Engineering Applications of Artificial Intelligence 61 (2017) 4756

Contents lists available at ScienceDirect

Engineering Applications of Articial Intelligence

journal homepage: www.elsevier.com/locate/engappai

A relative value trading system based on a correlation and rough set analysis MARK
for the foreign exchange futures market

Sukjun Leea, David Enkeb, Youngmin Kimc,
a
Business School, Kwangwoon University, 20 Kwangwoon-ro, Nowon-Gu, Seoul 01897, Republic of Korea
b
Laboratory for Investment and Financial Engineering, Department of Engineering Management and Systems Engineering, Missouri University of Science
and Technology, 221 Engineering Management, 600 W. 14th Street, Rolla, MO 65409-0370, USA
c
Laboratory for Investment and Financial Engineering, Department of Engineering Management and Systems Engineering, Missouri University of Science
and Technology, 205 Engineering Management, 600 W. 14th Street, Rolla, MO 65409-0370, USA

A R T I C L E I N F O A BS T RAC T

Keywords: This paper describes the conceptual framework of a relative value (RV)-based trading system focused on the
Correlation analysis data characteristics of the foreign exchange futures market using a correlation and rough set analysis. RV
Foreign exchange futures trading is an investment strategy that can generate potential prots based on the RV of two securities, regardless
Relative value trading system of market direction. We select pairs with a positive correlation, negative correlation, or no correlation based on
Rough set analysis
the correlation coecients between foreign exchange futures contracts. To implement and experiment with the
proposed system, trading rules are generated using a rough set analysis that employs technical indicators
derived from the RVs of the pairs. The performance of the proposed trading system is analyzed using the
momentum and buy-and-hold trading strategies as benchmarks. The experimental results and analyses
demonstrate that the level of the correlation of the pairs must be considered when developing stable and
protable RV trading systems in a foreign exchange futures market.

1. Introduction tion, appropriate modeling and forecasting of time series data in

Capital market liberalization and globalization have led to oppor-
tunities for investment in the foreign exchange (FX) market for
commercial banks, international companies, individual traders, and
government organizations. Moreover, the FX market has become
increasingly competitive and unstable as circulating foreign currency-
denominated assets have become volatile worldwide (Chang et al.,
2013). The FX market is inuenced by many co-integrated, micro-
economic, macro-economic, political, and even psychological factors.
As a result, both modeling and forecasting the FX markets pose an
important challenge because the FX market behaves like a random
walk in a global economy (Ni and Yin, 2009). For several years,
researchers and investors have dedicated considerable eort to devel-
oping models for forecasting the FX markets, which is one of the more
important international monetary market indices (Meese and Rogo,
1983).
Researchers have determined that classical econometric and time
series models are inadequate to overcome the random walk eect
because market prices move in a purely random and unpredictable
manner and are partially based on unrealistic assumptions applied to
traditional methods (Kilian and Taylor, 2003). To resolve this limita-
combination with both linear and non-linear methods have been
studied in various markets (Adhikari, 2015; Bas et al., 2015). For
example, Chen et al. (2016) proposed a multi-factor time series model
based on an adaptive network-based fuzzy inference system (ANFIS)
for both the Taiwan Stock Exchange Capitalization Weighted Stock
Index (TAIEX) and Hang Seng Index (HSI) forecasting. These authors
selected a technical indicator using a stepwise regression method
combined with ANFIS to develop their forecasting model. Bas et al.
(2015) proposed a new hybrid forecasting model that combined an
autoregressive model and a fuzzy time series model with their proposed
model, which is evaluated using criteria such as the root mean square
error (RMSE) and mean absolute percentage error (MAPE) for the
TAIEX and Istanbul Stock Exchange markets. In addition, Chen et al.
(2014) developed a forecasting model that integrated fuzzy time series
models that based on tting functions to forecast the TAIEX and HSI.
Their model applied correlation coecients to select prominent input
variables, such as technical indicators. In general, a growing number of
studies combine statistical analysis and machine learning techniques
for application to the nancial market domain.
Several studies have demonstrated that in an inecient FX market,
machine learning techniques outperform linear models, such as the

Corresponding author.
E-mail addresses: sjlee@kw.ac.kr (S. Lee), enke@mst.edu (D. Enke), kimym@mst.edu (Y. Kim).

http://dx.doi.org/10.1016/j.engappai.2017.02.014
Received 26 September 2016; Received in revised form 14 February 2017; Accepted 27 February 2017
0952-1976/ 2017 Elsevier Ltd. All rights reserved.
S. Lee et al. Engineering Applications of Artificial Intelligence 61 (2017) 4756

