Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Fuzzy time series based on frequency density-based partitioning and k-

means clustering for forecasting exchange rate
To cite this article: U S Mukminin et al 2021 J. Phys.: Conf. Ser. 1943 012119

View the article online for updates and enhancements.

This content was downloaded from IP address 158.46.170.233 on 15/07/2021 at 04:13

ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Fuzzy time series based on frequency density-based

partitioning and k-means clustering for forecasting exchange
rate

U S Mukminin1, B Irawanto2, B Surarso2, Farikhin2

1
Undergraduate Program of Mathematics, Faculty of Science and Mathematics,
Diponegoro University, Semarang, Indonesia
2
Department of Mathematics, Faculty of Science and Mathematics, Diponegoro
University, Semarang, Indonesia

Corresponding author : [email protected],

[email protected]

Abstract. Fuzzy Time Series (FTS) has been growing rapidly in recent years. There are many
models that were developed. In this paper, we propose a new method to forecast exchange rate
data by combining some models. Firstly, we use the average-based interval to make optimal
interval numbers. Secondly, we use frequency density-based partitioning for optimal
partitioning. In this part, we divide the three highest frequency of intervals into four, three, and
two sub-intervals, respectively, and discarding intervals if there is no data distributed. And
thirdly, we use k-means clustering to construct the Fuzzy Logical Relationship Group (FLRG).
We divide Fuzzy Logical Relationship (FLR) into 16 initial clusters. Then we evaluate model by
calculating the error value using MSE (Mean Squared Error) and AFER (Average Forecasting
Error Rates). The study case of this paper is daily exchange rate data (USD to IDR) started from
January until May with its unstable fluctuation caused by Pandemic Covid-19. The study aims
to obtain a forecasting model of exchange rate data as the preparation and evaluation for future
conditions.

1. Introduction
Fuzzy time series method was firstly introduced by Song and Chissom [1] in 1993. They introduced a
method for forecasting time series data in the future. The data that is used should be collected
periodically based on days, months, years.
This method is growing rapidly in recent years. Some researches that were using fuzzy time series
for forecasting TAIEX are [2] that using particle swarm optimization algorithm for optimal partition,
and k-means clustering algorithm to build fuzzy logical relationship group. Another method for
forecasting TAIEX is [3] Another developing application of fuzzy time series researched for forecasting
stock index [4].
In this paper, we modify fuzzy time series for the forecasting exchange rate (USD to IDR) using
frequency density-based partitioning for optimal partitions [5], using average-based interval for making
an optimal interval [6], and modify the model by constructing the Fuzzy Logical Relationship Group
(FLRG) using k-means clustering method [2].
Study case of this paper is the daily exchange rate of USD to IDR from January-May 2020 that we
got from www.bi.go.id. We also evaluate the forecasting data by calculating MSE (Mean Squared Error)
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

value and AFER (Average Forecasting Error Rate) value. It aims to obtain a forecasting model of
exchange rate data as preparation and evaluation for future conditions.

2. Fuzzy Time Series

The roughly fuzzy set can be interpreted as a number class with vague boundaries. Let U be a universe
of discourse, where U = u1 , u2 , u3 , , un  , then a fuzzy set Ai in U with the membership function
generally represented as follows :
Ai =  Ai (u1 ) / (u1 ) +  Ai (u2 ) / (u2 ) + +  Ai (u p ) / u p (1)

Where  Ai is a membership function of fuzzy set Ai ,  Ai ( ui )  0,1 and 1  i  p .

Based on FTS that proposed by Song and Chissom [1], fuzzy time series is following concepts:

Definition 2.1 [7]:

A fuzzy set A of exchange rate X, where X =  x1 , x2 , , xn  defined as A =  f A ( xi ), i = 1,2, , n where
f A is a membershp function of fuzzy set A, f A : X →  0,1 , f A ( xi ) denotes membership degree
of xi in fuzzy set A, 1  i  n .

Definition 2.2 [5] :

Let F ( t ) be a forecasted data of t time that is caused by F ( t − 1) which is noted by F ( t − 1) → F ( t )
so, there is fuzzy relationship that is denoted F ( t ) = F ( t − 1) R ( t , t − 1) . Where R ( t , t − 1) is a
fuzzy relationship and “ ” is the max min composition operator.

