A Discretization Method for the Detection of Local Extrema and
Trends in Non-discrete Time Series
Konstantinos F. Xylogiannopoulos
1
, Panagiotis Karampelas
2
and Reda Alhajj
1
1
Department of Computer Science, University of Calgary, Calgary, Alberta, Canada
2
Department of Informatics and Computers, Hellenic Air Force Academy, Dekelia Air Base, Athens, Greece
Keywords: Moving Linear Regression Angle, Linear Regression, Pattern Detection, Trend Detection, Local Extrema,
Local Minimum, Local Maximum, Discretization.
Abstract: Mining, analysis and trend detection in time series is a very important problem for forecasting purposes. Many
researchers have developed different methodologies applying techniques from different fields of science in
order to perform such analysis. In this paper, we propose a new discretization method that allows the detection
of local extrema and trends inside time series. The method uses sliding linear regression of specific time
intervals to produce a new time series from the angle of each regression line. The new time series produced
allows the detection of local extrema and trends in the original time series. We have conducted several exper-
iments on financial time series in order to discover trends as well as pattern and periodicity detection to fore-
cast future behavior of Dow Jones Industrial Average 30 Index.
1 INTRODUCTION
The study of time series is a very important research
area for many different applications and scientific do-
mains. Any variable that changes over time can be de-
fined as a time series. The study of such variables and
their change over time can be very important for var-
ious reasons, e.g., to understand past behavior and
based on that predict future behavior. Such studies are
very important since they can be applied to a wide
spectrum of scientific fields such as psychology, eco-
nomics, physics, meteorology, geology, biology, etc.
Usually, a variable and its representation as a time
series involve real values. Therefore, a direct analysis
over these values can be extremely difficult since, for
example, if we want to analyze temperatures in Can-
ada the values may vary from -50 degrees Celsius up
to 40 degrees Celsius. Having also one decimal digit
for every observation it means that we have 901 dis-
crete values to be analyzed. Due to this wide range of
values in order to proceed with their analysis, a dis-
cretization of the time series must first be conducted.
For this purpose, many discretization techniques have
been developed (Yang et al. 2005). Discretization
groups values that are close (the closeness depends on
the discretization method and its parameters), and
then the new time series can be analyzed, e.g., detect-
ing patterns that occur often. For the discretization, a
predefined alphabet is used and a specific letter from
the alphabet is assigned to each group of data values.
By applying this method continuous (real) values can
be transformed to discrete values and, therefore, pat-
tern, periodicity or trend detection can be performed.
In this paper, we present a new discretization
method that allows us to directly identify local min-
ima/maxima and trends inside a time series. By ap-
plying a mathematical transformation on the original
time series values we use the outcome to perform
sliding linear regression analysis of short time inter-
vals. We have named this method Moving Linear Re-
gression Angle (MLRA) because for each linear re-
gression analysis we use the angle of the regression
line (calculated from its slope) in order to create a new
time series. Using this new time series we can detect
fast the turning points of the time series, i.e., the local
minima and maxima. Having such information we
can detect all sub-trends that exist in a time series
since the discretization method uses the same alpha-
bet letter for up or down trends. The conducted testing
demonstrates the applicability and effectiveness of
the proposed approach.
The rest of the paper is organized as follows: Sec-
tion 2 is a review of discretization and trend detection
methods. Section 3 presents the proposed MLRA
based approach. Section 4 reports the experimental
results obtained using financial data and more
346
Xylogiannopoulos K., Karampelas P. and Alhajj R..
Discretization Method for the Detection of Local Extrema and Trends in Non-discrete Time Series.
DOI: 10.5220/0005401203460352
In Proceedings of the 17th International Conference on Enterprise Information Systems (ICEIS-2015), pages 346-352
ISBN: 978-989-758-096-3
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
specifically Dow Jones Industrial Average 30 Index.
Section 5 is conclusions and future work.
