A Model for Trading the Foreign Exchange Market

Nwokorie, E. C.1, Nwachukwu, E. O.2

1
Department of Computer Science, Federal University of Technology, Owerri
[email protected]
2
Department of Computer Science, University of Port Harcourt
[email protected]

Abstract

The electronic Foreign Exchange (FOREX) market where currencies are bought and sold has
become more complex and dynamic with characteristics of high volatility, nonlinearity, and
irregularity. Some important factors such as economic growth, trade development, interest rates,
inflation rates, etc. have significant impacts on the exchange rate fluctuation. Existing foreign
exchange (FOREX) trading models have been found inadequate. They have tended to use only
past price data and appear to be too stochastic or too deterministic. In this work an improved
model that provides a wide set of dynamic process information has been developed. Technical
and fundamental methods of analysis of FOREX market data were modeled with neural
networks. The predictions from the networks are integrated to get the direction of price
movement. Market sentiment and volatility values are combined with the neural network
prediction to develop trading strategies using Marcov chain. Finally, an application of the model
in FOREX trading is demonstrated and implemented with the Meta-Quote scripting Language
(MQL) of the meta-Trader platform. The historical test of the robot for the last 12 months
resulted in a range of significant profitability with annual returns between 40% to700% and with
maximum relative draw down risk between 8% to 52%

Keywords: FOREX, marcov chain, model, neural network, trading robot

Introduction
Electronic currency trading in the end product will have a direct impact on
Foreign Exchange (FOREX or FX) market volumes, and explained in a comment to
is now a very popular activity. FOREX is Forex Magnates: “Sustaining high standards
the single largest market in the world in technology, execution, transparency, and
accessible to anyone. Its volumes are greater support have proven to significantly drive
than all stock, commodities and debt volumes up.”
markets combined [1]. Average daily The market is open day and night from
foreign-exchange volume rose to $3.873 Sunday night to Friday night. The
trillion in October 2013, from $3.604 development of trading models is
trillion in the previous October, according to challenging as the problem of figuring out
reports published by the Bank of proper input and output values that derive
International Settlements in the first phase long term profitable results is not simple.
of its triennial findings for its three yearly Consistently predicting FX markets has
global FX survey. [2] attributes the sharp seemed like an impossible goal but recent
increase in FX activity to several factors advances in financial research now suggest
including; market volatility, acceptance of otherwise. With newly developed
FX as an asset class and global distribution computational techniques and newly
of the product. Cyril Tabet, Partner & CEO available data, the development of
at JFD Brokers, believes the quality of the successful trading models is now possible.

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 3
Many economic variables (including interest components like stochastic trend. The model
rates) have little explanatory power at high is uncertain when faced with the volatility of
frequencies. Many traders have therefore the market prices as measured in terms of
used technical analysis. Technical analysis annualized standard deviation of the market
attempts to predict markets by identifying prices. The notion of volatility as measured
patterns (Channel, Head and shoulder, Cup by standard deviation or variance has
and handle, etc.) in the price actions. become a focus of attention in quantitative
Developing optional trading models by finance, which led to a plethora of research
combining a number of technical, streams such as implied volatility, volatility
fundamental, sentiment and informational smile and volatility surface in option
indicators to predict volatility and pricing, ARCH and GARCH type of models
directionality reduces failure in changing for volatility dynamics, and value at risk as a
conditions. The complexity of the financial measure of risk exposure under any given
price behaviour is often so high as if it looks volatility distribution [8].
like intractable, or one could assume no The Log-Periodic Power Law (LPPL) is a
such complexity exists as expressed by the nonlinear regression model for aggregate
Efficient Market Hypothesis (EMH). The market prices developed by [9] and his
debate on the market efficiency has a long associates since 1996. The LPPL is used for
history which has resulted in no financial bubbles and its counterpart – anti-
confirmative conclusions. Nonetheless, with bubbles. The duality of financial bubble and
so much empirical observations and anti-bubble corresponds to a far-from-
numerical tests conducted on the financial equilibrium state process of the market,
market price distributions, quantitative which cannot be captured by the GBM-type
financial researchers and professional models under the EMH.
analysts have reached a consensus view that These two models – GBM and LPPL –
the markets are dynamic, and always swing represent the two extreme types of financial
between the efficient and inefficient regimes market price models: the GBM is extremely
and modes [3]. Therefore, dynamic stochastic, while the LPPL is extremely
structural patterns do appear existent from deterministic. This duality of over stochastic
time to time with a temporary and fleeting versus over-deterministic models reminds us
nature and perhaps also with highly abstract of the existence of a big gap in modeling the
invariances. Separating different aspects of financial market prices. It naturally points
the market when making predictions is out a need to find or devise a modeling
important as it leads to profitable trading structure which is useful to explore more
setups. Thus a co- integration (COINT) general stochastic and dynamic patterns of
model that incorporates technical, the market prices between these two
fundamental, sentiment and news is extremities. The model proposed in this
proposed to optimize profit subject to risk work is intended to serve this purpose.
considering real world trading constraints. However relying on just financial time
series price data may not be sufficient as
Literature Review there are several other factors that might
Many new models have been developed, contribute to market fluctuations. Financial
such as agent-based simulation models of markets are sensitive to economic news but
financial markets [4], evolutionary are also influenced by a wide variety of
computing – neural networks and genetic unanticipated events. While the market can
algorithms [5] as well as swarm intelligence react strongly to some news, it could remain
[6] and fractals of complex dimensions. completely insulated from other financial
Geometric Brownian Motion (GBM) by [7]) news.
is one of the earliest most popular models.
It is a model of the Efficient Market Trading Models
Hypothesis (EMH) with little dynamic The ARGARCH Model