autoregressive moving average and random walk models (Hann and
Steurer, 1996; Koskela et al., 1997; Kodogiannis and Lolis, 2002; Cao,
2003; Chen and Leung, 2004). Recently, trading system models for FX
have been developed based on trading rules that generate prot using
historical data. Ince and Trafalis (2006) analyzed the EUR/USD, GBP/
USD, JPY/USD, and AUD/USD exchange rates using multilayer
perceptron and support vector regression methods. Olson (2004)
studied a moving-average crossover system with 18 currencies between
1971 and 2000 and observed prots for the early part of the sample
period when assuming the normality of trading prots and ignoring
interest dierentials. Hirabayashi et al. (2009) introduced a forecasting
optimization model based on a genetic algorithm (GA) to automatically
create trading rules based on technical indicators (i.e., the moving
average and relative strength index). Deng et al. (2012) developed
trading rules in a GA model using technical indicators as input
variables and generated a prot in simulated trading with real FX
market data. However, these trading models do not consider the
correlations between FX nancial instruments, which are aected by
the interactions between currencies, as they only considered correla-
tions for constructing a well-diversied portfolio. This oversight is a
limitation. Given that a single FX determines the system payo, such
trading systems may yield unstable prots when the correlations are
ignored. Therefore, the relative value (RV) arbitrage must be an
investment strategy that exploits price dierentials between related
nancial instruments such as stocks, futures, options, and bonds by
simultaneously buying and selling the dierent securities and allowing
investors to potentially prot from the RV of the two securities as a
result. In this study, the RV is dened as the price of security A divided
by the price of security B.
In addition, a trading system based on data-driven approaches
considers data characteristics. The FX market is generally acknowl-
edged to be less ecient than the stock market due to Central Bank
intervention (Neely, 2002). Engel and Hamilton (1990) proved that
there is a long-term mean reverting pattern in the FX market. Chiang
and Jiang (1995) found that FX returns are highly positively correlated
over a short-time horizon (i.e., 1252 weeks) and negatively correlated
over a longer time horizon (i.e., 3 years). Similarly, Okunev and White
(2003) nd that a momentum strategy in the FX market is appropriate
for obtaining prot over the short-term. Notably, a momentum strategy
consists of buying currencies (winners) that have obtained high returns
and selling currencies (losers) that have yielded low returns in the
recent past (Serban, 2010). Thus, an FX trading system model must be
built that considers the characteristics of the data among FXs based on
their relative value.
This study proposes a conceptual framework for an RV-based trading
system (RVTS) that is designed to generate protability based on the
correlation and a rough set analysis of the FX futures market. The
proposed trading system employs a correlation analysis to form pairs
and a rough set analysis to generate trading rules. Rough set theory has
been successfully applied to the nancial domain (Lee et al., 2012; Kim
and Enke, 2016; Yeh et al., 2016) to identify potential information from
vague and uncertain data. Moreover, no studies of trading systems based
on integrating the correlation and a rough set analysis of the FX futures
market were found. This approach relies on rough set theory to generate
decision rules that determine buying, selling, and holding a position
when RV movements and certain pair formations occur. We conduct a
correlation analysis among several foreign exchange futures and select
pairs with a positive correlation (POC), negative correlation (NEC), or no
correlation (NOC). The RVTS generates trading rules using a rough set
The remainder of this paper is organized as follows. Section 2
briey introduces the concepts of the rough set, whereas Section 3
presents the construction procedure for the RVTS. Section 4 presents
an empirical study that was performed to verify the performance of the
RVTS. Finally, conclusions are presented in Section 5.
2. The basic concepts of the rough set
Rough set theory was introduced by Pawlak (1982) to address
uncertain, noisy, or incomplete information in various industry do-
mains (Wang, 2005; Shyng et al., 2007). Recently, various approaches
have implemented rough sets as a data mining tool (Liao and Chen,
2014; Jiang and Chen, 2015; Zhang et al., 2015). In general, rough set
theory utilizes indiscernibility relationships that allow examples to be
partitioned based on the application-specic denition of an equiva-
lence class (Chou et al., 2007). The partition follows two steps. In the
rst step, the data objects are classied into elementary sets, which are
also referred to as indiscernible sets. In the second step, the elementary
sets are inducted to generate the reducts. The rules generated from the
reducts provide the most precise method of detecting the data objects
and establishes the minimal subset, which includes the attributes of the
data objects. The classication based on this subset is the same as that
used for the universal set of attributes.
Rough set theory can assess the degree of approximation from
certain data objects, and the approximation set is called the boundary
set. To obtain a boundary set, the lower and upper approximations are
calculated with respect to the addressed data objects. Let U be a
nonempty set of objects, called the universal set, and let R be an
equivalence relation to U , called the indiscernibility relation, which is
based on a set of available attributes V (i.e., R V ). [x ]R is the
equivalence class of R containing object x for each x U . If X
represents a subset of U , the lower and upper approximations and
the boundary set of X under R are dened as follows:
RX = {x U [x ]R X} (1)
R X = {x U [x ]R X } (2)
RNB (X ) = R X RX (3)
Consider the following example for the initial classication result of
the stock price in Table 1, where V is an available set of attributes of U ,
including the 1st, 2nd, and 3rd indicator. Equivalent relations are
designated based on the value of the attributes in R . Let
X = {x : Signal (x ) = Buy}, as shown in Table 1, and the set X is
approximated by the set of conditional attributes
R = {1st indicator , 2nd indicator , 3rd indicator}. xt3 to xt6 belong to
the lower approximation of the signal set because all times with the
same equivalence relation are classied as signals. However, it is
impossible to classify the objects between xt1 and xt2 as the values
of the conditional attributes are equivalent, while the decision value is
dierent. As a result, the following approximations are obtained:
RX = {xt 3, xt 4 , xt 5, xt 6}, R X = {xt 1, xt 2 , xt 3, xt 4 , xt 5, xt 6}, and
RNB (X ) = {xt 1, xt 2}.
The decision rule introduced by Pawlak (2002) describes the
approximations and was represented in the form IF condition(s),
Table 1
Example of stock price attributes.
Time 1st indicator 2nd indicator 3rd indicator Signal

analysis with the technical indicators derived from the RV of the pairs xt6 1020 1530 520 Buy
and simulates an RV for trading. The performance of the RVTS is then xt5 1020 1530 520 Buy
compared to that of a simple momentum strategy and a buy-and-hold xt4 1020 1530 1030 Buy
trading strategy as benchmarks. This study also utilized an ANOVA xt3 1020 1530 1030 Buy
xt2 1020 3545 015 Buy
analysis to identify the performance of the RVTS in terms of correlation
xt1 1020 3545 015 Sell
strength and the training window width for generating trading rules in a xt 4060 5565 3545 Sell
given exchange rate data period.