Definition 2.3 [7]:

The membership function of fuzzy triangular is represented by 3 fuzzy parameters a = ( a1 , a2 , a3 )
where a1 , a2 , a3  . It curve is a combination of functions of the equations :
 x − a1
 a − a , a1  x  a2
 3 1
 a −x
a ( x) =  3 , a2  x  a3 (2)
 a3 − a2
 0 else



3. K-Means Clustering Algorithm

K-means clustering that firstly proposed by MacQueen [8] is one of the non-hierarchical data grouping
method that can partition data into two or more groups. This method will partition the data into a group
where the data with the same characteristics will be included in the same group, while the data that has
different characteristics will be grouped into other groups. The K-Means algorithm is one of the most
widely used algorithms in grouping because of simplicity and efficiency [9] and is recognized as one of
the top 10 data mining algorithms by IEEE International Conference on Data Mining (ICDM) [10].

K-Means Clustering firstly was proposed by [8] with the following concepts :

2
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Definition 2.4 [2] :

Let X =  x1 , x2 , x3 , , xn  be a set of data points and let C = c1 , c2 , c3 , , ck  be a set of cluster
centers, where k is a number of cluster. Then, it aims to minimizing an objective function J (Squared
Error function) given by equation :
nj

( )
k
J ( C ) =  xi − c j
2
(3)
i =1 j =1

where xi − c j is Euclidean distance, n j is the number of data points of cluster j, and k is the number
of clusters.

Here the steps for forecasting data using frequency density based partitioning [7] and k-means clustering
[2] :

Step 1 : Determine universe discourse U. Then calculate the interval data using average-based
interval [9]. Let X min and X max be the minimum and maximum datum of historical data, respectively.
Then we define universe of discourse U as  X min − D1 , X max − D2  where D1 and D2 are two proppers
possitive numbers. Then partition the universe of discourse U into several intervals u1 , u2 , , un using
average-based interval [6]. It follows several steps as follows :
1. Calculate the absolute mean of data


n
i
X = i =1
(4)
n
Where,
X : Absoluted mean value of data
i : Absoluted difference value of i
2. Calculate range value
Range(r ) = X max − X min
3. Calculate basis by following form :
X
Basis =
2
4. Calculate the number of intervals
Range
I=
basis
Step 2 : Divide into sub-interval based on frequency density. Interval with the three highest frequency
are divided into four, three, and two sub-intervals equal length, respectively. If there’s no data distributed
in the intervals, it will be discarded [5].
Step 3 : Based on step 2, define fuzzy sets A1 , A2 , , A46 and define the membership function of
triangular fuzzy as shown as eq. 1.
Step 4 : Construct FLR and determine FLRG by using k-means algorithm [8]. Each group cluster
has some fuzzy set members. We assume that FLR are shown as follows :

3
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Am1 → An1
Am2 → An2
Am3 → An3

Am p → An p
In FLR above we know that Am1 , Am2 , Am3 , , Amp are called current state and An1 , An2 , An3 , , An p are
called next state. Then we cluster the subscripts of current state fuzzy sets that shown as
m1 , m2 , m3 , , m p into several groups/ clusters G1 , G2 , G3 , Gk
Step 5 : Calculate the distances d1 , d 2 , d3 , , d k between y and center clusters by the following
equation :
di = y − Ci (5)
Where 1  i  k , y is a subscript of fuzzy set Ay , and C is cluster center. Minimum distances from each
y to the center cluster show that y is the member of the cluster. Then re-calculate the center of new
clusters with new members until there is no data is reassigned.
Step 6 : Calculate the forecasted fuzzy set Ay using cluster centers as a subscript of the final
forecasted fuzzy set Ay . Then calculate the result forecasted data using centroid method by the following
forms [7]:

 1 + 0,5
 , if j =1
 1 0,5
  +
 j j +1

 0,5 + 1 + 0,5
tj =  , if 2  j  n −1 (6)
 0,5 + 1 + 0,5
  j −1  j  j +1

 0,5 + 1
, if j=n
 0,5 1
  +
 j −1 j

Where t j is forecasted data, and  j −1 , j , j +1 are the middle value of interval forecasted fuzzy based
on k-means clustering A j −1 , A j , A j +1 , respectively.
Step 7 : Calculate error value using MSE (Mean Square Error) and AFER (Average Forecasting Error
Rates) by the following forms [7] :
 ( X i − Fi )
n 2

MSE = i =1
n

X i − Fi

AFER =
 100%
Xi

n
Where X is an actual data, F is a forecasted data, and 1  i  99 with criteria of AFER is shown as
table 1.

