability of .
mutation are used as the genetic operators of GNP. The
o Change parameters: each or in node i is
steps of evolution are as follows [12, 14]:
changed to the other value with the probability
1. Initialize the first population and calculate the
of . (we explain about in section 4 )
fitness of the population.
2. Generate new individuals for the next generation (3) Generated new individual becomes the new one of the
by tournament selection and genetic operators. next generation.
3. Calculate the fitness of the new individuals.
4. Repeat steps 2–3 until the terminal condition III. THE PROPOSED GNP MODEL
meets. Previous studies generated binary trading rules, i.e.
they could only indicate two signals of buying or selling
for trading decisions. Sometimes, do nothing is the
optimal decision in real world decision making. Hence,
no trade can be an optimal decision in stock trading,
which is not indicated by binary trading rules. Therefore,
we have extended the structure of previous trading rules
to trinary ones which can produce three signals of
buying, selling and no trade.
A. Extending GNP model to three signals
When the current activated node i is a judgment node, the
corresponding function ( ) is one of the technical
indexes. In this paper, we use four technical indexes
Fig. 2. The basic structure of GNPcn. named relative strength index (RSI), rate of change

www.iiec2014.com
(ROC), volume ratio (VR) and stochastics. Each
technical index has its own judgment function. As an
example the judgment function of Rate of Change (ROC)
is shown in Fig. 3, where x axis shows the value of each
technical index and y axis shows the importance index
(IMX) output of the judgment function which is used at a
processing node. Importance index tells GNP whether or
not the buying or selling signals are likely to appear at the
current day. Chen et al., use these importance index
functions in their studies on stock market for generating
trading rules [13]. The shape of Importance Index
functions is fixed in our research based on the knowledge Fig.4. Judgment node structure in our proposed model
of the real stock market. We use the IMX function of the In binary signals (buy and sell) model, when the current
related studies. node is a processing node, the computation is based on
one technical index in each judgment node from the
previous processing node to the current processing node.
However, in our study in a processing node, we have two
computations based on: 1) the first importance indexes
(IMXs) obtained at the judgment nodes executed in the
node transition from the previous processing node to the
current processing node, 2) the second importance
indexes (IMXs) obtained at the judgment nodes executed
in the node transition. Fig. 5 shows structure of a
Fig.3. Importance index (IMX) as Judgment functions of ROC [13]. processing node our trinary signal model.
When we are going to make a decision of selling or
buying stocks (in binary trading rules model), we use
only one technical index in each judgment node, so the
node gene of judgment nodes contains , , , and
connection gene contains , , , . is a threshold
for technical index value that shows the judgment result.
For example, if the judgment node i judges the Rate of
Change and the value of this rate is more than ,
judgment result becomes A and otherwise B. This
parameter ( ) is randomly determined at the beginning Fig.5. Processing node structure in our trinary signal model
of the first generation, and changes appropriately by When the current node at time t is a processing node, the
evolution. following process will be done:
But one of the most important points in our study is that 1) is the threshold for determining buying or
we do not force the node to make the decision in only selling stocks in models with two signals (Buy
two manners, buying or selling because there is also and Sell) and in this paper with three signals
another way in stock trading model named no trade (Buy, Sell and No trade).
which is optimal decision in some situations. Therefore, 2) Calculate an average of the IMXs obtained at
we have extended the structure of previous trading rules the judgment nodes executed in the node
to trinary ones which can produce three signals of transition from the previous processing node to
buying, selling, and no trade. the current processing node by using equations 1
In our trinary trading rules model, we use two technical and 2.
indexes in each judgment node, so the node gene of each
∑′ ′ ′
(1)
judgment nodes contains , , , , , , ′

