This article has been accepted for inclusion in a future issue.

IEEE TRANSACTIONS ON NEURAL NETWORKS 1

Stock Trading Using RSPOP: A Novel Rough

Set-Based Neuro-Fuzzy Approach
Kai Keng Ang, Student Member, IEEE, and Chai Quek, Member, IEEE

Abstract—This paper investigates the method of forecasting interpreted to imply that the technical approach to forecasting
stock price difference on artificially generated price series data stock price is invalid, but recent literature presented from a
using neuro-fuzzy systems and neural networks. As trading profits behavioral finance perspective [5] and statistical inference from
is more important to an investor than statistical performance, this
paper proposes a novel rough set-based neuro-fuzzy stock trading computational algorithms [6] further exemplified the evidence
decision model called stock trading using rough set-based pseudo on the predictability of financial market using technical anal-
outer-product (RSPOP) which synergizes the price difference fore- ysis. Thus technical analysis has recently enjoyed a renaissance
cast method with a forecast bottleneck free trading decision model. and most major brokerage firms publish technical commentary
The proposed stock trading with forecast model uses the pseudo on the market and individual securities.
outer-product based fuzzy neural network using the compositional
rule of inference [POPFNN-CRI(S)] with fuzzy rules identified The main approach in financial forecasting is to identify
using the RSPOP algorithm as the underlying predictor model trends at an early stage in order to maintain an investment
and simple moving average trading rules in the stock trading strategy until evidence indicates that the trend has reversed.
decision model. Experimental results using the proposed stock Predictability of security from past real-world data using
trading with RSPOP forecast model on real world stock market two of the simplest and most popular trading rules, namely
data are presented. Trading profits in terms of portfolio end values
obtained are benchmarked against stock trading with dynamic moving average and trading range break-out rules, were first
evolving neural-fuzzy inference system (DENFIS) forecast model, investigated in [3] on the Dow Jones Index. Other techniques
the stock trading without forecast model and the stock trading used include regression methods and the ARIMA models [7],
with ideal forecast model. Experimental results showed that the but these models fail to give satisfactory forecast for some
proposed model identified rules with greater interpretability and series because of their linear structures and some other inherent
yielded significantly higher profits than the stock trading with
DENFIS forecast model and the stock trading without forecast limitations [8]. Although there are also ARCH/GARCH models
model. [9] to deal with the nonconstant variance, these models also fail
to give satisfactory forecast for some series [8] (please refer
Index Terms—Forecasting theory, fuzzy neural networks, rough
set theory, stock market, time series. to [10]–[12] for a modern guide to the statistical approach to
technical financial forecasting). Increasingly applications of
artificial intelligence (AI) techniques, mainly artificial neural
networks, have been applied to technical financial forecasting
I. INTRODUCTION [13]–[15] as they have the ability to learn complex nonlinear
HERE are two major approaches to the analysis of stock mapping and self-adaptation for different statistical distribu-
T market price prediction: Fundamental and technical
analyses. Fundamental analysis is the approach of studying the
tions (please refer to [16] for a review and evaluation of neural
networks in technical financial forecasting). An investigation
overall economy, industry, financial conditions and manage- on the nonlinear predictability of security returns from past
ment of companies to measure the intrinsic value of a particular real-world returns using single layer feed-forward neural
security (please refer to [1] for a classical guide to fundamental network and moving average rules was presented in [17] on
analysis). This approach uses revenues, earnings, future growth, the Dow Jones Index. The results showed that evidence of
return on equity, profit margins, and other data to determine a nonlinear predictability in stock market returns can be found
company’s underlying value and the potential for future growth by using the past buy and sell signals of the moving average
rules. Application of AI techniques in financial forecasting is
of its security. Technical analysis, on the other hand, does not
attempt to measure a security’s intrinsic value. This approach not restricted only to the technical analysis approach, but has
evaluates securities by analyzing statistics generated by market also been applied to the fundamental approach. For example,
activity, such as past prices and volume (please refer to [2] for in the work of [18], a genetic algorithm based fuzzy neural
network is trained with additional political, financial, economic
a modern guide to technical analysis). The pioneering technical
analysis technique is attributed to C. Dow back in the late 1800s factors etc. to formulate trading decisions. A number of re-
[3]. The efficient market hypothesis (EMH) [4] is generally search investigations have been published on the application
of AI techniques in the technical analysis approach of fore-
casting stock price, but only a few presented quantitative results
Manuscript received May 12, 2005; revised August 11, 2005. This work was on trading performance using real world stock market data
supported by the Singapore Millennium Foundation. [19]–[24]. Saad et al. [21] performed analysis of predictability
The authors are with the Centre for Computational Intelligence, School of based on a history of closing price of a number of high volatility
Computer Engineering, Nanyang Technological University, Singapore 639798,
Singapore (e-mail: [email protected]). stocks and consumer stocks using time delay, recurrent and
Digital Object Identifier 10.1109/TNN.2006.875996 probabilistic neural networks. Leigh et al. [22] used neural

1045-9227/$20.00 © 2006 IEEE

 This article has been accepted for inclusion in a future issue.