48
S. Lee et al.
Fig. 1. The construction procedure of the proposed RVTS.
THEN decision(s) by Slownski et al. (1997). Please see Pawlak, (1982,
1997), Jackson et al. (1998), and Dimitras et al. (1999) for more
information on rough set theory.
3. Development of the RVTS
This section provides a detailed description of the framework,
which includes the three stages of the proposed RVTS. The rst stage
identies the pairs. The second stage generates the technical indicators
(i.e., the input variables) for the pairs. The third stage creates the
trading rules using rough set modeling. Fig. 1 presents the overall
structure of the RVTS.
3.1. Stage 1: formation of pairs
At this stage, the correlation coecient measures the strength and
direction of the correlation between the foreign exchange futures
{X1, X2, , Xn}. Pearson's product-moment correlation coecient, com-
monly denoted by r , is calculated by dividing the covariance of the two
currency futures by the product of their standard deviations; this
coecient varies from 1 to 1. A value of +1 (1) indicates a perfect
linear relationship with a positive (negative) slope between the two
futures, whereas a correlation coecient of 0 implies that there is no
linear relationship between the two futures. At the heart of the RV of
the RVTS is the notion that if the two futures are suciently correlated,
any changes in the correlation may be followed by a reversion to the
mean of the RV. Thus, the values of r between {X1, X2, , Xn} are
calculated using a correlation analysis, and a pair (i.e., a portfolio of
two futures) is determined by the strength and direction of the
correlation.
3.2. Stage 2: generation of prominent input data
The input data for each l = 1, , L relationship (with a dierent
strength and direction of r ) are generated. Assume that the L relation-
ships are dened by a distinct r for the pairs. For instance, in our
empirical study, L = 3 is used, where each L is characterized by its own
distinct correlation (i.e., POC, NEC, or NOC).
49
Engineering Applications of Artificial Intelligence 61 (2017) 4756
The input data are generated independently for each relationship,
and the training data for the lth relationship are created from the RV.
The RV is the closing price of one futures contract divided by the
closing price of the other futures contract. The application of dierent
technical indicators of I (that is, 1, , I ) is performed to obtain the RV
(see the Appendix A). The technical indicators used in this study are
oscillators, which are a type of leading indicator that moves up and
down within a price range. Because the oscillating indicator follows a
wavelike pattern, it yields a buy or sell signal before the start of either a
new pattern or a pattern reversal. Achelis (1995) demonstrated that
oscillating indicators perform better than other types of indicators. An
application of the trading rule designed for i to the RV for i = 1, , I
yields I return rates (please see the Appendix A for the specic trading
rules designed for each oscillating indicator). The return rate denotes
the annual prot rate based on Lee et al. (2010). Next, one can select
the best I0 (I ) from among the I indicators in terms of the calculated
return rate. Let (1) | l, , (I0) | l . (i ) | l denote the ith best indicator in terms
of protability when the trading rules of I indicators are applied to the
given training data. The input data are the ith best indicator values, and
the process is completed by repeating the above steps for l = 1, , L .
3.3. Stage 3: creation of the trading rules
3.3.1. Discretization of the data
In this step, the data transformation is generally conducted by
discretization, which essentially reduces the value set of the attribute.
Various discretization methods can be used, including equal frequency
binning, a naive algorithm, Boolean reasoning, and manual cuts (hrn,
1999). In our study, an equal frequency-binning method to transform
the input data was utilized. This method divides the data into k groups,
with each group containing approximately the same number of values.
3.3.2. Creation of the reducts
This step produces the reducts using the discretized data. The
creation of the reducts is an important step in the rough set analysis
because the core information of the discretized data is used to generate
the reducts, which are necessary for generating a specic rule. The
corresponding m technical indicators are randomly selected among I0
S. Lee et al.
Fig. 2. The movement of six futures (09/18/200109/20/2010).
indicators (1) | l, , (I0) | l to produce the corresponding reduct Rk
(k = 1, , m1). In this step, there are m1 reducts out of a possible
I0 !
reducts (for example, R1, , R m1). Reducts can be created using
m ! (I0 m ) !
a variety of techniques, including manual reducers, genetic algorithms,
the Johnson algorithm, and the dynamic reduct method (hrn, 1999).
In this study, the manual reducer method, which enables an analyst to
manually choose a subset from among the set of attributes, is employed
to generate all the possible reduct combinations from the I 0 most
protable oscillators.
3.3.3. Generation of the rules
In the nal step, the rules are developed for the given relationship.
Based on the reducts created in Step 2, the rules are expressed in an
IF-THEN form, which combines the condition values with the
decision values. To apply the generated decision rules, some factor(s),
such as a successive application of the rules or the number of positions
to hold, must be considered during trading. Thus, the following trading
rule is employed to limit the number of positions held to one contract.
IF the signal at time t is BUY (SELL) and the position is none, then
IF the 1st indicator value belongs to the range of the
corresponding indicator
AND the 2nd indicator value belongs to the range of the
corresponding indicator
AND the n th indicator value belongs to the range of the
corresponding indicator,
THEN BUY (SELL)
IF the signal at time t is BUY (SELL) and the position at time t 1
is SHORT (LONG), then
BUY (SELL), ELSE HOLD
The moving window method introduced by Jang et al. (1993) and
Hwarng (2001) is employed in the trading simulation to assess the
RVTS. The trading rules created for each training period are applied for
each testing period.
50
Engineering Applications of Artificial Intelligence 61 (2017) 4756
4. Empirical studies
4.1. Data description
The data were obtained from the CHECKExpert terminal of the
Korea Securities Computing Corporation (KOSCOM). To develop and
evaluate the RVTS, six foreign exchange futures contracts were selected
and their daily prices investigated. Specically, the data included the
Australian Dollar (AUD), British Pound (GBP), Canadian Dollar (CAD),
Japanese Yen (JPY), and Swiss Franc (CHF) futures contracts traded
on the Chicago Mercantile Exchange (CME), along with the US Dollar
(USD) futures contract traded on the Korea Exchange (KRX). The USD
was intentionally included because the USD futures contract shows
dierent movements than the other contracts (see Fig. 2). The strength
and direction of the relationship between two markets is determined by
the correlation coecient that reects the coinciding change in value
for a pair of numeric series over time. Table 2 provides the specica-
tions for the foreign exchange futures considered in this study.
Each foreign exchange futures contract can be traded for a specic
period, which ends on the maturity date. The maturity date is the
specic date on which the contract expires and the underlying assets
are delivered. The maturity date of the futures contracts is the third
Thursday of every March, June, September, and December. In this
study, the rst date of the period represents the initial date of the
futures contracts during the third futures contract period on Sept. 18,
2001. The last date of the period is the maturity date of the third
futures contract period on Sept. 20, 2010.
A moving-window scheme based on the maturity date was designed
to establish the training and testing periods. The training window
width increased incrementally as each of the four futures contract
periods elapsed over a three-month interval during the year (i.e., 3, 6,
9, and 12 months). The testing window width was xed at one specic
contract period (i.e., 3 months). The moving-window scheme needed to
be established to develop and assess the RVTS. Table 3 describes the
moving window scheme. The training window of each relationship
included 32 experimental sets.
4.2. RVTS experimental results
At Stage 1, the pairs X1, X2, , X6 were selected by the correlation
S. Lee et al. Engineering Applications of Artificial Intelligence 61 (2017) 4756

