Forecasting Currency Exchange Rates: Neural Networks and

the Random Walk Model

Eric W. Tyree and J. A. Long

City University

Presented at the Third International Conference on Artificial Intelligence Applications on

Wall Street
New York, 1995
 Forecasting Currency Exchange Rates: Neural Networks
and the Random Walk Model
Eric W. Tyree J. A. Long

Dept. of Business Computing Dept. of Business Computing

City University City University
London EC1V 0HB London EC1V 0HB
United Kingdom United Kingdom
Tel: 017 - 477 - 8413 Tel: 017 - 477 - 3404
E-Mail: [Link]@[Link] E-Mail: [Link]@[Link]

ABSTRACT quantitative methods used to forecast the behaviour of

financial markets often produce unsatisfactory if not
This work provides an evaluation of the use of neural dismal results given the complex interactions between
networks as a technique for forecasting currency a given market's behaviour and other economic
exchange rates. Recently, successful attempts at phenomena. Part of the problem lies in the fact that
forecasting exchange rates such as the US$ - DM the relationships existing between financial markets
and US$ - SF have been reported in the literature (i.e. and the economy as a whole are often poorly
Refenes et al (1993, Weigend et al (1992))) but their understood. On top of this there are also a variety of
methodologies have been less than stringent leaving political and psychological factors influencing the
them open to accusations of data mining. The work dynamics of the markets over time. Neural networks
presented here will attempt to replicate some of this may provide some hope of producing a suitable
previous work and then subjugate the resulting neural methodology for overcoming some of these
network forecasts to a more stringent level of difficulties.
analysis. More specifically, standard backpropagated
feedforward networks will be used to forecast the US$ A number of successful claims of using neural
- DM exchange rate 1, 5, and 20 trading days into the network based market forecasting systems have been
future with the resulting performances compared to published. Unfortunately, much of this work suffers
the random walk forecasting model and to an from inadequate documentation regarding
autoregressve forecasting model. The experimental methodology (Binks and Allinson (1991), Collard
techniques used here are also proposed as a general (1991), Lee and Park, (1992)) or claims of positive
framework which should be followed when making results not backed up by comparisons with other
claims of the successful application of neural relevant forecasting techniques (Binks and Allinson
networks to financial time series generally seen as (1991), Lee and Park, (1992), Collard (1991)
unforecastable. Weigend et al (1992)). This makes it difficult to both
replicate previous work and obtain an accurate
assessment of just how well connectionist techniques
INTRODUCTION really perform in comparison to other forecasting
techniques. What previous work has been done using
One of the more difficult problems in economics is connectionist approaches to market forecasting can be
the forecasting of financial markets. Traditional roughly categorised based on how a forecast is being
extracted from the input data with the neural network
model. Most have attempted to extrapolate the future current price changes are independent of past price
behaviour of a market with a neural network based changes. In other words, univariate forecasting should
times series analysis by having the network output be impossible as past price changes do not offer any
some value representing the future behaviour of the clues to what form the future behaviour of prices
market (i.e. forecasting the price, expected return or might take. Since price changes in efficient markets
degree of change etc...). This is usually done by giving such as exchange rates are assumed to be a random
the network information about the market's past distribution with 0 mean (see Pindyck and Rubinfeld
behaviour (Refenes et al (1993)) or information about (1991)) the best forecast one can make for any amount
its past behaviour in conjunction with the dynamics of of time in the future is to assume the future price will
a variety of other economic variables used as be the same as today's price. As previous claims of
additional input (Weigend et al (1992), Lee and Park success with univariate forecasting of the US$ - DM
(1992) and Hutchinson (1994)). Others have tried to exchange rate contradict the random walk model, the
train the network to recognise known market patterns random walk model is the most appropriate to base a
(Binks and Allinson, (1991)) or attempt to train the comparison with neural networks with1 in this
network to learn an optimised trading strategy instance.
(Collard (1991) and Kimoto et al (1990)).
Second, when claiming positive results steps should
This report will attempt to apply a connectionist be taken to guard against accusations of data mining.
approach to the forecasting of a notoriously It will be shown here that spurious results are not
"unpredictable" financial market - currency exchange difficult to obtain in some instances. To help
rates. Some relatively straight foreword methods of circumvent this problem, the approach taken here is
using standard backpropagated feedforward neural to run our simulations on multiple portions of the data
networks to forecast the US$ - DM exchange rate will set to guard against the possibility of chance results.
be analysed and compared with other forecasting
models. These experiments will include univariate
forecasting of the exchange rate at 1, 5 and 20 days in METHODS
advance and multivariate forecasts at 1 and 5 days in
advance. This study consists of five main experiments intended
to examine the relative performance of neural
In addition, a methodological framework is also networks and the random walk model in forecasting
proposed for the use of neural networks in financial the US$ - DM exchange rate. The first experiment
forecasting. The framework is quite simple and attempts to use a feedforward backpropagated
consists of two basic techniques. First, the network to forecast the US$ - DM exchange rate one
performance of neural networks should be compared trading day in advance using input consisting solely
with other relevant forecasting models. Simply of daily US$ - DM data in much the same way as
demonstrating that neural network based methods Refenes et al (1993). The second experiment attempts
"work" is not enough as this does not shed any light to fit the random walk model to the exchange rate data
on their relative performance to potentially simpler to more accurately ascertain the appropriateness of the
and more accurate forecasting methods. For this work, random walk model as an explanation for the
the random walk forecasting model will be use as the behaviour of the price changes in the exchange rate.
primary comparison model as currency exchange rates The third experiment will use an identical technique as
are widely viewed to be best explained as random
1Note that the random walk model only refers to univariate or
walks (Diebold and Nason, (1990)). The random walk "technical" forecasting. It does not state that price changes in
model simply states that due to market efficiency, particular markets that follow random walks are also operating
independent of other variables.
the first experiment to forecast 5 trading days (one
week) and 20 trading days (one month) in advance. 0.008 S LP AR (2)