2 IEEE TRANSACTIONS ON NEURAL NETWORKS

networks and genetic algorithm to perform pattern recognition curacy. In practice, one of the two properties prevails. The fuzzy
of the bull flag pattern and to learn the trading rules from price modeling research field is divided into two areas: linguistic fuzzy
and volume of the NYSE Composite Index. Results showed that modeling that is focused on interpretability, mainly the Mam-
the forecasting method yielded statistically significant returns dani model [37] given in (1); and precise fuzzy modeling that is
that are better than the overall average 20-day horizon price focused on accuracy, mainly the Takagi–Sugeno–Kang (TSK)
increase. Moody et al. [23], [25] used recurrent reinforcement model [38] given in (2) and (3) (please refer to [39] for a re-
learning without forecasting to train a trading system to trade cent comprehensive coverage on interpretability issues of fuzzy
using past prices of S&P500 stock index while accounting for modeling) [31], [39], [40]. Both of these models are investigated
the effects of transaction costs (please refer to [26] for details in this paper, specifically the Dynamic Evolving Neural-Fuzzy
on reinforcement learning). Chen et al. [24] used a probabilistic Inference System (DENFIS) (please refer to Section III in [36]
neural network (PNN) to forecast the direction of price moment for a description of DENFIS) which is based on the precise
of the Taiwan Stock Index and presented two PNN-guided fuzzy model and the pseudo outer-product based fuzzy neural
investment strategies to translate the predicted direction to network (POPFNN) (please refer to [41, Sec. II] for a brief de-
trading signals. Field and Singh [20] used Pareto evolutionary scription of POPFNN) which is based on the linguistic fuzzy
neural network (Pareto-ENN) to forecast 37 different interna- model
tional stock indexes.
Although neural networks possess the properties required for
technical financial forecasting, they cannot be used to explain IF is AND
the causal relationship between input and output variables be- is THEN is (1)
cause of their black box nature. Neuro-fuzzy hybridization syn- IF is AND
ergizes neural networks and fuzzy systems by combining the
human-like reasoning style of fuzzy systems with the learning is THEN (2)
and connectionist structure of neural networks. Neuro-fuzzy hy-
(3)
bridization is widely termed as fuzzy neural networks (FNNs) or
neuro-fuzzy systems (NFSs) in the literature [27]. NFSs incor-
porates the human-like reasoning style of fuzzy systems through where , are the input vector and output
the use of fuzzy sets and a linguistic model consisting of a set of value, respectively; , are the linguistic labels with fuzzy
IF–THEN fuzzy rules. Thus the main strength of NFSs is that sets associated defining their meaning; number of inputs; and
they are universal approximators [28]–[30] with the ability to are tbe number of rules.
solicit interpretable IF–THEN rules [31]. In recent years, in- The consequents in (1) are linear functions of the inputs
creasing number of research applied NFSs in financial engi- whereas the consequents in (2) are simply linguistic labels.
neering [32]. Some works that applied NFSs in forecasting stock Therefore, the TSK model has decreased interpretability but
price are [8], [21], [33]–[35]. increased representative power compared against the Mamdani
This paper proposes a novel rough set-based neuro-fuzzy model [39]. Recently, a number of research work addressing
stock trading decision model called stock trading using rough the issues of the TSK model have been reported [42]–[44].
set-based pseudo outer-product (RSPOP). Section II reviews In contrast, an increased number of fuzzy rules is needed in
the two main NFSs and outlines the proposed rough set-based the Mamdani model to yield the same representative power of
neuro-fuzzy approach. Section III reviews the commonly used the TSK model. Recently, Rough Set methods [45] have been
time-delayed price forecast approach and the time-delayed shown to significantly reduce pattern dimensionality and proven
price difference forecast approach in forecasting stock prices. to be viable data mining techniques in [46]. This motivated
Section IV presents experimental results of forecasting stock the investigation of the rough set-based neuro-fuzzy approach,
price difference using various neuro-fuzzy systems and neural which uses the RSPOP algorithm [41] to identify the fuzzy
networks on artificially generated price series data. Section V rules in POPFNN. The use of the RSPOP algorithm reduced
reviews existing trading models with and without forecast and computational complexity, improved the interpretability by
presents the proposed forecast bottleneck free stock trading identifying significantly fewer fuzzy rules as well as improved
with RSPOP forecast model. Section VI presents extensive the accuracy of the POPFNN (please refer to [41, Sec. IV] for
experimental results using the proposed stock trading with a detailed description of the RSPOP algorithm).
RSPOP forecast model on real world stock market data. The The RSPOP algorithm consists of three parts. The first part
trading profits in terms of portfolio end values are presented Rule Identification identifies only one most influential rule in-
and compared against the stock trading with dynamic evolving stead of all possible rules from one instant of the training data.
neural-fuzzy inference system (DENFIS) [36] forecast model, The second part Attribute Reduction performs feature selection
the stock trading without forecast model and the stock trading through the reduction of redundant attributes using the concept
with ideal forecast model. Finally, Section VII concludes this of knowledge reduction from rough set theory [45]. As there are
paper. many possible reducts for a given rule set, this part uses an ob-
jective measure to identify reducts that improve rather than de-
II. ROUGH SET-BASED NEURO-FUZZY APPROACH teriorate the inferred consequence after attribute reduction. The
The strength of neuro-fuzzy systems involves two contradic- third part Rule Reduction performs partial feature selection by
tory requirements in fuzzy modeling: interpretability verses ac- extending the reduction to rules without redundant attributes.
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 3

TABLE I
FUZZY RULES IDENTIFIED USING RSPOP ON NAKANISHI DATA SET (A) FROM
[41]

Fig. 1. Time-delayed price forecast approach.

If is just a sequence of unspecified exogenous variables

represented by a stochastic variable instead, then (4) re-
sembles an autoregressive equation [10] and time series pre-
diction is formulated as: Given values
; predict as . Some called this time
series prediction using neural networks as time-delayed neural
network [21]; some called it recurrent neural network [33]. In
the case of nonlinear time series prediction, it is known as the
An example from [41] is used to help illustrate the RSPOP time delay embedding technique [47] and denotes the em-
rule identification and reduction process. Table I shows the ini- bedding dimension [48], [49]. The linear time series prediction
tial and reduced rules identified using RSPOP on the Nakan- in (4) is henceforth referred to as the time-delayed approach to
ishi example (a) data set [35]. The sets discern it from recurrent networks with feedback connections.
and represent the condition and consequent at- Several works applied this approach using neural networks or
tributes, respectively, where , , neuro-fuzzy systems [21], [33], [50]–[52]. As a benchmark, the
, , and repre- best estimation of (4) is the Random Walk model, which sug-
sent the membership functions. The first part Rule Identification gests that the day-to-day price change of a stock should have
identified an initial 22 rules. The second part Attribute Reduc- a mean value of zero [10]. Formally, the Random Walk model
tion identified one reduct that resulted in asserts that the value of the time series should follow the sto-
no deterioration of the objective measure. Thus, the attribute chastic difference equation in (5) [10]. Therefore, the predicted
is redundant and the number of rules is reduced from 22 to 20 value using the Random Walk model is simply as given in
after removing duplicated rules. The third part Rule Reduction (6)
identified groups of reducible rules that resulted in no deterio-
ration of objective measures and reduced the number of rules
from 20 to 17. (5)
(6)
III. STOCK PRICE FORECAST
Time series prediction has a diverse range of applications [7], where is a normal random deviate with an expected mean
and forecasting stock price is one such application. As time se- of zero.
ries prediction is concerned with the estimation of equations Fig. 1 illustrates the forecasting of stock price using the time-
containing stochastic components, it is modeled by an th-order delayed approach. This architecture employs previous values
linear difference equation with constant coefficients given in (4) of the series as input to the network and the predicted value as
[10] an output from the network. Each node represents a lag
operation for one time instant. The network is trained using a
series of known prior to its use as a time series predictor.
Specifically, assume that are known, the
(4)
training data set is formed using of these input–output
data tuples.
where is the time series where represents a value at time Incidentally, quite a number of research publications using
instant ; is one of the arbitrary parameters that do not de- neural networks based on the time-delayed price forecast ap-
pend on any values of or ; and is the term called the proach did not compare against the random walk model (see
forcing process where it can be any function of time, current and [16, App. A ]), except a recent work in [24]. This is because
lagged values of other variables and/or stochastic disturbances. the time-delayed approach performs direct forecast on a stock
 This article has been accepted for inclusion in a future issue.

