ISSN 2520-4750 (Online) & ISSN 2521-3040 (Print)

Volume: 3, Issue: 2
Page: 146-166
2019 International Journal of Science and Business

Comparing Clustering Algorithms

using Financial Time-series data
Duangrux Tangsirisakul
Abstract:
Data clustering is one of the most popular unsupervised machine learning
approaches. Clustering data can help identify the pattern of what seems to be
similar data and leads to the best solution for all commercial problems. For
example, taxi booking application, customer’s data can be clustered to match
supply with demand, to detect fraud pattern of an e-commerce transaction or
clustering customers in dating application, etc. In order to carry out the best
calculation of clustering certain requirement is needed in each method and
approach such as the basic assumption of data. When analyzing data with a
wrong assumption, it results in low-quality outcomes. So we would like to study
and compare this type of data in an in-depth manner. Time-series analysis is
used in many future prediction tasks based on previously observed values, IJSB
Accepted 27 March 2019
mixing cluster analysis and time-series data to serve the initial purpose that Published 31 March 2019
researcher would like to share to the public for better understanding of the DOI: 10.5281/zenodo.2617341
clustering, researcher would also like following researchers to refer to this work
and develop this theory and apply in wider issues in future. In this paper, the
focus is on comparing time-series clustering algorithm with financial time-series
data, which is common data such as cryptocurrency, exchange rate currency, the
Shanghai Stock Exchange (SSE50), and the stock exchange of Thailand 50
(SET50). The paper introduces the importance of data mining, machine learning,
and time-series clustering and some related methods, which lays a theoretical
foundation for the formal research of this paper. By analyzing the structure of
time-series clustering, that consists of several parts, including distance
measurement, time-series prototype, a clustering algorithm, and clustering
evaluation. From research result, the hierarchical algorithm is the most efficient
algorithm for unequal length of cryptocurrency series and SSE 50. In another
hand, the partitional algorithm is the most efficient for an equal length of
exchange rate currency and SET 50.

Keywords: time-series clustering, machine learning, dynamic time warping, crypto-

currency

About Author

Duangrux Tangsirisakul, Department of Mathematics, China University of Mining and

Technology, Jiangsu, China.

146 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

1. Introduction
Time-series data is a kind of data that was collected from the data points. It is a continuous
sequence of time such as daily stock data, daily currency exchanged rate, daily temperature
data or cancer growth rate etc. Current time-series data plays an important in research of
various disciplines, such as bioinformatics, robotics, medicine, chemistry, gesture recognition,
speech recognition, tracking, finance, biometrics, astronomy, manufacturing, etc. Data mining
of time-series data is an interesting topic. To analyze and drilling the time-series data of the
relationships, models or insights, which are hidden, useful information based on the
principles of mathematics, statistics, database, recognition and learning of the machine
(Machine Learning such as association rule, classification, prediction, clustering, anomaly
detection, and visualization). This paper raises the clustering analysis to experiment by
comparing 3 scenarios of clustering algorithm with various time-series data (crypto-
currency, exchange rate currency, the Shanghai Stock Exchange and the Stock Exchange of
Thailand 50) using the principle of distance measurement which is Dynamic Time Warping
techniques (DTW), however there are still problems caused by DTW techniques calculation
that is very dynamic, so it would take time to calculate and could not speed up easily. After
finished data clustering, clustering evaluation is used to deciding clustering algorithm
scenarios quality and indicated the dataset that is fit and suitable to which clustering
algorithm scenario.

2. Related work
Recently, the big data, machine learning and AI topic are interested in many industries. This
increasing amount of digital data consequently effects to the role of data analysis as well. In
which reflected in the continuously rising amount to researches related to the clustering
algorithm in part years, (Liao T. W. (2005), Rokach, L. (2009), Ling H E et al. (2007), Nayak J,
Naik B and Behera H S. (2015), Sasirekha, Sasirekha, K. and Baby, P. (2013) and Rui, X., and D.
Wunsch. (2005)). Especially for time-series clustering which is reviewed in-depth on and
processing detail of the theory by Aghabozorgi, S. et al. (2015). All process of time-series
clustering which are distance measurement, time-series prototype, a clustering algorithm,
and cluster evaluation. For distance measurement in time-series clustering, Berndt DJ and
Clifford J proposed dynamic time warping (DTW) to find patterns in time-series data which
time-series clustering by approximate prototypes are propose by Ville Hautamäki et al (2008)
to decide cluster presenter. The other important point of this research is cluster validity
indices (CVIs) that will be explained by Arbelaitz, O. et al. (2013). And in order to complete
this research R software was applied using by Sardá-Espinosa (2018)’s manual. Due to much
interesting in time-series clustering, therefore there are studies on this topic；Tsay, R. S.
(2010) did cluster analysis with time-series data of American unemployment rate
Niennattrakul, V. , & Ratanamahatana, C. A. . (2006) applied time series clustering to compare
the efficiency of multimedia data using representation method and multimedia data using
traditional processing, so this research has demonstrated the profit of time-series
representation method. Saikhamwong N and Rimcharoen S. (2002)did cluster analysis
applied with stock data.

