Neurocomputing 34 (2000) 79}98

A case study on using neural networks to perform technical

forecasting of forex
Jingtao Yao *, Chew Lim Tan

Department of Information Systems, Massey University, Palmerston North, New Zealand

School of Computing, National University of Singapore, Singapore 119260, Singapore

Received 15 November 1997; accepted 12 April 2000

Abstract

This paper reports empirical evidence that a neural network model is applicable to the
prediction of foreign exchange rates. Time series data and technical indicators, such as moving
average, are fed to neural networks to capture the underlying `rulesa of the movement in
currency exchange rates. The exchange rates between American Dollar and "ve other major
currencies, Japanese Yen, Deutsch Mark, British Pound, Swiss Franc and Australian Dollar are
forecast by the trained neural networks. The traditional rescaled range analysis is used to test
the `e$ciencya of each market before using historical data to train the neural networks. The
results presented here show that without the use of extensive market data or knowledge, useful
prediction can be made and signi"cant paper pro"ts can be achieved for out-of-sample data
with simple technical indicators. A further research on exchange rates between Swiss Franc and
American Dollar is also conducted. However, the experiments show that with e$cient market it
is not easy to make pro"ts using technical indicators or time series input neural networks. This
article also discusses several issues on the frequency of sampling, choice of network architecture,
forecasting periods, and measures for evaluating the model's predictive power. After presenting
the experimental results, a discussion on future research concludes the paper.  2000 Elsevier
Science B.V. All rights reserved.

Keywords: Foreign exchange rate; Neural network; Forecasting; Time series

1. Introduction

The foreign exchange market is the largest and most liquid of the "nancial markets,
with an estimated $1 trillion traded every day. Foreign exchange rates are amongst

* Corresponding author.
E-mail address: [email protected] (J. Yao).

0925-2312/00/$ - see front matter  2000 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 5 - 2 3 1 2 ( 0 0 ) 0 0 3 0 0 - 3
80 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

the most important economic indices in the international monetary markets. The
forecasting of them poses many theoretical and experimental challenges. Given the
abandonment of the "xed exchange rates, the implementation of the #oating ex-
change rate system in the 1970s, and the recent GATT talks towards further liberaliza-
tion of trade, understanding the relationship among major currency exchange rates
is of utmost importance. Foreign exchange rates are a!ected by many highly corre-
lated economic, political and even psychological factors. The interaction of these
factors is in a very complex fashion. Therefore, to forecast the changes of foreign
exchange rates is generally very di$cult. Researchers and practitioners have been
striving for an explanation of the movement of exchange rates. Thus, various kinds of
forecasting methods have been developed by many researchers and experts. Technical
and fundamental analysis are the basic and major forecasting methodologies which
are in popular use in "nancial forecasting. Like many other economic time series,
forex has its own trend, cycle, season, and irregularity. Thus to identify, model,
extrapolate and recombine these patterns and to give forex forecasting is the major
challenge.
Traditionally, statistical models such as Box}Jenkins [2] models dominate the
time series forecasting. White [20] suggested that the relationship between neural
networks and conventional statistical approaches for time series forecasting are
complementary. Refenes et al. [11] also indicated that traditional statistical tech-
niques for forecasting have reached their limitation in applications with nonlinearities
in the data set such as stock indices. Neural Network technology has seen many
application areas in business especially when the problem domain involves classi-
"cation, recognition and predictions. According to a survey research conducted by
Wong et al. [21], more than 127 neural network business applications had been
published in international journals up to September 1994. The number rose to 213
after a year [22].
This paper describes the application of neural networks in foreign exchange rates
forecasting between American Dollar and "ve other major currencies, Japanese Yen,
Deutsch Mark, British Pound, Swiss Franc and Australian Dollar. Technical indi-
cators and time series data are fed to neural networks to capture the underlying
`rulesa of the movement in currency exchange rates. The rescaled range analysis is
conducted to test the `e$ciencya of each market before using historical data to train
the neural networks. The results presented in this paper show that without the use of
extensive market data or knowledge, useful prediction can be made and signi"cant
paper pro"t can be achieved for out-of-sample data with simple technical indicators.
A further research on exchange rates between Swiss Franc and American Dollar is
also conducted. However, the experiments show that with the e$cient market it is not
easy to pro"t using technical indicators or time series input neural networks. This
article discusses several issues on the frequency of sampling, choice of network
architecture, forecasting periods, and measures for evaluating the model's predictive
power. After presenting the experimental results, a discussion on future research
concludes the paper.
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 81