Mean Sq. Error

The fourth experiment will conclude the univariate
0.006 R .W. MLP
forecasting by taking a multistep approach to
forecasting. In multistep forecasting the output from 0.004
the network after presentation of the final training
pattern is taken as input to the network for the next 0.002
forecast step. This process is then repeated for the
0
entire length of the forecast lead period. Finally, the
last experiment will attempt one and five trading day Fig. 1: Results of univariate single day forecasts.
forecasts using multivariate input incorporating
interest rates and other currencies.
RESULTS
In all the experiments, the networks consisted of
standard feedforward networks with full connectivity Experiment 1: Single day univariate forecasting
between layers. Connections between units were
restricted to being from one layer to the next. Learning In this first experiment the notion that a previous
was conducted with the standard backpropagation sequence of the US$ - DM exchange rate Wi =
algorithm with momentum term and utilised the y(t), y(t - 1), ... , y(t - n) can be used to forecast the
standard sum squared error cost function: value of the exchange rate one day in advance Wo =
N

∑ (A
y(t + 1) with a neural network was examined. The
E = 0.5 p − Dp )2
training data used consisted of the first 850 trading
p =1
days of the exchange rate data and the test data
where A is the network output for output pattern p, D
consisted of the following 50 days. Various sizes of
is the desired output for pattern p and N is the number
Wi were tried, ranging from 2 to 20 trading days, of
of patterns. The results of the neural network model
which none provided satisfactory results. Increasing
were then analysed with respect to random walk
the number of hidden units did not produce any
forecasting model. The random walk forecasting
improvement either.
model was defined as today's price being the best
forecast for any point in the future. More formally, y(t
The results are shown in fig. 1 which compares the
+ N) = y(t - 1) + ε (t) where y(t + N) is any point in
results from the random walk model, an
the future and ε (t) is an error term2. The relative
autoregressive model, a multilayer perceptron with 10
performances of the models used in this work was
inputs and 5 hidden units and a single layer
analysed by comparing the mean squared errors of the
perceptron with 10 inputs. The network results
models over the test sets. All currency and interest rate
displayed were typical of that found in this experiment
data used was daily data from the period beginning
regardless of the hidden or input layer size. Clearly,
Jan. 1, 1990 and ending May 31, 1994.
the random walk model is producing more accurate
forecasts. An autoregressive process used to fit the
training data produced a second order linear model of
the form:

y(t) = 1.0426y(t - 1) - 0.053y(t - 2) + ε (t)