3. Time-series clustering
Time-series clustering consists of several parts, including distance measurement, time-series
prototype, clustering algorithm, and clustering evaluation. Table 1 shows the overall of each
step of time-series clustering. This section briefly describes basic time-series clustering,
which used in this work. This paper, the time-series data has 2 types, equal length and non-
147 International Journal of Science and Business Published By
Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

equal length. Each type, we will compare the clustering algorithm in 3 scenarios. The
cryptocurrency data and the Shanghai Stock Exchange (SSE 50) are non-equal lengths, so the
algorithm which clustering time-series is hierarchical clustering, partitional clustering with k-
medoid and partitional clustering with k-shape. In another data type which is exchange rate
currency and the stock exchange of Thailand (SET 50) that use hierarchical clustering,
partitional clustering with k-medoid and partitional clustering with TADPole. To compare
these scenarios of cluster algorithm whether they are suitable for this dataset or not. We use
clustering evaluation approaches such as Silhouette index, COP index, DB index, DB* index
and CH index. The flow chart of this research framework shown in Figure 1.

Table 1. The overall each step of time-series clustering

Time-series Distance Time-series Clustering Clustering evaluation
representati measuremen prototypes algorithm
ons t
Data adaptive Dynamic Time Mean and Hierarchical Internal evaluation
Warping median clustering Crisp partitions
(DTW) - Silhouette index
- Global DTW - Dunn index
constraints - COP index
- Lower bounds - Davies-Bouldin index
for DTW - Modified Davies-Bouldin
Non-data Global Partition Partitional index
adaptive alignment around cluster - Calinski-Harabasz index
kernel distance medoids (PAM) - k-medoids - Score Function
- TADPole
- k-Shape Fuzzy partitions
Model-based Soft-DTW DTW Fuzzy - MPC index
barycenter clustering - K index
averaging - T index
(DBA) - SC index
Data dictated Shape-based Soft-DTW - PBMF index
distance (SBD) centroid
Shape
extraction
Fuzzy based
prototype External evaluation
Crisp partitions
- Rand Index
- Adjusted Rand Index
- Jaccard Index
- Fowlkes-Mallows
- Variation of Information

Fuzzy partitions
- Soft Rand Index.
- Soft Adjusted Rand Index
- Soft Variation of
Information
- Soft Normalized
Mutual Information

148 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Figure 1. Flow chart of this research framework

3.1 Time-series representation

Time-series representation method is dimension reduction, which was the raw represented
the raw time-series in another space by transforming data to the lower space dimension or by
extracting features. Its advantages are decreasing of the time series dimensionality,
emphasizing on essential shape characteristics, noise management, reduce mandatory
memory and calculation complexity of machine learning reactions.
Definition: Time-series representation, given a time-series data ,
representation transformed the time-series to another dimensionality reduced vector
where and if two series are similar in the initial space, their

149 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

representations will be similar in the transformed spaces. Time-series representation

methods are broken down to four groups (Saeed Aghabozorgi (2015)), such as data adaptive,
non-data adaptive, model-based, and data dictated (clipped data). Data adaptive
representation method is used to minimize the reconstruction error of time-series dataset
using non-equal length segment. Non-data adaptive representation method is suited for time-
series dataset that has equal length segment, while the comparison of representations of
several time-series is very direct. Model-based representation is suitable with time-series
with stochastic results, Statistical models, and Time-series models. Finally, the representation
method is data dictated, the proposal that the condensation-ratio is described based on
original time-series data like in the related work Ratanamahatana, C. (2005)
In this paper, we used the non-data adaptive representation method. Perceptually Important
Point (PIP) (Aghabozorgi, S. et al. (2015) and Tak Chung Fu et al. (2001)) was used for SET50
and exchange rate currency dataset because time-series dataset is equal length segment. This
data adaptive representation method, Piecewise Linear Approximation (PLA) was used for
cryptocurrency and SSE 50 dataset, which each cryptocurrency born time is different so time-
series length is also different.