analysis. Table 4 presents the correlation coecients based on the daily Table 3
percentage yields calculated over the same overall experimental period. Moving window scheme and the experimental set for the RVTS.
The long time scales provide an obvious theoretical benet because
Starting date Training Testing Ending dates Number of
averaging over long periods restrains windowing errors and noise, thus of the window window of the testing experimental sets
revealing underlying long-term correlations (Katsanos, 2008). As training set width width set
shown in Table 4, the AUD had a high positive correlation with the
Sept. 18, 3 months 3 months Dec. 21, 2009 32
CAD (r = 0.955), whereas the GBP had a high negative correlation with
2001 6 months Mar. 15,
the USD (r = 0.835). The JPY had a low correlation with the CAD 2010
(r = 0.016 ) and the AUD (r = 0.175). AUD-CAD, GBP-USD, and CAD- 9 months Jun. 21,
JPY were formatted as pairs and designated as POC, NEC, and NOC, 2010
respectively, for Stage 2. 12 months Sept. 20,
2010
In Stage 2, the prominent input data were created. The application
of I = 8 indicators to each relationship (i.e., the application of 1, , 8 in
the Appendix A with its own trading rule) allowed us to choose the
I0 = 5 best indicators from 1, , 8 in terms of prot. Tables 5 and 6
provide the ve indicators of the two best trading performances of the Table 4
RVTS among all experiments (i.e., four types of the training window Pearson correlation analysis results.
widths (3, 6, 9, and 12 months) * three types of the FX relationships
USD CHF JPY GBP CAD AUD
(POC, NEC, and NOC)=12 experiments). For instance, we obtained
(1) |1 = ROC (13.6%), (2) |1 = Momentum (11.7%), (3) |1 = TRIX (11.6%), USD 1 0.284 0.465 0.835 0.573 0.426
(4) |1 = Band %b (10.3%), and (5) |1 = MACD (6.9%) for l = 1 (POC) when CHF 1 0.453 0.603 0.76 0.844
JPY 1 0.281 0.016 0.175
the training window width was 3 months (see Table 5). We obtained
GBP 1 0.746 0.645
the training window width for each relationship from the input data CAD 1 0.955
(the ve best technical indicator values). AUD 1
A set of decision rules for the training window width for each
relationship was generated from the input data using a rough set
analysis (Stage 3). To this end, m = 3(I0 = 5) indicators were ran-
5!
domly selected to create m1 = 10( 3 ! 2 ! = 10) reducts using the manual NOC in Table 7(c), the return rate was almost 0, regardless of the
reducer method for the training window width of each relationship. In window width, indicating that the level of the correlation coecient
turn, a set of rules was produced in the form of trading rules. ROSETTA and training window width for generating trading rules should be
software (hrn, 1999) was used to implement this stage. considered when developing the RVTS.
We were able to measure the return rates of the RVTS for each To construct the momentum strategy portfolio, it was necessary to
relationship by applying the moving-window method over the entire select each winner and loser from among foreign currencies over the
period. We dened an initial cash balance for the initial futures value of previous 3- and 9-month periods. The buy-and-hold strategy is
the testing period. This study considered an RVTS for real-world implemented using an equal weighted method for the POC (AUD-
transaction costs and the slippage of the foreign exchange futures. As CAD) and NEC (GBP-USD). For the buy-and-hold strategy, the average
the default settings for trading, the initial cash position is $100,000, return rate and Sharpe ratio are 2.21% and 0.32 for the POC and 0.40%
the transaction cost is 3 pips, and the slippage is 0.5 ticks per trade. and 0.19 for the NEC, respectively (see Table 8). According to the
Table 7 compares the return rates, standard deviation (SD), and previous 3- and 9-month periods, momentum strategies yield average
Sharpe ratio (Sharpe, 1994) of the testing periods of the RVTS using an return rates of 1.63% and 0.83%, respectively. The 3- and 9-month
increasing training window width for each relationship. To select the Sharpe ratios obtained are 0.36 and 0.14, respectively (see Table 8).
training window width in a practical manner, the estimated Sharpe The best results of the RVTS (i.e., POC and NEC) are compared to the
ratio was used to evaluate the performance of the RVTS for the window momentum and buy-and-hold strategies. These results yield Sharpe
width. For the POC in Table 7(a), the 3-month training window width ratios that are approximately 1.31.8 times higher for the RVTS than
generated the highest average return rate (4.56%) and Sharpe ratio for the momentum and buy-and-hold strategies. As a result, the RVTS
(0.59). For the NEC in Table 7(b), the 9-month training window width outperforms the benchmark strategies, both in terms of the average
yielded the highest average return rate (3.37%) and Sharpe ratio (0.50), return and Sharpe ratio.
but this average return rate was lower than that obtained for the POC. An analysis of variance (ANOVA) was performed to estimate the
Thus, the results demonstrate that a 3-month training window width is error variance and to determine the relative importance of the 2-
appropriate for generating trading rules, where a POC between foreign dimensional factors, i.e., the types of correlation+training window
exchange futures can generate the highest return rate. However, for the width. A 12 (34) full-factorial design was considered with the above

Table 2
Specifications of FX futures on the CME and KRX.

Trade Unit Minimum Price Fluctuation (Tick) Contract Months

a
GBP/USD 62,500 GBP $0.0001 per GBP ($6.25) 1st 6 months in March quarterly cycle (March, June, September and December)
CHF/USDa 125,000 CHF $0.0001 per CHF ($12.50)
CAD/USDa 100,000 CAD $0.0001 per CAD ($10)
AUD/USDa 100,000 AUD $0.0001 per AUD ($10)
JPY/USDa 12,500,000 JPY $0.000001 per JPY ($12.50)
USD/WONb 10,000 USD 1000 per USD ($1.0)

a
CME Group (http://www.cmegroup.com).
b
KRX (http://www.krx.co.kr).

51
 Table 5
The five most profitable indicators and their return rates (%) for POC (AUD-CAD) using a 3-month trading window width.
S. Lee et al.