2 RELATED WORK
Due to the importance of analyzing time series and
especially those produced by continuous values,
many different discretization methods have been de-
veloped so far. Variables can be categorized as quali-
tative or quantitative (Yang et al. 2005). Each cate-
gory can be sub-categorized to nominal and ordinal
for qualitative and to interval or ratio for quantitative
variables. We study the second category of quantita-
tive data because of its importance and wide spectrum
of applications in various scientific domains. Differ-
ent taxonomies can be applied for the discretization
of quantitative values such as univariate or multivar-
iate, disjoint or non-disjoint, ordinal or nominal fuzzy
or non-fuzzy, etc. (Yang et al. 2005) Some of the most
common discretization methods are (a) equal-width
where each range has the same width, (b) equal-fre-
quency where the data are classified to ranges that
have the same amount of data, (c) clustered-based by
grouping data values together based on specific parti-
tions, (d) fuzzy discretization which applies its rules
based on a membership function, etc. (Bao, 2008)
Many methods have been introduced in the past
decades for forecasting purposes based on historical
data of a given time series. Esling and Agon (2012)
summarized many data mining techniques for the
analysis of time series, while White and Granger
(2011) provided a deep analysis of trends in financial
time series. Especially in finance some of the methods
can be classified as (a) numerical linear models like
ARIMA (Bao, 2008; Bao et al., 2013; De Gooijer and
Hundman, 2006; Kovalerchuck and Vityaev, 2000;
Qin and Bai, 2009; Xi-Tao, 2006), (b) rule-based
models like decision tree, naïve Bayesian classifier,
hidden Markov model etc. (Bao, 2008; Kovalerchuck
and Vityaev, 2000 ), non-linear models such as artifi-
cial networks (Balkin and Ord, 2000; Bao et al., 2013;
Qin and Bai, 2009; Selvarantnam and Kirley, 2006)
and (d) fuzzy system models and support vector ma-
chines (Muller et al., 1997; Qin and Bai, 2009).
Moreover, more financial forecasting tools have
been introduced for over a century based on technical
analysis. Such methods are Moving Average for dif-
ferent time spans, Relative Strength Index, Moving
Average Convergence Divergence for different time
spans, Momentum, etc. (Bao, 2008; Chen et al., 2014;
Edwards et al., 2007; Pring 2002) Furthermore, many
theories depending on specific pattern shapes have
also been introduced such as Elliot Waves of 1-2-3-
4-5 uptrend and A-B-C downtrend formation (Ed-
wards et al. 2007) or simpler like Resistant and Sup-
port Lines, Head-And-Shoulders, Triangles, Flags,
Rectangles, Double or Triple Bottom or Top for-
mation, Island formation etc. (Bao, 2008; Edwards,
2007; Pring, 2002) All these methods and patterns are
based on the detection of local extrema and how the
prices change over specific points and time intervals
in order to produce such formations. Although such
formations are very well known for many decades,
new methods are introduced very often to propose
new methodologies for detecting trends (Bao, 2008;
Bao et al. 2013; Chen et al., 2014).
For detecting trends in time series and especially
financial time series, many methods have been intro-
duced that apply techniques coming from different
data mining, mathematical and financial fields. Qin
and Bai (2009) have introduced a method that uses a
new Association Rules Algorithm in order to predict
trends in derivatives’ prices time series. Guerrero and
Galicia-Vazquez proposed in 2010 a new method that
decomposes a financial time series using exponential
smooth filtering into two different parts, i.e., the trend
and the noise of the time series. A more complex tech-
nique has been introduced by Chen at al. in 2014 that
uses advanced fuzzy logic approach in combinations
with the minimal root mean square root error crite-
rion. Another advanced method has been introduced
by Muhlbayer et al. in 2009 that uses advanced linear
regression methods to estimate trends. The specific
methodology has been used on meteorological and
precipitation time series, however, it can be applied
also in finance. Moreover, Gardner and McKenzie
(1985) have developed an exponential smoothing
model that damps erratic trends in order to provide
more accurate trend detection.
