JID: INS

ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

Information Sciences xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Information Sciences
journal homepage: [Link]/locate/ins

An intelligent pattern recognition model for supporting

investment decisions in stock market
Q1 Tai-liang Chen a,∗, Feng-yu Chen b
a
Department of Digital Context Application and Management, Wenzao Ursuline College of Languages, 900 Mintsu, 1st Road, Kaohsiung
807, Taiwan, ROC
b
Computer and Network Center, National Kaohsiung University of Applied Science, Chien Kung Campus 415 Chien Kung Road, Kaohsiung
807, Taiwan, ROC

a r t i c l e i n f o a b s t r a c t

Article history: For many years, how to make stock market predictions has been a prevalent research topic.
Received 10 December 2012 To carry out accurate forecasting, stock analysts and academic researchers have tried var-
Revised 2 April 2015
ious analysis techniques, algorithms, and models. For example, "technical analysis” is a
Accepted 29 January 2016
popular approach used by common stock investors to analyze market trend, and Artifi-
Available online xxx
cial Intelligence (AI) algorithms such as genetic algorithms (GAs), neural network (NN),
Keywords: and fuzzy time-series (FTS), were proposed by researchers to forecast the future stock in-
Stock market forecasting dex. Although the daily forecasts are very useful for professional investors who implement
Perceptually important point (PIP) intraday trading, we argue that forecasting a bullish turning point is a more interesting
identification matching issue than the future stock index for common investor because an accurate forecast will
Template matching technique bring a huge amount of stock return. Therefore, this paper proposes an intelligent pattern
Chart pattern recognition model, based on two new stock pattern recognition methods, “PIP bull-flag
Technical indicator pattern matching” and the “floating-weighted bull-flag template,” to recognize a bull-flag
stock pattern. The bull-flag pattern is a stock’s turning point with proper timing, which
can enable a stock investor to profit. To promote recognition accuracy, the proposed model
employs chart patterns and technical indicators, simultaneously, as pattern recognition fac-
tors. In the model verification, we evaluate the proposed model with stock returns by fore-
casting two stock databases (TAIEX and NASDAQ), and comparing the returns with other
advanced algorithms. The experimental results indicate that the proposed model outper-
forms the published algorithms, such as rough set theory (RST), genetic algorithms (GAs)
and their hybrid model, and gives a high-level of profitability. Additionally, the trading
strategies, provided by the proposed model, also help investors to make beneficial invest-
ment decisions in the stock market.
© 2016 Published by Elsevier Inc.

1 1. Introduction

2 Stock market is a very complex and changeable system influenced by many factors such as economic environment, po-
3 litical policy, industrial development, and market news, etc. To make profit from the capital market, stock investor has been
4 looking for right tools and techniques to analyze stock market. As we known, the technical analysis [9] is a popular ap-
5 proach used by investment professionals and common investors. The principles of technical analysis are derived from the

Q2 ∗
Corresponding author: Tel.: +886 0920975168; fax: +886 073425785.
E-mail address: 97007@[Link], [Link]@[Link] (T.-l. Chen).

[Link]
0020-0255/© 2016 Published by Elsevier Inc.

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

2 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

Fig. 1. Distance measure methods for PIP identification: (a) Euclidean distance based: PIP-ED, (b) perpendicular distance based: PIP-PD and (c) vertical
distance based: PIP-VD [11].

6 observations of stock markets over hundreds of years and many evidences have shown that the technical analysis can help
7 stock investor to make right judgment on stock market [2,3,19,21,22].
8 In academic research, many researchers also applied this approach in their forecasting models. For example, the study of
9 charting patterns analysis for investment decision [19], investigating price charting patterns with kernel regression for the
10 identification of ten patterns [16–18], and implementing a variation of the bull-flag stock chart with a template matching
11 technique based on pattern recognition [31]. The stock chart patterns of these studies used for pattern recognition are
12 established in fixed patterns given by experts and researchers, which are not very similar to the actual fluctuations in
13 historical stock data, due to possible chart patterns contained historical stock data are not considered to forecasting.
14 Additionally, many advanced models and algorithms were proposed and these have achieved considerable results in fore-
15 casting accuracy. According to theories of model building, stock index forecasting models can be summarized in two cat-
16 egories [28]: (1) linear models based on statistical theories, such as General Autoregressive Conditional Heteroskedasticity
17 (GARCH) and Stochastic Volatility model [12]; and (2) non-linear models based on artificial intelligence, such as the fuzzy
18 time-series [5,29,30], artificial neural network [4], the support vector machine [32], and the particle swarm optimization
19 [20].
20 After reviewing the related research, we argued that two issues are worth to be discussed further. Firstly, most re-
21 searchers employed forecasting error as performance indicator, such as RMSE and MAPE, to evaluate their models [1,5,6,
22 28–30]. Although the daily forecasts are very useful for professional investors who implement intraday trading, we think
23 that forecasting a bullish turning point is a more interesting issue than the future stock index because an accurate forecast
24 for bullish stock pattern will bring a huge amount of stock return. Secondly, most stock pattern recognition models employed
25 one matching template with many fixed weights assigned by researcher and it is not objective approach [8]; [16–18,31].
26 In this paper, we propose three new approaches to refine past methods: (1) a new definition for a bullish turning point
27 (see Fig. 2); (2) a new weighted method, based on the occurrence of observations, to produce dynamic weights for match-
28 ing template; and (3) a new hybrid model based on PIP bull-flag pattern matching [11] and bull-flag template [18,31] for
29 recognizing stock pattern.
30 To evaluate the profitability of the proposed model, we will give several trading criteria in experiments such as thresh-
31 olds, stock holding period and stopping loss. In model verification, the NASDAQ composite index and Taiwan stock market
32 weighted index (TAIEX) are taken as experimental datasets, and four advanced algorithms as benchmarks.

33 2. Related works

34 This section gives a brief review of methodology for two stock pattern identification methods: perceptually important
35 point (PIP) identification matching and template matching.

