Professional Documents
Culture Documents
Formation of Seasonal Groups and Application of Seasonal Indices
Formation of Seasonal Groups and Application of Seasonal Indices
0160-5682/14
www.palgrave-journals.com/jors/
Buckinghamshire New University, High Wycombe, UK; The University of Salford, Manchester, UK; and
Cardiff University, Cardiff, UK
Estimating seasonal variations in demand is a challenging task faced by many organisations. There
may be many stock-keeping units (SKUs) to forecast, but often data histories are short, with very few
complete seasonal cycles. It has been suggested in the literature that group seasonal indices (GSI)
methods should be used to take advantage of information on similar SKUs. This paper addresses
two research questions: (1) how should groups be formed in order to use the GSI methods? and
(2) when should the GSI methods and the individual seasonal indices (ISI) method be used?
Theoretical results are presented, showing that seasonal grouping and forecasting may be unied,
based on a Mean Square Error criterion, and K-means clustering. A heuristic K-means method is
presented, which is competitive with the Average Linkage method. It offers a viable alternative to
a companys own grouping method or may be used with condence if a company lacks a grouping
method. The paper gives empirical ndings that conrm earlier theoretical results that greater
accuracy may be obtained by employing a rule that assigns the GSI method to some SKUs and the
ISI method to the remainder.
Journal of the Operational Research Society (2014) 65, 227241. doi:10.1057/jors.2012.126
Published online 13 March 2013
Keywords: forecasting; seasonality; grouping; clustering
1. Introduction
The estimation of seasonal demand patterns is often a
challenging task. For some organisations, it is a task that
must be accomplished for many hundreds or thousands
of stock-keeping units (SKUs). Demand histories may be
short for these items, with few complete seasonal cycles.
As product life cycles shrink, such short data histories
are becoming increasingly common. Consequently, making accurate seasonal estimates for individual SKUs is
becoming much more difcult for many products. The
implications of inaccurate forecasts may be severe, with
detrimental effects on service levels and stock-holdings,
ultimately leading to obsolescent stock if remedial action is
not taken.
As complete seasonal cycles for individual SKUs may
be few, but SKUs themselves may be common, there is
an opportunity to take advantage of the abundance
of time series by pooling data from many SKUs to
estimate seasonal patterns. This idea is not new. Duncan
et al (1993) suggested that the use of information on
similar time series should benet forecasting accuracy
for a particular series. This argument seems plausible,
228
2. GSI methods
Two main approaches to GSI have been proposed in
the literature: one by Dalhart (1974) and the other by
Withycombe (1989).
Withycombe (1989) assumed that whatever causes the
seasonal uctuation in demand operates the same on all
products within the line (authors own emphasis). Dalhart
(1974) made the same assumption that all subaggregate
series had a consistent underlying seasonal behaviour. This
assumption led both authors to believe that estimating
seasonal indices from the group would be better than from
the individual series.
Dalhart (1974) proposed a group seasonal estimation
method by averaging the individual seasonal indices (ISI).
Let Si [ai1, ai2, . . . , aiq], where aih is the ISI for item i at
season h, Si is the multiplicative seasonal index vector for
item i and P
q is the length of the seasonal cycle. Then
SDGSI 1/m m
i 1Si where SDGSI is the group seasonal
vector of indices estimated by Dalharts Group Seasonal
Index (DGSI) method and m is the number of series in the
group. Therefore, Dalharts method is a simple average of
the ISI.
Withycombe (1989) proposed a different method to
obtain GSI, known as Withycombes Group Seasonal
Index (WGSI). He totalled all the series in the group and
then estimated combined seasonal indices from this single
time series. Therefore, Withycombes method is a weighted
average of the ISI.
While there have been many papers published on
seasonal forecasting, there are fewer on aggregation
approaches to seasonal forecasting. Rose (1977) investigated the properties of aggregated independent AutoRegressive Integrated Moving Average (ARIMA) processes, including processes with seasonality. Kim and
Moosa (2005) applied seasonal ARIMA models to
forecasting international tourist ows to Australia. They
found that indirect forecasting of aggregate ows was
more accurate than direct forecasting of aggregates. They
obtained similar results using regression-based models and
structural time series models (Harvey, 1989).