3 PROPOSED METHODOLOGY
Our discretization method that will help detecting
trends in a time series and identifying possible perio-
dicities is based on the detection of the local minima
and maxima. When a function is known, we can find
the local minimum/maximum by applying the second
derivative test. In this case, assuming that the function
is twice differentiable at a critical point where the first
derivative is equal to 0, we have to examine if the sec-
ond derivative is negative or positive, which means
that the critical point is a local maximum or mini-
mum, respectively, (we cannot determine if the sec-
ond derivative is equal to zero too). However, such a
process cannot be applied in a time series unless we
use first interpolation in order to produce a
DiscretizationMethodfortheDetectionofLocalExtremaandTrendsinNon-discreteTimeSeries
347
realfunction based on the data points of the time se-
ries. With the interpolation we try to fit the data points
on a polynomial that can emulate the time series
based on the given discrete data values. Yet, this is
one of the most difficult problems in Numerical Anal-
ysis, especially when the polynomial that we want to
fit on the data points of the time series can be of a very
large degree. Moreover, as we can observe from
“Fig.1.b”, in which we have the daily percentage
changes of the DJIA30 Index, due to very small up
and down fluctuations of the stock market we have
extreme noise and it is very difficult to find meaning-
ful turning points (minima/maxima) that will signal a
trend reversal and a possible opportunity for buying
or selling stocks.
Our method, Moving Linear Regression Angle
analysis (MLRA), is based on the continuous execu-
tion of sliding linear regression analysis over time.
We perform continuous regression analysis of spe-
cific time interval-sliding window (width-data points)
and in each loop we calculate the angle of the regres-
sion line with the x-axis (the time axis of the time se-
ries) from the slope of the regression line. Assuming
that we have a time series of data points we start at
the beginning of time
0. Then for a specific
time interval, e.g. for stock prices this can be charac-
terized by 10 days (if the time series is ex-
pressed in days), we perform a linear regression anal-
ysis for data points up to
9 (sliding window 0
to 9). Then we increase the starting point by one, i.e.,
1 and the ending point will become
10
(sliding window 1 to 10). We continue this process
until we reach the end of the time series (sliding win-
dow 10 to 1, assuming the length of the time
series is ). In each loop we calculate the slope of the
regression line, and based on this the angle of the line
with respect to the x-axis in radius /2, /2.
With this process we construct a new time series of
 points and with starting point at
and ending
point at

of the original time series. In the new
time series, the value of the angles can show us how
the segments of the original time series behave re-
garding their monotony. If the angle of each part is
larger than the previous then the specific part of width
w has an uptrend while if it is smaller it has a down-
trend. When the values change from larger to smaller
we have a local maximum while when they change
from smaller to larger we have a local minimum
“Fig.1.a”.
Table 1: Identicative Results of Repeated Patterns in DJIA30 Transformation for MLRA10.
Index Pattern Start Period Occ. Length Positions
1
ZZZZZZZZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
155 783 2 49 155,938
2
ZZZZZZZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
219 720 2 48 219,939
3
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAZZZZZZZZZ
251 700 2 45 251,951
4
ZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
225 524 2 39 225,749
5
ZZZZZZZZZZZAAAAAAAAAAAAAAAAAAA
435 505 2 30 435,940
6
AAAAAAAAAAAAAAAAAAAZZZZZZZZZAA
268 379 2 30 268,647
7
AAAAAAAAAAAAAAAAAAAAZA
59 820 2 22 59,879
8
ZZZZZZZZZZZZZZZZZZZZZZ
347 557 2 22 347,904
9
AAAAAAAAAAAAZZZZZZZZZZZZZZZZ
0 578 2 28 0,578
10
AAAAAAAAAAAAAAAAAAAA
267 231 4 20 267,498,729,960
Figure 1: Dow Jones Industrial Average 30 Prices and Daily Percentage Changes for 2010-2013.
ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems
348
Figure 4: Discretized Time Series for DJIA30 for 2010-2013 using MLRA for 10 days interval.
Figure 2: Microsoft Stock Prices and MLRA Transfor-
mations for years 2010-2013.
Figure 3: Dow Jones Industrial Average 30 Prices an
d
MLRA Transformations for 2010-2013.
Figure 5: DJIA30 Transformed Time Series and Trend De-
tection Examples.
Figure 6: DJIA30 Transformed Time Series and Trend De-
tection Examples.
However, there is a significant obstacle when we deal
with time series having their values very small and
close to the slope of the regression line, and based on
this the angle of the line with respect to the x-axis. In
“Fig.2.a” we have the stock prices for Microsoft from
January 4th, 2010 till December 31st, 2013. The val-
ues vary between $20 and $40. As we can observe in
“Fig.2.b” the values in the new time series created by
applying the proposed MLRA change very smoothly.