2. Neural networks as a forecasting tool for foreign exchange rate

Foreign exchange rates were only determined by the balance of payments at the
very beginning. The balance of payments was merely a way of listing receipts and
payments in international transactions for a country. Payments involve a supply of
the domestic currency and a demand for foreign currencies. Receipts involve a de-
mand for the domestic currency and a supply of foreign currencies. The balance was
determined mainly by the import and export of goods. Thus, the prediction of the
exchange rates was not very di$cult at that time. Unfortunately, interest rates and
other demand}supply factors had become more relevant to each currency later on. On
top of this the "xed foreign exchange rates was abandoned and a #oating exchange
rate system was implemented by industrialized countries in 1973. Recently, proposals
towards further liberalization of trades are discussed in General Agreement on Trade
and Tari!s. Increased Forex trading, and hence speculation due to liquidity and
bonds, had also contributed to the di$culty of forecasting Forex.
The application of forecasting method includes two basic steps: analyzing data
series and selecting the forecasting method that best "ts the data series. Generally,
there are three schools of thought in terms of the ability to pro"t from the "nancial
market. The "rst school believes that no investor can achieve above average trading
advantages based on the historical and present information. The major theory
includes the Random Walk Hypothesis and Ezcient Market Hypothesis [12]. The second
school's view is that of fundamental analysis. It looks in depth at the "nancial
condition of each country and studies the e!ects of supply and demand on each
currency. Technical analysis belongs to the third school of thought which assumes
that the exchange rates move in trends and these trends can be captured and used for
forecasting. It uses such tools as charting patterns, technical indicators and specialized
techniques like Gann lines, Elliot waves and Fibonacci series [14].
To maximize pro"ts from the liquidity market, more and more `besta forecasting
techniques are used by di!erent traders. Nowadays, assisted with powerful computer
technologies, traders no longer rely on a single technique to provide information
about the future of the markets but rather use a variety of techniques to obtain
multiple signals. Classical time series analysis, based on the theory of stationary
stochastic processes [4] does not perform satisfactorily on economic time series.
Economic data are not simple autoregressive-integrated-moving-average(ARIMA)
processes; they are not described by simple linear structural models; they are not
simple white noise or even random walks. Hence the major challenge ahead is the
development of new methods, or the modi"cation or integration of existing ones, that
are capable of accurately forecasting series whose patterns or relationships change
over time.
Because of the high volatility, complexity, and noise market environment neural
network techniques are prime candidates for prediction purpose. Refenes et al. [10]
applied a multi-layer perceptron network to predict the exchange rates between
American Dollar and Deutsch Mark, and to study the convergence issues related to
network architecture. Poddig [15] studied the problem of predicting the trend of the
American Dollar}German Mark exchange rates, and compared results to regression
82 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

analysis. Pi [13] proposed tests for dependence among exchange rates. Other exam-
ples using neural networks in currency applications include [3,7,9,18,25,23].
Neural networks are an emerging and challenging computational technology and
they o!er a new avenue to explore the dynamics of a variety of "nancial applications.
The backpropagation algorithm [16] has emerged as one of the most widely used
learning procedures for multilayer networks. They have been shown to have great
potential for "nancial forecasting [1]. Neural networks can make contributions to the
maximization of returns, while reducing costs, and limiting risks. They can simulate
fundamental and technical analysis methods using fundamental and technical indi-
cators as inputs. Consumer price index, foreign reserve, GDP, export and import
volume, etc. could be used as inputs. For technical methods, the delayed time series
data, moving average, relative strength index, etc. could be used as inputs.
Levich et al. [6] stated that active currency risk management based on a simple
application of technical trading signal ("lter rule and moving average crossover rule)
can lead to signi"cant pro"ts in the foreign exchange future market. Currencies have
a reputation as `momentum tradinga vehicles in which technical analysis has more
validity than usual [12]. Thus technical methods are studied in this paper.
The work discussed in this paper would represent a violation of the e$cient market
hypothesis postulated by the "rst school. The inclusion of fundamental factors
proposed by the second school is under study as part of our continuing work.

3. Rescaled range analysis and market e7ciency testing

The most well known, widely tested but little believed hypotheses to support the
"rst school of thought mentioned in Section 2 are the Random Walk Hypothesis and
E$cient Market Hypothesis. The Random Walk hypothesis states that the market
prices wander in a purely random and unpredictable way. The E$cient Market
Hypothesis states that the markets fully re#ect all of the available information and
prices are adjusted fully and immediately once new information become available. In
the actual market, some people do react to information immediately after they have
received the information while other people wait for the con"rmation of information.
The waiting people do not react until a trend is clearly established.
The Hurst exponent [5] is a measure of the bias in fractional Brownian motion.
The method could be used in economic and "nancial market time series to see
whether these series are also biased random walks which indicates the possibility of
forecasting.
The rescaled range analysis (or R/S analysis hereafter) [5,12] is able to distinguish
a random series from a fractal series, irrespective of the distribution of the underlying
series (Gaussian or non-Gaussian). It is a robust statistics for measuring the amount of
noise in a system. It can be used to determine the persistence of trends and the average

 The erratic path followed by a particle suspended in a #uid. The distance of a random particle covered
increases with the square root of time used.
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 83

length of non-periodic cycles. In this paper, it is used to detect the long-memory e!ect
in the foreign exchange rate time series over a time period. R captures the maximum
and minimum cumulative deviations of the observations x of the time series from its
R
mean (k), and it is a function of time (the number of observations is N):

R " max [x ]! min [x ], (1)

, R, R,
XRX, XRX,
where x is the cumulative deviation over N periods and de"ned as follows:
R,
R
x " (x !k ), (2)
R, S ,
S
where k "Average of x over N periods.
, S
The R/S ratio of R and the standard deviation S of the original time series can be
estimated by the following empirical law: R/S"N& when observed for various
N values. For some value of N, the H exponent can be calculated by
H"log(R/S)/log(N), 0(H(1 (3)
and the estimate of H can be found by calculating the slope of the log/log graph of R/S
against N using regression. The R/S used in the regression between log(R/S) and
log(N) for various N, is the average R/S value for each N. By regressing over the range
of di!erent values of N, the highest value obtained indicates the average length of the
non-periodic cycle (mean orbital period) of the observed system.
The Hurst exponent H describes the probability that two consecutive events are
likely to occur. There are three distinct classi"cations for the Hurst exponent: (1)
H"0.50, (2) 04H(0.50, and (3) 0.50(H(1.00. The type of series described by
H"0.5 is random, consisting of uncorrelated events. However, value H"0.50
cannot be used as an evidence of a Gaussian random walk, it only proves that there is
no evidence of a long memory e!ect. A value of H di!erent from 0.50 denotes that the
observations are not independent. When 04H(0.5, the system is an antipersistent
or ergodic series with frequent reversals and high volatility. Despite the prevalence of
the mean reversal concept in economic and "nancial literature, few antipersistent
series have yet been found. For the third case (0.5(H41.0), H describes a persistent
or trend-reinforcing series which is characterized by long-memory e!ects. The
strength of the bias depends on how far H is above 0.50. However, even in the case
that the Hurst process describes a biased random walk, the bias can change abruptly
either in direction or magnitude. Therefore, only the average cycle length of observed
data can be estimated. In this range, the lower the value of H, the more noise there is in
the system and the more random-like the series is. The higher the H value, the more
e$cient the market is since it has less noise than the others.
The datasets of "ve currencies studied in this paper comprise 2910 daily rates for
a sampling period between 18 May 1984 and 7 July 1995. In a real situation, there is
no closing price for Forex. Forex trading takes place 24 h a day. The 24 h data should
be used in order to capture the underlying rules of the movement in Forex rates. The
more data one uses the more rules one may get. However, in view of the di$culties in
84 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