and connection gene contains , ,…, , . is ′

∑′ ′
′
(2)
the threshold for the value of the first technical index and ′
Where, shows a set of suffixes of the
is the threshold for the value of the second technical
judgment nodes executed in the node transition
index, and are time delays spent on primary and
from the previous processing node to the current
secondary technical index, respectively. For example, if ′
processing node. shows an IMX output
the value of the first technical index is more than and
the value of the second technical index is more than , at node ′ ′
.
3) Determine the action (buying, selling or no
the judgment result is A. Fig.4 shows a judgment node
structure based on the value of two technical indexes. trade). After calculating and the trinary
 trading rule signal is determined according to Among the investigated risk adjusted measures, here the
Table 1. conditional Sharpe ratio is selected to apply in our GNP
4) Transfer to the next node from the current node. model. Conditional Sharpe ratio is the ratio of the
expected excess return over risk free rate and its
TABLE 1. The way for determining of signals using
conditional value at risk (CVaR). This ratio can be
trinary trading rules calculated by the Eq. (4):
Trading rule Conditional Sharpe ratio = (4)
signal
Where r is the return of the fund, is risk free rate;
True True
Buy CVaR1-α is the conditional value at risk over the given
time horizon with 100(1-a) % confidence level. CVaR is
No trade
True False the expected loss under the condition that the value at
No trade risk is exceeded. The reasons behind selection of the
False True conditional Sharpe ratio as a suitable risk-adjusted
Sell measure in our GNP model are as follows: (1) it uses
False False
CVaR as downside risk measure; (2) it uses CVaR which
At1 and At2: average IMXs of the first and the second technical indexes
ai1and ai2: thresholds for determining the actions from technical indexes
is an optimal coherent risk measure [16]; (3) it is a
suitable measure for evaluation of assets with non normal
B. Determining of traning and testing periods and fat tailed distributions of return [17]; and (4) this
measure includes the subjective risk aversion of an
Since the data in stock markets are time-varying and investor through confidence level selection [18].
changing every minute, we use kind of GNP method In this model, the excess risk adjusted return is evaluated
named time adapting genetic network programming (TA- as fitness function. Since we included transaction costs,
GNP) which considers the time-related fluctuation of dividend and splits in the fitness function of our model, it
stock prices well. In the traditional GNP method, is expected that the more realistic returns will be
individuals are evolved using the initial data of stock computed.
prices in a single period, and then the best GNP The risk adjusted return of the buy and hold strategy
individual in the last generation of the training period is ( & ) is calculated with conditional Sharpe ratio as
used for all future trading in the testing period. As shown Eq. (5).
in Fig. 6, this method uses recent historical data for &
&
(5)
training with a fixed data window shift in each iteration
[12]. Where & is the expected return of the buy and hold
strategy. It is calculated considering dividends, splits and
transaction costs. is calculated as Eq. (6).
(6)
Where X is the profit/loss of the stock over a given risk
horizon and is the worst expected loss over the
given time horizon at 1 confidence level.
is usually approximated via historical simulation method.
The cumulative return for a trading rule in a period is the
sum of all transaction returns considering dividends and
splits minus the related transaction costs. In risk adjusted
version of our model, the cumulative risk adjusted return
for a trading rule is the sum of risk adjusted returns of all
transactions in the period and is calculated as Eq. (7).
& ∑ (7)
Where n is the number of transactions in the period
advised by trading rule, is the return of th transaction
Fig.6. comparison between GNP and TA-GNP [12]. considering dividends, splits and transaction cost.
C. Applying risk-adjusted measures for fitness function The fitness function is simply calculated as Eq. (8), where
& is the risk adjusted return of the generated
Risk adjustment is very important and the most important trading rule (i.e., the buy and sell strategy).
reason is that these measures make it possible for us to (8)
& &
compare different investment opportunities with different
levels of expected returns and risks. There are a lot of
risk adjusted measures to adjust return for risk.