36 2.1. Perceptually important point (PIP) identification matching

37 Reducing the dimension (i.e. the number of data point) by preserving the salient points is a promising method for time
38 series representation [10]. These points are called as “perceptually important points (PIP)”. The PIP identification process is
39 introduced by Chung et al. [7] and used for matching of technical (analysis) patterns in financial applications.
40 In the PIP identification process, the vertical distance measure method is an automatic algorithm to recognize a specific
41 pattern [11]. Three parameters are defined in the method: fitting-window, holding-period, and distance-threshold. The fitting-
42 window is a window size of trading day for a stock pattern within a specific period of trading day such as 20-day, 40-day
43 and 60-day. The holding-period is a certain period of trading day for investor to hold a stock. The distance-threshold is a
44 specific price distance, d, for the method to define a stock pattern.
45 Fig. 1 illustrates the vertical distance measure method how to produce a price distance, d, for the “head-and-shoulder”
46 pattern. In Fig. 1, three stock data, p1 , p2 and p3 , are defined as three graphic points of two-dimension coordinates(x rep-
47 resents “a series of trading days” and y represents “stock price”) contained in a specific period of fitting-window, where p1

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx 3

0.5 0 -1 -1 -1 -1 -1 -1 -1 0
1 0.5 0 -0.5 -1 -1 -1 -1 -0.5 0
1 1 0.5 0 -0.5 -0.5 -0.5 -0.5 0 0.5
0.5 1 1 0.5 0 -0.5 -0.5 -0.5 0 1
0 0.5 1 1 0.5 0 0 0 0.5 1
0 0 0.5 1 1 0.5 0 0 1 1
-0.5 0 0 0.5 1 1 0.5 0.5 1 1
-0.5 -1 0 0 0.5 1 1 1 1 0
-1 -1 -1 -0.5 0 0.5 1 1 0 -0.2
-1 -1 -1 1 -1 0 0.5 0.5 -2 -2.5
Consolidaon Breakout

Fig. 2. A 10 × 10 grid of weights to represent a variation of the “bull-flag” charting pattern, which used price and trading volume as fitting values, advanced
by Leigh et al. [17,18].

-.25 -.4 -.45 -.7 -15. -1.6 -1.6 -1.6 -1.6 -.7
-.25 -.4 -.45 -.6 -.75 -1.4 -1.4 -1.4 -.8 1
-.25 -.4 -.45 -.55 -.5 -.75 -.75 -.5 -.5 .4
-.25 -.4 -.45 -.55 -.25 .9 .9 .9 -.15 -.35
-.25 -.5 -.6 -.25 .9 1 1 1 1 -.55
-.3 -.6 -.25 .8 1 .9 .9 .9 .8 -.45
.35 .1 .8 1 .65 .6 .6 .4 .75 -.15
.1 .8 1 .5 .3 .5 .5 .3 0 .1
.8 1 .5 .35 .15 0 0 0 .3 .35
1 .8 .35 0 0 0 0 .1 .25 .3

Fig. 3. A 10 × 10 grid of weights to represent a variation of the “bull-flag” charting pattern, which used price as fitting values, advanced by Wang and Chan
[31].

48 and p2 represent the first and last points respectively; and p3 represents the point with the highest stock price between the
49 period of fitting-window (fitting-window = x1 − x2 + 1). The price distance, d, is defined as the vertical distance from p3 to the
50 line connecting p1 and p2 . The two lines, p1 , p2 and p3 pc , intersect vertically at point pc . The formula to produce vertical
51 distance, d, is defined as Eq. (1), where xc = x3 .
 x c − x1

d = V D ( p3 , p c ) = y c − y 3 = y 1 + ( y 2 − y 1 ) × − y3 (1)
x 2 − x1
52 With the given parameters, x and y, belong to p1 , p2 and p3 , the above equation can be refined as Eq. (2).
 x 3 − x1

d = V D ( p3 , p 1 , p 2 ) = y c − y 3 = y 1 + ( y 2 − y 1 ) × − y3 (2)
x 2 − x1

53 2.2. Template matching

54 As developing in computer technology and arising in cross-domain research between computer sciences and financial
55 forecasting, many researchers have increasingly paid attention to using pattern analysis based on computer-based algorithms
56 for investment decisions. For example, Lo et al. [19] tested price charting patterns using kernel regression for the identifica-
57 tion of ten patterns; Leigh et al. [18] advanced a novel forecasting model with technical analysis, pattern recognizer, neural
58 networks, and genetic algorithm to predict the NYSE composite index; Leigh et al. [17], discoveryed stock market trading
59 rules using technical charting heuristics; Leigh et al. [16] provided a computational implementation of stock charting to
60 produce a signal for movement in New York Stock Exchange Composite Index. These researches implemented a variation
61 of the “bull-flag” charting pattern (shown in Fig. 2) and used a template matching technique [8] for pattern recognition.
62 Although Wang and Chan [31] also used the template matching method to examine the potential profit for forecasting the
63 Nasdaq Composite Index (NASDAQ) and Taiwan stock market weighted index (TAIEX), the two researchers concentrated on
64 identifying an increasing price trend and proposed another version of template (shown in Fig. 3) for the bull-flag charting
65 pattern.

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

4 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

Q3 Fig. 4. A bull-flag pattern based on PIP-VD. (For interpretation of the references to color in the text, the reader is referred to the web version of this
article.)

66 The template matching is a pattern recognition technique used to match a given stock pattern with a pictographic image
67 for object identification [8,31,18]. In the matching process, a selected period of stock pattern (a period of price time series
68 defined as a “fitting window”) is converted to a matrix based on the percentile order of stock price or volume and matched
69 with a given grid template, such as a “bull flag” stock chart, to produce a “fit” value. The fit value is computed by a matrix
70 multiplication operation with the selected period of stock pattern and the given grid template.
71 The matrix operation is called a cross correlation computation, in which the percentile value that falls in each cell of a
72 column is multiplied by the weight in the corresponding cell of the given temple. If the amount of the fit values [8,31,18] is
73 high, it represents that the selected period of stock pattern is close to “bull flag” stock pattern and denotes a right time to
74 invest in the stock market.
75 Take Fig. 2 as example, the template used for the bull flag pattern is a 10-by-10 grid with weights ranging from −2.5
76 to +1.0 in the cells. The weighting is used to define areas in the template for the “descending” consolidation and for the
77 “upward-tilting” breakout portions of this bull flag heuristic pattern [18]. For example, “0.5” means that the stock pattern
78 falling in the grid has a better similarity to the shape of the full flag than “0” and less than “1”. And “−0.5” means that the
79 stock pattern falling in the grid has a worse similarity to the shape of the full flag than “0” and better than “1”.
80 The 10-by-10 grid is applied to the time series of price or volume data one trading day at a time, with the leftmost time
81 series data point being the values for the trading day which precedes the current day by 59 trading days, and the rightmost
82 time series data point being the trading day which is currently being analyzed. Values for the earliest 10% of the trading
83 days (6 days of the 60 in the rolling window) are mapped to the first column of the grid, values for the next-to earliest 10%
84 of the trading days are mapped to the second column of the grid, and so on, until the most recent 10% of the trading days
85 are mapped to the rightmost column. For example, there will be 6 trading days represented in each column of a single 60
86 trading day window. If all 6 of these trading days have price values which are in the lowest decile of the 60 price values
87 for the day, then 100% (6 values out of a total of 6 in the column) will be the value in the lowest cell of the 10 cells in the
88 column. If this column is the leftmost of the columns in the window, then this 100% will be multiplied by the value in the
89 corresponding cell in the bull flag template (which is the one in the lowest left-hand corner), which has the value of “−1.0”
90 (see Fig. 2), to result in a cell fit value of “−1.0” × 100% ="−1.0”. This is done for the 10 cells in the column and summed,
91 resulting in a fit value for the column of “−1.0”, since there will be 0.0% in the other nine cells of the column [18].
92 In this way, 10-column fit values for price and 10-column fit values for volume are computed for each trading day. The
93 summation of all 20 values for a trading day represents a total fit for the trading day [8,18].