No. set I0 No. set I0

1 ROC (13.6) Mom (11.7) TRIX (11.6) Band%b (10.3) MACD (6.9) 17 TRIX (11.8) MACD (2.7) RSI (9.8) PO (10.6) Band%b (13.2)
2 RSI (31.1) Band%b (21) MACD (0) TRIX (0.0) ROC (5.4) 18 TRIX (34) ROC (22.6) Mom (22.6) PO (22) Band%b (22.0)
3 MACD (23.8) PO (22.1) TRIX (7.4) RSI (1.3) Band%b (9.4) 19 NCO (16.5) Band%b (16.5) PO (9.4) RSI (9.4) MACD (3.8)
4 RSI (23.5) Band%b (16) TRIX (0.0) NCO (11.1) PO (12.9) 20 RSI (30.4) PO (28.7) MACD (28.5) NCO (25.1) ROC (25.1)
5 MACD (31.3) PO (31.3) Band%b (30.8) RSI (21) TRIX (0) 21 TRIX (28.2) Band%b (22.9) RSI (16.2) PO (17.9) MACD (19.3)
6 Band%b (80.4) NCO (74.4) ROC (74.4) Mom (74.4) MACD (71.7) 22 Band%b (38.5) TRIX (20.7) RSI (8.3) PO (11.1) NCO (15.6)
7 TRIX (26.3) RSI (1.3) Band%b (0.6) ROC (9.7) Mom (9.7) 23 TRIX (49.4) NCO (42.9) PO (24.8) ROC (17.1) Mom (17.1)
8 MACD (76.5) PO (76.5) Band%b (70) RSI (57.6) ROC (9) 24 NCO (35.1) MACD (30.7) TRIX (15.6) Band%b (13.3) ROC (6.0)
9 ROC (11.9) Mom (11.9) MACD (10.6) PO (10.6) NCO (2) 25 Band%b (63.3) RSI (53.0) NCO (43.8) ROC (41.0) Mom (41.0)
10 Band%b (41.2) NCO (43.1) TRIX (43.1) ROC (53.6) Mom (53.6) 26 PO (42.9) ROC (22.5) Band%b (22.5) Mom (22.5) RSI (12.0)
11 NCO (18.6) ROC (18.6) Mom (18.6) Band%b (4.2) MACD (6.2) 27 ROC (97.2) Mom (97.2) PO (92.6) NCO (87.0) MACD (78.0)
12 MACD (56.2) PO (55.7) RSI (55.7) Band%b (53) NCO (3.8) 28 ROC (61.9) Mom (61.9) NCO (57.4) PO (46.6) TRIX (0.0)
13 MACD (23.3) PO (23.3) Band%b (22.2) RSI (22.2) ROC (15.4) 29 PO (17.9) TRIX (15.2) MACD (12.2) ROC (4.9) Mom (4.9)
14 Band%b (18.2) ROC (13.5) Mom (13.5) TRIX (13.2) MACD (10.6) 30 Band%b (135.2) RSI (99.7) ROC (10.5) Mom (10.5) NCO (24.8)
15 RSI (8.6) ROC (6.6) Mom (6.6) NCO (6.3) Band%b (7.5) 31 RSI (139.3) TRIX (104) Band%b (102.1) MACD (33.6) PO (47.1)
16 NCO (10.6) MACD (1.9) PO (0.6) TRIX (0.0) ROC (6.4) 32 RSI (6.4) MACD (19.8) TRIX (31.7) Band%b (41) NCO (69.1)

52
Table 6
The five most profitable indicators and their return rates (%) for NEC (GBP-USD) using a 9-month trading window width.

No. set I0 No. set I0

1 Band%b (72.1) TRIX (71.8) NCO (71.0) ROC (70.7) Mom (70.7) 17 ROC (52.7) Mom (52.7) NCO (19.7) TRIX (19.4) MACD (6.6)
2 Band%b (45.2) RSI (40.2) MACD (7.9) TRIX (4.0) PO (0.4) 18 NCO (25.9) ROC (25.9) Mom (25.9) Band%b (24.0) RSI (10.9)
3 MACD (55.8) RSI (48.3) PO (47.9) Band%b (38.3) TRIX (23.4) 19 NCO (69.0) ROC (69.0) Mom (69) Band%b (68.4) RSI (58.4)
4 ROC (66.1) Mom (66.1) TRIX (53.8) RSI (47.7) Band%b (40.4) 20 MACD (76.3) PO (70.0) Band%b (51.1) RSI (41.9) TRIX (11.5)
5 RSI (135.1) Band%b (133.5) TRIX (0.0) ROC (22.9) Mom (22.9) 21 ROC (108.7) Mom (108.7) TRIX (107.9) PO (100) RSI (100.0)
6 NCO (114.2) Band%b(113.6) RSI (101.6) TRIX (18.7) ROC (33.7) 22 Band%b (59.8) RSI (59.8) TRIX (57.1) ROC (25.6) Mom (25.6)
7 Band%b (158.5) NCO (153.7) ROC (153.7) Mom (153.7) RSI (142.4) 23 MACD (61.5) PO (58.0) ROC (53.0) Mom (53) Band%b (51.1)
8 TRIX (104.9) Band%b (86.8) RSI (76.1) MACD (13.8) PO (36.9) 24 TRIX (132.0) Band%b (79.8) RSI (66.8) MACD (65.5) ROC (47.7)
9 Band%b (32.3) RSI (19.8) MACD (0.6) PO (26.8) ROC (48.3) 25 Band%b (112.7) RSI (105.4) MACD (35.3) ROC (3.6) Mom (3.6)
10 ROC (14.3) Band%b (14.3) Mom (14.3) NCO (19.4) RSI (35.3) 26 NCO (93.0) ROC (91.8) Mom (91.8) MACD (77.1) TRIX (65.7)
11 Band%b (44.8) TRIX (30.8) RSI (28.6) MACD (21.6) PO (21.6) 27 NCO (211.0) ROC (211.0) TRIX (211.0) Mom (211) RSI (219.5)
12 ROC (103) Mom (103.0) Band%b (102.3) NCO (95.6) RSI (95.6) 28 MACD (233.1) PO (187.1) ROC (175.9) Mom (175.9) NCO (121.3)
13 PO (88.5) Band%b (88.5) RSI (88.5) MACD (85.8) TRIX (10.9) 29 MACD (119.3) ROC (28.8) Mom (28.8) PO (10.1) NCO (0.0)
14 NCO (16.2) ROC (16.2) Band%b (16.2) Mom (16.2) TRIX (16.0) 30 MACD (261.1) RSI (254.5) PO (251.6) Band%b (251.6) NCO (146.9)
15 NCO (43.5) TRIX (3.4) ROC (10.6) Mom (10.6) MACD (11.4) 31 Band%b (231.8) RSI (207.0) TRIX (74.4) MACD (167.2) NCO (167.4)
16 ROC (48.6) Mom (48.6) NCO (44.0) TRIX (5.3) RSI (31.3) 32 Band%b (125) ROC (122.6) Mom (122.6) TRIX (116.0) RSI (7.7)
Engineering Applications of Artificial Intelligence 61 (2017) 4756
S. Lee et al. Engineering Applications of Artificial Intelligence 61 (2017) 4756

Table 7 Table 7 (continued)