Table 1
Weekly foreign exchange rates statistics: number of observations: 510; period: May 1984}July 1995

Exchange Mean Standard deviation Variance Max Min

AUD 0.7424 0.0605 0.003642 0.9035 0.5959

CHF 1.6203 0.3828 0.1466 2.9245 1.1145
DEM 1.9450 0.4770 0.2263 3.4520 1.3529
GBP 1.6025 0.1978 0.03938 2.0040 1.0520
JPY 152.240 42.2177 1782.4335 262.80 80.83

Table 2
Hurst exponent and correlation for the experimented "ve
currencies

Exchange Hurst exponent Correlation

AUD 0.532681 0.046347

CHF 0.553941 0.077645
DEM 0.554672 0.078737
GBP 0.544408 0.063497
JPY 0.540706 0.058053

managing the huge data, "xed sampled time series are often used to model the
behavior of exchange rate changes. Daily data are easily available compared with
tick-by-tick data. In addition, trading too frequently will cost much on transaction.
Weekly forecasts over 7 daily data were adopted to reduce the amount of transaction.
In this research, the weekly data are used assuming that they have enough informa-
tion to capture the `rulesa. The weekly closing prices are used as the prediction target
of our experiment. They refer to Friday's closing prices in the Singapore forex market
which ranks the fourth largest in the world. In the event of Friday being a holiday, the
most recently available closing price for the currency was used. The dataset for each
currency in this study consists of 510 weekly data.
The statistics summary of the weekly data used in this study are shown in Table 1.
As shown in Table 2, the value of Hurst exponent for the logarithmic returns of daily
exchange rates data is higher than 0.5 for all the observed time series. This shows that
all the markets are fractal, and not random walk. The highest value is 0.554 for the
exchange rate of DEM/USD. In other words, it is the most e$cient market amongst
the "ve studied markets. Peters [12] found that US stock market is the most e$cient
market, followed by Germany, UK and Japan. The foreign exchange rates markets
e$ciency happens to be in the same sequence, i.e., Mark, Pound and Yen, with Peters'
"nding in stock markets. The results obtained in this study imply that all the "ve
markets are not random walk and are not highly e$cient thus forecasting of the
exchange rates is possible.
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 85

4. Building a neural network forecasting model

To build a neural network forecasting model, historical data are divided into three
portions: training, validation and testing sets. The training set contains 70% of the
collected data, while the validation and the testing sets contain 20% and 10%,
respectively. The division is based on a rule of thumb derived from the experience of
the authors. A model is considered good if the error of out-of-sample testing is the
lowest compared with the other models. If the trained model is the best one for
validation and also the best one for testing, one can assume that it is a good model for
future forecasting. The data are chosen and segregated in time order. In other words,
the data of the earlier period are used for training, the data of the later period are used
for validation, and the data of the latest time period are used for testing. The objective
is to discover the underlying `structurea of the mechanism generating the data, i.e., to
discover the relationship between present, past and future observations.

4.1. Measurement of neural networks

A usual measure to evaluate and compare the predictive power of the model is the
normalized mean squared error (NMSE) which was used to evaluate entries into the
Santa Fe Time Series Competition [19]. Given a set P comprising pairs of the actual
values (or targets, x ) and predicted values (x( ), the NMSE which normalizes the MSE
I I
by dividing it through the variance of respective series can be de"ned as
(x !x( ) 1 1
NMSE" IZ. I I " (x !x( ), (4)
(x !x ) p N I I
IZ. I I . IZ.
where p is the estimated variance of the data, x being the mean, and N is the size of
I
a set P. If the estimated mean of the data is used as predictor, NMSE"1.0
is obtained. The normalized mean squared error is related to R which measures
the (linear) dependence between pairs of desired values and predictions by
NMSE"1!R. Additional evaluation measures include the calculation of correct
matching number of the actual and predicted values, x and x( , respectively, in the
R R
testing set with respect to the sign and directional change (expressed in percentages).
Sign statistics can be expressed as
1
S " a , (5)
  N I
IZ.
where a "1 if x x( '0 or x "x( "0, and a "0 otherwise. Similarly, directional
I R R R R I
change statistics can be expressed as
1
D " b , (6)
  N I
IZ.
where b "1 if (x !x )(x( !x )50, and b "0 otherwise.
I R> R R> R I
These statistics are desirable because the normalized mean square errors measure
prediction only in terms of levels. Hence, the quality of the forecast can be measured
by R, the correctness of gradient predictions (D ) and the sign changes.
 
86 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

Table 3
The out of sample forecasting results for neural network models (delay method) for weekly
foreign exchange data

Exchange Model Test. NMSE (R) Gradient (%) US Ret 1 (%)

AUD 5-3-1 0.0543 (0.9456) 55.00 1.09

CHF 5-3-1 0.1100 (0.8900) 56.00 8.40
DEM 6-3-1 0.3153 (0.6847) 51.00 4.36
GBP 6-3-1 0.1555 (0.8445) 54.74 2.30
JPY 5-3-1 0.1146 (0.8853) 53.40 3.00

Training 1984}1993, testing Nov 1993}July 1995 US Return1 refers to the return achieved by
USD for trading Strategy 1.