3.2 Distance measurement

Dynamic Time Warping (DTW)
Dynamic Time Warping is the dynamic programming method, which used for measuring the
similarity between the 2 time-series data. By the results, the distance and the alignment are
the best value between two data, which can stretch and shrink for accommodating variations
in the axis of time that is shown in Figure2. One point can calculate distance with many
points. For example, we have two time-series, a series Y of length a, and series Z of length b.
The gab of two time-series is defined as:

Figure 2. Comparing Dynamic Time Warping distance measurement and Euclidean distance
measurement
The features of the time-series data that is different from generic data, causes many
problems, which are very high dimensional indices (difficulty to identify the locations of data
in high dimension) and other problems that cause by Dynamic Time Warping techniques that
is dynamic leads to long–time of calculation and could not speed up.
Shape-based distance (SBD)
The Shape-Based Distance (SBD) was taken a component of K-Shape clustering algorithm. It
depends on the Cross-Correlation with Coefficient Normalization ( ) sequence between

150 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

two series and is sensitive in scale, so normalized data is the suggestions; the distance
formula is defined as:

3.3 Time-series prototype

Partition Around Medoid (PAM)
The prototyping function could be called time-series averaging also. It is used for solidifying
the concert or to perform time-series categorization. This step is very important for time-
series clustering. Partition Around Medoid approach is the common type of time-series
prototype. This approach is suitable with time-series data which structure is not altered. In
the implementing process, k series data is randomly chosen as initial centroids. Then the
initial centroids and other series calculate distance and series will be appointed to the cluster
of its nearest centroid. And the total minimize distance of series is assigned to the new
centroid. This iteratively calculates for new centroid until no series could change clusters.
Shape extraction
Another part of k-Shape clustering algorithm is shaping extraction. This centroid function
should apply with z-score series data, furthermore, this centroid function can be done
between series with the different length, the shape extraction also applied to multivariate
series or each variable of all series.

3.4 Clustering algorithm

Hierarchical clustering
Hierarchical clustering(Han, J. and Micheline Kamber (2012)) is an algorithm of cluster
analysis. The approaches for hierarchical clustering fallen into 2 types, such as agglomerative
and divisive. The agglomerative approach is a bottom to top direction, all observations start
as one pair of clusters will be combined as together and moving to top hierarchy. The divisive
approach is top to bottom direction, all observations start as single cluster and split. Then
executed in circular moving up the hierarchy. This method can display the result as the
“Dendrogram” graph. Hierarchical algorithm implementation shows as Table 2
Table 2. Hierarchical algorithm
Hierarchical algorithm
Input: D is a dataset containing n objects.
Output: a set of k clusters
Method:
1. Compute the distance of the object. Input data in the distance matrix. 

2. Repeat
3. Examine for two most related clusters or objects.
4. Combine the two clusters or objects to create a cluster which has 2 objects at
least.
5. Update the matrix recalculate the distances between this new cluster and all
other clusters again to until no change. 


Partitional clustering, k-Medoid

Partitional clustering is different from hierarchical that strategically used to create partitions.
The k-shape and k-medoids are the most popular partitional algorithms. For this algorithm,
we must assign the number of the cluster as “k” value for calculation. The optimal k value can
be indicated by using the cluster evaluation procedure. The partitional algorithm with k-
medoid implementation shows as Table 3.
151 International Journal of Science and Business Published By
Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Table 3. The partitional algorithm with k-medoid

The partitional algorithm with k-medoid
Input: k is the amount of clusters, D is a dataset containing n objects.
Output: a set of k clusters
Method:
1. Promptly select k objects from D as the starting representative object or seed;
2. Repeat
3. Designate remaining objects onto the cluster with the closest representative
object;
4. Casually choose one non-representative object, ;
5. Calculate total cost, S, of switching representative object, , with ;
6. If then switch with to the new set of the representative
object until no change

Partitional clustering, k-Shape

Paparrizos, J. , and Gravano, L. . (2016) researched the k-Shape clustering algorithm was by in
2016. This partitional algorithm needs the custom distance measure as SBD and the custom
centroid function as shape extraction. It also requires z-normalization to default data. The
partitional algorithm with k-shape implementation shows as .

Table 4.