4 IEEE TRANSACTIONS ON NEURAL NETWORKS

price series that is usually non-stationary. A time series is sta-

tionary, which is an important property in time series analysis,
when the mean value of the series remains constant over the time
series. Since many econometric time series are non-stationary,
the time-delayed price forecast approach tend to include linear
trends and thus deterministic shifts in the out-of-sample forecast
period tend to exacerbate forecast errors [12] when compared
against the random walk model.
Although financial analysts often adopt trading practices that
rely on forecasting the price levels of financial instruments,
Sanders and Ritzman [53] showed that people focus on the
use of judgemental approach as a basis for forecasting and this
approach is about as accurate as the best statistical approaches.
O’Connor et al. [54] examined the way people forecast trends
in time series and how they respond to changes in the trend and
presented studies of laboratory experiments that showed indi-
viduals have different tendencies and behaviours for upward
and downward series. A recent study in [24] suggested that
trading strategies guided by forecasts on the direction of price Fig. 2. Time-delayed price difference forecast approach.
change, which is also the forecast of trends, is more effective
and generate greater profits. Empirical results on the index of
Taiwan Stock Exchange in [24] showed that this approach using This approach uses the difference operator compared to
probabilistic neural network obtained higher returns than other time-delayed price forecast approach in Fig. 1, so we called this
investment strategies. In addition, the recent development in the the time-delayed price difference forecast approach. This ap-
theory of economic forecasting in [11], [12] showed that trans- proach is illustrated in Fig. 2.
forming nonstationary econometric time series to stationarity
by differencing and cointegration helped robustify the price IV. EXPERIMENT ON ARTIFICIAL PRICE DATA
forecasts and avoid deterministic shifts in the out-of-sample
forecasts which exacerbate the forecast errors. Thus the notion In this section, an artificial experimental stock price data set
of the human inherent ability in forecasting price trends and is constructed given in (9) and (10) to investigate the predictive
the recent development in the theory of economic forecasting performance of the time-delayed price difference forecast ap-
motivate the investigation of forecasting of the stock price proach using various networks
difference instead of price level.
From (4), the approach in forecasting the price difference is (9)
formulated as: Given values , (10)
compute , and predict
the change in represented as where is where and are constants; and are normal random
the difference operator. deviates with zero mean and unit variance.
Subtracting from (4) gives The stochastically generated artificial experimental data set,
which consists of 10000 data samples, is shown in Fig. 3. This
data set is constructed similar to the simulation model used in
[25] with the parameters , and initial condi-
tions and using (9) and (10). In this experi-
ment, the first 3000 data samples of the differenced price series
are used to train the following networks and the remaining data
samples of the differenced price series are used for testing.
1) Feed-forward neural network trained using back-propaga-
tion (FFNN-BP) [55].
2) Radial basis function networks (RBFN) [56].
(7) 3) Adaptive-network-based fuzzy inference systems (AN-
FISs) [57] with subtractive clustering [58].
where is another arbitrary parameter; is a stochastic 4) Evolving fuzzy neural networks (EFuNNs) [59].
random error term; and . 5) Dynamic evolving neural-fuzzy inference system
Clearly, (7) is a modified version of (4) and can then (DENFIS) [36].
be obtained using (8) 6) Pseudo outer-product fuzzy neural network using com-
positional rule of inference and singleton fuzzifier [60]
using RSPOP [41] to identify fuzzy rules (abbreviated as
(8) RSPOP-CRI).
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 5

Fig. 3. Artificially generated data set.

FFNN-BP is configured with 10 input neurons, 30 hidden

neurons and 1 output neuron using bipolar sigmoid activation
function and is trained for 1000 training iterations using a
learning rate of 0.025. RBFN is configured with a default
spread of 0.1. ANFIS is configured with a range of influence
of 0.3 for subtractive clustering and a maximum epoch of
100. EFuNN and DENFIS are both configured with default
parameters. To remove heuristic tuning of training parameters
in fuzzy membership learning algorithms of RSPOP-CRI, its
fuzzy membership functions are constructed using three input
fuzzy sets and five output fuzzy sets that span the range of
from the training data. The fuzzy sets are constructed
with open tails so that they cover values that are outside the
training range. The fuzzy sets constructed are illustrated in
Fig. 4. Fig. 4. (a) Input fuzzy sets. (b) Output fuzzy sets of RSPOP-CRI using
This method of constructing the fuzzy sets facilitates the the time-delayed price difference forecast approach on the artificial data set.
semantic interpretation of each fuzzy set (please refer to [61]
and [62] for more details on fuzzy sets) and also adheres
more with the fuzzy concept of expressing observations and with random walk is a more suitable performance measure for
measurement uncertainties. In this case, the magnitude of the the prediction of price difference. The results show that the
change in price value of the inputs to is predicted price difference values using neuro-fuzzy systems
represented as three linguistic concepts: decrease, no change, such as ANFIS, EFuNN, DENFIS and RSPOP-CRI yielded
increase as labelled in Fig. 4(a). Similarly, the magnitude of better results than FFNN-BP and RBFN. The results also show
the predicted change in price value of the output that among the networks, EFuNN and RSPOP-CRI yielded
is represented by five linguistic concepts: decrease, small the best recall and predictive results respectively. Although
decrease, no change, small increase, increase as labelled EFuNN yielded the best recall results, its predictive results
in Fig. 4(b). The fuzzy rules are derived using the RSPOP suffered, demonstrating an over-trained phenomenon. Among
algorithm [41] from the training data. RSPOP is used instead all the networks, RSPOP-CRI and DENFIS yielded recall and
of POP specifically for its rapid learning and feature selection prediction results that are superior to the Random Walk model
capability on large dimension input without heuristic selection in terms of MSE.
of training parameters. The experimental results of using the An examination of the recall and prediction results using
aforementioned networks for both recall and prediction of the only 200 time instances in Figs. 5 and 6 show that the time-de-
training and test data sets respectively are presented in Table II layed price difference forecast approach using DENFIS and
in terms of the commonly used mean square error (MSE) RSPOP-CRI are suitable for predicting the price difference of
and Pearson product-moment correlation coefficient . the artificial price series compared against a zero price differ-
Table II shows that RBFN yielded a higher , but RBFN ence from the Random Walk model. However, the predictive
yielded almost the same MSE as random walk. These results results from just one artificially generated data set is not suffi-
show that the value of alone is not directly indicative of ciently convincing to conclude that DENFIS and RSPOP-CRI
predictive performance and the value of MSE in comparison are superior choices compared against the other networks.
 This article has been accepted for inclusion in a future issue.

6 IEEE TRANSACTIONS ON NEURAL NETWORKS

Fig. 5. (a) Magnified recall results and (b) predictive results of forecasting Fig. 6. (a) Magnified recall results and (b) predictive results of forecasting
price difference on the artificial data set using DENFIS. price difference on the artificial data set using RSPOP-CRI.