Average return rates (%) and Sharpe ratios for the RVTS.
No. set Training window widths
No. set Training window widths
3 months 6 months 9 months 12 months
3 months 6 months 9 months 12 months Return rates Return rates Return rates Return rates
Return rates Return rates Return rates Return rates
SD 8.70 8.17 6.74 8.36
(a) POC (AUD-CAD) Sharpe ratio 0.23 0.10 0.50 0.20
1 16.77 7.75 4.25 4.11
2 9.06 8.11 1.21 4.55 (c) NOC (CAD-JPY)
3 10.92 0.11 0.35 12.14 1 0.61 0.76 1.08 0.05
4 2.12 3.34 6.31 2.85 2 1.21 1.13 2.82 1.47
5 8.13 9.11 6.03 9.36 3 0.01 0 0.39 0
6 13.03 10.68 14.52 0.89 4 0 0 3.34 0.67
7 5.40 6.95 5.60 0.10 5 0.67 3.35 0.11 2.54
8 15.24 0.72 0.99 0.05 6 0.38 0.02 0.42 0.47
9 1.914 3.38 0.01 4.01 7 0.63 1.98 0.67 1.29
10 16.36 1.63 8.97 1.32 8 2.82 0.80 0.15 1.97
11 5.36 8.27 4.60 1.15 9 0.79 0.69 5.69 0.01
12 13.73 5.07 4.02 2.21 10 0 0 1.79 0.02
13 3.25 4.24 1.36 0.51 11 0.30 2.08 0.10 1.27
14 8.36 4.08 1.85 0.89 12 1.79 0 0 0
15 9.37 3.95 3.34 3.23 13 0.14 0 4.13 1.72
16 6.15 6.68 0.23 1.60 14 0 0 0.08 1.13
17 2.95 5.13 0.77 1.60 15 5.03 0.68 0 0.81
18 13.48 4.89 4.02 7.60 16 0.66 1.15 1.21 0.08
19 2.64 3.33 0.84 0.16 17 0.43 0.49 0 1.10
20 1.02 3.95 9.16 4.07 18 0.81 0.12 1.10 1.04
21 3.41 5.40 2.47 4.98 19 1.30 1.31 0.98 3.07
22 3.07 4.30 4.63 0.30 20 1.10 0.56 3.03 1.38
23 10.12 0.09 0.05 4.05 21 0.55 3.03 0.04 1.74
24 10.2 0.83 4.03 9.97 22 10.58 1.38 1.30 2.32
25 16.65 5.11 3.25 0.34 23 2.19 0 0.44 1.47
26 7.77 25.9 14.35 3.88 24 1.80 0.44 2.98 0
27 2.91 14.11 12.06 13.05 25 0.12 1.17 5.08 5.40
28 7.94 23.2 3.43 4.48 26 1.48 0 0 0
29 1.33 10.11 1.28 0.32 27 0 0 0 0
30 6.63 6.16 7.71 0.60 28 0 0 0 0
31 1.99 3.00 5.12 8.21 29 0 4.68 0.20 0
32 4.48 5.65 13.57 8.77 30 14.02 2.61 0.88 2.39
Average 4.56 2.36 2.63 1.37 31 1.67 0 0.67 0
SD 7.74 8.19 5.71 5.08 32 0 0 0 0.56
Sharpe ratio 0.59 0.29 0.46 0.27 Average 1.07 0.14 0.34 0.18
SD 3.21 1.45 1.94 1.58
(b) NEC (GBP-USD) Sharpe ratio 0.33 0.10 0.18 0.11
1 1.56 2.60 19.35 2.02
2 0.03 8.61 0.43 0.39
3 1.94 4.73 0.50 3.74
4 3.03 3.47 3.64 14.70 two factors and multiple levels for each factor. To evaluate the eects of
5 4.14 0.87 1.78 8.27 the two factors and their cross eects, a multi-factor ANOVA was
6 1.18 1.91 6.60 20.63
conducted with the following three hypotheses:
7 2.49 0.70 1.48 4.21
8 1.23 10.24 20.35 14.25
9 1.05 6.94 1.77 4.02
10 17.89 0.64 4.96 4.78 H0 (1): All three types of correlations result in equal performance.
11 10.41 7.83 0.29 8.06 H0 (2): All four training window widths result in equal performance.
12 4.18 0.08 1.09 2.65 H0 (3): There are no interactions between the types of correlations
13 0.92 6.61 8.85 3.43
and the training window widths.
14 0.50 5.91 0.64 4.19
15 4.81 0.13 12.60 6.29 The experimental results were analyzed using a full-factorial mixed-
16 3.78 11.56 3.50 2.15
17 27.51 8.68 2.50 4.07
model ANOVA in SPSS for Windows (Release 18.0), as shown in
18 9.26 2.80 3.51 1.01 Table 9. The analysis indicated that the type of correlation (A) has a
19 6.19 9.25 6.17 0.21 signicant eect on the performance at a 5% level of signicance. There
20 1.04 0.22 1.13 0.70 was no interaction between the type of correlation (A) and the training
21 5.91 2.19 1.85 4.18
window width (B).
22 7.59 1.36 0.22 2.61
23 2.38 0.71 1.97 6.26
24 0.65 1.41 0.30 4.49
25 3.87 1.15 3.54 18.07 5. Concluding remarks
26 3.69 5.37 10.31 19.68
27 19.49 2.00 11.84 9.41 In this study, we proposed an RVTS that focuses on the data
28 15.36 21.36 13.10 1.06 characteristics of the FX futures markets using correlation and rough
29 3.91 29.47 12.30 13.12
30 11.90 3.72 6.22 5.41
set analysis. Selected pairs were applied to the RVTS based on the level
31 5.16 7.63 1.77 4.52 of their correlation coecients. The RVTS generated trading rules
32 14.88 0.25 0.64 8.34 through a rough set analysis and using technical indicators derived
Average 2.04 0.82 3.37 1.70 from the RVs of the selected pairs. The performance of the RVTS was
measured by its protability that considered the training window width

53
S. Lee et al. Engineering Applications of Artificial Intelligence 61 (2017) 4756

Table 8
Average return rates (%) and Sharpe ratios for the benchmark strategies.