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 4
 The AR(p)-GARCH(1,1) process is The real-time trading model traces the
written as market trend and opens position when a
certain threshold is reached. The model
(1.0) identifies an overbought/oversold situation
during market movement and recommends
taking a position against the current trend.
Where N(0,1) and
This strategy is governed by rules that take
. The estimated the recent trading history of the model into
parameters of the AR(p)-GARCH(1,1) account. The RTT model goes neutral only
processes together with the simulated to save profits or when a stop-loss is
residuals are used to generate the simulated reached. The gearing function for the RTT is
returns from these processes [10]. Half of
the average spread is subtracted (added)
from the simulated price process to obtain (1.4)
the simulated bid (ask) prices.
where
Exponential Moving Averages with
Robust Kernels
The moving average indicators are used (1.5)
to summarize the past behavior of a time
series at a given point in time. In many Where is the logarithmic price and
cases, they are used in the form of a
momentum or differential, the difference
between two moving averages. The moving
averages can be defined with their weight or
kernel function. The choice of the kernel (1.6)
function has an influence on the behavior of
the moving average indicator. A particular and
type of moving average called exponential
average plays an important role in the
technical analysis literature. Exponential (1.7)
moving average (EMA) operator is a simple
average operator with
where measure the size of
(1.1) the signal and acts as a contrarian
strategy.

an exponential decaying kernel. r determines and are functions of current position,

the range of the operator and t indexes the volatility and trading frequency. is a
time. An EMA is written as function of position in, previous position,
sign of the return of the previous position.
(1.2) The model is subject to the open-close and
holiday closing hours. The model has
where maximum stop-loss and maximum gain
limits set by the environment.
(1.3)
Support Vector Machine (SVM)
The Real Time Trading (RTT) Model Regression Model
[11] used SVM based models for
predicting currency rates. SVMs formulate