TABLE II TABLE III

SUMMARY OF RESULTS OF FORECASTING PRICE DIFFERENCE ON THE AVERAGED RESULTS OF FORECASTING PRICE DIFFERENCE ON 20
ARTIFICIAL DATA SET ARTIFICIALLY GENERATED DATA SET

Therefore, further experimental results are collected from addi-

tionally generated artificial stock price data sets using (9) and number of fuzzy rules than RSPOP-CRI. However, DENFIS
(10). Table III presents the accuracy of the recall and predictive is based on the TSK model [38], [40] used for precise fuzzy
results in terms of average MSE and the interpretability of the modeling, which generally yield decreased interpretability
neuro-fuzzy systems used in terms of average number of fuzzy but increased accuracy [39]. In contrast, POPFNN-CRI(S)
rules on 20 artificially generated data sets. Standard deviations [60] is based on the Mamdani model [37] used for linguistic
of the average MSE and number of fuzzy rules are also included. fuzzy modeling, which is focused on interpretability. Thus, in
Table III shows that DENFIS and RSPOP-CRI yielded a terms of interpretability, RSPOP-CRI is a superior choice than
MSE that is lower than the Random Walk model. It also shows DENFIS since both networks used almost the same number
that DENFIS yielded a slightly lower MSE and slightly higher of fuzzy rules. In terms of accuracy, both RSPOP-CRI and
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 7

A number of techniques can be used to generate buy and sell

signals using technical analysis techniques. One of the simplest
and popular trading rules for deciding when to buy and sell
in a security market is the moving average rule [3], [17]. A
widely used variant of the moving average rule is the moving
average convergence/divergence (MACD) oscillator originally
developed by Gerald Appel [5]. MACD uses the crossovers of
a Fast signal given in (12) and a Slow signal given in (13) to
indicate a buy or sell signal. The exponential moving average
(EMA) of a price series is given in (14) [5]. The Fast signal in
(12) is computed from the difference between the EMA
Fig. 7. Trading decision model on technical analysis approach. and the EMA of using the closing price where
. The Slow signal in (13) is computed from the
EMA of the Fast signal
DENFIS yielded superior predictive performance in terms
of averaged MSE compared against the random walk model.
Therefore, these results motivate the use of RSPOP-CRI in the (12)
time-delayed price difference forecast approach. (13)
This section presented the results based on artificial stock
(14)
price data set using the time-delayed price difference forecast
approach with several networks. Results showed that the use of
DENFIS [36] and a novel rough set-based neuro-fuzzy system, where ; is the number of time instance of the
namely POPFNN-CRI(S) [60] with fuzzy rules identified using moving average; is the price at the current time instance
the RSPOP algorithm [41] (abbreviated as RSPOP-CRI), on ar- ; and is the EMA of time instance .
tificially generated data sets yielded superior results than the Instead of using the crossover of the fast and slow signal to
Random walk model. A trading model and experimental trading generate the buy/sell signal, simpler moving average trading
results on real market data is presented in Sections V and VI rules can be generated using just the slow signal given in (15),
to assess the trading performance of RSPOP-CRI and DENFIS where the slow signal as given in (13) is a moving average of the
using the time-delayed price difference forecast approach. fast signal and the fast signal is the two moving averages of the
price level: a long-period average and a short-period average [3]
V. FINANCIAL TRADING SYSTEMS
In order to assess the trading performance of the proposed (15)
approach, a trading decision model where profits or losses are
computable from trading decisions is required. A realistic yet
simple computation of profits or losses of a trading decision where is the action from the trading module for time in-
model based on recurrent reinforcement learning is presented in stant ; and is given in (13).
[25]. Based on this model, two commonly used trading decision Fig. 8 shows the box plots summarizing the multiplicative
models based on the technical analysis approach will be dis- profits using just simple moving average trading rules without
cussed; they are, namely: with price forecast and without price using RSPOP-CRI price forecast for 100 experiments on ar-
forecast. tificially generated price series data generated using (9) and
A block diagram for a generic trading decision model based (10). The multiplicative profits are computed using (11) and the
on the technical analysis approach is shown in Fig. 7. The tech- trading rules are constructed using (12)–(14) and (15) with a
nical trading analysis in this model is based on the premise that heuristically chosen , and
the market price patterns are assumed to recur in the future [17]. (please refer to [3], [63] on the choice of moving average pa-
In Fig. 7, the price value of a security is represented as a time rameters). Comparing the results in Fig. 8 against similar ex-
series , where represents a value at time instant . The periments using recurrent reinforcement learning (RLL) neural
action of the trading system is assumed to be one of short, neu- network in [23], the results in the former yielded a median of
tral or long, and this action is, respectively, represented by about 25 and 6 times while the latter yielded a median of about
where [25]. The trading system return is sub- 6 and 4 times for transaction costs of 0.2% and 0.5% respec-
sequently modeled by portfolio end value using a multiplicative tively. These results further exemplified the high returns ob-
return given in (11) where the transaction cost is assumed to be tainable using simple moving average trading rules compared
a fraction of the transacted price value [25] against other trading strategies as shown in [3].
A block diagram for a generic trading decision model based
on the technical analysis approach with forecast is shown in
(11) Fig. 9 [25]. In such a system, the forecasting module is trained
using supervised learning on a training data set and is then used
where ; is the action from to forecast an out-of-sample data set. The forecasts are then
the trading system; and is the transaction rate. used as input to a trading module to generate trading signals.
 This article has been accepted for inclusion in a future issue.

8 IEEE TRANSACTIONS ON NEURAL NETWORKS

Fig. 9. Trading decision model on technical analysis approach with forecast

adapted from [25].

Fig. 8. Box plots of portfolio end values against transaction costs over 100
artificial data set using just simple moving average trading rules.

This common practice of using only the forecasts as input to the

trading module results in loss of information or a forecast bottle-
neck that may lead to suboptimal performance [25]. An example
on the use of this model is the work of [24] where trading de-
cision signals are generated using the degree of certainty on the Fig. 10. Stock trading using RSPOP: A forecast bottleneck free trading deci-
directional forecast of price movements obtained from a trained sion model and RSPOP predictive model.
PNN. Comparing the models in Figs. 7 and 9, the former allows
the generation of trading signals from simple trading rules like
the moving average trading rules while the latter only allows (17)
the generation of trading rules based on the forecasts. Despite (18)
the high profitable returns obtainable through the use of simple (19)
moving average rules as shown previously, this forecast bottle-
neck as indicated in [25] impedes the use of the price history of where ; is the number of time instance of the
the input series in the generation of trading signals. Hence, moving average; is the action from the trading module for
this forecast bottleneck encumbers the trading performance of time instant ; and is the forecast price value from the
the forecast module. This motivates the design of a forecast bot- forecast module.
tleneck free trading decision model that in turn facilitates the
research using a synergy of various forecasting methods and VI. EMPIRICAL RESULTS OF STOCK TRADING USING RSPOP
trading decision methods. This section presents experimental results in the trading of
A novel rough set-based neuro-fuzzy forecast bottleneck stock in real world market using the proposed stock trading
free stock trading decision model is proposed and shown in using RSPOP model shown in Fig. 10. The trading per-
Fig. 10. We called the proposed trading decision model stock formances of Stock Trading using RSPOP (labelled with
trading using RSPOP, which is based on the technical analysis RSPOP forecast) are benchmarked against Stock Trading using
approach. Comparing the models in Figs. 9 and 10, the latter DENFIS (labelled with DENFIS forecast), trading decision
simply includes an additional signal of the input series to model without forecast (labelled without forecast) and trading
the trading module and uses the time-delayed price difference decision model with ideal forecast (labelled with ideal forecast)
forecast approach. Although this model appears to be an overly on 2 sets of real world market data; namely, Neptune Orient
simplistic solution to resolve the forecast bottleneck, this model Lines (NOL) and Development Bank of Singapore (DBS). In
facilitates the synergy of the time-delayed price difference these experiments, the trading signals for stock trading without
forecast approach with the use of the moving average trading forecast model is computed using (12)–(15). The trading
rules. The computation of the trading signal using both signals for stock trading with DENFIS and RSPOP forecast
the forecast value and is given in (19). The models are computed using (16)–(19) where is the
computation of the moving average signals is given in (16)–(18) forecast obtained using DENFIS and RSPOP respectively. The
trading signals for stock trading with ideal forecast model are
also computed using (16)–(19), but the forecast is
replaced with the actual price value of the next time instance
(16) . All portfolio end values are computed using (11).
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 9

Fig. 11. Price and trading signals on NOL using the stock trading with RSPOP forecast model.