No. set Momentum strategy Buy-and-hold strategy

3 months 9 months POC (AUD-CAD) NEC (GBP-USD)

Return rates Return rates Return rates Return rates

1 3.13 1.81 1.62 0.31

2 3.78 4.16 7.47 0.90
3 0.51 2.75 2.35 1.34
4 1.28 12.54 4.23 1.41
5 2.75 0.89 6.70 0.78
6 5.76 7.20 15.32 1.91
7 0.01 1.46 0.75 3.16
8 3.78 2.96 9.93 7.56
9 2.47 2.48 0.18 0.55
10 1.18 8.69 4.59 0.43
11 2.83 1.62 4.43 1.15
12 2.07 0.15 8.44 0.94
13 4.10 1.19 2.58 2.80
14 0.15 0.65 1.06 1.61
15 0.12 2.19 1.54 0.17
16 1.97 1.61 1.82 1.51
17 0.61 0.34 0.73 2.56
18 2.11 3.30 3.67 2.40
19 1.20 1.05 1.16 1.09
20 3.19 1.06 1.08 0.28
21 1.05 3.54 0.58 0.02
22 4.74 0.25 7.05 0.07
23 3.54 4.25 0.86 0.28
24 0.11 0.57 1.34 0.48
25 3.31 2.95 4.72 4.42
26 1.36 17.97 0.22 0.71
27 6.17 6.18 5.73 3.11
28 18.39 16.54 18.93 0.42
29 2.40 2.07 4.81 0.87
30 3.54 1.61 17.38 2.78
31 12.13 1.33 13.73 1.65
32 1.61 6.14 2.40 2.21
Average 1.63 0.83 2.21 0.40
SD 4.49 5.74 6.69 2.14
Sharpe ratio 0.36 0.14 0.32 0.19

designed for the stock index futures market. Previously, the trading
Table 9 strategies applied in one market have not been successfully applied in
Multi-factor ANOVA results for the test problem. other markets since there are fundamental or data characteristic
dierences among the various markets (Serban, 2010). To overcome
Source of Sum of Degrees of Mean F P-values
variation squares freedom square this limitation, the proposed trading system reects the data char-
acteristic of the FX market by applying the correlation analysis for pair
Model 3747.334 12 312.278 4.332 0.000 formation and technical indicators based on RV for a rough set analysis
Types of 874.707 2 437.354 6.067 0.003
instead of the originally designed technical indicators. According to the
correlation (A)
Training window 544.255 3 181.418 2.517 0.058
overall results, it can be concluded that the RVTS may be used as a
widths (B) complementary trading system in the FX market by constructing a
AB interaction 583.646 6 97.274 1.349 0.234 portfolio combined with both a short- and long-term investment
Error 26815.617 372 72.085 strategy. Even though this strategy does not always guarantee the best
Total 30562.951 384
trading results, it can help traders and investors in this eld consider
an alternative perspective when developing a trading system or a
of each relationship. In an empirical study, we found that co-movement decision support system.
pairs are appropriate for short periods (e.g., 3 months) when extracting The limitations of this study mainly inhere in the fact that the
trading rules, whereas decoupled pairs are appropriate for long periods correlation analysis was applied to identify pairs throughout the overall
(e.g., 9 months) when applying the RVTS to the FX futures market. experiment period. Moreover, although there are many other technical
Thus, negatively correlated pairs should be considered for RV trading indicators that are available to generate trading rules, the proposed
along with positively correlated pairs. Using an ANOVA, we veried system only considered eight technical indicators. The current work
that the dierences depended on the strength and direction of the can be extended in future studies. For instance, an alternative method
correlations. might be considered, such as a cointegration analysis to the RVTS to
This study contributes to the literature by not only proposing a new identify pairs dynamically. The proposed system might also be
trading system based on the relative value in the FX futures market but improved by considering an existing forecasting model or hybrid
also by applying a correlation analysis and rough set trading system system, as presented in the Introduction.

54
S. Lee et al. Engineering Applications of Artificial Intelligence 61 (2017) 4756

Appendix A. Technical Indicators (Oscillators)

Notes: The closing value is RV, which is calculated as the closing price of one futures contract divided by the closing price of another futures
contract. Ct is the day's closing value, and Ct1 is the previous day's closing value. n represents the period, with s representing the short-term and l
representing the long-term.

Technical Indicators Formula Trading Rules

NCO NCOt (n ) = Ct Ct n If NCOt (12) crosses over 0, then

Buy
(Net Change Oscillator) If NCOt (12) crosses under 0,
then Sell

ROC ROCt (n ) = ( C
Ct
1) 100 If ROCt (12) crosses over 0, then
tn
Buy
(Rate of Change) If ROCt (12) crosses under 0,
then Sell

PO POt (s, l ) =
MAt (s ) MAt (l ) If POt (5, 10) crosses over 0,
MAt (s )
then Buy
(Price Oscillator) where MAt (s ) is the moving average of s and MAt (l ) is the moving average of l . If POt (5, 10) crosses under 0,
then Sell

Band %b Band %bt (n ) =

Ct L t (n ) If Band %bt (9) crosses over
Ut (n ) L t (n )
Band %bt1 (9), then Buy
in=1 (Ct n + i Ct (n ))2

Ut (n ) = Ct (n ) +
n
n
i =1 (Ct n + i Ct (n )) 2 If Band %bt (9) crosses under
L t (n ) = Ct (n ) Band %bt1 (9), then Sell
n
where is a pre-assigned number.

TRIX EMAt3 (C , n ) EMAt31 (C , n ) If TRIXt (18) crosses over 0, then

TRIXt (n ) =
EMAt31 (C , n ) Buy
3
where EMAt3 is the day's triple exponential moving average, and EMAt1 is the If TRIXt (18) crosses under 0,
previous day's triple exponential moving average. then Sell

Momentum Momt (n ) =
Ct
100 If Momt (10) crosses over 0, then
Ct n
Buy
(Mom) If Momt (10) crosses under 0,
then Sell

RSI RSIt (n ) = 100

100 If RSIt (14) crosses over 70, then
1 + RSt (n )
n n
Buy
RSt (n ) = i =1 Ut n + i / i =1 Dt n + i

(Relative Strength Index) C Ct 1, Ct Ct 1, C Ct , Ct Ct 1, If RSIt (14) crosses under 30,

Ut = t Dt = t 1 then Sell
0, otherwise 0, otherwise

MACD MACDt (m, n) = EMAt (C , m ) EMAt (C , n ) If MACDt (12, 26) crosses over
0, then Buy
(Moving Average If MACDt (12, 26) crosses under
Convergence- 0, then Sell
Divergence)

References
Achelis, S.B., 1995. Technical Analysis from A to Z. Probus Publishing, Chicago.
Adhikari, R., 2015. A mutual association based nonlinear ensemble mechanism for time
series forecasting. Appl. Intell. 43, 233250.
Bas, E., Egrioglu, E., Aladag, C.H., Yolcu, U., 2015. Fuzzy-time-series network used to
forecast linear and nonlinear time series. Appl. Intell. 43, 343355.
Cao, L., 2003. Support vector machines experts for time series forecasting.
Neurocomputing 51, 321339.
Chang, C., Gonzlez-Serrano, L., Jimenez-Martin, J., 2013. Currency hedging strategies
using dynamic multivariate GARCH. Math. Comput. Simul. 94, 164182.
Chen, A., Leung, M.T., 2004. Regression neural network for error correction in foreign
exchange forecasting and trading. Comput. Oper. Res. 31, 10491068.
Chen, Y., Cheng, C., Chiu, C., Huang, S., 2016. A study of ANFIS-based multi-factor time
series models for forecasting stock index. Appl. Intell. 45, 277292.
Chen, Y., Cheng, C., Tsai, W., 2014. Modeling tting-function-based fuzzy time series
patterns for evolving stock index forecasting. Appl. Intell. 41, 327347.
Chiang, T.C., Jiang, C.X., 1995. Foreign exchange returns over short and long horizons.
Int. Rev. Econ. Financ. 4, 267282.
Chou, H., Cheng, C., Chang, J., 2007. Extracting drug utilization knowledge using self-
organizing map and rough set theory. Expert. Syst. Appl. 33, 499508.