2Other definitions of the random walk model exist which also
accommodate systematic "drift" in the data. As no such long term
drift was found, the simpler version was used here.
where y(t) is the value of the US$ - DM at time t Experiment 2: Fitting the random walk model
indicating that the value at time t is almost entirely
dependant on the value at time t - 1. Although the The superior performance of the random walk model
networks did slightly better than the AR(2), none in the previous experiment necessitates
outperformed the random walk model. Even further, investigating more formally how well the random
the best performing networks seemed to be walk model can explain the prince changes in the US$
implementing something approaching a random walk - DM exchange rate. If the price changes in the US$ -
as can be seen in fig. 2. The top of fig. 2 displays the DM exchange rate do in fact follow a random walk,
random walk forecast along with the test set while the the differences between one day's rate and the next
bottom displays the output of the 10 - 5 - 1 network should be random. In more precise terms, if the daily
on the test data. Clearly, the network is simply using changes in the exchange rate can be explained by the
the previous value of the exchange rate as a forecast random walk model, the residuals left over from
for the next day - a random walk. fitting the random walk model should be random
noise. The residuals are defined as Y(t) - M(t) where
S ingle day random walk Y(t) is the actual value of the time series at time t and
M(t) is the value at time t given by the model.
0.53
US$/DM (normalized)

0.51
0.49 In this experiment, the random walk model is used to
0.47 fit the US$ - DM exchange rate. As mentioned before,
0.45 this model assumes that a value at a given time t + N
0.43
is equal to the value at time t plus some noise. If this is
0.41
0.39 true, then the k - day differenced US$ - DM exchange
0.37 rate series should be random noise as these values
0.35 would simply be the random series left over from
11
16
21
26
31
36
41
46
1
6

fitting the random walk model to the k - day price

T IME (days )
changes. It is also good practice to look at the squares
of the differenced series as this helps prevent any
cyclic component of the series from cancelling out and
making the series appear random when in fact it is not.
Univariate s ingle day forecas t

0.55 Two tests for randomness were run to test the fit of
US$/DM (normalized)

random walk model on the differenced and

0.5 differenced squared US$ - DM exchange rate data.
These tests consisted of the difference-sign test and a
0.45 serial correlation test. The difference sign test simply
looks at the differenced series and counts the number
0.4 of times a positive change is found in the series. It can
be shown (see Kendall and Stuart (1968)) that a truly
0.35
random series will have (n - 1)/2 positive changes in
11
16
21
26
31
36
41
46
1
6

T IME (days )
value and a variance of (n + 1)/12 with the resulting
distribution tending rapidly towards normality
(Moore and Wallis (1943)). The serial correlation test
Fig. 2: Random walk (top) and neural network (bottom) simply tries to find a correlation between successive
single day forecasts. The thick lines are the actual exchange rate values. If a given series has structure beyond random
and the thin lines represent the forecasts.
fluctuations, there will be some degree of correlation
between one value and the next (Kendall and Stuart Looking first at the difference-sign test, the top part of
(1968)). table gives the expected number of turning points and
the expected standard deviation for each lag time. The
next four rows give the results of the differenced and
RANDOM WALK MODEL RESIDUALS
differenced squared exchange rates along with p
Lag Times
which indicates the probability that the number of
1 Day 2 Day 3 Day 4 Day 5 Day
positive sign changes found in the exchange rate is
DIFFERENCE SIGN indicative of it being a random series using a simple z
exp p. d. 575.0 288.0 192.0 144.0 115.0 test.
exp s.d. 9.80 6.94 5.67 4.92 4.40
In short, these results are quite marginal except in two
diff 559, 278, 185, 144, 116, cases. A 1 day lag it can be said with greater than 95%
p= 0.10 0.15 0.22 1.0 0.82
diff sq. 544, 298, 194, 141, 117, certainty (p = 0.002) that the difference squared series
p= 0.002 0.15 0.73 0.54 8.82 in not random. Conversely, at 4 days it can be said
SERIAL CORRELATION that the differenced series is random with greater than
diff 0.030 0.005 -0.036 -0.049 -0.056
95% certainty (p = 1.0). The rest of the results fail the
diff sq. 0.063 0.112 0.208 -0.005 0.183 95% percent certainty criteria for either accepting or
±sig @
0.059 0.083 0.102 0.118 0.132
rejecting that the exchange rate is random. In these
0.95%
3
cases the probability that the number of positive
Table 1: Random walk model residuals. changes observed indicate that the exchange rate is
random ranged from 0.10% (p = 0.1) and 82% (p =
In this experiment the entire data set used in this work 0.82). A less stringent criteria for accepting the
was run through these tests at time lags of 1, 2, 3, 4, hypothesis that the changes in the US$ - DM are
and 5 days. In other words, the adequacy of the random could be adopted in which any number of
random walk model is being tested for the changes in positive changes found within one standard deviation
the US$ - DM exchange rate for periods of 1 to 5 (i.e. p <= 68%) of the expected number of positive
days. It should be noted that when testing at time lags changes would be accepted as indicative of
greater than 1 day the series must be differenced such randomness. In this case, the results are still mixed
that dn = ytk - y(t - 1)k where d is the differenced value with the differences series indicating non randomness
and k is the lag. The reason for this is that if one were at lags of 1 (p = 0.01), 2 (p = 0.15) and 3 (p = 0.22)
to simply difference every value in the series from the and the differenced squared series indicating non
value k time steps in the past, one would artificially randomness at lags of 1 (p = 0.002), 2 (p = 0.15) and 4
induce correlation in the series that did not initially (p = 0.54).
exist as the various values resulting from the
subtraction process would contain common terms. The serial correlation test resulted in slightly more
Therefore, a series of N values will produce a consistent results. For the differenced series, all five
differenced series of size N/k. For this reason, only time lags indicated randomness p >= 95% while the
lags of up to 5 days were tested for the random walk differenced squared series indicated significant non
model as lags of more than 5 would produce too small randomness at 95% certainty at all lags except 4.
of samples. The results are displayed in table 1.
As much as these results are mixed, they do seem to
3This is the probability that the correlation found is significantly indicate that the US$ - DM exchange rate is not
different from 0 using the general heuristic of 2/ N to determine strictly random. In other words, there is some structure
95% certainty (Chatfield, (1975)). to be found in the 1 - 5 day price changes albeit small
and probably very subtle. The random walk model demonstrated that there may be some subltle structure
can probably be rejected as the most appropriate in the US$ - DM exchange rate, this next experiment
model explaining the changes in the US$ - DM attempted to use a network to find a relationship
exchange rate. Nonetheless, because the structure that between a segment of the time series consisting of
exists in the changes is so small (and possibly data up to and including the rate at time t and the
complex) forecasting these changes will most value of the exchange rate at time t + 5 and t + 20 that
probably be anything but trivial. can be used for a forecast.