Table 4
The testing results for neural network models using indicators for weekly foreign exchange data

Exchange Model Test. NME Gradient Ret1 Ret1 Ret2 Ret2

31 31
(%) (%) (%) (%) (%)

AUD 5-3-1 0.035105 73.86 8.82 12.19 12.43 15.90

AUD 6-4-1 0.032362 76.14 8.97 12.34 12.67 16.16
CHF 5-3-1 0.068819 65.91 28.49 9.99 22.49 4.85
CHF 6-4-1 0.065962 64.77 32.36 13.31 21.64 4.15
DEM 5-3-1 0.063462 61.36 22.86 8.86 15.20 2.07
DEM 6-4-1 0.061730 64.77 27.84 13.27 18.00 4.55
GBP 5-3-1 0.061370 73.86 7.22 2.87 14.78 10.13
GBP 6-4-1 0.053650 72.73 10.62 6.13 16.48 11.76
JPY 5-4-1 1.966195 46.59 19.71 3.47 0.00 !13.57
JPY 6-4-1 1.242099 46.59 23.42 6.67 0.00 !13.57
SGD 5-3-1 0.268410 39.77 7.70 2.42 5.95 0.75
SGD 6-4-1 0.376993 38.64 7.00 2.26 5.28 0.62

Training 18 May 1984}12 July 1991, Validation 19 Nov 1991}29 Oct 1993, Testing 5 Nov 1993}7 July
1995. Ret1: Return using the di!erence between the predictions this and next period as a criterion for
trading; Ret2: Return using di!erence between the prediction next period and the actual level this period;
Ret1 denotes the starting currency which is USD.
31

The real aim of forecasting is the trading pro"ts or "nancial gains based on
prediction results. It does not matter whether the forecasts are accurate or not in
terms of NMSE or gradient. After the forecast results were obtained from the neural
network models, a program simulating the real trading was developed to test the
possible monetary gains. Since this is not the real trading, we name it as paper pro"ts.
The paper pro"ts are calculated by the return one can expect if one starts with
either the USD or the currency under consideration. Assume that a certain amount of
seed money is used in this program. The seed money is used to buy a certain amount
of another currency when the prediction shows a rise in that currency. At the end of
the testing period, the currency should be converted to the original currency of the
seed money using the exact direct or cross rate of that day. The results obtained are
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 87

shown in Tables 3 and 4. The paper pro"t is calculated as follows:


MoneyObatined U
Return" !1, (7)
SeedMoney
where MoneyObtained is the amount of the money obtained on the last day of testing;
SeedMoney is amount of money used for trading on the the "rst day of testing; and
w the number of weeks in testing period.

4.2. Trading strategies

Trading is an art. As there is no perfect forecasting technique, trading pro"t is

ensured only by a good trading strategy taking `fulla advantage of a good forecasting
method. There are two kinds of trading strategies used in this study. One uses the
di!erence between predictions, and another uses the di!erence between the predicted
and the actual levels to trade.
Strategy 1:
if (x( !x( )'0 then buy else sell. (8)
R> R
Strategy 2:
if (x( !x )'0 then buy else sell. (9)
R> R
Here x is the actual level at time t, where x( is the prediction of the neural networks.
R R
The reason for the di!erent results of the di!erent strategies lies in the noise. Assume
that the noise of the prediction at time t is d , then the following equation holds:
R
x "x( #d . (10)
R R R
Which strategy is more accurate depends on "d !d " and "d ". If d and d have
R> R R> R> R
di!erent signs, then Strategy 1 is better than Strategy 2. Otherwise Strategy 2 is better
than Strategy 1.
Transaction cost will be considered in real trading. In this paper, 1% of transaction
cost was included in the calculation. The transaction cost of a big fund trading, and
thus a!ecting the market rates was not taken into consideration. To be more realistic,
a speci"c amount of transaction cost has to be included in the calculation.
Using one currency as an example, say JPY, if the forecast shows that JPY is going
up and the currency at hand is Dollar, then the strategy is to buy Yen. If the forecasts
show that Yen is going down and the currency at hand is Yen, then the strategy is to
sell Yen. Otherwise, no action is taken.
What are the criteria for determining whether the currency is going up or down? If
the output of neural network is given in percentage of changes, we can use positive or
negative output to show that the currency is going up or down. The returns of the two
di!erent strategies based on the best results so far are as follows: The GBP trades that
depend on the di!erence between the prediction of this period and the prediction of
the next period (Strategy 1) achieved 14.78% in annual return. The GBP trades that
depend on the di!erence between the prediction next period of the neural network and
the actual level of this time period (Strategy 2) received 7.22% in annual return. In
88 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

actual trading, practitioners may choose one of the strategies. A conservative trading
strategy would require a trader to act only when both strategies recommend the same
actions.