4
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Table 1. Criteria AFER

AFER Value Criteria
<10% Very good
10% - 20% Good
20% - 50% Good enough
>50% Poor

4. Numerical Simulation
In this session, we demonstrate the proposed method for exchange rate data of IDR and USD from
January 2020 until May 2020 which can be accessed in www.bi.go.id. Here the step-by-step to forecast
data.

Step 1 : Define the universe of discourse U. From the data, we can see that X min = 13680,06 and
X max = 16824,71 with D1 and D2 are 680,06 and 175,29 respectively. So, we get U = (13000,17000 ) .
From the calculation to find the optimal interval based on average-based interval [6], we get the universe
of discourse U partition into 71 intervals.

Step 2 : We divide intervals into 46 sub-interval based on frequency density-based partitioning [5].
Table 2 shows sub-intervals of data.
Table 2. Sub-interval
Sub-Intervals
u1 = [13672;13686] u2 = [13686;13700] u3 = [13700;13714] u4 = [13714;13728]
u5 = [13728;13764,67] u6 = [13765;13784] u7 = [13784;13812] u8 = [13812;13840]
u9 = [13840;13896] u10 = [13896;13952] u11 = [13952;13980] u12 = [13980;14008]
u13 = [14008;14064] u14 = [14064;14120] u15 = [14232;14288] u16 = [14288;14344]
u17 = [14344;14400] u18 = [14400;14456] u19 = [14456;14512] u20 = [14512;14568]
u21 = [14792;14848] u22 = [14848;14904] u23 = [14904;14960] u24 = [14960;15016]
u25 = [15016;15072] u26 = [15072;15128] u27 = [15128;15184] u28 = [15184;15240]
u29 = [15296;15352] u30 = [15464;15520] u31 = [15520;15576] u32 = [15576;15632]
u33 = [15632;15688] u34 = [15688;15744] u35 = [15744;15800] u36 = [15800;15856]
u37 = [15856;15912] u38 = [15912;15968] u39 = [16304;16360] u40 = [16360;16416]
u41 = [16416;16472] u42 = [16472;16528] u43 = [16528;16584] u44 = [16584;16640]
u45 = [16640;16696] u46 = [16808;16864]

Step 3 : Define fuzzy sets A1 , A2 , , A46 and define membership function of triangular fuzzy. For
Example, membership function of u1 can be made from interval [13672;13686] shown as figure 1.
 x − 13672
13679 − 13672 , if 13672  x  13679

 13686 − x
a ( x) =  , if 13679  x  13686 (7)
13686 − 13679
 0, else



5
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Figure 1. Membership function of triangular fuzzy A1

Step 4 : Construct FLR and determine FLRG by k-means clustering algorithm [8]. The FLR are
divided into 16 initial clusters. For example, we have FLRG of Group G1 , which has membership FLR
as follows :
Group G1 : A11 → A11
A11 → A13
A13 → A12
A12 → A12
A12 → A10
A10 → A19
By the same way, we can determine FLRG based on k-means clustering of 16 initial clusters.

Step 5 : Calculate the distances of each subscript of the current state of fuzzy sets. Re-calculate the
center of new clusters with new members of each cluster until there is no data is reassigned. Table 3
shows the final cluster centers.
Table 3. Final cluster centers and its members
Clusters C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16
Centers 11 4 2 5 7 7 14 17 24 41 44 38 33 29 26 21
Number
of 8 9 4 7 4 5 5 7 5 5 5 7 9 5 6 8
members

Step 6 : Calculate the forecasted fuzzy set Ay using cluster centers as a subscript of the final
forecasted fuzzy set Ay using form 6.

