American Journal of Theoretical and Applied Statistics

2013; 2(4): 94-101

Published online July 10, 2013 (http://www.sciencepublishinggroup.com/j/ajtas)
doi: 10.11648/j.ajtas.20130204.11

Evaluation of artificial neural networks in foreign

exchange forecasting
Akintunde Mutairu Oyewale
Department of Statistics, University of Botswana, Botswana

Email address:
[email protected]

To cite this article:

Akintunde Mutairu Oyewale, Evaluation of Artificial Neural Networks in Foreign Exchange Forecasting. American Journal of Theoretical
and Applied Statistics. Vol. 2, No. 4, 2013, pp. 94-101. doi: 10.11648/j.ajtas.20130204.11

Abstract: This study investigates the modeling, description and forecasting of exchange rates of four countries (Great
Britain Pound, Japanese Yen, Nigerian Naira and Batswana Pula) using Artificial Neural Network, the objective of this
paper is to use ANN to predict the trend of these four currencies. ANN was used in training and learning processes and
thereafter the forecast performance was evaluated or measured making use of various loss functions such as root mean
square error (RMSE), mean absolute error (MAE), mean absolute error (MAE), mean absolute precision error (MAPE) and
Theill inequality coefficient (TIC). The loss functions used are good indicator of measuring the forecast performance of
these series, the series with the lowest function gave a best forecast performance. Results show that the ANN is a very
effective tool for exchange rate forecasting. Classical statistical methods are unable to efficiently handle the prediction of
financial time series due to non-linearity, non-stationarity and high degree of noise. Advanced intelligence techniques have
been used in many financial markets to forecast future development of different capital markets. Artificial neural network is
a well tested method for financial markets analysis.

Keywords: Artificial Neural Networks, Foreign Exchange, Loss Functions, Training and Learning Processes

forecast exchange rates. ANNs are nowadays used in a

1. Introduction large variety of modeling and forecasting problems. The
Forecasting is a very important activity and essential merits of using artificial neural networks in this study has
factor in the financial markets that is useful for various to do with adaptive learning nature of ANNs that is ability
players like investors, academia, practitioners, regulators to learn how to perform an operation on data given based
and policy makers. Forecasting with a very week tools on the training and self Organization. It provides promising
has an adverse effect on the economic development due to alternatives tools for forecasters. The inherently non linear
its negative effect on international trade and investment, structure of neural networks is useful for capturing the
using a weak model to forecast will lead to taking a wrong complexity underlying relationship in many real world
judgment. For this reason forecasting is very important to problems It is a more reliable methods for forecasting
all categories listed above. An artificial neural network applications because not only that they can find linear
(ANNs), as an emerging discipline emulates the structure in a problem they can also model linear processes.
information processing capabilities of neurons of the However, the weakness or limitations of neural networks in
human brain. It uses a distributed representation of the any study includes: Black-box problem which is a situation
information stored in the network, and thus resulting in where all information are contained in weight matrices for
robustness against damage and corresponding fault their output and hidden layers, and these do not make
tolerance (Shadbolt and Taylor, 2002). To develop a feed apparent identities of salient predictor variables. This is in
forward artificial neural network for forecasting exchange contrast with parametric variables which do not only
rate purpose, the specification of its architecture in terms of identify influential variables but also provides the degree of
number of inputs, number of hidden layer, number of their contributions.
neurons and output is very important. Many years ago neural networks have found their ways
In the last four or five decades many different non-linear into financial analysis as well, as evidenced by conferences
models have been proposed in the literature to model and such as Neural networks in the Capital markets’ and a large
 American Journal of Theoretical and Applied Statistics 2013; 2(4): 94-101 95

number of books (Trippi and Turban.1993: Azoff, 1994; Where:

Refenes, 19955; Gately, 1996) and articles in scientific h j is the vector of hidden neurons,
journals dealing with financial applications of neural yk is the output-layer of neurons
networks. A casual literature search suggests that many
articles are published dealing with modelling and f h ( x ) and f k ( x ) are logistic sigmoid activation functions
forecasting stock prices (Gencay,1996; ) Haefke and
Hlmenstein, (1996a) and (1996b), Gencay and Stengos, from input layer to the hidden layer and from hidden layer
1998; Qi and Maddala, 1999), exchange rates (Kuan and
to output layer respectively, and λh and λk are the
Liu,1995, Franses and van Griensven, 1998; Franses and
Van Homelen, 1998; Gencay,1999). Neural Networks have variables which control the slope of the sigmoid function.
successfully been used for exchange rate forecasting. The output of each neurons is obtained by applying an
However, due to a large number of parameters to be
estimated empirically, it is not a simple task to select the activation functions f h ( x ) and f k ( x ) . The nodes are used
appropriate neural networks architecture for an exchange
rate forecasting problem. Researchers often overlook the to perform the non-linear input/output transformation using
effect of neural networks parameters on its performance.
a sigmoid activation function.
STEP III: The calculations of errors in the output and
2. Structure of Ann Used in this Study hidden layers can be expressed as follows:
This study used the Feed forward back-propagation The output layer error between the target and the observed
Artificial Neural Network and the procedure used is based output is expressed as:
on the previous work by Erims, et.al 2007. δ k = − ( d k − yk ) f k1
STEP I: Evaluation of net input to the j th node and k th
node in the hidden layer as below:
f k1 = yk (1 − yk ) for sigmoid function
n
net j = ∑ wij xi − θ j Where δ k is the vector of errors for each output neurons
i =1
yk and d k are the target activation of output layer. The
n
netk = ∑ w jk x j − θ k term δ k depends only on the error ( d k − yk ) and f k1 is the
j =1
local slope of the node activation function for output nodes.
Where i is the input node, j is the hidden layer node, The hidden layer error term is expressed as

k is the output layer, wij is the weights connecting the i th n

δ j = f h1 ∑ wkjδ k ,
input node to the j th hidden layer node, w jk is the weights h =1

f h1 = h j (1 − h j ) , for sigmoid function

th th
connecting the j hidden layer node to the k output
layer, θ j is the threshold between the input and hidden
layers θ k the threshold connecting the hidden and output
δ j is the vector of errors for each hidden layer neurons,
layers.

STEP II: Evaluation of the j th node in the hidden layer δ k is a weighted sum of all nodes and f h1 is the local

and the output of the k th node in the output layer as: slope of the node activation for hidden nodes.
STEP IV: The adjustment of weights and thresholds in the
 n 
h j = f h ∑ wij xi − θ j  , output layer and hidden layer is obtained as follows:
 i =1 
(
wkjt +1 = wkjt + αδ k h j + η wkjt − wkj(
t −1)
),
n 
yk = f k ∑ w jk x j − θ k 
 j =1  (
wtji+1 = wtji + αδ j hi + η wtji − w(ji
t −1)
),
Where
θ kt +1 = θ kt + αδ k ,θ tj +1 = θ tj + αδ j
1
and f k ( x ) = , Where α is the learning rate, η is the momentum factor,
(1 + exp ( −λ x ) )
k
and t the time period.
96 Akintunde Mutairu Oyewale: Evaluation of Artificial Neural Networks in Foreign Exchange Forecasting