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 5
the regression problem as a quadratic From equation (3.32),Stis a log-normally
optimization problem. They perform a distributed random variable
nonlinearity mapping of the input data into a
high dimensional feature space by means of  S eµt , S 2e2 µt eσ 2t − 1 
a kernel function and then do the linear
St
 0 0 ( ) 
(2.2)
regression in the transformed space. The
whole process results in nonlinear regression The GBM as expressed in equation (2.0)
in the low dimensional space. Given a data provides a simplistic model for financial
set. asset price returns as it only contains a linear
trend – the percentage drift component
N
G = {( xi , yi )}i =1 (1.8) µwhile the rest is a Brownian motion. In
terms of the original asset prices, the trend
modeled is rather geometrical (exponential)
A function f is determined that approximates
with fixed parameters –µandσ.
g(x) based on the knowledge of G as shown
The forex rate S is defined by two
in (1.8)
currencies – foreign versus domestic. The
GBM model for a forex rate takes the form:
f ( x ) = ∑ωi ∅i ( x ) + b (1.9
i
dS = ( rd − rf ) St dt + σ St dW (2.3)
Where ∅i ( x ) are the features, coefficients
whererd and rf denote the (continuous)
ωi and b have to be estimated from data. foreign and domestic interest rate
The limitation in SVM based models is that respectively, and σdenotes the volatility.
the choice of kernel function is based on Similar to equation (2.1), the solution for the
pattern behavior of historical rates of forex rate S can be derived as
individual currency.
 σ2  
Geometric Brownian Model St = S 0 exp  rd − rf −  t + σ Wt  (2.4)
Financial investment returns are  2  
geometric in terms of percentage instead of
arithmetic change. According to [7] the The GBM as a mathematical model of
expected percentage return required by the financial market prices does not express
investors from a financial asset is the reality of the financial market price
independent of the asset’s price. Let S behaviours as we know it today. It suffers
denote the price of an asset, the geometrical from two major shortcomings: (1) it is too
Brownian motion (GBM) as a stochastic simplistic and incapable to express the rich
model of the asset price is defined as variety of stochastic-dynamic patterns in the
price behaviours such as trendlines, trend
dS= µS dt +σSdW (2.0) channels, support and resistance, waves and
bubbles on multiple time scales; (2) it does
where µ is the drift rate (percentage drift) not provide information variables to connect
and σis variance rate (percentage volatility) . to market and economic information sources
Whenµand σare constant, for an arbitrary
initial value S0, the differential equation of Modeling the FOREX process
GBM (2.0) has the analytic solution Trading models emulate the trading
conditions of the real foreign exchange
 σ2   market as closely as possible. They do not
st = s0 exp  µ −  t + σ Wt  (2.1) deal directly but instead instruct human
 2   foreign exchange dealers to make specific
trades. The central part of a trading model is
the analysis of the past price movements

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 6
which are summarized within a trading indicator computations to generating trading
model in terms of indicators. The indicators recommendations.
are then mapped into actual trading A currency P can be modeled in terms of
positions by applying various rules. For a technical information set αP extracted from
instance, a model may enter a long position its own timeseries (P(t), P(t−1),...), an
if an indicator exceeds a certain threshold. economic information set βP including the
Other rules determine whether a deal may be interest rate rp and other selected
made at all. Among various factors, these economic/sentiment indicators and financial
rules determine the timing of the indexes which are influential to P, and an
recommendation. A trading model thus information set extracted from impacting
consists of a set of indicator computations news µP including scheduled snP and
combined with a collection of rules. The unscheduled unP. For any 2 given
former are functions of the price history. currencies P, Q we have a minimal set of
The latter determine the applicability of the dynamic models as

P(t+λ) = FP (αP (P(t), P(t-1), βP (γP,…), µP(snP , unP)) (2.5)

Q(t+λ) = FP (αP (Q(t), Q(t-1), βP (γP,…), µP(snP , unP))

Where λ is how far into the future the stronger currencies which are linked to
prediction is supposed to be. If [PQ]is the countries viewed as stable. A country whose
exchange rate of P to Q, which is actually inflation levels are high will be viewed as a
observable from the forex market, then we poor prospect for forex trading because
have the equation future economic growth is likely to be
hampered by high prices. Investors’
P(t)/Q(t) = [PQ](t) + ԐPQ (2.6) perception of an economy and interpretation
of various economic indicators determine
Where Ԑ is the error term. The two sets of the overall market sentiment for a currency.
equations 2.5- 2.6 provide a framework for
modeling currencies integrating technical, Neural Network Based Forex Model
fundamental, sentiment and news Neural networks are of particular interest
information and observable forex market because they offer a means of efficiently
data. There are many possible component modeling large and complex problems in
models for each of technical, fundamental, which there may be hundreds of predictor
sentiment and news aspects and also for the variables that have many interactions.
interplay of these aspects. Artificial Neural Neural networks (figure 1.1) starts with an
Networks was used to analyze and predict input layer, where each node corresponds to
the technical aspect with special interest on a predictor variable. These input nodes are
volatility and market sentiment. connected to a number of nodes in the
hidden layer. Each input node is connected
Market Sentiment to every node in the hidden layer. The nodes
Market sentiments play an important role in the hidden layer may be connected to
in determining currency values. These nodes in another hidden layer, or to an
directly influence demand and supply within output layer. The output layer consists of
the market [12]. During times of global one or more response variables
economic unrest, values will increase for
.