In “Fig.2.c” we can see how the values fluctuate very
close to 0. So far the new time series behaves exactly
like the original time series and it is very difficult to
detect the local minima/maxima and the change in
trends. In order to make this process easier algorith-
mically, we will use a transformation on the original
time series. For the transformation, we will multiply
the original time series with a constant, which we will
name Sharpness Transformation Factor, denoted

, in order to move away the time series from the
. Doing this we will not lose any information
of the original time series, however, the regression
lines of each MLRA phase will become much steeper.
As we can see in “Fig.2.d”, we have the transformed
Microsoft stock prices and in “Fig.2.e” we have the
AAAAAAAAAAAAZZZZZZZZZZZZZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAZAZZZZZZZZZZZZZZZZZZZZZZZAZZZZZZAAAA
AAAAZZZZZZZZZZAAAAAAAAZZZAAAAAAAAAAAAAZZZZZZZZZZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAZZZZZZZZ
ZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAZZZZZZZZZAAZZZZZZZZZAAAAAAAAAAAAAAZZZZZZZAAAAAAAAAA
AZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZAAAAAAAAAAAAAAAAZZZZZAAAAAZZZZZZZZZZZZZZAAAZZZZZAAAAAAAZZZZZZAAAAAZZZZZZZZZZZAAAAAAAAAAA
A
AAAAAAAZAAZZAAAZZZZZZZZZZAAAAAAAAZZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAZZAAAAAAAAAZAAAAAAAAAAAZZZZZAAAAAAAAZZZZZAAAAZ
ZZZZZZZAAAAAAAAAAAAAZZZZZZZZZZZZZZZZAAZZZZZZAAAAAAAAAAAZZZZAAAAAAZZZZZAAAAAZZAAAAAAAAAAAAAAAAAAAZZZZZZZZZAAAAAAAAAA
AAAAZZZZZZZAAAZZZZZZAAAZZZZZZZZAAZZZZZZZZZZAAAAAAAAAAAAAAAAAAAAAZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAZZZZZAAAAA
AAAAAAAAAAAAZAAAAAAAAAAAAAAAAZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAZZZZZZZZZZZAAAZZZZZZZZAAAAAAAAAAAAAAAAAAAAZAAAAZZZZZ
ZZZZZZZZZZZZZZZZZAAAAAAAAAAAAZZZZZZZZZZZZZAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAZZZZZZZZZZAAAAAAAAAAAAA
DiscretizationMethodfortheDetectionofLocalExtremaandTrendsinNon-discreteTimeSeries
349
MLRA transformation for

10,000. The choice
of the specific value for the

is not critical since
we have used this for two reasons (a) we have exper-
imentally observed that when the values of a time se-
ries is above 100,000 then the regression lines of the
MLRA are very steep and the identification of the
slopes is more obvious and accurate; and (b) it is pre-
ferred to use multiples of 10 (or power of 10) as

because this transforms the original value to a multi-
ple of 10 and the values remain recognizable. For ex-
ample, with Microsoft’s stock prices in “Fig.2.a” the
values are between 20 and 40 while in the trans-
formed time series with

10,000 the values are
between 200,000 and 400,000 as shown in “Fig.2.d”.
It is easy to translate a transformed value of 278,800
for January 4
th
, 2010 to 27.88 which is the actual
value of the original time series. Moreover, the anal-
ogy between values has not changed since for in-
stance between January 4
th
and January 5
th
, 2010 the
percentage change is 0.036% (from $27.88 to
$27.89), while in the transformed time series the
change is also 0.036% (from 278,800 to 278,900). We
can observe that the time series diagram is exactly the
same except that if we apply a linear regression anal-
ysis in both lines we have a slope and intercept of
10,000 times larger for the transformed time series.
However, the important outcome of the transfor-
mation can be observed in “Fig.2.c” and “Fig.2.e”. In
“Fig.2.c” we have the original MLRA time series
which fluctuates very smoothly around 0 while in
“Fig.2.e” we have the new time series constructed by
the MLRA on the transformed time series. The sec-
ond MLRA time series gives extreme values for the
angles which are mainly close to /2 and /2 with
very few exceptions. In this case, having the values
close to /2 means that we have a positive slope and,
therefore, an uptrend while being close to /2
means a negative slope and, therefore, a downtrend.