In regards to the 5 day forecasts, various sizes of input

ML P(20)
window were attempted. An input window size of 20
0.014 trading days Wi = y(1), y(t - 1) ... y(t - 20) is
0.012 R .W.(20) displayed in the results here as originally it was
Mean Sq. Error

0.01 thought that a month's worth of trading days would be

0.008 sufficient for the network to derive a weekly forecast.
0.006 Other window sizes did not yield any better results.
0.004 The number of hidden units was also varied from 1 -
R .W.(5) ML P(5)
0.002 30 none of which lead to an improved performance.
0 The results for the 5 day forecast are displayed in fig.
3.
Fig. 3: Results of the 5 and 20 day forecasts.
Again, the random walk model is producing the
lowest error performance. As with the one day
Univariate five day forecas t forecasts, the networks seem to be implementing a
random walk type of forecast which can be seen in
0.55 fig. 4. Again, none of the networks have improved on
US$/DM (nomalized)

the mean error performance of the random walk

0.5
model. Also, looking at the graphs of the actual output
of the networks (fig. 5) it can be seen that the
0.45
networks forecasting 20 days in advance are simply
0.4 outputting previous input.

0.35 Fig. 3 also gives the results of the 20 day forecasts in

11
16
21
26
31
36
41
46
1
6

which a 60 trading day (3 month) input window was

T IME (days ) used. Fig. 5 displays a typical result from the
networks. Clearly, the networks are simply giving the
last seen input value (albeit somewhat degraded) as a
Fig. 4: The 5 day univariate forecast. The thick line is the actual forecast of the future course of the exchange rate. As
US$ - DM and the thin line is the forecast.
with the 5 day forecasts, the number of hidden units
was systematically varied. Due to long learning times,
Experiment 3: 5 and 20 day univariate forecasting
though, the input window size was pegged at a value
of 60 trading days.
The previous experiment attempted to forecast the
US$ - DM exchange rate at time t + 1 with a moving
window of univariate data up to and including the rate
at time t. Given the results of experiment 2 which
 and the next 50 days as a test set. Apparently, the
20 day random walk forecas t networks are modelling the data as being cyclical in
nature whose dynamics are largely determined by the
0.8
US$/DM (normalized)

0.75
in the previous input although what exactly the
0.7 networks are modelling in the data is unclear. All that
0.65 can be said, though, is that the initial "positive" results
0.6 in fig. 6 were probably spurious. It should also be
0.55
0.5 noted that a similar result was found on the 20 day
0.45 forecasting where a single positive result could not be
0.4 replicated on other parts of the data set. Nonetheless,
0.35
100 these results underscore the need for more care to be
1
10
19
28
37
46
55
64
73
82
91

taken when analysing results to ensure they are not

T IME (days ) spurious.