Fig. 12. Portfolio end values R(T ) on NOL.

A. NOL ified by the introduction of a 1% moving average band. The in-

troduction of this hysteresis band reduces the number of trading
In this experiment, the trading performance of using the pro- signals by eliminating the “whiplash” signals when the short-pe-
posed stock trading with RSPOP forecast model on the daily riod and long-period moving averages are close [3].
closing price of the NOL is compared against the stock trading Figs. 11 and 12 show the out-of-sample price series, the
with DENFIS forecast model, the stock trading without fore- long and short EMAs, the trading signals and the portfolio end
cast model and the stock trading with ideal forecast model. The values . Starting with a portfolio value of ,
forecast models are trained with previous values of the at the end of the experiment, the stock trading without forecast
differenced price series as inputs. The experimental price series model yielded a portfolio end value of , the
consists of 5917 price values obtained from the Yahoo Finance proposed stock trading with RSPOP forecast model yielded
website on the counter N03.SI from the period of January 2, , the stock trading with DENFIS forecast
1980 to March 1, 2005. The in-sample training data set is con- model yielded and the stock trading with
structed using the first 4000 data points and the out-of-sample ideal forecast model yielded . The results show
test data set is constructed using the more recent 1917 data that the stock trading with forecast models yielded higher re-
points. Trading signals are generated using heuristically chosen turns than the stock trading without forecast model and yielded
moving average parameters , , lower returns than the stock trading with ideal forecast model.
and the portfolio end values are calculated with a transac- Overall, the use of the proposed stock trading with RSPOP
tion cost of . The moving average trading rule is mod- forecast model yielded an increase of in
 This article has been accepted for inclusion in a future issue.

10 IEEE TRANSACTIONS ON NEURAL NETWORKS

Fig. 13. Magnified results of Fig. 11 from time T = 4400 to T = 4600.

Fig. 14. (a) Fixed input fuzzy membership functions. (b) Fixed output fuzzy membership functions. (c) Fuzzy rules identified using RSPOP in the stock trading
with RSPOP forecast model on NOL.

portfolio end value trading NOL stock compared with the stock with RSPOP forecast model and the stock trading with DENFIS
trading without forecast model. The stock trading with DENFIS forecast model yielded a much higher portfolio end value than
forecast model yielded a lower increase of in the stock trading without forecast model. However, since the
portfolio end value trading NOL stock compared with the stock stock trading with forecast models are unable to forecast with
trading without forecast model. exact accuracy, these models yielded a lower portfolio end value
Fig. 13 shows the magnified results of Fig. 11 from time than the stock trading with ideal forecast model. This result is
to . The moving average trading signals are intuitive since it is impossible to obtain higher returns than the
generated from the Slow EMA of the difference stock trading with ideal forecast model, which uses
between the Long EMA and Short EMA the actual future price value.
of the NOL price values. When the Short EMA crosses the Fig. 14 shows the fixed fuzzy membership functions and
Long EMA from below as indicated in Fig. 13, it indicates a fuzzy rules identified using RSPOP in the stock trading with
rising price trend and, thus, the a buy signal is RSPOP forecast model. A total of eight fuzzy rules are iden-
generated from the Slow EMA. The stock trading with RSPOP tified by RSPOP in contrast to nine fuzzy rules identified by
forecast model generated the buy signal earlier through the use DENFIS. Since POPFNN-CRI(S) neuro-fuzzy system [60]
of the forecast value . Hence the proposed stock trading using RSPOP [41] is a Mamdani fuzzy model [37] whereas
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 11

Fig. 15. Price and trading signals on NOL using the stock trading with RSPOP forecast model but using another set of moving average trading parameters.

Fig. 16. Portfolio end values R(T ) on NOL using another set of moving average parameters.

DENFIS neuro-fuzzy system [36] is a TSK fuzzy model [40], computing an appropriate embedding dimension compared
the rules identified in the former are more interpretable than against the stock trading with DENFIS forecast model.
DENFIS model [39]. Moreover, Fig. 14 shows that only the
three of the time-delayed price values ( , and B. NOL—Different Trading Parameters
) are used in the stock trading with RSPOP forecast The experiment is repeated using another set of moving av-
model instead of the all the time-delayed price values to erage parameters , , , and the
forecast the future price values. Since the heuristically chosen portfolio end values are calculated with the same transac-
value of does not represent the embedding dimension tion cost of . Figs. 15 and 16 show the out-of-sample
of the NOL price series, the results show that the use of the price series with long and short EMAs, the trading signals and
RSPOP algorithm enabled the computation of an appropriate the portfolio end values using this set of moving average
value of the embedding dimension. In contrast, the fuzzy rules parameters.
in the stock trading with DENFIS forecast model used all ten Starting with a portfolio value of , at the
time-delayed price values and the fuzzy rules are too complex end of the experiment, the stock trading without forecast
to be listed. Therefore, the results showed that the proposed model yielded a portfolio end value of , the
stock trading with RSPOP forecast model not only yielded a proposed stock trading model RSPOP forecast model yielded
higher portfolio end value than the stock trading with DENFIS , the stock trading with DENFIS forecast
forecast model, but also yielded more interpretable rules by model yielded and the stock trading with
 This article has been accepted for inclusion in a future issue.

12 IEEE TRANSACTIONS ON NEURAL NETWORKS

Fig. 17. Price and trading signals on DBS using the stock trading with RSPOP forecast model.

Fig. 18. Portfolio end values R(T ) on DBS.

ideal forecast model yielded . Overall, the trading with ideal forecast model. Most significantly, the results
use of the proposed stock trading with RSPOP forecast model showed that the proposed stock trading with RSPOP forecast
yielded an increase of in portfolio end value model yielded a higher portfolio end value than the stock
trading NOL stock compared with the stock trading without trading without forecast model as well as the stock trading with
forecast model. The stock trading with DENFIS forecast model DENFIS forecast model, despite using a different set of moving
yielded a lower increase of in portfolio end average parameters.
value trading NOL stock compared with the stock trading
without forecast model. Comparing the results from this experi- C. Development Bank of Singapore (DBS)
ment using moving average parameters , ,
against the results from the previous experiment In this experiment, the trading performance of using the
using moving average parameters , , proposed stock trading with RSPOP model on the daily closing
; the results showed that the use of different moving price of the DBS is compared against the stock trading with
average parameters yielded different portfolio end values for DENFIS forecast model, the stock trading without forecast
the stock trading without forecast model, the stock trading with model and the stock trading with ideal forecast model. The
forecast models as well as the stock trading with ideal forecast forecast models are trained with previous values of the
model. The results also showed that the stock trading with differenced price series as inputs. The experimental price series
forecast models yielded higher returns than the stock trading consists of 6222 price values obtained from Yahoo Finance
without forecast model and yielded lower returns than the stock web site on the counter D05.SI from the period of January 2,
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 13

Fig. 19. (a) Fixed input fuzzy membership functions. (b) Fixed output fuzzy membership functions. (c) Fuzzy rules identified using RSPOP in the stock trading
with RSPOP forecast model on DBS.