In order to verify that the time series characteris-
tics have not changed we can check how the actual
values are changing. This specific transformation of
type 

∗ does not alter the time series in
a way to produce false outcome. The only noticeable
change is the absolute Euclidean distance between the
points. For example, if we have the points (1,1) and
(2,2) they form a line with a slope of 45 degrees with
the x-axis (). If we multiply the y-coordinates
by 10 then we have two new points (1,10) and (2,20)
that form a new line with approximately 84 degrees
slope with the x-axis (10). The only change
is the Euclidean distance between the points which
now is
101
instead of
2. However, when analyz-
ing time series we care mostly about the relative
positions, i.e., how the analogies between the points
stand. In the specific example the change in the first
case is 100% (from 1 to 2) and the same is in the sec-
ond case (from 10 to 20).
Our method although gives direct information
about the trends of the segments of a time series it can
also provide more information. For example, when a
trend changes the specific point has to be either a lo-
cal minimum or a local maximum. Based on this we
can find the actual points in the time series and calcu-
late the time lag  between two changing points
(min-max or max-min) and find also the value change
 (difference of the two points on the y-axis). Based
on these two observations we can calculate the inten-
sity of the trend, i.e., how fast or slow it changes and
towards which direction. For example an upward
change of 100% in 10 days is more intense and im-
portant than the same change over 100 days (“Fig.1”).
Based on the above method, we can discretize the
new MLRA time series using a three letters alphabet,
e.g., A for values in 1, /2 , Z for values in
/2, 1 and O for values in 1,1. Type O val-
ues are very rare and we can eliminate them if we use
a different

value which will create even steeper
linear regression lines. After we have created the new
MLRA time series we will apply ARPaD Algorithm
(Xylogiannopoulos et al., 2014), which is an im-
provement of COV Algorithm (Xylogiannopoulos et
al., 2012; 2014) and allows the detection of all re-
peated patterns in a time series. The ARPaD Algo-
rithm is the only algorithm that can detect all repeated
patterns in a very efficient time. This has been proven
experimentally with the analysis of 100 million deci-
mal digits for each one of the four most famous math-
ematical constants (π, e, φ,
2
) and for which ARPaD
managed to detect all repeated patterns (Xylogi-
annopoulos et al., 2014). After detecting the repeated
patterns we can use a periodicity detection algorithm
(Rasheed et al., 2010) in order to check for periodici-
ties in the previously detected repeated patterns.
4 EXPERIMENTS
For our experiments we used a PC with a double core
CPU at 2.6GHz and 4GB RAM. We have conducted
experiments on the Dow Jones Industrial Average 30
Index for the period from January 4th, 2010 until De-
cember 31st, 2013. We have performed 4 different
experiments using different time intervals and more
specifically we have used MLRA for 10, 20, 30 and
60 days. In “Fig.3.a” we can see the actual DJIA30
time series while in “Fig.3.b” through “Fig.3.d” we
ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems
350
have the different MLRA time series. From the dia-
grams we can make two observations. First, we can
see that indeed the MLRA detects the trends of the
original DJIA30 time series, e.g., in the three shaded
regions we have marked on the diagram. The first two
show an uptrend while the third a downtrend. We can
see that MLRA(10) shows exactly the trends and ad-
ditionally for the second time period (shaded region)
it detects also smaller downtrends (short-term analy-
sis) in the main uptrend. These fluctuations have been
eliminated in MLRA(60) because by taking a larger
time interval we actually eliminate the noise of the
small fluctuations of the index in the specific time pe-
riod (long-term analysis). The second thing we can
observe is that we have a lag between the actual turn-
ing point (local minimum or maximum) and therefore
the change of the trend and the change of the MLRA
values. Actually the larger the MLRA interval the
larger the lag. This is normal since MLRA is the lin-
ear regression of past values in time. The more data
points we use for the linear regression analysis in
MLRA the later the change will be observed. How-
ever, the lag is always the same for each analysis (de-
pending on the time interval). Therefore, when we
conduct a pattern and periodicity detection we have
just to move the turning point detected by the MLRA
specific data points back, according to the length of
the MLRA analysis.
In “Fig.4” we have the transformed time series
constructed by the MLRA(10) process (ten days in-
terval for the DJIA30). With A we have values close
to /2 while with Z we have values close to /2.