94 3. Proposed model

95 In this paper, we proposed an intelligent recognition model, which integrated two new methods in identification process.
96 The first method is the “PIP bull-flag pattern matching”. It is refined out of two past pattern matching methods: (1) template
97 matching method [8,18]; and (2) PIP identification matching method [7,10,11]. We crystallize the advantages of the meth-
98 ods into a new stock pattern recognition method which can provide an intelligent approach to recognize possible bull-flag
99 patterns. The proposed concept to explain how they work simultaneously is introduced in the next paragraphs.
100 As Leigh et al. [17,18] defined, a bull-flag pattern is a horizontal or sloping flag of “consolidation” followed by a sharp
Q4
101 rise in the positive direction, the “breakout” (see Fig. 2). Based on the definition, the bull-flag pattern can also be defined
102 as a PIP pattern (Fig. 4 demonstrates the PIP patterns), where the black line denotes an actual stock pattern; the red line
103 between P1 and P3 denotes consolidation; and the red line between P3 and P2 denotes breakout. Therefore, a new approach
104 to recognize the bull-flag pattern from stock market is proposed in this paper and it is named as “PIP bull-flag pattern
105 matching”. To detail the method, the pseudo codes for it are given and shown in Fig. 5 (where the input variable P is a set
106 that contains candidate stock time-series patterns; the output variable, SP, is a set that contains selected “bull-flag” patterns
107 from P; the “distance-threshold” is a given value to define the height of the “bull-flag” pattern; the “fitting-window” is a given

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx 5

Fig. 5. Pseudo codes for PIP bull-flag pattern matching method.

108 value to define the window size of stock time-series pattern in P; the “holding-period” is a given period of time for holding
109 the bought stock).
110 The second method is “floating weighted” method to build a “bull-flag template” for pattern recognition. From the first
111 pattern recognition method, the stock patterns with bull-flag shape are selected and they can be used as the basis for
112 recognizing the bull-flag pattern. In this method, a bull-flag template is built based on the selected bull-flag patterns and
113 each cell of the template is assigned a “floating” weight according to its occurrence. Additionally, to extract the characteristics
114 of the bull-flag pattern as much as possible by technical analysis approaches, two types of templates, charting pattern and
115 technical indicator, are proposed to create the bull-flag template.
116 In the pattern recognition process of the proposed model, the template matching technique is applied to calculate a
117 “fitting value” for the present stock pattern with the bull-flag template. The value can be used to judge whether it is a good
118 timing (i.e. bull-flag pattern) to invest in stock market or not. For example, if the fitting value for the present stock pattern
119 reaches a certain high-level, then there is a good chance to make profit on investing the stock because the stock pattern is
120 highly similar to the bull-flag pattern.
121 Lastly, to examine model performance, five experimental parameters are taken to evaluate the proposed model: (1) pat-
122 tern fitting window; (2) PIP-distance, d, for stock pattern; (3) holding period; and (4) stock return indicator.

123 3.1. The research framework of the proposed model

124 The research framework for the proposed model contains two phases, model training (Fig. 6) and testing (Fig. 7), and
125 each phase contains four steps. In the first phase, the “bull-flag template” and “threshold” are produced from historical stock
126 data and defined as the criteria for “bull-flag pattern”. In the second phase, the criteria are used to forecast testing stock
127 data with various factors to examine model performance. Each step in the proposed model is described in detail as follows.

128 3.2. Data preparation

129 The initial experimental stock database contains six basic attributes: date, opening index, closing index, highest index,
130 lowest index, and trading volume. Because two types of template grids, charting patterns and technical indicators, are em-
131 ployed for representing the “bull-flag template”, stock patterns and technical indicators should be produced in advance as
132 forecasting attributes. Stock pattern is composed of daily closing index. Nine popular technical indicators are generated by
133 the six basic attributes and they are MA, RSI, STOD, OBV, ROC, VR, PSY, AR, and DIS. These indicators are selected because
134 they are highly related to the future stock index [5,15]. Additionally, to reduce computing complexity, cumulative probability
135 distribution approach (CDPA) [29,30] is employed to granulate the value of each technical indicator into linguistic values
136 (the initial linguistic values are set as three values : low (L1 ), medium (L2 ) and high (L3 )).

137 Phase I: Model training

138 Step 1: Bull pattern recognizing

139 This step uses the “PIP-bull pattern matching” (the pseudo codes are listed in Fig. 5) method to recognize the “bull-flag
140 pattern” from training stock data. Three parameters are used as experimental factors and those are fitting-window (a window

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

6 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

Fig. 6. The framework of the proposed model in training phase.

141 size of trading days for a stock pattern), holding-period (a certain period of trading days for investor to hold stock), and
142 distance-threshold (a given value to define a “bull-flag pattern”). An example for a bull-flag pattern, based on the template
143 matching method, is demonstrated in Table 1 (charting patterns) and Table 2 (technical indicators).

144 Step 2: Template building

145 This step is to build two types of templates (charting patterns and technical indicators) for representing the “bull-flag
146 template”. The weighted averages method is used to produce the weight for each cell of the template. The pseudo codes for
147 the “bull-flag template” of charting pattern are illustrated as follows.