Table 4. The partitional algorithm with k-shape

Partitioning algorithm with k-shape
Input: k is the number of clusters, D is a dataset containing n objects.
Output: a set of k clusters
Method:
1. Promptly select choose k objects from D as the starting representative object or
seed;
2. repeat
3. Base on the mean value of the objects in the cluster, designate remaining
objects onto the cluster with a similar object
4. Recomputed the mean value of the objects in each cluster to update the
cluster until no change

Partitional clustering, TADPole

The TADPole clustering was proposed by Begum, N. et al. (2015), this time-series clustering
use DTW as the distance measurement and use PAM as the centroid. TADPole depends on the
DTW bounds defining for time-series with equal length only.

3.5 Cluster evaluation

The cluster evaluation can be split into two types such as external and internal evaluations. If
we know the group truth, we can use external evaluation to recheck the clustering with the
group truth and re-measure without the group truth, we can use the internal evaluation to
measure separating of cluster. In this paper, we will use Silhouette index, COP index, Davies-
Bouldin index, Davies-Bouldin star index, and Calinshi-Harabasz index of internal evaluation
to compare time-series clustering.

152 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Silhouette index
Silhouette index is the internal validation of consistency within cluster data. Some algorithm
needs to know k-value before, therefore silhouette is used to measure k-value which control
the amount of clusters in a dataset and relationship between objects in the dataset. The
silhouette coefficient value should always be the maximum number after measurement.
Definition: The dataset , of n objects, suppose is segregated to k clusters, . For
each object , we compute as the average distance of and other objects in the
cluster where belongs. Likewise, is the minimum average distance from to call
clusters to where does not belongs. Formally, suppose ; then

and

The Silhouette coefficient (Han, J. (2005)) of is then defined as

COP index
It was first presented to adopted in the partnership of a cluster hierarchy post-processing
algorithm, Traditional cluster validity indices could also adopt this. It is a ratio-type index,
where the coherence is indicated by the distance from the position in a cluster to its centroid.
The partition establish based on the furthest border width is defined as:

Davies-Bouldin index (DB)

This index is commonly used in cluster validity comparison studies. It assesses the coherence
formed on the distance from the positions in a cluster to centroid and partition based on the
distance of centroids. It is defined as

where

Davies-Bouldin star index (DB star)

This index is the variation of the Davies-Bouldin index. It is defined as

Calinshi-Harabasz index (CH)

This ratio index cohesion is predicted from the distance from the point in a cluster to its
centroids. Its partition is a measurement of the distance from the centroids to the global
centroid. This can be defined as:
153 International Journal of Science and Business Published By
Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

4. Experiment result
4.1 Dataset
This paper has 4 dataset experiments, which are cryptocurrency dataset, SSE 50 dataset,
exchange rate currency dataset, and SET 50 dataset. Dataset structure detail shows as Table
5.
Table 5. Dataset structure detail
Dataset Length No. No. Data Points
Currency / Stock
1,866
Cryptocurrency, Zone A 154 234,949
(Unequal length)
1,096
Cryptocurrency, Zone B 297 130,663
(Unequal length)
522
Cryptocurrency, Zone C 1,192 265,547
(Unequal length)
2,435
SSE50 43 99,353
(Unequal length)
92
Exchange rate currency 146 13,432
(Equal length)
244
SET50 50 12,200
(Equal length)

Cryptocurrency dataset
Cryptocurrency dataset is all historical closing price of all cryptocurrencies from the Kaggle
website (www.kaggle.com/jessevent/all-crypto-currencies/home). All datasets have 631,159
observations, the used variables are currency, data, closing price and period of dataset
between 28, April 2013 to 21, May 2018. But each cryptocurrency has a different length,
therefore we split cryptocurrency to 3 zones; zone A has 154 cryptocurrencies that have
1,866 lengths of 234,949 data points, zone B has 297 cryptocurrencies that have 1,096
lengths of 130,663 data points, and zone C has1,192 cryptocurrencies that have 522 lengths
of 631,159 data points. Time-series dataset of each zone shows as Figure 3.

154 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Figure 3. Time-series of all cryptocurrency and breaking down as zone A, B, and C

Shanghai Stock Exchange 50 Index (SSE 50)

Shanghai Stock Exchange, SSE 50 Index, is a capitalization-weighted index based on the top
50 stocks listed in China SSE index. It has high market capitalization and high liquidity. SSE 50
dataset is the historical closing price for 50 stock indices from finance yahoo.com website
(https://finance.yahoo.com). But some stock indices have too much of null value, so we cut 7
stock indices off. All datasets have the unequal length of 99,353 data points of 43 stock
indices; the used variables are currency data, closing price and period of dataset between 15,
December 2008 to 15, December 2018, shows as Figure 4.