Step 7 : Calculate error value using MSE (Mean Squared Error) and AFER (Average Forecasted
Error Rate). From the error calculation, we get MSE value 9197,613 and AFER value 0,414755% that
represented by table 4. Fig. 2 shows a comparison graphic between actual data and forecasted data.

6
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

18000.00
16000.00
14000.00
12000.00
10000.00
8000.00
6000.00
4000.00
2000.00
0.00
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97

Actual data Forecasted data

Figure 2. Graphic of forecasting result

Table 4. Forecasting result and error value

Forecasted Data X i − Fi
Actual Data ( X i ) ( X i − Fi )
2
Day
( i)
F Xi
1 13964,48
2 13968,50 13962,46 36,4816 0,000432
3 14030,81 14119,5 7865,916 0,006321
4 13988,60 13962,46 683,2996 0,001869
5 14003,67 13962,46 1698,264 0,002943
6 13929,30 13962,46 1099,586 0,002381
7 13881,06 13962,46 6625,96 0,005864
8 13776,54 13799,14 510,76 0,00164
9 13722,27 13723,82 2,4025 0,000113
10 13774,53 13799,14 605,6521 0,001787
11 13726,29 13723,82 6,1009 0,00018
12 13716,24 13723,82 57,4564 0,000553
13 13722,27 13723,82 2,4025 0,000113
14 13726,29 13723,82 6,1009 0,00018
15 13746,39 13747,06 0,4489 4,87E-05
16 13694,13 13692,99 1,2996 8,32E-05
17 13700,16 13692,99 51,4089 0,000523
18 13680,06 13692,99 167,1849 0,000945
19 13715,24 13723,82 73,6164 0,000626
20 13702,17 13692,99 84,2724 0,00067
21 13720,26 13723,82 12,6736 0,000259
22 13730,31 13747,06 280,5625 0,00122
23 13794,63 13799,14 20,3401 0,000327
24 13828,80 13799,14 879,7156 0,002145

7
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Forecasted Data X i − Fi
Day Actual Data ( X i ) ( X i − Fi )
2
( Fi ) Xi
25 13785,59 13799,14 183,6025 0,000983
26 13730,31 13747,06 280,5625 0,00122
27 13715,24 13723,82 73,6164 0,000626
28 13776,54 13799,14 510,76 0,00164
29 13754,43 13747,06 54,3169 0,000536
30 13727,30 13723,82 12,1104 0,000254
31 13747,40 13747,06 0,1156 2,47E-05
32 13775,54 13799,14 556,96 0,001713
33 13761,47 13747,06 207,6481 0,001047
34 13744,38 13747,06 7,1824 0,000195
35 13785,59 13799,14 183,6025 0,000983
36 13803,68 13799,14 20,6116 0,000329
37 13845,89 13962,46 13588,56 0,008419
38 13932,32 13962,46 908,4196 0,002163
39 13962,47 13962,46 0,0001 7,16E-07
40 14035,83 14119,5 7000,669 0,005961
41 14088,09 14119,5 986,5881 0,00223
42 14305,17 14371,89 4451,558 0,004664
43 14485,07 14371,89 12809,71 0,007814
44 14293,11 14371,89 6206,288 0,005512
45 14241,86 14119,5 14971,97 0,008592
46 14238,84 14119,5 14242,04 0,008381
47 14338,34 14371,89 1125,602 0,00234
48 14413,71 14371,89 1748,912 0,002901
49 14483,06 14371,89 12358,77 0,007676
50 14394,62 14371,89 516,6529 0,001579
51 14562,45 14762,82 40148,14 0,013759
52 14889,08 14762,82 15941,59 0,00848
53 14892,09 14762,82 16710,73 0,00868
54 15158,42 15099,9 3424,59 0,003861
55 15299,12 15337,35 1461,533 0,002499
56 15790,56 15659,9 17072,04 0,008275
57 16354,37 16022,02 110456,5 0,020322
58 16691,04 16611,91 6261,557 0,004741
59 16569,43 16611,91 1804,55 0,002564
60 16409,64 16443,9 1173,748 0,002088
61 16311,15 16022,02 83596,16 0,017726
62 16417,68 16443,9 687,4884 0,001597
63 16448,84 16443,9 24,4036 0,0003
64 16495,07 16443,9 2618,369 0,003102
65 16824,71 16611,91 45283,84 0,012648