(a) X = {x1 , x2 , x3 , ..., xk } is a set of daily closing index containing a possible bull market pattern, where k is the fitting
window size.
(b) Rank the index values in set X at a decrement.
(c) Calculate It,i for trading day t (1 ≤ t ≤ k), and i is 1 ≤ i ≤ 10.
It,i = 1 if ( i − 1)  k / 10 < Rank (xt ) ≤ ik/10
It,i = 0 otherwise
(d) Repeat the procedure (c), until the last It,i of possible bull flag pattern is calculated.
(e) Calculate cp_wt,i in template girds for all possible bull flag patterns n, where n is the number of possible bull flag
patterns. 
cp_wt,i =
n
b=1
Bull_ f lag_patternb (It,i )
n

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx 7

Fig. 7. The framework for the proposed model in testing phase.

148 For the template of technical indicator, the pseudo codes are listed in the followings.

(a) X = {x1,1 , x1,2 , ..., x1,9 , x2,1 , ..., xk.m } is a set of technical indicator linguistic values containing a possible bull market
pattern, where k is the fitting window size, and m is the number of technical indicators.
(b) Calculate It,i,Lp for trading day t (1 ≤ t ≤ k), i is the technical indicator number (1 ≤ i ≤ 9), and Lp is the linguistic value
(where p = 1, 2, 3).
It,i,Lp = 1 if indicator(i)’s linguistic xt,i = Lp
It,i,Lp = 0 otherwise
(c) Repeat the procedure (b), until the last It,i,Lp of possible bull flag pattern is calculated.
(d) Calculate ti_wt,i,Lp in template girds for all possible bull flag patterns n, where n is the number of possible bull flag
patterns. 
ti_wt,i,Lp =
n
b=1
Bull_ f lag_patternb (It,i,Lp )
n

149 Tables 3 and 4 illustrate the examples of “bull-flag template” of charting pattern and technical indicator, respectively,
150 based on a 20-day pattern fitting window from 1995/01/06 to 1995/02/06.

151 Step 3: Template matching

152 This step is to produce a fitting value for each “bull-flag pattern” selected from training data (step 1) and the value is
153 used, in the next step, to define the thresholds for judging whether a stock pattern is the bull-flag pattern. The template

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

8 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

Table 1
An example for a bull-flag pattern based on template matching method (charting pattern).

154 matching technique [19] is applied to calculate the fitting value and the pseudo codes to produce a fitting value, based on
155 the template of charting pattern and technical indicator, are illustrated in the following.

(a) X = {x1 , x2 , x3 , ..., xk } is a set of daily closing index containing a candidate stock pattern, where k is the fitting window
size.
(b) Rank the index values in set X at a decrement.
(c) Calculate Jt,j for trading day t (1 ≤ t ≤ k), and j is 1 ≤ j ≤ 10.
Jt,j = 1 if ( j − 1)k/10 < Rank (xt ) ≤ jk/10
Jt,j = 0 otherwise
(d) Calculate a fitting value (CP_Fit) for a stock pattern based on the template of charting pattern.
k  10
CP_F it = cp_wt j × Jt j
t=1 j=1

(a) X = {x1,1 , x1,2 , ..., x1,9 , x2,1 , ..., xk.m } is a set of technical indicator linguistic values containing a candidate stock pattern,
where k is the fitting window size, and m is the number of technical indicators.
(b) Calculate Jt,j,L for trading day t (1 ≤ t ≤ k), j is the technical indicator number
(1 ≤ j ≤ 9), and Lp is the linguistic value (where p = 1,2,3).
Jt,j,Lp = 1 if indicator(j)’s linguistic xt,j = Lp
Jt,j,Lp = 0 otherwise
(c) Calculate a fitting value (TI_Fit) for a stock pattern based on the template of technical indicator

k  10
T I_F it = ti_wt j,Lp × Jt j,Lp
t=1 j=1

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx 9

Table 2
An example for a bull-flag pattern based on template matching method (technical indicator).

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

10 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

Table 3
The template grids of weights for charting pattern analysis.

Relative high value of stock index

1 0.22 0.24 0.1 0.08 0.06 0.08 0.06 0.06 0.08 0.05 0.06 0.08 0.02 0.03 0.02 0 0.08 0.14 0.24 0.3
2 0.06 0.1 0.21 0.21 0.11 0.08 0.06 0.08 0.13 0.1 0.06 0.05 0.1 0.08 0.06 0 0.03 0.19 0.14 0.16
3 0.08 0.08 0.1 0.1 0.21 0.06 0.1 0.16 0.08 0.06 0.03 0.16 0.13 0.16 0.1 0.03 0.13 0.06 0.11 0.08
4 0.14 0.1 0.06 0.03 0.03 0.19 0.17 0.16 0.03 0.11 0.13 0.11 0.13 0.05 0.08 0.11 0.11 0.08 0.11 0.06
5 0.06 0.03 0.08 0.1 0.1 0.11 0.14 0.13 0.22 0.16 0.17 0.13 0.1 0.14 0.08 0.1 0.03 0.03 0.06 0.03
6 0.02 0.1 0.06 0.03 0.06 0.06 0.14 0.05 0.21 0.24 0.19 0.14 0.14 0.11 0.08 0.06 0.06 0.05 0.06 0.13
7 0.06 0.05 0.05 0.05 0.1 0.13 0.08 0.19 0.08 0.06 0.16 0.06 0.16 0.13 0.17 0.02 0.06 0.24 0.06 0.1
8 0.08 0.03 0.06 0.14 0.14 0.14 0.13 0.05 0.08 0.05 0.06 0.14 0.14 0.13 0.05 0.13 0.19 0.06 0.11 0.08
9 0.05 0.1 0.14 0.24 0.08 0.08 0.06 0.05 0.03 0.13 0.08 0.08 0.03 0.11 0.19 0.13 0.21 0.13 0.08 0.02
10 0.22 0.19 0.14 0.03 0.11 0.06 0.05 0.08 0.06 0.05 0.05 0.05 0.06 0.06 0.17 0.43 0.1 0.02 0.02 0.05
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Stock index time-series.

Table 4
The example for the “bull-flag template” of technical indicator.