Figure 4. Time-series of SSE 50 stock index

Exchange rate currency dataset

Exchange rate currency dataset is the historical closing price for all currencies in the world
that has 92 currencies from finance yahoo.com website (https://finance.yahoo.com). All
datasets have the same length of 13,432 data points; the used variables are currency, data,
closing price, and period of dataset between 1, July 2018 to 30, September 2018, shows as
Figure 5.

Figure 5. Time-series of all exchange rate currency

155 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

The Stock Exchange of Thailand 50 (SET 50) dataset

The stock exchange of Thailand, SET 50 Index, is a capitalization-weighted index based on the
top 50 stocks listed in Bangkok. Thailand SET index has high market capitalization and high
liquidity. SET 50 dataset is the historical closing price of 50 stock indices from investing
website (www.investing.com). All dataset has the same length of 12,200 data points; the used
variables are currency data, closing price and period of dataset between 1, October 2017 to
30, September 2018, shows as Figure 6.

Figure 6. Time-series of SET 50 index stock

4.2 Cluster algorithm

This paper will break down 3 scenarios for each data type of experiment. The scenario of the
clustering algorithm detail describes in Table 6. The hierarchical and partitional algorithm
was needed to normalize data and cluster evaluating to find the best number of cluster (k-
value).
Table 6. The scenario of the clustering algorithm by a column of distance measurement and
centroid
Distance Clustering
Scenario Centroid Data type
measurement algorithm
hc DTW PAM Hierarchical Equal or non equal length
Partitional with
pc_dtw DTW PAM Equal or non equal length
k-medoid
Shape Partitional with
pc_sbd SBD Non-equal length only
extraction k-shape
Partitional with
pc_tp DTW PAM Equal length only
TADPole

To evaluate the clustering algorithm, we compare 3 scenarios of the clustering algorithm,

time series representation, a common image clustering algorithm, and clustering by the color
of currency values.

4.3 The time-series clustering experiment result

The cryptocurrency clustering result
Before running the clustering algorithm, we have to prepare time-series data by doing
representation methods to reduce noise and normalize time-series. To evaluate the number
of clusters (k-value), we calculate Silhouette coefficient value to identify k-value of each
clustering algorithm of zone A, B, and C that show as Figure 7, Figure 8, and Figure 9
respectively. In zone A, the k-value of hierarchical and partitional with k-medoid scenario
156 International Journal of Science and Business Published By
Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

resulted in for 2 clusters as best, while the partitional with k-shape is 4 clusters. The series
and centroid result of the hierarchical scenario, the partitional with the k-medoid scenario,
and the partitional with k-shape scenario are shown as Figure 10, Figure 11, and Figure 12
respectively.

Figure 7. Cryptocurrency, zone A of Silhouette index of each scenario algorithm and each k
clusters

Figure 8. Cryptocurrency, zone B of Silhouette index of each scenario algorithm and each k
clusters

Figure 9. Cryptocurrency, zone C of Silhouette index of each scenario algorithm and each k
clusters

Figure 10. The series (left) and centroid (right) result of the hierarchical scenario of
cryptocurrency time-series, zone A

157 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Figure 11. The series (left) and centroid (right) result of partitional with the k-medoid
scenario of cryptocurrency time-series, zone A

Figure 12. The series (left) and centroid (right) result of partition with the k-shape scenario of
cryptocurrency time-series, zone A

In zone B, the k-value of hierarchical and partitional with k-shape scenario resulted in 2
clusters as best, while the partitional with k-medoid is 4 clusters. The series and centroid
result of the hierarchical scenario, the partitional with the k-medoid scenario, and the
partitional with k-shape scenario are shown as Figure 13, Figure 14, and Figure 15
respectively.

Figure 13. The series (left) and centroid (right) result of the hierarchical scenario of
cryptocurrency time-series, zone B

158 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Figure 14. The series (left) and centroid (right) result of partitional with the k-medoid
scenario of cryptocurrency time-series, zone B

Figure 15. The series (left) and centroid (right) result of partitional with the k-shape scenario
of cryptocurrency time-series, zone B.

In zone C, The k-value of hierarchical, partitional with k-shape and partitional with k-medoid
scenario resulted in 2 clusters as best. The series and centroid result of the hierarchical
scenario, the partitional with the k-medoid scenario and the partitional with k-shape scenario
are shown as Figure 16, Figure 17, and Figure 18 respectively.