20 Day univariate forecas t Multis tep forecas t for data s et A

US$/DM (normalized)

0.85 0.55
0.53
US$/DM (normalized)

0.75
0.51
0.65 0.49
0.55 0.47
0.45
0.45 0.43
0.41
0.35
0.39
1
9
17
25
33
41
49
57
65
73
81
89
97

0.37
T IME (days ) 0.35
1
5
9
13
17
21
25
29
33
37
41
45
49
T IME (days )
Fig. 5: 20 day random walk forecast and the univariate 20 day Fig. 6: Multistep forecast using first 850 days as training.
forecast.
Multis tep forecas t for data s et B

Experiment 4: Multistep forecasting 0.55

US$/DM (normalized)

0.5
An attempt was also made to look at the multistep 0.45
predictive abilities of the networks used in experiment
0.4
1. Initially the results looked surprisingly promising
given the previous results. Fig. 6 shows the results of 0.35
the single day multistep forecast made over the same 0.3
part of the data set as was used before. The network
0.25
does seem to have forecast three of the major turning
1
5
9
13
17
21
25
29
33
37
41
45
49

points. Attempts to replicate this on other parts of the T IME (days )

data set were unsuccessful though. Generally, these
other attempts produced results such as can be seen in Fig. 7: Multistep forecast using days 200 - 750 as training.
fig. 7 which displays a single day multistep forecast
using 550 days training staring after the first 200 days Experiment 5: Multivariate forecasting
 For the 5 day multivariate forecasts the networks were
In this experiment, the data used to forecast the US$ - given identical information as before except that a 20
DM one and five days in advance was expanded to day window of each variable was used giving a total
include the US$ - DM exchange rate, the US$ - Brit. of 260 inputs to the networks. The first 850 days were
Pound exchange rate and the US$ - Yen exchange used for training and the following 100 days for
rate. In addition, one month and one year testing. The results are displayed in fig. 9.
Eurocurrency interest rates for each of the above
currencies were also given as input to the networks
along with the one month and one year London Multivariate 5 day forecas t

Interbank interest rates.

UD$/DM (normalized)
0.8
0.7
For the single day forecasts the networks were given 0.6
10 days of each of the above variables. Thus the 0.5
networks had a total of 130 input units with the 0.4
hidden units being varied from 0 to 30. The networks 0.3
were trained on the first 850 days of the data set and 0.2

1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
tested on the following 100. The results of the single
T IME (days )
day multivariate forecasts with a 5 hidden unit
network are displayed in fig. 8. Actual US $/DM F orecas t

Multvariate s ingle day forecas t

Fig. 9: 5 day multivariate forecast.
US$/DM (normalized)

0.8
0.7 The best performance was found with the 20 hidden
0.6 unit MLP although none of the networks used were
0.5 able to outperform the random walk model in terms of
0.4 mean squared error. Nonetheless, the network is not
0.3
strictly outputting previous input and in fact does to
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99

be picking up some degree of the general direction of

T IME (days ) change (fig. 9). Similar results were found on other
parts of the data set.
Actual UD$/DM F orecas t