4.3. Neural network forecasting model

Basically, three experiments are conducted in this research. First, a purely time-
delayed model is built to capture the relationship between next week's exchange rate
and the historical exchange rates. The purely time-delayed forecast method is one of
the simplest technical analysis methods. The normalized exchange rates of the pre-
vious periods are fed to a neural network to forecast the next period exchange rate in
this model. For example, the inputs to a 5-x-1 neural network are FX , FX ,
G\ G\
FX , FX and FX while the output of the neural network is FX , the next
G\ G\ G G>
week's exchange rate where FX stands for the current week's exchange rate. The
G
architecture of the neural network is denoted by i}h}o. The i}h}o stands for a neural
network with i neurons in input layer, h neurons in hidden layer, and o neurons in
output layer. In our experiments, up to eight weeks of time-delayed data are used. We
"nd that there is no signi"cant improvement in terms of NMSE for the networks with
more than six inputs. Sometimes the results are even worse. This shows that increasing
the number of inputs does not necessarily increase the accuracy of predictions. One
reason may be that the information provided by others factors are already included in
the essential factors. In the case of worse performance, it may be due to that
introducing of extra new delayed inputs brings in nodes that learnt on speci"c noises.
In the interest of space, only the best of the models obtained in the experiment are
presented in this paper.
The purely time-delayed method for prediction sometimes leads to prediction that
seems to generate a time-delayed time series of the original time series. With the
inclusion of some popular indicators used by traders, it might help to remove some of
the time-delay characteristics of the prediction.
The father of Dow-Jones, Charles Dow, categorized the trend into three di!erent
levels, namely the primary or major trend, the secondary or intermediate trend, and
the minor trend [17]. He likened the three levels of trend to the tides, waves and
ripples of the sea. The advantage of moving average is that it tends to smooth out
some of their irregularities that exist between market days. One disadvantage of this
method is that the calculation always lags behind the current market. The primary
trend lasts more than one year or longer. The secondary trend usually lasts from three
weeks to three months. The minor trend is meaningless in itself. However, a series of
three or more minor trends can make up the primary trend or the corrective
secondary trend. Following Dow-Jones theory, moving averages are used as inputs to
the neural network in the second experiment. MA5, MA10, MA20, MA60, and
MA120 are fed to the neural networks to predict the following week's rate. They refer
to moving averages for one week, two weeks, one month, one quarter and half a year,
respectively. MAs are calculated on the trading days. MA5 stands for "ve trading
days' moving average. The "ltered "gures may provide more information to the model
than the purely delayed data. Further, the forecasts for each of the currencies were
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 89

repeated, but with a hybrid of indicators and one time-delay term. The results are
presented in Table 4 where the con"guration 5-x-1 stands for indicator model while
6-x-1 stands for the hybrid model.
Finally, a more in depth research on one of the currencies, namely CHF, is
conducted as the third experiment. Instead of "nding just one good solution, data are
partitioned equally using bootstrapping rules to "nd models in di!erent time periods.
Speci"cally, 12 data sets are generated and each contains six years training and
validation data while half a year data are used for out of sample testing.

5. Prediction results and discussion

5.1. Weekly exchange rate forecast using purely time delayed time series

The measurements of the forecasting results for purely time-delayed method are
shown in Table 3. Figs. 1(a)}(e) are the diagrams showing the predicted and the actual
time series of the exchange rates of AUD/USD, CHF/USD, DEM/USD, GBP/USD
and JPY/USD for the period of November 1993 to July 1995 (out of sample).
From the table and diagrams, we "nd that the "tness between the two curves and
the NMSE level are quite good. However, the gradients are only a little above 50%
which means the forecasts are only slightly better than the chance in tossing a coin.
No doubt such results would not be accepted by the practitioners. Thus the second
approach is necessary.

5.2. Weekly exchange rate forecast using indicators

The measurements of the forecasting results are shown in Table 4. An additional

term of time delay is introduced as input to the 6-x-1 model. Our results however
show that this does not contribute signi"cantly to the improvement of hit rate, or
D , as shown in Table 4. For example as shown in Figs. 2(a) and (b), for Australian
 
Dollar, the results show that the hit rate of 6-4-1 is slightly higher than that of the pure
indicator forecasting method with a con"guration of 5-3-1. The hit rate of the former
is 76.14% and that of the latter is 73.86%. Hence, the diagrams with additional time
delay term for the other currencies are not shown in this paper.
The comparative diagrams showing the predicted (NN output) and the actual (XXX
target) time series for the period of Nov 1993}July 1995 (out of sample) are shown in
Figs. 2(a)}(f) for AUD/USD, AUD/USD (with an additional time-delay input),
CHF/USD, DEM/USD, GBP/USD, and JPY/USD. From the results presented

 The big lap between the actual level and predicted level of JPY/USD is due to the normalization
method used in this study and the market changes. The normalized data are calculated using the actual data
and the maximum and minimum of training data. The minimum value of JPY/USD before November of
1993 is just below 110, but most of levels after April 1995 are lower than 85. The result of this research shows
that the neural networks have only learned the rule of trend and not the level of the exchange rate. So the
neural networks can only get better paper pro"t using Strategy 1. And the graph for JPY/USD is very bad if
we only consider the levels.
90 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

Fig. 1. Purely time-delay model prediction on weekly exchange rates, Nov 1993}July 1995.

here, we can safely conclude that hit rates of approximately 70% can be achieved
consistently for AUD, GBP, and somewhat lower for CHF and DEM.
Also, from the graphs, one can conclude that the forecasts for the "rst 20 weeks of
the pure testing period look `impressivea. The performance on the "ve currencies
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 91

Fig. 2. Indicator model prediction on weekly exchange rates, Nov 1993}July 1995.

shows a degradation in forecasting in about the third quarter after training. This
means that, the neural network needs to be retrained, probably every 20 weeks (or half
a year) with the latest data to increase the chance of achieving a better forecast. This
indicates the presence of &recency' problem of the network, namely, the network did
92 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

retain some memory of the history. A half a year forecast period is thus recommended
based on the present study. Notice that, in actual application, only validation data
sets would be required together with the training data sets. So, we manage to
experiment to retrain the neural network model every half a year. In Section 7, the
results for Swiss Franc will be presented.