www.iiec2014.com
 IV. APPLICATION OF OUR GNP MODEL FOR IRANIAN values, where the ‘‘window shift’’ in Table 3 indicates
COMPANIES the shifting period of data windows. From the results, it is
Our extended genetic network programming model has obvious that the best fitness value obtains when window
been applied on Ten Iranian Companies listed on Tehran size is 180 days and window shift is 20 days. We use the
Stock Exchange (TSE). We intend to examine the related work for determining the range of window size
profitability of our trading rules and their risk adjusted and window shift.
excess returns in these companies.
TABLE 3. Average fitness value of portfolio with different setting of
time windows [day].
TABLE 2. Investigated Iranian companies in this study Widow size
Activity sector Company Symbol 360 180 120 90
Motor vehicles and Iran Khodro Industrial 30 6.987 7.014 6.901 6.215
IKCO1
auto parts Group Window shift 20 6.513 7.595 7.229 6.348
Motor vehicles and 10 6.328 6.968 6.812 6.751
Saipa Co. SIPA1
auto parts
Motor vehicles and
auto parts
Pars Khodro Co. PKOD1 Table 3 shows when window size is large, the window
Refined petroleum shift with higher value shows better results, and when
products & nuclear Behran Oil Co. NBEH1 window size is smaller window shift with smaller value,
fuel shows better results. In fact, it seems there is a mutual
Chemicals & by-
products
Arak Petrochemical Co. SHARAK1 relation between window size and window shift.
Chemicals & by- Other parameters of this model are shown in table 4.
Abadan Petrochemical Co. SHPETRO1
products Our GNP model uses judgment nodes which judge the
Chemicals & by-
Khark Petrochemical Co. SHKHARK1 technical indices. There are four judgment functions
products which are the four selected technical indexes. The
Electric machinery
& apparatus
Jooshkab Yazd Ind. Co. JOSH1 number of processing function is three: buy, sell and no
Banking EN Bank VNOV1 trade. Four control nodes are assigned to each stock. The
Pharmaceuticals SobhanPharmacutical Co DSOB1 functions of each node are determined randomly at
the first generation, and then changed by evolution. The
A. Data
initial connections between nodes are also determined
Ten Iranian companies listed on Tehran Stock Exchange randomly at the first generation.
(TSE) have been chosen for testing of our GNP model.
Since Esfahanipour et al., in 2009, obtained better rules TABLE 4. Parameter setting for our GNP model
for companies having frequent trading in the market [19]; Population size 300
we also select companies with the highest liquidity for Number of generation 500
Number of judgment nodes 15
testing our GNP model. For this purpose, ten active Number of processing nodes 6
companies with the highest liquidity in the recent years Number of control node 40
have been selected in different activity sectors. Selected Selection method Tournament
companies and their sectors have been shown in Table 2. Mutation rate 0.03
Transaction cost is 0.515 percent of trade volume for Crossover rate 0.2
CVaR confidence level 95%
buying and 1.0665 percent of trade volume for selling the
stocks in TSE according to TSE’s regulation. The C. Results
historical data includes the dividend and splits adjusted
Genetic Network programming model is applied on the
closing price and transaction volumes for each working
portfolio of the ten mentioned Iranian companies. We
day (Research and Development Islamic Studies, 2009;
divided our simulation into two groups, first, risk-
Tehran Securities Exchange Technology Management,
unadjusted models and second, risk-adjusted models. The
2009).
trading rules were generated over the training period for
We use the data of the stock market from September 22,
the portfolio. In each group (risk adjusted and risk
2008 to September 21, 2012 in the simulations. Our data
unadjusted), Model 1 is the portfolio optimization by
is divided into two groups: training data and testing data.
generating binary trading rules from historical data and
B. Parameter settings Model 2 is the portfolio optimization by generating
Some of the best parameter settings for our GNP are trinary trading rules from historical data.
determined through the related studies [12, 13 and 14] as The results of implementing risk-unadjusted models
well as our preliminary experiments. on the portfolio of 10 stocks in Tehran Stock Exchange
Table 3 shows average of fitness function values with are shown in table 5. In this table, returns of the buy and
different setting of time windows. According to Fig. 6, hold strategy (B&H), cumulative returns of trading rules
Time Adapting GNP can adapt to the change of stock (B&S) and excess returns of trading rules over the buy
conditions with time windows. Thus, we test different and hold strategy are presented for each model in column
periods and shifts of the windows to obtain the optimal 2-4 of table 5, respectively.