Relative technical indicator linguistic value

MA_L1 0.38 0.38 0.38 0.38 0.38 0.38 0.38 0.38 0.37 0.37 0.37 0.37 0.37 0.37 0.37 0.37 0.37 0.38 0.38 0.37
MA_L2 0.32 0.32 0.32 0.3 0.3 0.3 0.3 0.3 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.3 0.32 0.3 0.3 0.32
MA_L3 0.3 0.3 0.3 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.32 0.33 0.32 0.32 0.32 0.32
RSI_L1 0.25 0.22 0.27 0.27 0.29 0.32 0.27 0.32 0.3 0.32 0.3 0.29 0.29 0.32 0.3 0.33 0.32 0.22 0.21 0.22
RSI_L2 0.4 0.44 0.43 0.41 0.38 0.37 0.38 0.35 0.35 0.33 0.37 0.37 0.37 0.32 0.35 0.4 0.32 0.37 0.41 0.43
RSI_L3 0.35 0.33 0.3 0.32 0.33 0.32 0.35 0.33 0.35 0.35 0.33 0.35 0.35 0.37 0.35 0.27 0.37 0.41 0.38 0.35
STOD_L1 0.37 0.38 0.38 0.32 0.32 0.35 0.32 0.35 0.41 0.37 0.35 0.33 0.38 0.43 0.52 0.67 0.22 0.06 0.03 0.08
STOD_L2 0.22 0.17 0.24 0.27 0.33 0.3 0.3 0.27 0.14 0.25 0.27 0.27 0.25 0.19 0.22 0.25 0.48 0.3 0.19 0.17
STOD_L3 0.41 0.44 0.38 0.41 0.35 0.35 0.38 0.38 0.44 0.38 0.38 0.4 0.37 0.38 0.25 0.08 0.3 0.63 0.78 0.75
OBV_L1 0.13 0.16 0.08 0.08 0.06 0.1 0.1 0.1 0.14 0.19 0.16 0.13 0.16 0.13 0.19 0.17 0.14 0.08 0.08 0.1
OBV_L2 0.51 0.46 0.56 0.57 0.57 0.56 0.52 0.56 0.48 0.49 0.56 0.54 0.59 0.67 0.6 0.62 0.57 0.56 0.49 0.46
OBV_L3 0.37 0.38 0.37 0.35 0.37 0.35 0.38 0.35 0.38 0.32 0.29 0.33 0.25 0.21 0.21 0.21 0.29 0.37 0.43 0.44
ROC_L1 0.19 0.25 0.25 0.21 0.24 0.22 0.27 0.21 0.27 0.27 0.19 0.24 0.24 0.29 0.27 0.35 0.17 0.03 0.02 0
ROC_L2 0.49 0.38 0.38 0.37 0.41 0.44 0.44 0.49 0.37 0.4 0.49 0.48 0.44 0.43 0.49 0.52 0.62 0.65 0.57 0.33
ROC_L3 0.32 0.37 0.37 0.43 0.35 0.33 0.29 0.3 0.37 0.33 0.32 0.29 0.32 0.29 0.24 0.13 0.21 0.32 0.41 0.67
VR_L1 0.21 0.24 0.29 0.19 0.17 0.13 0.21 0.22 0.25 0.29 0.24 0.27 0.35 0.21 0.21 0.22 0.16 0.08 0.1 0.08
VR_L2 0.35 0.33 0.32 0.44 0.44 0.49 0.33 0.43 0.3 0.4 0.44 0.38 0.37 0.54 0.59 0.57 0.46 0.51 0.35 0.41
VR_L3 0.44 0.43 0.4 0.37 0.38 0.38 0.46 0.35 0.44 0.32 0.32 0.35 0.29 0.25 0.21 0.21 0.38 0.41 0.56 0.51
PSY_L1 0.25 0.27 0.24 0.19 0.22 0.16 0.17 0.16 0.19 0.25 0.22 0.21 0.24 0.22 0.27 0.29 0.19 0.17 0.1 0.11
PSY_L2 0.38 0.4 0.41 0.43 0.37 0.44 0.41 0.44 0.37 0.33 0.32 0.37 0.33 0.33 0.29 0.37 0.44 0.38 0.52 0.43
PSY_L3 0.37 0.33 0.35 0.38 0.41 0.4 0.41 0.4 0.44 0.41 0.46 0.43 0.43 0.44 0.44 0.35 0.37 0.44 0.38 0.46
AR_L1 0.3 0.29 0.32 0.32 0.32 0.32 0.35 0.32 0.3 0.33 0.3 0.33 0.35 0.33 0.35 0.37 0.29 0.29 0.29 0.27
AR_L2 0.37 0.43 0.38 0.33 0.38 0.37 0.35 0.38 0.38 0.33 0.35 0.37 0.37 0.37 0.41 0.41 0.41 0.41 0.41 0.43
AR_L3 0.33 0.29 0.3 0.35 0.3 0.32 0.3 0.3 0.32 0.33 0.35 0.3 0.29 0.3 0.24 0.22 0.3 0.3 0.3 0.3
DIS_L1 0.24 0.22 0.24 0.19 0.21 0.22 0.24 0.21 0.27 0.24 0.22 0.21 0.25 0.27 0.25 0.37 0.1 0 0 0.02
DIS_L2 0.46 0.33 0.38 0.52 0.48 0.48 0.44 0.44 0.33 0.44 0.43 0.48 0.41 0.38 0.56 0.56 0.65 0.43 0.33 0.22
DIS_L3 0.3 0.44 0.38 0.29 0.32 0.3 0.32 0.35 0.4 0.32 0.35 0.32 0.33 0.35 0.19 0.08 0.25 0.57 0.67 0.76
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Stock index time-series.

156 Step 4: Threshold determining

157 In this step, average and minimum thresholds are applied to define the criteria to judge whether a stock pattern is the
158 bull-flag pattern or not. They are generated with the fitting values of the “bull-flag patterns” selected from training stock data.
159 For the template of charting pattern, the thresholds are defined by Eqs. (3) and (4), where n is the number of possible bull
160 market patterns.
n
b=1 Bull_market_patternb (CP_F it )
CP_T hresholdavg = (3)
n
CP_T hresholdmin = min (Bull_market_patternb (CP_F it )) (4)
b=1 to n

161 For the template of technical indicator, the thresholds are defined by Eqs. (5) and (6), where n is the number of possible
162 bull market patterns.
n
b=1 Bull_market_patternb (T I_F it )
T I_T hresholdavg = (5)
n
T I_T hresholdmin = min Bull_market_patternb (T I_F it ) (6)
b=1 to n

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx 11

163 From Eq. (3) to Eq. (6), we can obtain the “thresholds” based on the templates of charting pattern and technical indicator.
164 With these thresholds, a trading rule is provided as follows: “If the two fitting values (charting pattern and technical indicator)
165 for a stock pattern are both “greater than” or “equal to” the thresholds, the stock pattern should be defined as a “bull-flag pattern”
166 and we should buy the stock at the next day to make profit in the future.”