2.1. Input Data training is to attain a global minimum. This school of

thought emphasized the need/danger of not getting trapped
The inputs data are fed in through a file already created in a local minimum. The other school of thought advocates
in excel software and same is imported into the SPSS a series of train-test interruptions; here the training is
which is the software used in analyzing the data. The study stopped after a pre-determined number of iterations and the
has four variables of length 444 each, each of the four network ability to generalize on the testing set.
variables are analyzed one after the other from where the Generalization refers to the idea that a model based on
forecast values are obtained and compared with sample of the data is suitable for forecasting the entire
observed/collected data. The selection of hidden layers is population. It is the aim of the study to use the first school
largely due to the numbers that give a satisfactory model of thought as pre-determined the number of iteration ahead
performance. The number of hidden layers depend largely may lead to the bias of either over fitting or under fitting,
on the complexity or otherwise of problem theoretically. instead the network is allowed to determine the number of
However, it has been proved that two hidden layers are iteration that give optimal result.
sufficient and capable to solve any input-output problem. In The learning algorithm that controls back-propagation
the light of this, the study start from 2 hidden layers and follows a number of steps:
different hidden layers used until the optimum number of (i) Initialization: Initialize connection weights and
hidden layers emerged. The rule of thumb actually neurons to small random values.
recommended that the number of connections in a network (ii) Data introduction: Introduce the continuous
should be ten percent of the training data. This study will data set of inputs and actual outputs to the
use 1728 0f 1776 collected data for monthly forecasting; back-propagation network.
this implies that the nodes to be selected should be up to (iii) Calculations of outputs: Calculate the outputs,
173 connections. The above rule shall be taken seriously as and adjust the connection weights several times
a guide and the test shall be conducted for the optimum applying the current network error.
number of nodes so needed in each hidden layers by (iv) Adjustments: Adjust the activation thresholds
looking at the node with smallest mean square error. The and weights of the neurons in the hidden layers
search for the optimum number of nodes shall began by according to Delta rule.
first looking at 2 neurons in each hidden layer and there (v) Repeat steps (ii) to (iv) until desired network
after proceed to higher number of neurons until the accuracy (error) is reached.
optimum result is accomplished in terms of low mean Delta rule recursive algorithm re-adjusts connection
square error. The output here refers to the predicted values weights and biases starting at the output nodes and working
of the monthly data of exchange rate loaded into ANN back to the first hidden layer. Connection weights are
model. The output will then be compared with collected or adjusted by:
observed data where in the test of adequacy shall be
performed. γ ij ( t + 1) = iγ ij ( t ) + ηδ j x1i

Where
3. Train the Network and the Number of
Iteration for Training γ ij = the weight from hidden node I or from an input to

The objective of training is to find the set of weights node j at time t .

between the neurons that determine the global minimum of
error function. Unless the model is over fitted the set of x1i = either the output of node I or an input.
weights should provide good generalization. The back
propagation network uses a gradient descent training η = the positive constant that measures the speed of the
algorithm which adjusts the weights to move down the convergence of the weight vector (gain term or learning
steepest slope of the error surface. Finding the local
minimum is not guaranteed since the error surface can rate).
include many local minima in which the algorithm can
δ j = an error term for j (k goes over all nodes in the
become “stuck”. This section discusses when to stop
training a neural network and the selection of learning rate
layers before node j ):
and corresponding momentum values. There are two
schools of thoughts regarding the number of iteration for
 y j (1 − y j )( d j − y j ) , when j is an output node
training before stopping. The first school of thought 
δj = 
 x j (1 − x j ) ∑ δ k γ jk , when j is a hidden node
emphasized that the researcher can only stop training when 1 1

there is no improvement in the error function based on the  k

reasonable number of randomly selected starting weights.
The point at which the network error does not improve is Where
called point of convergence. The objective of convergence
 American Journal of Theoretical and Applied Statistics 2013; 2(4): 94-101 97

/ARCHITECTURE AUTOMATIC=YES
d j = the desired output of node j . (MINUNITS=1 MAXUNITS=50)
/CRITERIA TRAINING=BATCH
y j = the actual output of the node j . OPTIMIZATION=SCALEDCONJUGATE
LAMBDAINITIAL=0.0000005 SIGMAINITIAL=0.00005
The node biases are adapted through the similar iterative INTERVALCENTER=0 INTERVALOFFSET=0.5
algorithm by assuming they are connection weights on links MEMSIZE=1000
from constant-valued input /PRINT CPS NETWORKINFO SUMMARY
CLASSIFICATION
/PLOT NETWORK
4. Model Selection and Evaluation /STOPPINGRULES ERRORSTEPS= 1 (DATA=AUTO)
TRAININGTIMER=ON (MAXTIME=15)
Having obtained the outputs which are the forecast and
MAXEPOCHS=AUTO ERRORCHANGE=1.0E-4
the predicted values, these values are evaluated by
ERRORRATIO=0.0010
comparing the predicted values with the targets values
/MISSING USERMISSING=EXCLUDE.
using the mean square error and for the fewer pipes
deviation (the difference between the predicted from the 5.1. Case Processing Summary
actual in pipes).
The following criteria will be used to evaluate the Table 1 below shows that 320 cases were assigned to the
forecast performance of this model training sample and 124 to the testing sample (hold out),
this is in conformity with the standard rules that at least 70%