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 7
 Figure 1.1: Neural network architecture

The three layers feed forward network can • adder sums up all the inputs
be represented by the following expression: modified by their respective weights.
Y=λ0+Σλj×Φj(wj×x) • activation function controls the
amplitude of the output of the neuron
Where
y is the output, Model Application Design
x is a vector of inputs, The system design which is most creative
Φj is a series of functions (typically logistic and challenging refers to the technical
or hyperbolic tangents), specifications and procedures that will be
Wj is a series of weight vectors to ensure applied in implementing the model. It also
that each Φj receives a different input, λ0 a includes the construction of programs and
constant, and λ1 to λM a series of weights program testing. The architecture of the
that weight the outputs from Φ1 to ΦM. system from input through processing to
The neuron mathematical model is output is shown in figure 1.3 while the high
represented by three basic components as level design is shown in figure 1.4
shown in figure 1.2:
• weights,

Figure 1.2: Neuron components

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 8
 Figure 1.3: Trading system Architecture.

High level design

The proposed system consists of five compare its own output to see how close to
Phases: data collection and preprocessing, correct the prediction was, and go back and
the neural network prediction, the neural adjust the weight of the various
indicator, filtering/integration strategies and dependencies until it reaches the correct
the trading system decision. The road to a answer. Since the predictions from the
more reliable financial market analysis neural network most times produce
process begins with a more comprehensive unprofitable trades, filtering techniques
set of market data available through trading which include genetic algorithm are used to
servers in the internet. Technical data sets select the best. The generation of profitable
are extracted with a data collector program trading rule for currency trading is a difficult
and stored in excel file for preprocessing. problem. Integration and network
Data is normalized and fed to the neural combination strategies are used to produce
network for training and testing so that the profitable trading system which is tested
network can learn the dependencies and using the strategy tester of MetaTrader 4. It
apply those dependencies when presented is then deployed on a virtual demo account
with new data. From there the network can for live trading.
.

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 9
 Figure 1.4: High level design
Implementation
Automated trading systems have three you open and close your trades. A trading
functions, viz., money management, risk system operates on the signals generated by
control and market analysis. Money the neural network. The signals are
management is a part the of investment processed to determine if the trader should
strategy. In forex trading, money buy or sell a particular currency pair or
management is used to set the capital should close the existing position(s).
amount for each transaction. The risk Manage risk checks the value of the opened
control function helps traders set a stop and trades in relation to the account balance. It
loss level, and it is able to delete constantly monitors the equity of the trading
transactions that go in the opposite direction account and prevents costly drawdowns by
from the current market. The analysis closing half the volume of all loosing
function is used to capture trading accounts (22% drawdown or by closing all
opportunities, and it is always programmed loosing accounts completely (33%
by a variety of trading methods. However, drawdown). Profit protects the capital base
most trading robots cannot capture all the and the unrealized profits. The capital base
market sentiments, and the historical data is protected by ensuring that the trades are
cannot present all future movements. The exited with a fixed loss when the reasons for
major factor that causes trading robots to holding them are no longer valid. This is
make mistakes is the nature of the forex done by triggering a stop-loss order on the
market. forex brokerage account when the price
The principle of multi-aspects design of crosses the level which defined risk at the
trading systems contributes to the entry. The unrealized profits are protected
systematical and comprehensive solution of either by a take-profit order which is
the trading system creation task. Figure 1.5 triggered on the brokerage account when the
contains the Trading System block and its price reaches the level which defined profit
interactions. Open position and Close at the entry or with the help of the trailing
position controls when and at which price

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 10
stop-loss which gradually locks in more profits as the price moves in favor
.