By conducting pattern detection with the ARPaD al-
gorithm (Xylogiannopoulos et al., 2014) and perio-
dicity detection with the PDA algorithm (Rasheed et
al., 2010) we have found many long patterns that in-
dicate potential periodicities for forecasting purposes.
In Table 1, we have some indicative results for pat-
terns with periodicity confidence 1 and length equal
to or larger than 20 days. We have included the posi-
tions at which the patterns occur and also calculated
the period of the occurrences. As an example, for the
pattern
“AAAAAAAAAAAAZZZZZZZZZZZZZZZZ” (in-
dex 9) we can see that it starts at position 0 and repeats
with a period of 578 days. It has to be mentioned that
financial time series are referring to working and not
calendar days and, therefore, periods are calculated
over working days too. Moreover, in the specific ex-
periments we have few data prior to January 4
th
, 2010
and therefore the position 0 of MLRA(10) indicates
the angle of the linear regression for the nine last days
of 2009 and January 4
th
, 2010 of DJIA30. If we check
the diagram we can see that DJIA starts at 2010 with
an uptrend of 12 days followed by a downtrend of 16
days. The specific pattern repeats again on April 19th,
2012 (shaded regions (1) and (2) “Fig.5”) and it is ex-
pected to occur again at the end of July 2014. For the
longest repeated pattern we have discovered (index 1)
it occurs for the first time on August 16th, 2010 with
a downtrend for 13 days followed by an uptrend of 36
days. The specific pattern occurs again on September
26th, 2013 (shaded regions (3) and (4) “Fig.5”) with
a period of 783 days and it is expected to occur again
at the beginning of November 2016. Almost a similar
pattern of 48 days (instead of 49) with a downtrend of
12 days (instead of 13) followed by an uptrend of 36
days (index 2) occurs first on November 15th, 2010
(shaded region (5) “Fig.5”) and then again on Sep-
tember 26th, 2013 (shaded region (4) “Fig.5”) with a
period of 720 days and it is expected to occur again at
the middle of July 2016. Another interesting trend is
20 days of uptrend occurring 4 times with a period of
231 days and first occurrence on January 25th, 2011,
second on December 22nd, 2011, third on November
26th, 2012 and the last on October 25th, 2013 (index
10, shaded regions “Fig.6”). For the specific last out-
come, we expect to have the same trend occurring at
the end of September 2014. As we can see, our
method can be used not just for forecasting purposes
of the next data point, but also to make forecasts
longer in the future time.
5 CONCLUSIONS
In this paper, we introduce a new discretization
method for the real values of a time series that allows
detecting local extrema and trends inside the time se-
ries. The proposed method is based on the transfor-
mation of the values of the original time series and
then the construction of a new time series from the
angle of the regression lines that are produced each
time we run a linear regression analysis in a sliding
window of short time intervals. The specific method
can be applied on all kind of real values time series,
e.g., meteorological data, traffic, internet, economic,
etc. We have conducted experiments for different
time intervals of 10, 20, 30 and 60 days on the prices
of Dow Jones Industrial Average 30 Index from the
beginning of 2010 until the end of 2013. After the dis-
cretization and formation of the new time series, we
conducted pattern and periodicity detection. The ex-
perimental results have proven the correctness and
consistency of the method in order to detect trends in
time series and through them to perform forecasting
based on historical data.
Furthermore, as we have discussed, this method
DiscretizationMethodfortheDetectionofLocalExtremaandTrendsinNon-discreteTimeSeries
351
can detect the local minima and maxima and through
them perform deeper analysis of the trends. More spe-
cifically, we can find the intensity of the trend (i.e.
how fast or slow it changes) and the overall perfor-
mance of the trend (i.e., the percentage change from
the minimum to the maximum data point or the rever-
sal). The specific process needs, besides the trend de-
tection, the actual minima and maxima values over
the time series and more calculations on the trends’
data values. Such process will be extensively ana-
lyzed in future work.
REFERENCES
Balkin, S.D., Ord, K., 2000. Automatic Neural Network
Modeling for Univariate Time Series. International
Journal of Forecasting, (1)6,4, pp. 509-515.
Bao, D., 2008. A Generalized Model for Financial Time Se-
ries Representation and Prediction. Applied Intelli-
gence, (29), pp. 1-11, doi: 10.1007/s10489-007-00631.