167 Phase II: Model testing

168 Step 5: Template matching

169 This step is to produce two fitting values for each stock pattern selected from testing data. One is from the “bull-flag
170 template” of charting pattern and the other is from the template of technical indicator.

171 Step 6: Bull pattern judging

172 This step is to judge whether the present pattern is the “bull-flag pattern” or not, and the trading (judging) rule is listed
173 in step 4.

174 Step 7: Trading policies deciding

175 Two experimental factors, holding-period (a certain period of trading days for investor to hold the investing stock) and
176 “stop loss” are employed in the trading rule (step 4) to examine investing profit and avoid a heavy loss when stock market
177 trend is going in the other direction for a long time. In the proposed model, the holding-period is set as four vales: 5-days,
178 10-days, 15-days and 20-days and the “stop loss” is set as 10% of buying price for the stock.

179 Step 8: Profit evaluating

180 Two performance indicators, total index return percentage (TIR%) and total index return (TIR), are employed to evaluate
181 model performance because both are adopted in past research models [5,31]. TIR% is defined by Eq. (7) which measures a
182 gain rate of price index for total transactions, and TIR is defined by Eq. (8) which measures an absolute gain of price index
183 for total transactions. The major difference between the performance indicators is that TIR is more convenient to produce
184 actual profit return for investor when transaction cost is determined.

m
Closing_indexs − Closing_indexb
Total index return(% ) = (7)
Closing_indexb
n=1
185

m
Total index return = Closing_indexs − Closing_indexb (8)
n=1

186 4. Model verification

187 In this section, we conduct an experiment with 3 factors (pattern fitting window, PIP-distance for stock pattern, and
188 holding-period) for evaluating the proposed model, and employ 4 advanced models as benchmarks: Wang and Chan’s model
189 [31], the RST algorithms [23,27], the genetic algorithms [13], and Cheng et al.’s model [5].

190 4.1. Performance evaluation and model comparison

191 The experimental datasets of NASDAQ contains four testing-periods: (1)1989/01/07 to 1992/10/27; (2) 1992/10/28 to
192 1996/08/19; (3)1996/08/12 to 200/05/26 and (4) 2000/05/27 to 2004/40/30. The TAIEX contains two testing-periods: (1)
193 1990/08/15 to 1997/02/17 and (2) 1997/02/18 to 2004/03/24. The training–testing ratio is 7:1 (year) for the proposed model.
194 Because two “bull-flag” templates, charting pattern and technical indicator, are employed in forecasting process, the pro-
195 posed model can be classified as three models based on the usage of templates: (1) model A employs the template of
196 technical indicator to generate a total fit value, (2) model B employs the template of charting pattern to generate a total fit
197 value, and (3) model C employs the two templates simultaneously to a total fit value.
198 Initially, Wang and Chan’ [31] model is used as benchmark in forecasting performance. Table 5 shows the performance
199 data for the proposed and the comparison models. It is clear that the proposed model (models A, B, C) bears much better
200 TIR% than Wang and Chan’s for forecasting the NASDAQ and the TAIEX; model B performs better than the others in the
201 NASDAQ; and models A and C give the best TIR% for the TAIEX.
202 To examine the forecasting performance further, we use the RST algorithms, the GAs algorithms and Cheng et al.’s [5]
203 model as comparison in the second experiment. The experimental datasets contains 5 testing-periods, the last two months
204 of the year from 2001 to 2005. The testing periods for the proposed model are the same with Cheng et al.’s model. To
205 make the model comparison in fair condition, only the performance data of model A is provided because it employs the
206 same forecasting factors (the nine technical indicators: MA, RSI, STOD, OBV, ROC, VR, PSY, AR, and DIS) with the three algo-
207 rithms. Table 6 lists all performance data with TIR and it also has revealed that model A outperforms the listing advanced
208 algorithms.

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

12 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

Table 5
Q5 Performance comparison with the model advanced by Wang and Chan [31].

Model Wang and Chan’s Proposed model A Proposed model B Proposed model C

Dataset Testing period TIR (%) Average IR (%) N (trades) PIP-distance TIR (%) N (trades) TIR (%) N (trades) TIR (%) N (trades)

NASDAQ 1989/01/17 291 1.94 150 5 132.29 232 *429.12 241 −7.04 52
| 10 −31.30 150 253.10 170 −25.35 40
1992/10/27 15 −57.13 80 273.43 116 −22.14 24
1992/10/28 201.96 1.32 153 5 0.00 0 *338.94 278 0.00 0
| 10 0.00 0 217.04 224 0.00 0
1996/08/09 15 0.00 0 163.17 149 0.00 0
1996/08/12 444.27 2.51 177 5 99.94 32 *1084.17 446 46.65 23
| 10 94.16 34 1056.40 403 40.87 25
2000/05/26 15 93.51 41 937.34 356 51.22 28
2000/05/27 24.96 0.39 64 5 *168.24 70 −35.06 242 4.31 40
| 10 −15.15 115 −62.81 240 −166.18 62
2004/03/20 15 −227.94 169 −103.50 242 −336.86 86
TAIEX 1990/08/15 286.88 1.63 176 100 *1291.20 319 1156.96 395 1159.79 216
| 200 694.97 319 1190.92 443 787.83 210
1997/02/17 300 742.23 257 1216.23 430 592.98 162
1997/02/18 314.5 1.7 185 100 479.40 635 818.71 667 *940.98 269
| 200 340.95 633 495.73 604 752.13 239
2004/03/24 300 123.94 498 166.57 441 352.94 120

Table 6
Performance comparison with the models advanced by RST, GAs and Cheng et al. [5].

TAIEX Testing period Comparison model Proposed model A (pattern fitting window = 60;

PIP-distance = 200; minimum threshold)

Buy-and-hold Rough set Genetic Cheng et al.’s 5-day 10-day 15-day 20-day
theory algorithms holding holding holding holding

*
2001(Nov.–Dec.) 1612.16 923.9 1272.05 1683.51 4923.25 11203.01 16524.59 20303.3
2002(Nov.–Dec.) −144.24 304.28 353.56 421.82 −454.07 689.46 2085.70 *
3402.29
2003(Nov.–Dec.) −163.62 70.78 247 336.25 100.08 1045.00 3026.30 *
5369.40
*
2004(Nov.–Dec.) 414.04 196.06 481.65 780.26 1034.71 985.25 724.19 322.84
*
2005(Nov.–Dec.) 740.59 105.85 163.8 210.83 1982.12 4733.39 6355.68 7517.35
*
Average 491.79 320.17 503.61 686.53 1517.22 3731.22 5743.29 7383.04

Performance indicator: TIR.