Figure 16. The series (left) and centroid (right) result of the hierarchical scenario of
cryptocurrency time-series, zone C

Figure 17. The series (left) and centroid (right) result of partitional with the k-medoid
scenario of cryptocurrency time-series, zone C

Figure 18. the series (left) and centroid (right) result of partitional with the k-shape scenario
of cryptocurrency time-series, zone C

The Shanghai Stock Exchange 50 Index (SSE 50) clustering result

For Shanghai stock exchange 50 indices, the k-value of hierarchical, partitional with k-shape,
and partitional with k-medoid scenario resulted in 3, 2, and 4 clusters respectively as best,
the result showed as Figure 19. The series and centroid result of the hierarchical scenario, the
159 International Journal of Science and Business Published By
Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

partitional with the k-medoid scenario, and the partitional with k-shape scenario are shown
as Figure 20, Figure 21, and Figure 22 respectively.

Figure 19. The Shanghai Stock Exchange 50 Index (SSE 50) of Silhouette index of each
scenario algorithm and each k clusters

Figure 20. The series (left) and centroid (right) result of hierarchical scenario of SSE 50

Figure 21. The series (left) and centroid (right) result of partitional with k-medoid scenario of
SSE 50

Figure 22. The series (left) and centroid (right) result of partition with the k-shape scenario of
SSE 50

The exchange rate currency clustering result

For the exchange rate currency, the k-value of all scenarios resulted in 2 clusters as best, the
result shown as Figure 23. The series and centroid result of the hierarchical scenario, the
partitional with the k-medoid scenario, and the partitional with k-shape scenario are shown
as Figure 24, Figure 25, and Figure 26 respectively.
160 International Journal of Science and Business Published By
Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Figure 23. The exchange rate currency of Silhouette index of each scenario algorithm and
each k clusters

Figure 24. The series (left) and centroid (right) result of hierarchical scenario of the exchange
rate currency

Figure 25. The series (left) and centroid (right) result of partitional with the k-medoid
scenario of the exchange rate currency

Figure 26. The series (left) and centroid (right) result of partitional with TADPole scenario of
the exchange rate currency

The Stock Exchange of Thailand 50 (SET 50) clustering result

For SET 50, The k-value of hierarchical and partitional with k-medoid scenario resulted in 2
clusters as best, the partitional with k-shape is 4 clusters, the result showed as Figure 27. The
series and centroid result of the hierarchical scenario, the partitional with the k-medoid
scenario and the partitional with TADPole scenario are shown as Figure 28 Figure 29 and
Figure 30 respectively.

161 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Figure 27. SET 50 of Silhouette index of each scenario algorithm and each k clusters

Figure 28. The series (left) and centroid (right) result of the hierarchical scenario of SET 50

Figure 29. The series (left) and centroid (right) result of partitional with the k-medoid
scenario of SET 50

Figure 30. The series (left) and centroid (right) result of partitional with TADPole scenario of
SET 50

4.3 Comparing time-series clustering result

To evaluate our proposed method, we run 3 zones of cryptocurrency series using 3 clustering
algorithms such as the hierarchical scenario, the partitional with the k-medoid scenario, and
the partitional with k-shape which are evaluated by Silhouette index, COP index, DB index, DB
star index, and CH index.
Table 7,
Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.45 0.41 0.70 0.70 74.62
DTW 0.44 0.39 0.91 0.91 148.81
PC_k-Shape 0.23 0.24 1.30 1.39 39.93

162 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Table 8 8, 9 and 10 report the clustering evaluation value that represent in the same direction,
the hierarchical scenario is the most effective for the unequal dataset; cryptocurrency and
SSE 50 time-series. For exchange rate currency and SET 50 which are the equal length time-
series, their evaluation values are shown as
Table 11 and
Table 12 respectively. The most effective algorithm with exchange rate currency and SET 50
is partitional with TADPole and partitional with k-medoid respectively.