Figure 1.5: Trading system modules

Discussion of Results
Backtesting helps to see how the system the way it did during the backtesting. The
would have performed if it was run during closer the system's performance during the
some period in the past. Indicator forward testing is to its performance during
parameters are optimized using the price the backtesting the more robust the system
data in the backtesting period. It is important and the more confident we can be that it will
that the time period the system is backtested continue to trade in a similar manner during
on represents the currency pair to trade. It the real-time trading. Backtesting was done
should include all types of market with Metatrader 4 using EURUSD H1and
conditions (trending, range bound) and it the profitability graph is shown in figure 1.6
should be as recent as possible. Once the Implementation of the system was done on
performance of the system is satisfactory, MetaTrader 4 platform which was used to
the forward test is carried out – it is run on create the network and design the trading
the out-of-sample price data (the price data Robot. Testing was done on a demo account
that would be immediate future to the and the result presented in figure 1.7 and
backtesting period). This way the figure 1.8
performance of the system is compared with

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 11
 Figure 1. 6: Backtesting profitability Graph

Figure 1.7: sell signal

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 12
 Figure 1.8: Buy signal

Trading Strategies

The built-in Strategy Tester in on the Forex Markets. The trading robot
MetaTrader 4 Trading Platform is used to performs virtual transactions on historical
test how the Trading robot can perform in data of the financial instrument in
trading. This powerful tool allows simulated accordance to its trading algorithm. The
testing of the efficiency of an Expert result of virtual trading done from January –
Advisor and detects the best input December 2014 is summarized in table 1.2
parameters before it is used for real trading and table 1.2.

Table 1.1: EURUSD Live test Result summary

Initial deposit 10000.00
Total net 458.20 Gross profit 9629.55 Gross loss -9171.35
profit
Profit factor 1.05 Expected payoff` 4.36
Total trades 105 Short positions 105 Long positions 0(0.00%)
(Won%) (4286%) (won%)
Profit trades 45/42.86%) Loss trades 60(57.14%)
(% of total) (% of total)
Largest Profit trade 955.15 Loss trade -497.30
Average Profit trade 213.99 Loss trade 152.86
Maximum Consecutive wins 5(376.75) Consecutive losses (in 7(-661.75)
(profit in money) money)
Maximum Consecutive profit 1158.00 (3) Consecutive loss -893.15 (6)
(count of wins) (count of loss)
Average Consecutive wins 2 Consecutive losses 2

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 13
 Table 1.2:GBPUSD Live test Result
Initial deposit 10000.00
Total net 11017.24 Gross profit 24645.85 Gross loss -13628.62
profit
Profit factor 1.81 Expected payoff` 48.11
Total trades 229 Short positions 110 Long positions 119(73.11%
(Won%) (70.91%) (won%) )
Profit trades 165(72.05% Loss trades 64(27.95%)
(% of total) ) (% of total)
Largest Profit trade 1688.88 Loss trade -212.95
Average Profit trade 149.37 Loss trade -1760.35
Maximum Consecutive wins 12(5151.98) Consecutive losses (in 12(-
(profit in money) money) 5238.64)
Maximum Consecutive profit 5151.98 Consecutive loss -5238.64
(count of wins) (30) (count of loss) (12)
Average Consecutive wins 10 Consecutive losses 4

Comparative evaluation
The comparative evaluation of the model relative position size is RPS=0.8 lots /
with other models indicated significant $10,000, which performed with the best
improvement as shown in Table 1.3. This is Survival Risk Reward Ratio SRRR = 6.54. It
graphically depicted in figure 1.9. The means that for the worst downside risk, the
performance measure of the robot is reward – Annual Return of Investment ARI
summarized in table 1.4 and figure 2.0. is more than 6 times bigger than that
From the table it can be seen that the best maximum risk.