6. Comparison with ARIMA

To compare the neural network forecasting results with other models. The auto-
regressive integrated moving average (ARIMA) model is used as the "rst benchmark in
this research. ARIMA Model was introduced by George Box and Gwilym Jenkins [2]
in 1976. The Box}Jenkins methodology provided a systematic procedure for the
analysis of time series that was su$ciently general to handle virtually all empirically
observed time series data patterns. ARIMA(p, d, q) is the general form of di!erent
ARIMA models. Here p stands for the order of the autoregressive process, d represents
the degree of di!erencing involved, and q is the order of the moving average process.
To compare the forecasting results of the neural networks, a number of ARIMA
models were built. Table 5 is the results of ARIMA models with di!erent trading
strategies.
The entire data set was used as "tting data for the ARIMA models. In other words,
the data forecast by ARIMA were already used in the "tting stage of ARIMA model
building. The results obtained from the ARIMA models should be compared with
Ret1 in Table 4. In other words, the ARIMA models should deliver worse out-of-
sample forecasting returns than the ARIMA results indicated in Table 5.
Focusing on the gradients, the ARIMA method can achieve about 50% of
correctness while up to 73% of correctness can be achieved using neural network
models.

Table 5
The result of using ARIMA (AUD(101) stands for the ARIMA result of
AUD using ARIMA(1, 0, 1) model and the same rules are applied to other
currencies)

ARIMA model Gradient (%) Ret1 (%) Ret2 (%)

31 31
AUD(101) 52.27 1.43 1.36
AUD(202) 54.32 1.53 1.21
CHF(101) 38.64 6.94 !1.42
CHF(202) 55.86 5.43 0.64
DEM(101) 43.18 3.48 3.49
DEM(202) 44.62 3.48 3.22
GBP(101) 53.41 2.24 3.67
GBP(202) 51.77 2.63 1.32
JPY(101) 44.32 !1.47 !0.52
JPY(202) 44.32 !0.78 0.02
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 93

From practitioners' point of view, returns are more important than gradients. With
reference to Tables 5 and 4, the di!erences between ARIMA models and neural
network models are signi"cant. The best return for ARIMA models using Strategy 1
regardless of the devaluation and strategies is 6.94%, while for neural network models
is 32.36%.
The results show that irrespective of NMSE, gradient or pro"t, the neural network
models are much better than the traditional ARIMA model. In the case of forecasting
"ve major currencies exchange rates, neural network model can work as a viable
alternative forecasting tool.

7. Further research on CHF/USA rates forecasting

In this section, we will investigate the consistency of the neural network models. As
found in the previous diagrams, the good results can last about half a year. We
partition the whole data into 12 segments (namely CHF1 to CHF12) each covering an
overlapping period of 6 years. The twelve 6 year periods are progressively displaced
 
by  year. Thus, the "rst 6 year period (for CHF1) spans from January 1984 to June
 
1990. The next period is slid along the time horizon for half a year, i.e., it spans from
July 1984 to December 1990. The last time period spans from July 1989 to November
1995.
Amongst the 6 year data, the "rst 6 years' data, or 312 weeks, were used for

training and validation while the remaining half a year's data, or 26 weeks, were used
to test the performance of the neural network model. Of the 312 data, 260 were used as
training data and 52 are used as validation. For each data segment, a variety of
network architectures were experimented. The best architecture, in term of NMSE, for
each data segment, is presented in the Table 6.

Table 6
The technical details of chosen architectures and their forecasting potentials

Name Archit. a g NMSE Grad1 (%) Grad2 (%)

CHF1 6-2-1 0.9 0.0030 0.170480 64.00 84

CHF2 6-2-1 0.9 0.0050 0.030029 61.54 72
CHF3 6-2-1 0.9 0.0030 0.120893 53.85 76
CHF4 6-2-1 0.9 0.0006 0.123096 61.42 36
CHF5 6-3-1 0.9 0.0006 0.042554 65.38 84
CHF6 6-3-1 0.9 0.0007 0.144654 42.31 76
CHF7 6-4-1 0.9 0.0010 0.662526 53.85 68
CHF8 6-3-1 0.9 0.0008 0.051245 65.38 84
CHF9 6-2-1 0.9 0.0009 0.207124 53.85 72
CHF10 6-2-1 0.9 0.0200 0.106726 69.23 92
CHF11 6-3-1 0.9 0.0003 0.040063 61.53 68
CHF12 6-2-1 0.9 0.0008 0.145288 66.67 82

a: learning rate; g: momentum rate; NMSE: normalized mean-squared error; Grad1, Grad2: correct-
ness of gradients for testing data.
94 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

Table 7
Benchmark results for di!erent time period

Name Ret1 Ret1 Ret2 Ret2 Bench. I Bench. II

31" 31"
(%) (%) (%) (%) (%) (%)

CHF1 18.59 11.20 35.93 27.45 ! 6.24 12.59

CHF2 9.31 19.47 12.58 23.05 !3.00 12.21
CHF3 8.72 29.63 12.49 34.12 19.05 1.27
CHF4 !7.97 10.97 !4.96 14.60 32.70 !9.09
CHF5 47.72 !0.58 58.99 7.00 25.90 27.99
CHF6 !10.54 21.62 8.92 48.06 35.94 !2.45
CHF7 20.34 19.30 26.45 25.35 !0.87 !9.69
CHF8 31.85 12.14 67.74 42.66 !14.95 16.76
CHF9 !9.91 !11.57 23.31 15.46 !6.36 !13.66
CHF10 20.24 27.80 30.01 38.17 6.28 !2.11
CHF11 5.23 25.97 19.97 32.84 19.70 !8.38
CHF12 1.80 13.58 20.71 34.68 11.58 !3.69

Table 8
Analysis of results for di!erent models

Name Ret1 Ret1 Ret2 Ret2 Bench. I Bench. II

31" 31"
Median 9.31 19.30 20.71 27.45 11.58 !2.11
Average 11.28 14.96 26.01 28.62 9.98 1.81
Portfolio 12.43 11.00 27.03 29.45 9.88 1.88
Acceptability 69.23 84.62 84.62 100 61.54 30.77

Median: median of 12 models; Average: average pro"t from 12 models; Portfolio: Pro"t of the sum of
20% of Max, 20% of Min and 60% average of others; Acceptability: Percentage of pro"t greater than
5% annually.