Table 7. Cryptocurrency, zone A clustering evaluation of each scenario

Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.45 0.41 0.70 0.70 74.62
DTW 0.44 0.39 0.91 0.91 148.81
PC_k-Shape 0.23 0.24 1.30 1.39 39.93

Table 8. Cryptocurrency, zone B clustering evaluation of each scenario

Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.45 0.53 0.89 0.89 16.28
PC_DTW 0.18 0.28 1.96 2.20 103.47
PC_k-Shape 0.24 0.37 1.28 1.28 144.10

Table 9. Cryptocurrency, zone C clustering evaluation of each scenario

Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.51 1.13 0.31 0.31 3.25
PC_DTW 0.35 0.35 1.51 1.51 1,075.66
PC_k-Shape 0.27 0.30 1.85 1.85 550.91

Table 10. SSE 50 clustering evaluation of each scenario

Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.41 0.30 0.99 1.00 23.52
DTW 0.35 0.49 1.54 1.54 28.76
PC_k-Shape 0.10 0.59 2.42 2.88 11.56

Table 11. Exchange rate currency clustering evaluation of each scenario

Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.31 0.70 1.02 1.02 16.78
PC_DTW 0.28 0.46 1.15 1.15 111.73
PC_TADPole 0.42 0.98 0.72 0.72 13.38

Table 12. SET 50 clustering evaluation of each scenario

Sil COP DB DBStar CH
(max) (min) (min) (min) (max)
HC 0.36 0.43 1.33 1.33 26.59
PC_DTW 0.37 0.43 1.19 1.19 25.41
PC_TADPole 0.22 0.55 2.82 3.18 14.72

163 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

5. Conclusion
According to Table 13, which we have demonstrated the clustering time-series of some kind
of financial time-series. We use 4 time-series datasets, which can split into 2 data type (equal
and unequal length). This experiment, comparing time-series clustering using 3 scenarios of
cluster algorithm for each time-series data set and evaluating clustering algorithm using 5
indices to identify the validity of each clustering algorithm. From research result, the
hierarchical algorithm is the most efficient algorithm for unequal length of cryptocurrency
series and SSE 50. In another hand, the partitional algorithm is the most efficient for an equal
length of exchange rate currency and SET 50.

Table 13. Conclusion of research

Dataset Data type The Best No. of Cluster
Algorithm
Cryptocurrency, Zone A Unequal length Hierarchical 2
Cryptocurrency, Zone B Unequal length Hierarchical 2
Cryptocurrency, Zone C Unequal length Hierarchical 2
SSE50 Unequal length Hierarchical 3
Partitional with
Exchange rate currency Equal length 2
TADPole
Partitional with
SET50 Equal length 2
k-medoid

Exploring more on the source of the result, some element of the unequal length data might
contribute to the result such as the fact that original data is very large in both the data period
and type of currency dimensions. when we execute and cluster data, it is hard to do it in equal
length manner, where SSE50, includes 50 stock indices and data period is 10 years while
cryptocurrency, includes 643 currencies in 5 years. Therefore, the hierarchical algorithm is
suitable with the larger and high dynamic variant datasets. On another hand, SET50 includes
50 stock indices with fixed data only period of 1 year while exchange rate currency in the
study includes 92 currencies around the world within 3-month period. With these low
amount and much less variant dataset, a partitional clustering algorithm is more suitable. We
hope our research would be good motivation for researchers to study further on time-series
clustering and its application in much wider area including biometric clustering, exchange
rate currency and stock clustering to manage the trading portfolio.

Acknowledge
Firstly, I would first like to thank my thesis advisor Professor Li Jin Yu of the School of
Mathematics at China University of Mining and Technology. The door to Prof. Li office was
always open whenever I ran into trouble or had a question about my research. His opened mind
allows me to put my curiosity at its best and push this workpiece beyond its stand by steering me
on in the right direction and provide constructive comment to it. I would also like to
acknowledge Mr. Pakorn Leewasuthorn, my forever friend and Ms. Zhu Lin my senior apprentice
sister at faculty educational companion as the second reader of this thesis. I am gratefully
indebted to his/her for very valuable comments on this thesis. I would also like to acknowledge
Jiangsu Province and China University of Mining and Technology who supported scholarship and

164 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

this valuable opportunity to study in China. I would also like to acknowledge my senior
apprentice Zhu Lin, Zhao Xing Wei, Xu Yuan Yuan, Chinese classmates, Wang Pei, Dong Min Jie,
Yang Jia Hui, Cheng Li, Aommy, Jo Jo, Maylo, all my Chinese and International friends that gave
me long last friendship ， and emotional support, take care and made my life in China
meaningful. Finally, I must express my very profound gratitude to my parents and older
brothers, my grandma, and my aunt for providing me with unfailing support, give me freedom
and continuous encouragement throughout my years of study and through the processes of
researching and writing this thesis. This accomplishment would not have been possible without
them. Thank you.