1980 to March 1, 2005. The in-sample training data set is con- tation of an appropriate value of the embedding dimension. In
structed using the first 4000 data points and the out-of-sample contrast, the fuzzy rules in the stock trading with DENFIS fore-
test data set is constructed using the more recent 2222 data cast model used all 10 time-delayed price values and the fuzzy
points. Trading signals are generated using heuristically chosen rules are too complex to be listed.
moving average parameters , , The results obtained on the NOL and DBS stock prices show
with a 1% moving average band and the portfolio end values that the proposed stock trading with RSPOP forecast model
are calculated with a transaction cost of . yielded portfolio end values of and
Figs. 17 and 18 show the out-of-sample price series, the , respectively, using the same moving average parame-
long and short EMAs, the trading signals and the portfolio end ters , , and . This translates to
values . Starting with a portfolio value of , a profit of 16.0362 and 2.8582 times the initial capital invested
at the end of the experiment, the stock trading without forecast in NOL and DBS, respectively. Although a significant amount
model yielded a portfolio end value of , the of profit is obtained, the results once again exemplified the fact
proposed stock trading with RSPOP forecast model yielded that profit or loss strongly depends on the trading module and its
, the stock trading with DENFIS forecast parameters, as well as the stock counter selected. These results
model yielded and the stock trading with ideal obtained showed a fair representation of the real world market
forecast model yielded . The results show that because the effectiveness of the moving average trading rule re-
the stock trading with forecast models yielded higher returns lies heavily on the heuristic choice of moving average parame-
than the stock trading without forecast model and yielded ters for a particular data set. Hence, it is fair to generalize that
lower returns than the stock trading with ideal forecast model. profit or loss strongly depends on the trading module. Never-
Overall, the use of the proposed stock trading with RSPOP theless, the results obtained using the proposed stock trading
forecast model yielded an increase of in using RSPOP forecast model in all three experiments using real
portfolio end value trading DBS stock compared with the stock market data yielded significantly higher portfolio end values
trading without forecast model. The stock trading with DENFIS than the stock trading without forecast model. This showed that
forecast model yielded a lower increase of in using the proposed stock trading with RSPOP forecast model
portfolio end value trading DBS stock compared with the stock improved the portfolio end value yielded compared against a
trading without forecast model. Fig. 19. shows the fixed fuzzy stock trading without forecast model. The results also showed
membership functions and fuzzy rules identified using RSPOP that the use of RSPOP algorithm enables the identification of
in the stock trading with RSPOP forecast model. the embedding dimension of the stock price series, which pro-
Fig. 19 showed that a total of 5 fuzzy rules are identified duced a smaller number of interpretable rules.
by RSPOP whereas a total of 10 fuzzy rules are identified by
DENFIS. Fig. 19. shows that only two of the time-delayed price VII. CONCLUSION
values ( and ) are used in the stock trading with A novel rough set-based neuro-fuzzy stock trading decision
RSPOP forecast model instead of the all the time-de- model called stock trading using RSPOP is proposed in this
layed price values to forecast the future price values. Since the paper. The proposed stock trading model circumvents the
heuristically chosen value of does not represent the em- forecast bottleneck and synergizes the time-delayed price dif-
bedding dimension of the NOL price series, the results showed ference forecast approach with simple moving average rules for
again that the use of the RSPOP algorithm enabled the compu- generating trading signals. Experimental results on forecasting
 This article has been accepted for inclusion in a future issue.