( )
 T
 of the sample must be set for training and the remaining
RMSE = T −1 ∑ Yt − Yˆt  sample left is for testing.
 t =1 
T
Table 1. Case Processing Summary
MAE = T ∑ Yt − Yˆt
−1

t =1
N Percent
Training 320 72.1%

∑ (Y − Y )
T
−1 Sample
MAD = T t t Testing 124 27.9%
t =1
Valid 444 100.0%
T
Yt − Yˆt Excluded 0
MAPE = T −1 ∑ X 100
t =1 Yt Total 444

The smaller the values of these criteria the better the Table 2. Network Information
forecast performance; If the values are close to zero we say
it is better, if the results are not consistent among the first Input Factors 1 TIME
three we choose the MAPE to be the benchmark. Layer Number of Units a
37

Number of Hidden Layers 1

5. Data Analysis and Discussion of Hidden Number of Units in Hidden Layer 1 a
8
Results Layer(s) Hyperbolic
Activation Function
tangent
GET DATA
/TYPE=XLS 1 POUND
/FILE='C:\Users\Akintunde\Desktop\2 NORMALIZED Dependent 2 YEN
DATA OF 04012013.xls'
Variables 3 PULA
/SHEET=name 'NORMALIZED DATA OF 17122012'
/CELLRANGE=full 4 NAIRA
Output
/READNAMES=on Number of Units 4
/ASSUMEDSTRWIDTH=32767. Layer
Rescaling Method for Scale
EXECUTE. Standardized
DATASET NAME DataSet1 WINDOW=FRONT. Dependents

*Multilayer Perceptron Network. Activation Function Identity

MLP ZNAIRA (MLEVEL=S) ZPOUND (MLEVEL=S) Error Function Sum of Squares
ZPULA (MLEVEL=S) ZYEN (MLEVEL=S) BY DATE
/PARTITION TRAINING=7 TESTING=3
HOLDOUT=0
98 Akintunde Mutairu Oyewale: Evaluation of Artificial Neural Networks in Foreign Exchange Forecasting

5.3. Model Summary

Table 3 below shows information on the result of training

and applying the network to the testing (hold out) sample.
Sums of squares error is displayed because the output layer
has scale-dependent variables. This is the error function
that the network tries to minimize during training. It should
be noted that the sums of squares and all errors values are
computed for the rescaled values of the dependent variables
as shown in the table below.
The relative error for each scale-dependent variable is
the ratio of the sum of squares error for dependent variable
to the sum of squares error for the “null” model in which
the mean value of the dependent variable is used as the
predicted value for each case. There appeared to be more
errors in Pound (0.014), Pula (0.012)and Yen (0.018)
compared with Naira (0.001).
The average overall error is the ratio of sum of squares
for all dependent variable to the sum of squares error for
the “null” model, in which the mean values of dependent
variables are used as the predicted values for each case. In
this case the average overall error differs considerably for
Naira and somehow closes to other variables. The average
overall relative error and relative errors are fairly constant
for training and testing, which gives us some confidence
that the model is not over trained and that the error in future
cases scored by the network will be close to the error
reported in the table.

Table 3. Model Summary of the series used in the study

Sum of Squares Error 7.095

Average Overall Relative Error .011
POUND .014
Relative Error for YEN .018
Scale Dependents PULA .012
Training
NAIRA .001

1 consecutive
Stopping Rule Used step(s) with no
decrease in errora

Training Time 0:00:00.45

Figure 1. Network Architecture for The Study Sum of Squares Error 2.846
Average Overall Relative Error .012
5.2. Network Information
POUND .011
The network information table 2 displays information Testing
Relative Error for Scale YEN .020
about neural networks and its usefulness in ensuring that
the specifications are correct. We note here that the number Dependents PULA .016
of units in the input layer is the number of factors which is NAIRA .001
37 (years) in all, likewise, a separate output units is created
a. Error computations are based on the testing sample.
for each category of previously defaulted, for a total of 4
units in the output layer. Automatic architecture selection
has chosen 9 units in the hidden layer. All other network
information is default for the procedure,
 American Journal of Theoretical and Applied Statistics 2013; 2(4): 94-101 99

5.4. Parameter Estimates of the Series Used in the Study the data for the study was collected is 37. The predicted
values for the ninth hidden layers are seen in the column
Table 4 show the parameters estimates and the predicted titled output layer. This prediction we shall later used in
values as shown in the table below, the total number of chapter to determine if this model has actually predicted
hidden layers used for the architecture of the model under very well.
study are nine and the number of years (Time) upon which
Table 4. Parameter Estimates of the series used in the study