Instead of ARIMA model, two strategies serving as new benchmarks are introduc-
ed. Benchmark I uses a `buy-and-holda strategy while Benchmark II uses a `trend-
followa strategy. The Benchmark I's strategy is to buy the USD at the beginning of the
testing period and then sell it at the end of the testing period. The Benchmark II's
strategy is to buy when the market is continually up for two weeks and sell when it is
down for one week. Table 7 shows the di!erences between paper pro"ts and their
benchmark pro"ts. All trade on Fridays.
The study of the 12 data segments shows that the neural network models could be
applied to future forecasting. Compared with the two benchmarks, the neural network
model is better. As shown in Table 8, for the neural network, the worst percentage of
acceptable pro"t, assuming at least 5%, is 69.23 among the four strategies, while the
benchmarks can only achieve 61.54 and 30.77 respectively. Referring to Table 8 again,
another comparison was made for the forecasting results. Assume that we have
a portfolio pro"t based on 20% of maximum of pro"t, 20% of minimum pro"t, and
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 95

60% of average pro"ts from other data segments. This will indicate the general
performance of each strategy.
Figs. 3(a) and (b) are the graphical presentation of the pro"t gained for each
strategy.

8. Conclusion and further research

The present study has given rise to the following observations:

(1) The forecasting results are very promising for most currencies except Yen. In
using neural networks to perform technical forecasting, i.e., feeding simple tech-
nical indicators as inputs to neural network, better results are obtained
for Australian Dollar, Swiss Franc, and British Pound and perhaps Deutsch
Mark. The results for Japanese Yen are the worst especially when Strategy 2 is
used. This is similar to our forecasting results using ARIMA models. There
may be some factors related to Yen that need to be studied. One reason could be
that the market for Yen is bigger and more e$cient than the markets for
Australian Dollar, Swiss Franc, and British Pound. The traders of the Yen
market may depend more on technical analysis and they may act quickly
after any sign appears. Hence, technical analysis may not be a good tool for
forecasting the trends of Yen as everybody is well aware of the meaning of
technical signals.
(2) To take more advantage from neural network forecasting results, trading strat-
egies need to be considered in real trading. Strategy 2 has been found more
powerful in using the neural network forecasting results, the theoretical study of it
is under exploration. Between the two simple trading techniques, the `buy-and-
holda strategy (Benchmark I) is better than the `trend-followa (Benchmark II)
though they are generally worse than neural network models as shown in Fig. 3
and Table 7.
(3) There is a need to have an e!ective measure of neural network performance rather
than just the goodness of "t such as NMSE. A very small NMSE is not necessarily
a general indicator of good performance. The sum of the NMSE of the three parts
of data (training, validation and testing) must be kept small, not just the training
NMSE alone. Sometimes having small NMSEs for testing and validation is more
important than having small NMSE for training. Instead of just using NMSE
alone, three levels of judgment, NMSE, gradient and pro"ts are proposed and
studied in this research. The hit rate may be a better standard for determining the
quality of the forecast. After all, the return depends a lot on the trading strategies
and how the forecasting information are being used for trading advantages.
However, the level of hit rates and paper pro"ts also depend on the period of
forecast. Hence, periodic retraining of the neural networks is necessary. For the
practitioners, the levels of exchange rate and trend can be used depending on their
expectation of return and risk. In addition to the above-mentioned two strategies,
after di!erent trading strategies may be used by them.
96 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

Fig. 3. Each models pro"ts compared with benchmarks. The vertical axis is the annualized pro"t in
percentage, the horizontal axis is the names of di!erent time periods.

(4) A backpropagation network used in the present study has proved to be ade-
quate for forecasting. Simple technical indicators such as moving average are
enough.
The above observations have con"rmed the applicability of neural network for
"nancial forecasting albeit with some problems that need to be addressed. We are
seeking an automatic facility of model building to minimize the tedious trial-and-error
procedures. The above problems have enlightened us in our e!ort as our future work
in formulating a generic neural network model for forecasting as follows:
E The model should be self-adapting to di!erent situations. Parameters for such
adaption will need to be studied and incorporated into one model. Examples of
parameters include volatility of current measurement, data size and frequency, etc.
E To provide more statistical analyses in order to "nd out which kind of data
segments will best capture the underlying behavior of the market changes.
E The model should minimize the recency problem by self-detecting the possibility of
such a problem and provide its own measure to avoid the problem.
E The model should provide the frequency support. Sampling of data according to
volatility need to be applied to the raw data. As Zhou [24] stated, increasing
observation frequency does not always help to improve the accuracy of forecasting.
In this research, the weekly data are used assuming that they have enough
information to capture the `rulesa. Due to the volatility of the currency movement
[8], a di!erent frequency of data maybe needed than the weekly data. The data
could be sampled according to the market character, e.g. bullish, bearish, or
trading, etc. In other words, when the market is volatile, we sample more data for
training, and vice versa. Non-linear or volatile time scale will be taken into
consideration in our further research.
E The model should aim at the "nancial pro"t in addition to the usual "tness of
forecast and target series. Due consideration should be given to emphasize "nancial
pro"t, such as, by incorporating pro"t into neural network training. The forecast-
 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98 97

ing targets such as actual rate, percentage of changes, etc., may be decided auto-
maticly by the model.
E The model should recommend the trading rules based on the expectation of returns
and risk. More trading strategies are needed in addition the two studied here. To
bene"t more practitioners, risk-based trading strategies will be taken into consid-
eration in the model.

Acknowledgements

The authors are grateful to the anonymous referees for their helpful comments. We
thank Dr. Hean-Lee Poh, Yili Li and Teo Jasic for their friendly cooperation and
invaluable discussions.