14 IEEE TRANSACTIONS ON NEURAL NETWORKS

stock price difference on artificially generated price series [6] A. W. Lo, H. Mamaysky, and J. Wang, “Foundations of technical anal-
data showed that two neuro-fuzzy systems, namely DENFIS ysis: Computational algorithms, statistical inference, and empirical im-
plementation,” J. Finance, vol. 55, no. 4, pp. 1705–1765, Aug. 2000.
and a novel rough set-based neuro-fuzzy system called the [7] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis:
RSPOP FNN yielded superior predictive performance than the Forecasting and Control. Englewood Cliffs, NJ: Prentice-Hall, 1994.
well-established random walk model. As the trading profits [8] C.- S. Lin, H. A. Khan, and C.- C. Huang, “Can the neuro fuzzy model
predict stock indexes better than its rivals?,” in Proc. CIRJE, Aug. 2002,
is more important to an investor than statistical performance, CIRJE-F-165.
these neuro-fuzzy systems are incorporated as the underlying [9] R. Engle, “Autoregressive conditional heteroskedasticity with esti-
predictor model in forecasting stock price difference with a mates of the variance of united kingdom inflation,” Econometrica, vol.
50, no. 4, pp. 987–1007, Jul. 1982.
forecast bottleneck free trading decision model using moving [10] W. Enders, Applied Econometric Time Series. Hoboken, NJ: Wiley,
average trading rules. Experimental results based on real world 2004.
stock market data, namely the stock prices of NOL and DBS, [11] M. P. Clements and D. F. Hendry, Forecasting Economic Time Se-
ries. Cambridge, U.K.: Cambridge Univ. Press, 1998.
are presented. Experimental trading profits in terms of portfolio [12] ——, Forecasting Non-Stationary Economic Time Series. Cam-
end values of the proposed stock trading with RSPOP forecast bridge, MA: MIT Press, 1999.
model are benchmarked against the stock trading with DENFIS [13] Y. S. Abu-Mostafa, A. F. Atiya, M. Magdon-Ismail, and H. White,
“Introduction to the special issue on neural networks in financial en-
forecast model, the stock trading without forecast model and gineering,” IEEE Trans. Neural Netw., vol. 12, no. 4, pp. 653–656, Jul.
the stock trading with ideal forecast model. Experimental 2001.
results showed that the proposed stock trading with RSPOP [14] R. R. Trippi and E. Turban, Neural Networks in Finance and Investing:
Using Artificial Intelligence to Improve Real World Performance.
forecast model identified fewer fuzzy rules as well as yielded Chicago, IL: Probus, 1993.
higher portfolio end values compared against the stock trading [15] A.- P. Refenes, Neural Networks in the Capital Markets. New York:
with DENFIS forecast model. Thus results showed that using Wiley, 1995.
[16] M. Adya and F. Collopy, “How effective are neural networks at fore-
RSPOP as the underlying predictive model identified rules with casting and prediction? A review and evaluation,” J. Forecasting, vol.
greater interpretability and accuracy. Experimental results also 17, no. 5–6, pp. 481–495, Sep. 1998.
provided evidence that profitability strongly depends on the [17] R. Gencay, “The predictability of security returns with simple technical
trading rules,” J. Empir. Finance, vol. 5, no. 4, pp. 347–359, Oct. 1998.
trading module and its parameters as well as the stock counter [18] R. J. Kuo, C. H. Chen, and Y. C. Hwang, “An intelligent stock trading
selected. Despite the different yield in profits, the experimental decision support system through integration of genetic algorithm based
results consistently showed that the proposed stock trading with fuzzy neural network and artificial neural network,” Fuzzy Sets Syst.,
vol. 118, no. 1, pp. 21–45, Feb. 2001.
RSPOP forecast model yielded significantly higher profits than [19] T. Van Gestel, J. A. K. Suykens, D.-E. Baestaens, A. Lambrechts, G.
the stock trading without forecast model. Lanckriet, B. Vandaele, B. De Moor, and J. Vandewalle, “Financial
The potential of the proposed stock trading using RSPOP time series prediction using least squares support vector machines
within the evidence framework,” IEEE Trans. Neural Netw., vol. 12,
model discussed in this paper is notable as it demonstrates the no. 4, pp. 809–821, Jul. 2001.
feasible application of POPFNN in the area of financial predic- [20] J. E. Fieldsend and S. Singh, “Pareto evolutionary neural networks,”
tion. Although the proposed approach is not proven to assure IEEE Trans. Neural Netw., vol. 16, no. 2, pp. 338–354, Mar. 2005.
[21] E. W. Saad, D. V. Prokhorov, and D. C. Wunsch II, “Comparative
profits from real world markets, this paper presented a forecast study of stock trend prediction using time delay, recurrent, and proba-
bottleneck free trading decision model that opened new oppor- bilistic neural networks,” IEEE Trans. Neural Netw., vol. 9, no. 6, pp.
tunities for the synergy of this approach with various trading de- 1456–1470, Nov. 1998.
[22] W. Leigh, R. Purvis, and J. M. Ragusa, “Forecasting the NYSE com-
cision schemes. Design of intelligent trading decision schemes posite index with technical analysis, pattern recognizer, neural net-
that synergize with forecasting to yield optimal trading perfor- work, and genetic algorithm: A case study in romantic decision sup-
mance on real world stock market data now appears to be a port,” Decision Support Syst., vol. 32, no. 4, pp. 361–377, Mar. 2002.
[23] J. Moody and M. Saffell, “Learning to trade via direct reinforcement,”
promising research area. The research presented in this paper is IEEE Trans. Neural Netw., vol. 12, no. 4, pp. 875–889, Jul. 2001.
in line with the research direction of Centre for Computational [24] A.-S. Chen, M. T. Leung, and H. Daouk, “Application of neural net-
Intelligence (C2i) [64], formerly the Intelligent Systems Labo- works to an emerging financial market: Forecasting and trading the
Taiwan Stock Index,” Comput. Oper. Res., vol. 30, no. 6, pp. 901–923,
ratory, at the Nanyang Technological University, Singapore. C2i May 2003.
undertakes active research in intelligent neuro-fuzzy systems for [25] J. Moody, L. Wu, Y. Liao, and M. Saffell, “Performance functions and
the modelling of complex, nonlinear and dynamic problem do- reinforcement learning for trading systems and portfolios,” J. Fore-
casting, vol. 17, no. 5–6, pp. 441–470, Sep. 1998.
mains. Examples of neural and neuro-fuzzy systems developed [26] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduc-
at C2i are MCMAC [65], GenSoFNN [66], and POPFNN [67]. tion. Cambridge, MA: MIT Press, 1998.
[27] S. Mitra and Y. Hayashi, “Neuro-fuzzy rule generation: Survey in soft
computing framework,” IEEE Trans. Neural Netw., vol. 11, no. 3, pp.
748–768, May 2000.
REFERENCES [28] J. L. Castro, “Fuzzy logic controllers are universal approximators,”
IEEE Trans. Syst., Man, Cybern., vol. 25, no. 4, pp. 629–635, Apr.
[1] B. Graham and D. Dodd, Security Analysis: Principles and Tech- 1995.
niques. New York: McGraw-Hill, 1940. [29] H. Ying, Y. Ding, S. Li, and S. Shao, “Comparison of necessary con-
[2] R. D. Edwards, J. Magee, and W. H. C. Bassetti, Technical Analysis of ditions for typical Takagi-Sugeno and Mamdani fuzzy systems as uni-
Stock Trends. Boca Raton, FL: Saint Lucie Press, 2001. versal approximators,” IEEE Trans. Syst., Man, Cybern. A, Syst., Hu-
[3] W. Brock, J. Lakonishok, and B. LeBaron, “Simple technical trading mans, vol. 29, no. 5, pp. 508–514, Oct. 1999.
rules and the stochastic properties of stock returns,” J. Finance, vol. 47, [30] D. Tikk, L. T. Kóczy, and T. D. Gedeon, “A survey on universal ap-
no. 5, pp. 1731–1764, Dec. 1992. proximation and its limits in soft computing techniques,” Int. J. Approx.
[4] E. F. Fama, “Efficient capital markets: II,” J. Finance, vol. 46, no. 5, Reason., vol. 33, no. 2, pp. 185–202, Jun. 2003.
pp. 1575–1617, Dec. 1991. [31] S. Guillaume, “Designing fuzzy inference systems from data: An in-
[5] T. Plummer and A. Ridley, Forecasting Financial Markets: The Psy- terpretability-oriented review,” IEEE Trans. Fuzzy Syst., vol. 9, no. 3,
chology of Successful Investing. London, U.K.: Kogan, 2003. pp. 426–443, Jun. 2001.
 This article has been accepted for inclusion in a future issue.