Predictor Predicted

Hidden Layer 1 Output Layer

H(1:1) H(1:2) H(1:3) H(1:4) H(1:5) H(1:6) H(1:7) H(1:8) POUND YEN PULA NAIRA

(Bias) -0.265 -0.132 -0.257 0.316 -0.796 -0.321 0.344 0.198

[TIME=1] -0.275 -0.389 -0.058 -0.698 -0.725 -0.471 -1.155 0.257
[TIME=2] -0.247 -0.659 -0.026 -0.576 0.104 -0.062 -1.385 0.028
[TIME=3] 0.217 -0.113 0.458 -0.452 -0.082 -0.014 -0.849 0.32
[TIME=4] -0.316 0.422 -0.179 -0.001 -0.425 -0.603 -0.656 -0.355
[TIME=5] 0.354 0.232 0.161 -0.041 -0.528 -0.784 -0.383 0.121
[TIME=6] 0.554 0.022 -0.182 -0.336 -0.132 -0.634 -0.357 0.085
[TIME=7] 0.476 0.098 -0.28 -0.353 0.235 -0.197 -0.428 -0.051
[TIME=8] 0.056 -0.34 0.105 -0.032 -0.117 -0.381 -0.768 -0.369
[TIME=9] -0.31 -0.356 -0.036 0.18 -0.072 -0.395 -0.883 -0.369
[TIME=10] -0.074 -0.534 0.07 0.091 -0.028 -0.14 -0.793 0.178
[TIME=11] 0.185 -0.724 0.422 -0.034 -0.686 0.137 -0.611 0.199
[TIME=12] 0.319 -0.101 0.001 0.36 -0.567 -0.141 -0.172 0.455
[TIME=13] 0.314 0.51 -0.506 0.007 -0.055 0.178 -0.07 -0.247
[TIME=14] 0.508 0.164 -0.625 0.461 -0.285 -0.407 0.193 0.3
[TIME=15] 0.656 0.057 -0.166 0.458 -0.371 -0.236 0.205 0.159
[TIME=16] 0.605 -0.026 -0.621 0.5 -0.197 -0.527 0.211 -0.382
[TIME=17] 0.393 0.204 -0.485 0.41 -0.429 -0.367 0.152 -0.365
[TIME=18] -0.134 0.113 -0.355 0.638 0.639 0.462 -0.359 0.099
[TIME=19] -0.165 0.62 0.067 0.649 0.117 0.504 -0.153 -0.311
[TIME=20] -0.145 0.698 -0.008 0.304 0.001 0.745 -0.081 0.151
[TIME=21] -0.136 0.508 -0.11 0.578 -0.015 0.596 -0.028 0.3
[TIME=22] 0.16 0.039 -0.158 0.365 -0.126 0.789 0.111 0
[TIME=23] 0.076 -0.205 -0.249 0.1 -0.437 1.006 0.051 0.242
[TIME=24] 0.38 -0.788 -0.218 -0.027 -0.493 1.244 0.193 0.312
[TIME=25] -0.072 -0.022 0.262 0.04 0.44 0.169 0.279 -0.136
[TIME=26] -0.142 0.041 0.311 -0.259 0.167 0.276 0.471 0.321
[TIME=27] 0.191 -0.306 0.55 -0.31 -0.083 0.337 0.994 -0.168
[TIME=28] -0.392 -0.664 0.307 -0.172 -0.221 -0.134 0.601 0.413
[TIME=29] -0.519 0.521 0.846 -0.188 0.564 -0.17 0.204 -0.012
[TIME=30] -0.523 0.318 0.23 -0.14 1.288 -0.233 0.071 -0.255
[TIME=31] -0.184 0.167 0.676 -0.193 0.695 -0.131 0.38 0.301
[TIME=32] -0.237 0.34 0.773 -0.556 0.218 0.489 0.531 -0.193
Input [TIME=33] 0.036 -0.416 0.632 -0.169 0.153 0.184 0.896 -0.19
Layer [TIME=34] -0.491 -0.101 0.15 -0.52 -0.376 0.611 0.798 -0.169
100 Akintunde Mutairu Oyewale: Evaluation of Artificial Neural Networks in Foreign Exchange Forecasting