References

[1] M. Adya, F. Collopy, How e!ective are neural networks at forecasting and prediction? A review and
evaluation, J. Forecasting 17 (1998) 481}495.
[2] G.E.P. Box, G.M. Jenkins, Time Series Analysis: Forecasting and Control, Holden-Day, San Fran-
cisco, CA, 1976.
[3] H. Green, M. Pearson, Neural nets for foreign exchange trading, in: Trading on the Edge: Neural,
Genetic, and Fuzzy Systems for Chaotic Financial Markets, Wiley, New York, 1994.
[4] A.C. Harvey, Forecasting, Structural Time Series Models and the Kalman Filter, Cambridge Univer-
sity Press, Cambridge, 1989.
[5] H.E. Hurst, Long term storage capacity of reservoirs, Trans. Am. Soc. Civil Eng. 116 (1951) 770}808.
[6] R.M. Levich, L.R. Thomas, The merits of active currency risk management: evidence from interna-
tional bond portfolios, Financial Anal. J. 49 (5) (1993) 63}70.
[7] R. Manger, Using holographic neural networks for currency exchange rates prediction, Proceedings
of the 16th International Conference on Information Technology Interface, Pula, Croatia, June
1994.
[8] U. Muller, M. Dacorogna, R. Dave, G. Pictet, R. Olsen, R. Ward, Fractals and intrinsic time
} a challenge to econometricians, Proceedings of the 39th International Conference of the Applied
Econometrics Association on Real Time Econometrics, Luxembourg, October 1993.
[9] A.M. Rawani, Forecasting and trading strategy for foreign exchange market, Inform. Decision
Technol. 19 (1) (1993).
[10] A.N. Refenes, M. Azema-Barac, L. Chen, S. Karoussos, Currency exchange rate prediction and neural
network design strategies, Neural Comput. Appl. 1 (1) (1993) 46}58.
[11] A.N. Refenes, A. Zapranis, G. Francis, Stock performance modeling using neural networks: a com-
parative study with regression models, Neural Network, 5 (1994) 961}970.
[12] E.E. Peters, Chaos and Order in the Capital markets: A New View of Cycles, Prices, and Market
Volatility, Wiley, New York, 1991.
[13] H. Pi, Dependence analysis and neural network modeling of currency exchange rates, Proceedings of
First International Workshop on Neural Networks in Capital Markets, London, 1993.
[14] T. Plummer, Forecasting Financial Markets: A Technical Analysis and the Dynamic of Price, Wiley,
New York, 1991.
[15] A. Poddig, Short term forecasting of the USD/DM exchange rate, Proceedings of First International
Workshop on Neural Networks in Capital Markets, London, 1993.
[16] D.E. Rumelhart, J.L. McClelland, Parallel Distributed Processing: Explorations in the Micro-struc-
ture of Cognition, vol. 1, The MIT Press, Cambridge, MA, 1986, pp. 318}362.
98 J. Yao, C.L. Tan / Neurocomputing 34 (2000) 79}98

[17] R. Teweles, E. Bradley, T. Teweles, The Stock Market, 6th Edition, Wiley, New York, 1992, pp.
374}379.
[18] A.S. Weigend, Generalization by weight-elimination applied to currency exchange rate prediction,
Proceedings of the IEEE International Joint Conference on Neural Networks, Singapore, November
1991.
[19] A.S. Weigend, N.A. Gershenfeld (Eds.), Time Series Prediction: Forecasting the Future and Under-
standing the Past, Proceedings of the NATO Advanced Research Workshop on Comparative Time
Series Analysis, Santa Fe, NM, May 14}17, 1992, Addison-Wesley, Reading, MA, 1994.
[20] H. White, Learning in arti"cial neural networks: a statistical perspective, Neural Comput. 1 (1989)
425}464.
[21] B.K. Wong, A bibliography of neural network business applications research: 1988}September 1994,
Expert Systems 12 (3) (1995) 253}262.
[22] B.K. Wong, T.A. Bodnovich, Y. Selvi, Neural-network applications in business: a review and analysis
of the literature(1988}1995), Decision Support Syst. 19 (4) (1997) 301}320.
[23] J.T. Yao, H.-L. Poh, T. Jasic, Foreign exchange rates forecasting with neural networks, Proceedings of
the International Conference on Neural Information Processing, Hong Kong, September 1996, pp.
754}759.
[24] B. Zhou, Estimating the variance parameter from noisy high frequency "nancial data, MIT Sloan
School Working Paper, No. 3739, 1995.
[25] X.R. Zhang, Non-linear predictive models for intra-day foreign exchange trading, Int. J. Intelligent
Syst. Accounting, Finance Manage. 3 (4) (1994) 293}302.

J. Yao is a Senior Lecturer in the Department Information Systems, College of

Business, Massey University. ([email protected]) Jingtao Yao was a Teach-
ing Assistant in the Department of Information Systems, School of Computing,
National University of Singapore. His research interests include "nancial forecast-
ing, neural networks, software engineering, and business process reengineering.
His research has been published in international refereed journals and conference
proceedings. He obtained a B.Eng. (1983), a M.Sc. (1988) in Computer Science
from Xi'an Jiaotong University, and a Ph.D. (1999) in Information Systems from
the National University of Singapore.

Chew Lim Tan is an Associate Professor in the Department of Computer Science,

School of Computing, National University of Singapore. His research interests are
expert systems, neural networks, computer vision and natural language process-
ing. He obtained a B.Sc. (Hons) degree in Physics in 1971 from the University of
Singapore, an M.Sc. degree in Radiation Studies in 1973 from the University of
Surrey in UK, and a Ph.D. degree in Computer Science in 1986 from the
University of Virginia, USA.