∗
The best among comparison model.

209 4.2. Findings and discussions

210 In this section, we give a summarization for experimental results. Form the comparison results with the listing 4 ad-
211 vanced algorithms, 6 findings are discovered in the followings.

212 (1) The proposed model performs differently for forecasting the TAIEX and NASDAQ because of their stock patterns are
213 dissimilar. From Table 5, we can see that the PIP-distance values with better TIR (%) for the NASDAQ (5, 10 and 15)
214 are much smaller than TAIEX (100, 200 and 300). This finding is easy to be understood because the spread of TAIEX
215 is greater than NASDAQ and, therefore, the bull-flag patterns of TAIEX should be much larger than NASDAQ.
216 (2) Small stock pattern is better than lager pattern for forecasting stock market. From Table 5, it is shown that the TIR
217 (%) for a lower PIP-distance value is better than a higher value in the TAIEX and NASDAQ. It tells that the bull-flag
218 patterns are usually in the pattern with smaller height.
219 (3) Charting pattern analysis is more effectively than technical indictor for forecasting the NASDAQ. From Table 5, it is
220 clear that model B performs better than model A in 3 testing datasets of the NASDAQ. However, in the TAIEX, model A
221 outperforms model B in the first testing dataset. This phenomenon implies that charting pattern analysis can discover
222 more bull-flag patterns in the NASDAQ and the signal of the bull-flag pattern for technical indictor in the NASDAQ is
223 not very clear. However, in the TAIEX, the signal of the bull-flag pattern is easy to be found by using technical indictor
224 analysis.
225 (4) The hybrid model (model C) using the two templates simultaneously to generate the total fit value can reduce the
226 transaction cost and maintain a high TIR (%). From Table 5, it is found that the amount of trades (N) for model C is
227 much less than model A and B for the TAIEX. For the first testing period, the amount of trades for model C is 216,
228 model A is 319, and model B is 395. For the second testing period, model C is 269, model A is 635, and model B is
229 667. Besides, in the first testing period, the TIR (%) for model C is 1159.79 that is a little less than the best TIR (%),

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx 13

230 1291.20. And, in the second testing period, model C bears the best TIR (%), 940.98, with the least amount of trades
231 among three models. This evidence indicates that model C brings more profit when forecasting the TAIEX.
232 (5) A longer holding period can make higher stock return. From Table 6, the performance data has clearly shown that
233 model A gives excellent forecasting performance (the average TIR of model A is 2–10 times of Cheng et al.’s model)
234 and four best TIRs are produced by using 20-day holding period. From the evidence, it is strongly suggested that
235 investor should invest the TAIEX for a longer holding period, such as 20 days at least. This phenomenon was also
236 found in the research conducted by Wang and Chan [31].
237 (6) Overall, model B is a better profitable model among the testing models and especially for forecasting the TAIEX. From
238 Table 5, it is clear that model B performs well for forecasting the NASDAQ and the TAIEX except the testing period
239 of the NASDAQ (from 2000/05/27 to 2004/03/20). In practical applications for stock market forecasting, model A is a
240 traditional method refined by the framework of the proposed model. Model B is a “new” method based on technical
241 indicators provided by the framework of the proposed model. Model C is a “hybrid” model of model A and model
242 B. From the experimental results, model B is the most profitable one among three models but the trading number
243 is the largest and model C bears the smallest trading number with acceptable profit return for the TAIEX. Therefore,
244 model B is suggested when the investor has a “positive” or “bullish” view for the present stock market and model C
245 is employed when the investor has a “negative” or “bearish” view.

246 5. Conclusion

247 In this paper, we proposed an intelligent pattern recognition model to ascertain the bull-flag patterns contained in his-
248 torical stock patterns, and extract two types of bull-flag templates, the chart pattern and the technical indicator, as judging
249 criteria for the bull-flag pattern.
250 The contributions for the proposed model are: (1) address a new definition for the bull-flag pattern (see Fig. 4) beside
251 the subjective the bull-flag pattern defined by researchers [8,18,31]; (2) provide a new weighted method, based on the oc-
252 currence of observations, to effectively fuse information of stock pattern and indicators; (3) provide an effective granulating
253 method, the CPDA method [5], to discretize technical indicators to reduce computing complexity; (4) provide a new con-
254 cept which integrates the advantages of the past two pattern-matching methods, PIP identification-matching and template-
255 matching, in the recognition process beside the past models which devoted to find the bull-flag pattern only with the chart
256 pattern [8,18,31]. Therefore, it provided effective rules (bull-flag templates) to judge a stock pattern more accurately.
257 Based on the trading rule for the bull-flag pattern, the proposed model gave an unprecedented stock index return (TIR%
258 and TIR) in forecasting the NASDAQ and TAIEX. In the experiments for model verification, the performance data proved
259 superior to the proposed models by comparing them with the three advanced stock market forecasting algorithms, including
260 RST (LEM2), GAs and their hybrid model, issued by Cheng et al. [5]. The experimental results also demonstrate a robust stock
261 return, when used to forecast the TAIEX. Although the proposed model (model B) had a negative TIR% for forecasting the
262 fourth testing period of the NASDAQ, the average TIR% was still at a high, positive level (454.29%), and was much better
263 than Wang and Chan’s model (240.55%). Additionally, the proposed model, with two templates (model C), can give a high
264 level of stock index return (TIR) with a low number of trades.
265 The experimental data indicate that this model can provide a highly accurate forecast for the bull-flag pattern, thereby
266 allowing stock investors to accrue huge profits by employing it. Further, this model also allows investors to examine the
267 stock information of selected bull-flag patterns, including basic and technical indicators. This information can help stock
268 analysts or senior stock investors to check stock patterns more carefully.
269 In the future, in order to strengthen the proposed model, two suggestions are recommended, as follows: (1) use other
270 stock databases, such as the S&P 500 and the NYSE, to evaluate the proposed model; and (2) apply other advanced pattern-
271 searching methods to the proposed model, such as wavelet theory, to extract the bull-flag pattern to produce the bull-flag
272 template and evaluate stock return.
Q6
273 Uncited references

274 [14,24–26]