DISCUSSION
Fig:8: Multivariate single day forecast.
Although the US$ - DM price changes where shown
The results here are similar to the univariate 1 day to be not strictly random in a statistical sense, from a
forecasts. Increasing the number of hidden units did forecasting point of view what little structure actually
not lead to improved performance. As can be seen the is present may well be too negligible to be of much
networks have not been able to outperform the use. Given the performance of the network models in
random walk model. Similar to the univariate one day univariate portion of this study, the random walk
forecasts, the networks here were very roughly model appears to be the more accurate for daily,
approximating a random walk type performance. weekly and monthly forecasting of the US$ - DM
exchange rate. Does this mean that the random walk
model is the optimal technical forecasting technique daily data. This suggests that either the results of
for the US$ - DM exchange rate? Certainly not. previous work require some reevaluation or that there
However, in forecasting the US$ - DM exchange rate is something special about the nature of hourly
using only daily US$ - DM data, the random walk changes in the US$ - DM prices changes that can be
model has been the most effective of the models exploited for forecasting that cannot be found in the
examined here at all three forecast lead times. daily changes. Second, more research could be
conducted examining the use of additional types of
This has some implications regarding the work of input and architectures in neural network based
Refenes et al (1993). Essentially, the only difference financial market forecasting systems. Finally, given
between their study of single day forecasting and the the lack of cross model comparisons in previous
univariate experiments conducted here, is that they research, more work needs to be done examining the
used hourly data. It seems reasonable to argue that if relative forecasting abilities of connectionist and more
their model was truly robust, the use of daily data standard financial modelling techniques. An
should have worked just as well. If anything, the experimental framework incorporating cross
hourly data would have been even more noisy than comparisons between different forecasting models
the daily data thus making forecasting even more combined with multiple simulation runs is
difficult. Refenes et al also looked at multistep recommended in research claiming the superiority of
forecasting which was also attempted here. The initial neural networks in financial forecasting. This will
positive results we found with multistep forecasting in ensure that the advantages of connectionist forecasting
this work turned out to be spurious. In future work, the methods cannot be simply written off as data mining.
use of hourly data will be investigated to see if the
dynamics of hourly price changes are more conducive
to univariate forecasting than daily changes. REFERENCES

The multivariate 1 day experiment was not much Binks, D. L. and Allinson, N. M. (1991) "Financial
different than the univariate single day forecasting. data recognition and prediction using neural
Why this is so is unclear but perhaps the use of networks", Artificial Neural Networks, T. Kohonen,
additional currency and interest rate data here was not K. Makisara and J. Kangas (editors), Elsevier Science
sufficient to capture the dynamics of the daily price Publishers, B.V. (North-Holland), 1709 - 12.
changes.
Chatfield, C. (1975) The analysis of time series:
The 5 day forecast were a bit more interesting if only Theory and practice, Chapman and Hall, London, pg.
because the networks did seem to pick up on some of 25.
the general trend information. Future work will
explore the use of a larger variety of input variables Collard, J. E. (1991) "A B-P commodity trader",
and network architectures. Advances in Neural Information Processing Systems
III, B. M. Spatz (editor), Morgan Kaufmann
Publishers, San Mateo, CA., 551 - 56.
CONCLUSIONS
Diebold, F. X. and Nason, J. A. (1990)
The conclusions of this work are three fold: First, "Nonparametric exchange rate prediction?", Journal
previous work claiming good forecasting performance of International Economies, 28, 315 - 32.
of the US$ - DM using univariate hourly input to
feedforward networks could not be replicated with
Hutchinson, J. M. (1994) A radial basis function
approach to financial time series, Masters Thesis,
Department of Electrical Engineering and Computer
Science, Massachusetts Institute of Technology,
Mass, USA.

Kendall, M. G. and Stuart, A. (1969) The advanced

theory of statistics, Vol 3 Charles Griffin and
Company Ltd., London, 355 - 6.

Kimoto, T., Asakawa, K., Yoda, M. and Takeoka, M.

(1990) "Stock Market Prediction with Modular Neural
Networks", Procedings of the International Joint
Conference on Neural Networks, San Diego, CA.

Lee, C. H., Park, C. P. (1992) "Prediction of monthly

transitions of the composite stock price index using
recurrent back propagation", Artificial Neural
Networks 2, I. Aleksander and J. Taylor (editors),
Elsevier Science Publishers, B.V. (North-Holland),
1629 - 32.

Moore, G. H. and Wallis, A. (1943) "Time series

significance tests based on signs of differences",
Journal of the American Statistical Association, 32,
153, (cited from Kendall and Stuart (1969)).

Pindyck, R. S. and Rubinfeld, D. L. (1991)

Econometric Models and Economic Forecasts (third
ed), McGraw - Hill, Inc., New York, pg. 441.

Refenes, A. N., Azema-Barac, M., Chen, L. and

Karoussos, S. A. (1993) "Currency exchange rate
prediction and neural network design strategies",
Neural Computing and Applications, Vol. 1, 46 - 58.

Weigend, A. S., Huberman, B. A. and Rumelhart, D.

E. (1992) "Predicting sunspots and exchange rates with
connectionist networks", Nonlinear Modelling and
Forecasting, SFI Studies in the Sciences of
Complexity, M. Casdagli and S. Eubank (editors),
Addison-Wesley, 395 - 432.