55
S. Lee et al.
Deng, S., Sun, Y., Sakurai, A., 2012. Robustness test of genetic algorithm on generating
rules for currency trading. Procedia Comput. Sci. 13, 8698.
Dimitras, A.I., Slowinski, R., Susmaga, R., Zopounidis, C., 1999. Business failure
prediction using rough sets. Eur. J. Oper. Res. 114, 263280.
Engel, C., Hamilton, J., 1990. Long swings in the dollar: are they in the data and do
markets know it? Am. Econ. Rev. 80, 689713.
Hann, T.H., Steurer, E., 1996. Much ado about nothing? Exchange rate forecasting:
neural networks vs. linear models using monthly and weekly data. Neurocomputing
10, 323339.
Hirabayashi, A., Aranha, C., Iba, H., 2009. Optimization of the trading rule in foreign
exchange using genetic algorithm. In: Proceedings of the 11th Annual Conference of
Genetic and Evolutionary Computation. ACM, New York, pp. 15291536.
Hwarng, H.B., 2001. Insights into neural-network forecasting of time series
corresponding to ARMA(p,q) structures. Omega 29, 273289.
Ince, H., Trafalis, T.B., 2006. A hybrid model for exchange rate prediction. Decis.
Support. Syst. 42, 10541062.
Jackson, A.G., Pawlak, Z., LeClair, S.R., 1998. Rough sets applied to the discovery of
materials knowledge. J. Alloy. Compd. 279, 1421.
Jang, G.S., Lai, F., Jiang, B.W., Parng, T.M., Chien, L.H., 1993. Intelligent stock trading
system with price trend prediction and reversal recognition using dual-module
neural networks. Appl. Intell. 3, 225248.
Jiang, F., Chen, Y., 2015. Outlier detection based on granular computing and rough set
theory. Appl. Intell. 42, 303322.
Katsanos, M., 2008. Intermarket trading strategies. Wiley, Chichester.
Kilian, L., Taylor, M.P., 2003. Why is it so dicult to beat the random walk forecast of
exchange rates? J. Int. Econ. 60, 85107.
Kim, Y., Enke, D., 2016. Developing a rule change trading system for the futures market
using rough set analysis. Expert. Syst. Appl. 59, 165173.
Kodogiannis, V., Lolis, A., 2002. Forecasting nancial time series using neural network
and fuzzy system-based techniques. Neural Comput. Appl. 11, 90102.
Koskela, T., Varsta, M., Heikkonen, J., Kaski, K., 1997. Time series prediction using
recurrent SOM with local linear models. Research reports B15. Helsinki University of
Technology, Laboratory of Computational Engineering.
Lee, S.J., Ahn, J.J., Oh, K.J., Kim, T.Y., 2010. Using rough set to support investment
strategies of real-time trading in futures market. Appl. Intell. 32, 364377.
Lee, S.J., Oh, K.J., Kim, T.Y., 2012. How many reference patterns can improve
56
Engineering Applications of Artificial Intelligence 61 (2017) 4756
protability for real-time trading in futures market? Expert. Syst. Appl. 39,
74587470.
Liao, S., Chen, Y., 2014. A rough set-based association rule approach implemented on
exploring beverages product spectrum. Appl. Intell. 40, 464478.
Meese, R.A., Rogo, K., 1983. Empirical exchange rate models of the seventies: do they
t out of sample? J. Int. Econ. 14, 324.
Neely, C.J., 2002. The temporal pattern of trading rule returns and exchange rate
intervention: intervention does not generate technical trading prots. J. Int. Econ.
58, 211232.
Ni, H., Yin, H., 2009. Exchange rate prediction using hybrid neural networks and trading
indicators. Neurocomputing 72, 28152823.
hrn, A., 1999. Discernibility and Rough Sets in Medicine: Tools and Applications (Ph.D.
Thesis). Norwegian University of Science and Technology, ISBN 82-7984-014-1.
Okunev, J., White, D., 2003. Do momentum-based strategies still work in foreign
currency markets? J. Financ. Quant. Anal. 38, 425447.
Olson, D., 2004. Have trading rule prots in the currency markets declined over time? J.
Bank. Financ. 28, 85105.
Pawlak, Z., 1982. Rough sets. Int. J. Comput. Inf. Sci. 11, 341356.
Pawlak, Z., 2002. Rough sets and intelligent data analysis. Inf. Sci. 147, 112.
Pawlak, Z., 1997. Rough set approach to knowledge-based decision support. Eur. J. Oper.
Res. 99, 4857.
Serban, A.F., 2010. Combining mean reversion and momentum trading strategies in
foreign exchange markets. J. Bank. Financ. 34, 27202727.
Sharpe, W.F., 1994. The Sharpe ratio. J. Portf. Manag. 21, 4958.
Shyng, J., Wang, F., Tzeng, G., Wu, K., 2007. Rough set theory in analyzing the attributes
of combination values for the insurance market. Expert. Syst. Appl. 32, 5664.
Slownski, R., Zopounidis, C., Dimitras, A.I., 1997. Prediction of company acquisition in
Greece by means of the rough set approach. Eur. J. Oper. Res. 100, 115.
Wang, F., 2005. On acquiring classication knowledge from noisy data based on rough
set. Expert. Syst. Appl. 29, 4964.
Yeh, C., Chi, D., Lin, T., Chiu, S., 2016. A hybrid detecting fraudulent nancial statements
model using rough set theory and support vector machines. Cyber. Syst. 47,
261276.
Zhang, B., Min, F., Ciucci, D., 2015. Representative-based classication through
covering-based neighborhood rough sets. Appl. Intell. 43, 840854.