275 References

276 [1] S. Asadi, E. Hadavandi, F. Mehmanpazir, M.M. Nakhostin, Hybridization of evolutionary Levenberg-Marquardt neural networks and data pre-processing
277 for stock market prediction, Knowledge Based Syst. 35 (2012) 245–258.
278 [2] L. Blume, D. Easley, M. O’Hara, Market statistics and technical analysis: The role of volume, J. Finance 49 (1) (1994) 153–181.
279 [3] W. Brock, J. Lakonishok, B. Lebaron, Simple technical trading rules and the stochastic properties of stock returns, J. Finance 47 (5) (1992) 1731–1764.
Q7 280 [4] C.H. Chen, Neural networks for financial market prediction, in: IEEE World Congress Computational Intelligence, 2, 1994, pp. 1199–1202 .
281 [5] C.H. Cheng, T.L. Chen, L.Y. Wei, A hybrid model based on rough sets theory and genetic algorithms for stock price forecasting, Inf. Sci. 180 (9) (2010)
282 1610–1629.
283 [6] V. Cho, MISMIS – a comprehensive decision support system for stock market investment, Knowledge Based Syst. 23 (2010) 626–633.
284 [7] F.L. Chung, T.C. Fu, R. Luk, V. Ng, Flexible time series pattern matching based on perceptually important points, in: Paper presented at International
285 Joint Conference on Artificial Intelligence Workshop on Learning from Temporal and Spatial Data, Seattle, Washington, 2001.
286 [8] Duda R.O., Hart P.E., Pattern Classification and Scene Analysis, John Wiley & Sons, New York, 1973.
287 [9] R.D. Edwards, J. Magee, W.H.C. Bassetti, Technical Analysis of Stock Trends, ninth ed., CRC Press, Florida, 2007.

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]
 JID: INS
ARTICLE IN PRESS [m3Gsc;February 10, 2016;18:26]

14 T.-l. Chen, F.-y. Chen / Information Sciences xxx (2016) xxx–xxx

288 [10] T.C. Fu, A review on time series data mining, Eng. Appl. Artif. Intell. 24 (2011) 164–181.
289 [11] T.C. Fu, F.L. Chung, R. Luk, C.M. Ng, Stock time series pattern matching: Template-based vs. rule-based approaches, Eng. Appl. Artif. Intell. 20 (2007)
290 347–364.
291 [12] B.D. Garland, SV mixture models with application to S&P 500 index returns, J. Financ. Econ. 85 (3) (2007) 822–856.
292 [13] J.H. Holland, Adaptation in Nature and Artificial Systems, University of Michigan, 1975.
293 [14] K.J. Kim, I. Han, The extraction of trading rules from stock market data using rough sets, Expert Syst. 18 (4) (2001) 194–202.
294 [15] M.J. Kim, S.H. Min, I. Han, An evolutionary approach to the combination of multiple classifiers to predict a stock price index, Expert Syst. Appl. 31 (2)
295 (2006) 241–247.
296 [16] W. Leigh, N. Modani, R. Hightower, A computational implementation of stock charting: abrupt volume increase as signal for movement in New York
297 stock exchange composite index, Decis. Support Syst. 37 (4) (2004) 515–530.
298 [17] W. Leigh, N. Modani, R. Purvis, T. Roberts, Stock market trading rule discovery using technical charting heuristics, Expert Syst. Appl. 23 (2) (2002)
299 155–159.
300 [18] W. Leigh, R. Purvis, J.M. Ragusa, Forecasting the NYSE composite index with technical analysis, pattern recognizer, neural networks, and genetic
301 algorithm: a case study in romantic decision support, Decis. Support Syst. 32 (4) (2002) 361–377.
302 [19] A. Lo, H. Mamaysky, J. Wang, Foundations of technical analysis: computational algorithms, statistical inference, and empirical implementation, J. Fi-
303 nance 55 (4) (2000) 1705–1765.
304 [20] R. Majhi, G. Panda, Prediction of S&P500 and DJIA stock indices using particle swarm optimization technique, in: Paper presented at 2008 IEEE
305 Congress on Evolutionary Computation, Hong Ko n g, China, 2008.
306 [21] C. Neely, P. Weller, R. Dittmar, Is technical analysis in the foreign exchange market profitable? A genetic programming approach, J. Financ. Quant. Anal.
307 32 (4) (1997) 405–426.
308 [22] S.N. Neftci, Naive trading rules in financial markets and Wiener-Kolmogorov prediction theory: a study of “Technical analysis”, J. Bus. 64 (4) (1991)
309 549–571.
310 [23] Z. Pawlak, Rough sets, Int. J. Comput. Inf. Sci. 11 (5) (1982) 341–356.
311 [24] Z. Pawlak, Rough sets and intelligent data analysis, Inf. Sci. 147 (2002) 1–12.
312 [25] Z. Pawlak, A. Skowron, Rough sets: some extensions, Inf. Sci. 177 (2007) 28–40.
313 [26] Z. Pawlak, A. Skowron, Rudiments of rough sets, Inf. Sci. 177 (2007) 3–27.
314 [27] B. Predki, R. Slowinski, J. Stefanowski, R. Susmaga, S. Wilk, ROSE-software implementation of the rough set theory, Lecture Notes in Artificial Intelli-
315 gence, 1424th ed.„ Springer-Verlag, Berlin, 1998, pp. 605–608.
316 [28] W. Shen, X. Guo, Wu C., D. Wu, Forecasting stock indices using radial basis function neural networks optimized by artificial fish swarm algorithm,
317 Knowledge Based Syst. 24 (2011) 378–385.
318 [29] C.H. Su, T.L. Chen, C.H. Cheng, Y.C. Chen, Forecasting the stock market with linguistic rules generated from the minimize entropy principle and the
319 cumulative probability distribution approaches, Entropy 12 (2011) 2397–2471.
320 [30] H.J. Teoh, C.H. Cheng, Chu H.H., J.S. Chen, Fuzzy time series model based on probabilistic approach and rough set rule induction for empirical research
321 in stock markets, Data Knowl. Eng. 67 (2008) 103–117.
322 [31] J.L. Wang, S.H. Chan, Stock market trading rule discovery using pattern recognition and technical analysis, Expert Syst. Appl. 33 (2) (2007) 304–315.
323 [32] W. Wuang, Y. Nakamori, S. Wang, Forecasting stock market movement direction with support vector machine, Comput. Oper. Res. 32 (10) (2005)
324 2513–2522.

Please cite this article as: T.-l. Chen, F.-y. Chen, An intelligent pattern recognition model for supporting investment deci-
sions in stock market, Information Sciences (2016), [Link]