8
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

Forecasted Data X i − Fi
Day Actual Data ( X i ) ( X i − Fi )
2
( Fi ) Xi
66 16546,32 16611,91 4302,048 0,003964
67 16638,78 16611,91 721,9969 0,001615
68 16492,05 16443,9 2318,422 0,00292
69 16326,23 16022,02 92543,72 0,018633
70 16322,21 16022,02 90114,04 0,018392
71 15919,20 16022,02 10571,95 0,006459
72 15800,61 16022,02 49022,39 0,014013
73 15785,54 15659,9 15785,41 0,007959
74 15865,94 16022,02 24360,97 0,009837
75 15580,52 15659,9 6301,184 0,005095
76 15620,71 15659,9 1535,856 0,002509
77 15721,22 15659,9 3760,142 0,0039
78 15644,84 15659,9 226,8036 0,000963
79 15708,15 15659,9 2328,063 0,003072
80 15630,77 15659,9 848,5569 0,001864
81 15668,96 15659,9 82,0836 0,000578
82 15565,44 15337,35 52025,05 0,014654
83 15492,08 15337,35 23941,37 0,009988
84 15232,79 15337,35 10932,79 0,006864
85 15148,37 15099,9 2349,341 0,0032
86 15179,52 15099,9 6339,344 0,005245
87 15202,64 15337,35 18146,78 0,008861
88 15084,05 15099,9 251,2225 0,001051
89 15010,68 14987,9 518,9284 0,001518
90 15052,89 15099,9 2209,94 0,003123
91 14961,44 14987,9 700,1316 0,001769
92 15020,73 15099,9 6267,889 0,005271
93 14983,55 14987,9 18,9225 0,00029
94 14959,43 14987,9 810,5409 0,001903
95 14987,12 14987,9 0,6084 5,2E-05
96 14858,93 14762,82 9237,132 0,006468
97 14847,87 14762,82 7233,503 0,005728
98 14834,81 14762,82 5182,56 0,004853
99 14842,85 14762,82 6404,801 0,005392
100 14806,67 14762,82 1922,823 0,002962
MSE = 9197,613 AFER = 0,414755%

9
ISNPINSA 2020 IOP Publishing
Journal of Physics: Conference Series 1943 (2021) 012119 doi:10.1088/1742-6596/1943/1/012119

5. Conclusion
We have presented a new forecasting method for the forecasting exchange rate (USD-IDR) using
frequency density-based partitioning [5], calculate the length of average-based interval [6] and
modification FLRG as k-means clustering [2]. We divided fuzzy sets number of each fuzzy logical
relationship into 16 clusters and then do iterations by using k-means clustering. Final cluster centers can
be used to build a new Fuzzy Logical Relationship Group (FLRG). The result of forecasting data, as
shown in table 4, shows that it gives an error value MSE value 9197,613 and AFER value 0,414755%.

References
[1] Song Q and Chissom B S 1993 Fuzzy Sets Syst. 54 269
[2] Cheng S H, Chen S M, Jian W S 2016 Inf. Sci. 327 272
[3] Yolcu O C and Alpaslan F 2018 Appl. Soft Comput. J. 66 18
[4] Chen S M, Zou X Y, Gunawan G C 2019 Inf. Sci. 500 127
[5] Jilani T A, Burney S M A, and Ardil C 2007 Int. J. Comp. Elec. Auto. Cont. Inf. Eng. 4 112
[6] Saxena P, Sharma K, Easo S 2012 Int. J. Comput. Technol. Appl 3 957
[7] B Irawanto et al 2019 J. Phys.: Conf. Ser. 1321 022081
[8] MacQueen J 1967 Fifth Berkeley Symp. 54 281
[9] Harding J A et al 2006 J. Manuf. Sci. Eng. Trans. 128 969
[10] Wu X et al 2008 Knowl Inf Syst 14 1

10