Reference
Ville Hautamäki, Pekka Nykänen, & Pasi Fränti. (2008). Time-series Clustering by
Approximate Prototypes. International Conference on Pattern Recognition. IEEE.
Niennattrakul, V. , & Ratanamahatana, C. A. . (2006). Clustering Multimedia Data Using Time
Series. International Conference on Hybrid Information Technology. IEEE.
Gullo, F. , Ponti, G. , Tagarelli, A. , Tradigo, G. , & Veltri, P. . (2012). A time series approach for
clustering mass spectrometry data. Journal of Computational Science, 3(5), 344-355.
Izakian H, Pedrycz W, Jamal I. (2015). Fuzzy clustering of time series data using dynamic time
warping distance. Engineering Applications of Artificial Intelligence, 39, 235-244.
Liao, T. W. . (2005). Clustering of time series data—a survey. Pattern Recognition, 38(11),
1857-1874.
PA Wang, W. , & Zhang, Y. . (2007). On fuzzy cluster validity indices. Fuzzy Sets and Systems,
158(19), 2095-2117.
Berndt DJ, Clifford J. (1994). Using dynamic time warping to find patterns in time series. In
KKD workshop, 10(1), 359-370.
Rokach, Lior, and Oded Maimon.(2005). "Clustering methods." Data mining and knowledge
discovery handbook. Springer US, 321-352.
Begum, N. , Ulanova, L. , Wang, J. , & Keogh, A. E. . (2015). Accelerating dynamic time warping
clustering with a novel admissible pruning strategy.
Ratanamahatana, C., Keogh, E., Bagnall, A. J., & Lonardi, S. (2005). A Novel Bit Level Time
Series Representation with Implication of Similarity Search and Clustering. Pacific-asia
Conference on Advances in Knowledge Discovery & Data Mining.
HHan, J. (2005). Data Mining: Concepts and Techniques.
Paparrizos, J. , & Gravano, L. . (2016). K-shape: efficient and accurate clustering of time series.
ACM SIGMOD Record, 45(1), 69-76.
Aghabozorgi, S. , Shirkhorshidi, A. S. , & Wah, T. Y. . (2015). Time-series clustering - A decade
review. Elsevier Science Ltd.
Tak Chung Fu - F.L. Chung - Robert Wing Pong Luk - Vincent T. Y. NgVincent. (2001). Flexible
time series pattern matching based on perceptually important points. JT Conference on
Artificial Intelligence Workshop, 1-7
Arbelaitz, O. , Gurrutxaga, I. , Muguerza, J. , Jesús M. Pérez, & Iñigo Perona. (2013). An
extensive comparative study of cluster validity indices. pattern recognit. Pattern
Recognition, 46(1), 243-256.
Hubert, L., & Arabie, P. (1985). Comparing partitions. Journal of Classification, 2(1), 193-218.
Sardá-Espinosa. (2018). Comparing Time-Series Clustering Algorithms in R using the
dtwclust Package - Retrieved from https://cran.r-
project.org/web/packages/dtwclust/vignettes/dtwclust.pdf

165 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com
 Volume: 3, Issue: 2, Year: 2019 Page: 146-166 IJSB-International

Tsay, R. S. . (2010). Multivariate Time Series Analysis and Its Applications. Analysis of
Financial Time Series, Second Edition. John Wiley & Sons, Inc.
Rokach, L. . (2009). A survey of clustering algorithms. Data Mining & Knowledge Discovery
Handbook, 16(3), 269-298.
Ling H E , Ling-Da W U , Yi-Chao C . (2007). Survey of Clustering Algorithms in Data Mining.
Application Research of Computers, 24(1), 10-13.
Nayak J, Naik B, Behera H S. (2015). Fuzzy C-Means (FCM) Clustering Algorithm: A Decade
Review from 2000 to 2014. Computational Intelligence in Data Mining, 2(1).
Sasirekha, K. , & Baby, P. . (2013). Agglomerative hierarchical clustering. Electronic Design,
17-17.
Rui, X., & D. Wunsch. (2005). Survey of clustering algorithms. IEEE Transactions on Neural
Networks, 16(3), 645-678.
Saikhamwong N - Rimcharoen S.(2002).K-Mean Clustering of the Stock Exchange of Thailand
50 (SET50) for Portfolio Diversification. 11th International Conference on e-Business
(iNCEB 2013).

Cite this article:

Tangsirisakul, D. (2019). Comparing Clustering Algorithms using Financial Time-
series data. International Journal of Science and Business, 3(2), 146-166. doi:
https://doi.org/10.5281/zenodo.2617341

Retrieved from http://ijsab.com/wp-content/uploads/334.pdf

Published by

166 International Journal of Science and Business Published By

Email: [email protected] Website: ijsab.com