Bao, D.N., Vy, N.D.K., Anh, D.T., 2013. A hybrid method
for forecasting trend and seasonal time series. 2013
IEEE RIVF International Conference on Computing
and Communication Technologies, Research, Innova-
tion, and Vision for the Future (RIVF), pp.203-208.
Chen, Y.S., Cheng, C.H., Tsai, W.L., 2014. Modeling Fit-
ting-Function-Based Fuzzy Time Series Patterns for
Evolving Stock Index Forecasting. Applied Intelli-
gence, doi: 10.1007/s10489-014-0520-6.
De Gooijer, J.G., Hyndman, R., 2006. 25 Years of Time Se-
ries Forecasting. International Journal of Forecasting,
(22), pp. 443-473.
Edwards, R., Magee, J., Bassetti, W.H.C., 2007. Technical
Analysis of Stocks and Trends. CRC Press. 9
th
edition.
Esling, P., Agon, C., 2012. Time-Series Data Mining. ACM
Computing Surveys (CSUR), (45)1,12.
Gardner, E.S.Jr., McKenzie, E., 1985. Forecasting Trends
in Time Series. Management Science, (31)10, pp. 1237-
1246.
Guerrero, V.M., Galicia-Vasquez, A., 2010. Trend Estima-
tion of Financial Time Series, Applied Stochastic Mod-
els in Business and Industry, (26), pp. 205-223.
Kovalerchuck, B., Vityaev, E., 2000. Data Mining in Fi-
nance: Advances in Relational and Hybrid Methods.
Kluwer Academic Publishers, ISBN 0792378040.
Muhlbauer, A., Spichtinger, P., Lohmann, U., 2009. Appli-
cation and Comparison of Robust Linear Regression
Methods for Trend Estimation. Journal of Applied Me-
teorology and Climatology [1558-8424] Spichtinger,
Peter, (48)9, pp.1961-1970.
Muller, K.R., Smola, J.A., Scholkopf, B., 1997. Prediction
Time Series with Support Vector Machines[C]. In Pro-
ceedings of International Conference on Artificial Neu-
ral Networks, Lausanne, pp. 999-1004.
Pring, M., 2002. Technical Analysis Explained. McGraw-
Hill. New York, NY, 4
th
edition. ISBN 0071226699.
Qin, L.P., Bai, M., 2009. Predicting Trend in Futures Prices
Time Series Using a New Association Rules Algorithm.
16
th
International Conference on Management Science
& Engineering 2009, pp. 1511-1517.
Rasheed, F., Alshalfa M., Alhajj, R., 2010. Efficient Perio-
dicity Mining in Time Series Databases Using Suffix
Trees. IEEE TKDE, (22)20, pp. 1-16.
Selvaratnam, S., Kirley, M., 2006. Predicting Stock Market
Time Series Using Evolutionary Artificial Neural Net-
works with Hurst Exponent Input Windows[C]. Ad-
vances in Artificial Intelligence, pp.617-626.
Xi-Tao, W., 2006. Study on the Application of ARIMA
Model in Time-Bargain Forecast[J]. E-Commerce and
Logistics, 22(15)139-140H.
White, H., Granger, C.W.J., 2011. Consideration of Trends
in Time Series, Journal of Time Series Econometrics,
(3)1,2.
Xylogiannopoulos, K., Karampelas, P., Alhajj, R., 2012.
Periodicity Data Mining in Time Series Using Suffix
Arrays, in Proc. IEEE Intelligent Systems IS’12.
Xylogiannopoulos, K., Karampelas, P., Alhajj, R., 2014.
Exhaustive Patterns Detection In Time Series Using
Suffix Arrays. , manuscript in submission.
Xylogiannopoulos, K., Karampelas, P., Alhajj, R., 2014.
Analyzing Very Large Time Series Using Suffix Ar-
rays, Applied Intelligence, (41)3, pp. 941-955.
Xylogiannopoulos, K., Karampelas, P., Alhajj, R., 2014.
Experimental Analysis on the Normality of pi, e, phi
and square root of 2 Using Advanced Data Mining
Techniques. Experimental Mathematics, (23)2.
Yang, Y., Webb, G., Wu, X., 2005. Data Mining and
Knowledge Discovery Handbook, Chapter 6. Springer.
ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems
352