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 14
 Table 1.3: Comparative evaluation of the models
Parameter GBM GARCH EMA COINT
Total Net Profit 733.56 348.44 567.62 2658.29
Bal. Drawdown Abs. 0.00 164.22 399.50 534.36
Eq. Drawdown Max. 339.50(3%) 287.60 413.08 625.36 (6%)
Profit Factor 4.72 1.50 1.56 1.55
Recovery Factor 2.16 1.78 1.95 4.25
Expected Payoff 30.5 8.93 15.37 78.08
Sharp Ratio 0.79 0.46 0.52 0.15
Total Trades 24 39 176 329
Total Deals 48 94 275 658
Profit Trades (% total) 21(87%) 24(42%) 45(42%) 187 (56%)
Average Profit Trade 44.33 41.97 21.70 39.95
Average Consec. Wins 5 6 4 2

Figure 1.9: Comparative Evaluation with other models

Table 1.4: Performance measure of the Robot

RPS ARI MRDD MSR CRRR SRRR
0.1 29.9 6.96 7.48 4.30 4.00
0.2 72.46 13.53 15.65 5.35 4.63
0.3 128.23 19.73 24.58 6.50 5.22
0.4 187.12 25.6 34.41 7.31 5.44
0.5 251.06 30.19 43.25 8.32 5.81
0.6 318.97 37.95 61.16 8.41 5.22
0.7 446.92 41.21 70.10 10.84 6.38
0.8 531.13 44.8 81.23 11.85 6.54
0.9 635.73 49.33 97.36 12.89 6.53
1 706.76 55.25 123.46 12.81 5.73

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 15
 Figure 2.0: Performance measure of the Robot for EURUSD on H1

Conclusion
An improved model that includes more Using integrated neural networks to analyze,
set of dynamic process information about predict and trade currencies in the foreign
the market prices is proposed. This was exchange market also produced better
done through the addition of new parameters prediction than one. A new strategy using
on sentiments and volatility to the existing marcov chain that identified eight market
model. A comparative benchmark with phases instead of the usual three has resulted
models based on technical data alone in a range of significant improvement on
showed that the addition increased profit. profitability
.

References
[1] Iosebashvili, I., Connaghan, C., (2014), Global Currency Trading Volumes Show
October Increase, @
http://www.wsj.com/articles/SB100014240527023035532045793485
91235873208. Retrieved on 25th November, 2014.
[2] Saddiqui, A., (2013), Global FX Volumes Jump 30%, Latest BIS FX Triennial Survey, at:
http://forexmagnates.com/global-fx-volumes-jump-30-latest-bis-fx-triennial-
survey/#sthash.JTcJC0yE.dpuf. Retrieved on 25th November, 2014.
[3] Sornette D., (2006), Essay: A complex system view of why stock market crash, New Thesis
1(1): 5-17.
[4] Sher, G. I., (2012), Evolving Chart Pattern Sensitive Neural Network Based Forex Trading
Agents, retrieved on 18thfeb., 2012.
[5] Myszkowski, P. B. and Bicz, A., (2011), Evolution algorithm in forex trade strategy
generation, Proceedings of the International Multiconference on Computer Science and
Information Technology, pp. 81–88
[7] Black F. and Scholes M., (1973), The pricing of options and corporate liabilities, Journal of

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 16
 Political Economics, 81:637-654.
[8] Pan, H.P., (2011), A basic theory of intelligent finance, New Mathematics and Natural
Computation, accepted, Proc. International Conference on Financial Institutes and Risk
Management, organized by Zhongshan University, Global Finance Journal (US)
[9] Sornette D., (2004), Essay: A complex system view of why stock market crash, New Thesis
1(1): 5-17.
[10] Gencay, R., (1998), Linear, nonlinear and essential foreign exchange prediction with
simple technical rules, Journal of International Economics, forthcoming. [11] Joarder, K.
and Ruhul, A. S., (2004), ANN-Based Forecasting of Foreign Currency Exchange Rates,
Neural Information Processing, Vol. 3, No. 2
[12] Plummer tony, Forecasting Financial Markets: The Psychology of Successful Investing,
Sixth Edition" Kogan Page, @http://ebookee.org/Forecasting-Financial-Markets-The-
Psychology-of-Successful-Investing-
Sixthedition_505068.html#eEuLWKQ8uzGdDoHl.99, retrieved on 18th Feb., 2012

West African Journal of Industrial & Academic Research Vol.14 No.1 June 2015 17