www.iiec2014.com
 TABLE 5. Returns of GNP models in risk-unadjusted basis.
B&H (%) B&S (%) Excess return (%) REFERENCES
Model 1
67.359 73.8642 6.5052
(2 signals)
Model 2 [1] Mabu S, Hirasawa K, Hu J. A graph-based evolutionary algorithm:
67.359 90.0047 22.6457 Genetic network programming and its extension using reinforcement
(3 signals)
learning. Evolutionary Computation 2007; 15(3):369–398.
[2] Markowitz H. Portfolio selection: Efficient diversification of
From table 5, we can see the excess return for model investments. New York: Wiley; 1959
2, which generates three signals, is greater than model 1, [3] Holland JH. Adaptation in natural and artificial systems: an
which generates two signals of buying or selling stocks. introductory analysis with applications to biology, control and artificial
intelligence. University of Michigan Press; 1975.
The results of trading rules with risk-adjusted returns
are presented in Table 6. [4] Lin D, Wang S, Yan H. A multiobjective genetic algorithm for
portfolio selection problem. In: Proceedings of the ICOTA 2001, Hong
TABLE 6. Returns of GNP models in risk-adjusted basis Kong.
B&H (%) B&S (%) Excess return (%) [5] Oh KJ, Kim TY, Min SH, Lee HY. Portfolio algorithm based on
Model 1 portfolio beta using genetic algorithm. Expert Systems with
8.284 13.0265 4.7425
(2 signals) Applications 2006;30(3):527–534.
Model 2 [6] Yang X., Improving portfolio efficiency: a Genetic Algorithm
8.284 15.8796 7.5956 Approach,
(3 signals)
Computational Economics, Springer Link, 28(1), (2006), 1-14.
[7] Chang T, Yang S, Chang K. Portfolio optimization problems in
Although portfolio optimization model based on different risk measures using genetic algorithm. Expert Systems with
generating trading rules were profitable in all cases, we Applications 36 (2009) 10529–10537.
can observe that excess risk adjusted return in model 2 [8] Koza JR. Genetic programming, on the programming of computers
(generating trading rules with three signals) that we by means of natural selection. Cambridge, MA: MIT Press; 1992.
[9] Abraham, A., Nedjah, N., &Mourelle, L. M. Evolutionary
proposed in this paper is more profitable than the first Computation: from Genetic Algorithms to Genetic Programming. In J.
model. Kacprzyk (Ed.).Studies in computational intelligence (SCI), Genetic
systems programming 2006 (Vol. 13, pp. 1–20). Berlin Heidelberg:
V. CONCLUSION Springer Verlag
[10] Potvin J, Soriano P, Vallee M. Generating trading rules on the
In this paper, a genetic network programming (GNP) stock markets with genetic programming. Computers & Operations
model has been developed in order to optimize a stock Research 2004;31(7): 1033–1047
portfolio by generating technical trading rules. A GNP [11] Esfahanipour A, Mousavi S. A genetic programming model to
model has been developed considering the risk generate risk-adjusted technical trading rules in stock markets. Expert
System with Application 38(2011) 8438-8445
importance in trader’s decision making and the effect of [12] Chen Y, Mabu S, Hirasawa K. A model of portfolio optimization
transaction cost, dividends and splits on transaction using time adapting genetic network programming. Computer &
returns. Trading rule’s structure is also extended to three Operation Research 37 (2010) 1697-1707
signals, including buy, sell and no trade. [13] Chen Y, Mabu S, Shimada K, Hirasawa K. Trading rules on stock
markets using genetic network programming with sarsa learning.
For the risk adjusting of our model, risk adjusted returns Journal of Advanced Computational Intelligence and Intelligent
are used to evaluate fitness of the generated trading rules. Informatics 2008;12(4): 383–392.
Among many risk adjusted measures which were [14] Chen Y, Ohkawa E, Mabu S, Shimada K, Hirasawa K. A portfolio
reported in the literature, conditional Sharpe ratio has optimization model using genetic network programming with control
nodes. Expert System with Application 36 (2009) 10735-10745
been selected in our model because of its less [15] Eto S, Mabu S, Hirasawa K, Huruzuki T. Genetic Network
assumptions and more accuracy in risk measurement. Programming with Control Nodes considering Breadth and Depth.
We carry out the simulation by using stock data of 10 SICE Annual Conference 2007. 470-475
Iranian companies in Tehran Stock Exchange (TSE) in [16] Inui, K., & Kijima, M. (2005). On the significance of expected
shortfall as a coherent risk measure. Journal of Banking & Finance, 29 ,
four years. 853–864.
We implemented our model in risk-adjusted case and also [17] Agarwal, V., &Naik, N. Y. (2004). Risk and portfolio decisions
in risk-unadjusted case. Numerical results show that involving hedge funds. Review of Financial Studies, 17(1), 63–98.
extending model to three signals has a significant effect [18] Acerbi, C., &Tasche, D. (2002). On the coherence of expected
shortfall. Journal of Banking & Finance, 26, 1491–1507.
on the trading rule generating model. Our extended GNP [19] Esfahanipour A., Karimi B., Mousavi S.; "The Profitability of
model shows more efficiency than previously used model Technical Trading Rules in Tehran Stock Exchange: An Application of
with only two signals of buying or selling stocks. Our Genetic Programming"; International Symposium on Innovations in
GNP model is also appropriate for an investor who Intelligent Systems and Application; 2009.
doesn’t want to consider risk for generating trading rules.

www.iiec2014.com

View publication stats