[TIME=35] -0.674 0.035 0.282 -0.61 -0.24 0.277 1.572 -0.115

[TIME=36] -0.588 -0.001 0.185 -0.458 0.158 0.136 1.375 -0.308
[TIME=37] -1.019 0.328 0.249 -0.569 0.014 0.231 1.519 -0.29
(Bias) -0.058 0.511 -0.43 -0.047
H(1:1) 1.159 0.568 -0.94 -0.992
H(1:2) 0.504 -0.46 -0.61 -0.286
H(1:3) -0.735 0.291 0.393 0.559
H(1:4) -0.963 -0.77 -0.3 -0.57
H(1:5) -0.84 -0.42 0.184 0.721
H(1:6) -1.212 -0.39 0.276 -0.258
Hidden H(1:7) -0.827 -1.73 1.627 1.444
Layer 1 H(1:8) -0.147 -0.18 -0.13 0.011

6. Prediction by Observed Chart 0n the x-axis for the combined training and testing samples.
The network shows that all series appears to be reasonably
Figures 2-4 below, shows that for scale dependent good at predicting as they all shows show how upward
variables, the predicted –by-observed chart displays scatter movement of each series.
plots of predicted values on the y-axis by observed values

Figure 2. predicted-by observed chart for Pound

Figure 3. predicted-by observed chart for Yen

Figure 4. predicted-by observed chart for Pula

 American Journal of Theoretical and Applied Statistics 2013; 2(4): 94-101 101

Figure 5. predicted-by observed chart for Naira

109-114.

7. Conclusion [4] Franses, P.H and Van Homelin, P. (1998). On forecasting

exchange rates using neural networks. Journal of applied
From summary table there appeared to be more errors in financial Economics, 8, 589-596.
Pound (0.014), Pula (0.012) and Yen (0.018) compared [5] Gately, E. (1995). Neural Networks for financial forecasting.
with Naira (0.001). So also the average overall relative Wiley trader’s exchange publishing company
error and relative errors are fairly constant for training and
[6] Gencay, R.(1999). Linear non-linear and essential foreign
testing, which gives us some confidence that the model is exchange prediction with simple technical rules. Journals
not over trained and that the error in future cases scored by of international Economics, 47, 91-107.
the network will be close to the error reported in the table.
Figures 2-4 below, shows that for scale dependent variables, [7] Gencay, R. and Stengos, T. (1998). Moving average rules,
volumes and predictability of security returns with feed
the predicted –by-observed chart displays scatter plots of forward neural networks. Journals of forecasting, 17,
predicted values on the y-axis by observed values on the 401-414.
x-axis for the combined training and testing samples. The
network shows that all series appears to be reasonably good [8] Haefke, C. and Helmenstein, C., (1996a,1996b). Forecasting
Austrian IPOs: An Application of Linear and Neural
at predicting as they all shows show how upward Network Error-Correction Models
movement of each series. This implies that any forecast
made on the basis of this model will be very reasonable and [9] Kuan, C.M. and Liu, T. (1995).Forecasting exchange rates
produce excellent results. using feed forward and recurrent neural networks.
Journal of Applied Econometrics 10(4), 347-364.

[10] Qi, M. and Madalla, G.S. (1999). Economic factors and

References stock market: A new perspective. Journal of forecasting,18,
151-166.
[1] Azoff, E.M., (1994). Neural networks time series
forecasting of financial markets. Wiley publishing company [11] Refenes, A. P N. (1995). "Neural Networks in the Capital
(Chichester and New York). Markets", Wiley & Sons, Chichester, ISBN 0-471-94364-9.

[2] Erims, K., Midilli, A., Dincer, I. and Rosen, M. A. (2007). [12] Shadboth, J. and Taylor, J. (2002). neural networks and the
Artifical Neural Network analysis of World Green Energy financial markets. Springer publishing company Limited.
Use. Energy Policy, 35, no. 3, p. 1731-1743. [13] Trippi, R. R. and Turban, E. (1993). Neural Networks in
[3] Franses, P.H and Van Griensen, K. (1998). Forecasting finance and investing: using artificial intelligence to
exchange rates using neural networks for technical trading improve real world performances. Probus publishing
rules. Studies in non-linear dynamics and Econometrics, 4, company (Chicago III).