ANG AND QUEK: STOCK TRADING USING RSPOP: A NOVEL ROUGH SET-BASED NEURO-FUZZY APPROACH 15

[32] R. R. Trippi and J. K. Lee, Artificial Intelligence in Finance & In- [56] S. Chen, C. F. N. Cowan, and P. M. Grant, “Orthogonal least squares
vesting: State-of-the-Art Technologies for Securities Selection and learning algorithm for radial basis,” IEEE Trans. Neural Netw., vol. 2,
Portfolio Management. Chicago, IL: Irwin Professional, 1996. no. 2, pp. 302–309, Mar. 1991.
[33] K. N. Pantazopoulos, L. H. Tsoukalas, N. G. Bourbakis, M. J. Brun, [57] J.-S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference
and E. N. Houstis, “Financial prediction and trading strategies using system,” IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685,
neurofuzzy approaches,” IEEE Trans. Syst., Man, Cybern. B, Cybern., May 1993.
vol. 28, no. 4, pp. 520–531, Aug. 1998. [58] S. L. Chiu, “Fuzzy model identification based on cluster estimation,”
[34] T. Nishina and M. Hagiwara, “Fuzzy inference neural network,” Neu- J. Intell. Fuzzy Syst., vol. 2, no. 3, pp. 267–278, 1994.
rocomput., vol. 14, no. 3, pp. 223–239, Feb. 1997. [59] N. Kasabov, “Evolving fuzzy neural networks for supervised/unsuper-
[35] H. Nakanishi, I. B. Turksen, and M. Sugeno, “A review and comparison vised online knowledge-based learning,” IEEE Trans. Syst., Man, Cy-
of six reasoning methods,” Fuzzy Sets Syst., vol. 57, no. 3, pp. 257–294, bern. B, Cybern., vol. 31, no. 6, pp. 902–918, Dec. 2001.
Aug. 1993. [60] K. K. Ang, C. Quek, and M. Pasquier, “POPFNN-CRI(S): Pseudo outer
[36] N. K. Kasabov and Q. Song, “DENFIS: Dynamic evolving neural- product based fuzzy neural network using the compositional rule of
fuzzy inference system and its application for time-series prediction,” inference and singleton fuzzifier,” IEEE Trans. Syst., Man, Cybern. B,
IEEE Trans. Fuzzy Syst., vol. 10, no. 2, pp. 144–154, Apr. 2002. Cybern., vol. 33, no. 6, pp. 838–849, Dec. 2003.
[37] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis [61] G. J. Klir and B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory and Appli-
with a fuzzy logic controller,” Int. J. Man-Machine Stud., vol. 7, no. 1, cations. Upper Saddle River, NJ: Prentice-Hall, 1995.
pp. 1–13, 1975. [62] D. Dubois and H. Prade, Fundamentals of Fuzzy Sets. Boston, MA:
[38] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its ap- Kluwer, 2000.
plications to modeling and control,” IEEE Trans. Syst., Man, Cybern., [63] S. Taylor, Basics of moving averages 2001 [Online]. Available: http://
vol. SMC-15, no. 1, pp. 116–132, Jan. 1985. www.investopedia.com/articles/technical/052201.asp
[39] J. Casillas, O. Cordón, F. Herrera, and L. Magdalena, Interpretability [64] Centre for computational intelligence 2005 [Online]. Available: http://
Issues in Fuzzy Modeling. Berlin, Germany: Springer-Verlag, 2003. www.c2i.ntu.edu.sg
[40] M. Sugeno and G. T. Kang, “Structure identification of fuzzy model,” [65] K. K. Ang and C. Quek, “Improved MCMAC with momentum, neigh-
Fuzzy Sets Syst., vol. 28, no. 1, pp. 15–33, Oct. 1988. borhood, and averaged trapezoidal output,” IEEE Trans. Syst., Man,
[41] K. K. Ang and C. Quek, “RSPOP: Rough set-based pseudo outer- Cybern. B, Cybern., vol. 30, no. 3, pp. 491–500, Jun. 2000.
product fuzzy rule identification algorithm,” Neural Comput., vol. 17, [66] W. L. Tung and C. Quek, “GenSoFNN: A generic self-organizing
no. 1, pp. 205–243, Jan. 2005. fuzzy neural network,” IEEE Trans. Neural Netw., vol. 13, no. 5, pp.
[42] T. A. Johansen and R. Babuška, “Multiobjective identification of 1075–1086, Sep. 2002.
Takagi-Sugeno fuzzy models,” IEEE Trans. Fuzzy Syst., vol. 11, no. [67] R. W. Zhou and C. Quek, “POPFNN: A pseudo outer-product based
6, pp. 847–860, Dec. 2003. fuzzy neural network,” Neural Networks, vol. 9, no. 9, pp. 1569–1581,
[43] J. Yen, L. Wang, and C. W. Gillespie, “Improving the interpretability of Dec. 1996.
TSK fuzzy models by combining global learning and local learning,”
IEEE Trans. Fuzzy Syst., vol. 6, no. 4, pp. 530–537, Nov. 1998.
[44] Y. Jin, “Fuzzy modeling of high-dimensional systems: Complexity re-
duction,” IEEE Trans. Fuzzy Syst., vol. 8, no. 2, pp. 212–221, Apr.
2000.
[45] Z. Pawlak, Rough Sets: Theoretical Aspects of Reasoning About
Data. Boston, MA: Kluwer, 1991.
[46] R. W. Swiniarski and A. Skowron, “Rough set methods in feature selec-
tion and recognition,” Pattern Recogn. Lett., vol. 24, no. 6, pp. 833–849, Kai Keng Ang (S’05) received the B.A.Sc. (with
Mar. 2003. first-class honors) and M.Phil. degrees in computer
[47] D. Kugiumtzis, B. Lillekjendlie, and N. Christophersen, “Chaotic time engineering from Nanyang Technological Univer-
series part I: Estimation of some invariant properties in state space,” sity, Singapore, in 1997 and 1999, respectively. He
Model., Ident. Control, vol. 15, no. 4, pp. 205–224, 1994. is currently working toward the Ph.D. degree at
[48] N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, “Geom- the Centre for Computational Intelligence, School
etry from a time series,” Phys. Rev. Lett., vol. 45, no. 9, pp. 712–715, of Computer Engineering, Nanyang Technological
Sep. 1980. University.
[49] F. Takens, , D. A. Rand and L.- S. Young, Eds., “Detecting strange He was a Senior Software Engineer with Delphi
attractors in turbulence,” in Dynamical Systems and Turbulence. New Automotive Systems, Singapore, Pte, Ltd., working
York: Springer-Verlag, 1981, pp. 366–381. on embedded software for automotive engine con-
[50] L. P. Maguire, B. Roche, T. M. McGinnity, and L. J. McDaid, “Pre- trollers from 1999 to 2003.
dicting a chaotic time series using a fuzzy neural network,” Inform. Mr. Ang was awarded the Singapore Millennium Foundation Ph.D. Scholar-
Sci., vol. 112, no. 1–4, pp. 125–136, Dec. 1998. ship in 2005.
[51] T. Kimoto, K. Asakawa, M. Yoda, and M. Takeoka, “Stock market
prediction system with modular neural networks,” in Proc. IJCNN Int.
Joint Conf. Neural Networks, 1990, vol. 1, pp. 1–6.
[52] L.-X. Wang and J. M. Mendel, “Generating fuzzy rules by learning
from examples,” IEEE Trans. Syst., Man, Cybern., vol. 22, no. 6, pp.
1414–1427, Nov. 1992. Chai Quek (M’83) received the B.Sc. degree in elec-
[53] N. R. Sanders and L. P. Ritzman, “The need for contextual and trical and electronics engineering and the Ph.D. de-
technical knowledge in judgmental forecasting,” J. Behav. Decision gree in intelligent control, both from Heriot Watt Uni-
Making, vol. 5, no. 1, pp. 39–52, Jan. 1992. versity, Edinburgh, U.K., in 1986 and 1990, respec-
[54] M. O’Connor, W. Remus, and K. Griggs, “Going up-going down: How tively.
good are people at forecasting trends and changes in trends?,” J. Fore- He is an Associate Professor and a member of
casting, vol. 16, no. 3, pp. 165–176, May 1997. the Centre for Computational Intelligence, School
[55] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, , D. E. Rumelhart of Computer Engineering, Nanyang Technological
and J. L. McClelland, Eds., “Learning internal representations by error University, Singapore. His research interests include
propagation,” in Parallel Distributed Processing: Explorations in the intelligent control, intelligent architectures, artificial
Microstructure of Cognition. Cambridge, MA: MIT Press, 1986, vol. intelligence in education, neural networks, fuzzy
1, pp. 318–362. systems, fuzzy rule-based systems, and genetic algorithms.