An alternative approach, based on exponential
smoothing, was presented by Dekker et al (2004). They
proposed that the Holt-Winters method should be
adapted, allowing seasonal estimates at the group level
while the level and trend estimates remain at the
individual level. The researchers applied the new
method to the forecasting of sales at two wholesalers
(food and electrochemical products) and found the new
method to give better performance than the classical
Holt-Winters method.
As summarised above, research in this area has progressed by an examination of alternative methods for GSI,
with more complex models being addressed as the
forecasting eld has progressed. However, the fundamental
3. Research questions
In the situation outlined in the introduction, an
organisation may have hundreds or thousands of SKUs,
each with a short demand history, perhaps covering
no more than three complete seasonal cycles. If a
seasonal grouping approach is to be used, there are
some immediate questions.
First, how should the groups be composed? Both
Dalhart (1974) and Withycombe (1989) assumed such
groups were given, in order to calculate their proposed
methods. However, both authors identied the importance
of forming seasonally homogeneous groups. More recent
work, including that of the present authors, has maintained
this assumption, thereby limiting the scope of analysis.
Second, should the grouping method be used for all
series in the group, or would it be better to forecast
demand for some SKUs using individual seasonal proles?
This question has been addressed theoretically (Chen and
Boylan, 2007) but empirical evidence is lacking. This paper
will address both research questions.
where
i
sufx t
sufx h
Y
mi
Sih
eith
MSE
q X
m1
X
m1
1 X
Sih
Sjh
m1 j1
h1 i1
q X
m2
X
h1 i1
s2
MSEih s2i i
r
for h 1; . . . ; q
229
!2
m2
1 X
Sih
Sjh
m2 j1
!2
m
1 X
q 1
s2
qr i1 i
q 1 2
sA1 s2A2
mr
P 1
Pm2
Since Sih m
j1 Sih =m1
j1 Sih =m2 , Equation (5) can
be re-written as:
!2
q m1
m1
X
1 XX
MSE 2
Sih Sjh
m1 h1 i1 j1
!2
q m2
m2
X
1 XX
2
Sih Sjh
m2 h1 i1 j1
m
1 X
q 1 2
sA1 s2A2
q 1
s2i
qr i1
mr
230
q m1 X
m1
2
1 XX
Sih Sjh
2
m1 h1 i1 j1
q m2 X
m2
2
1 XX
Sih Sjh
2
m2 h1 i1 j1
m
1 X
q 1 2
q 1
s2i
sA1 s2A2
qr i1
mr
(ie using ISI for some and GSI for others) when there is
seasonal heterogeneity. This is examined later in the paper.
Conceptually, there are two possibilities when applying
ISI and GSI methods. Suppose there are m series, which
form a group. The rst possibility is that GSI is applied to
all of the series. The second is that GSI is calculated using
all of the series, but applied only to some of the series. ISI
is applied to the other series because they are less noisy.
These series contribute to the formation of the seasonal
group because their seasonal patterns are homogeneous to
the group. However, because they are less noisy, it would
be better to apply ISI to these series so that they do not
borrow weakness from the group.
To bring the applications of ISI and GSI together, we
propose the following formula:
!2
1 2
s
Di 1
qr i
q 1s2A
s2i
2
1
D
Di
i
i
m2 qr
r
m
1 X
Sih
Sih
m i1
MSEih Di
m1
X
m1
1 X
Sih
Sih
m1 i1
Di
i1
!2
!2
1 2
Di 1
s
qr i
i1
i1
m
m
X
q 1s2A X
s2i
2
8
s
Di
1
D
i
i
m2 qr
r
i1
i1
m2
X
Di
m2
1 X
Sih
Sih
m2 i1
m
X
2
2
2 3
m
m
P
P
1
1
P
P
Gi1 Sih Sih m
Gin Sih 7
6 Sih m Gi1
G
q X
m
X
i1
i1 in i1
6
7
i1
P
P
6
7
m
n
MinMSE
Di 6
P
m P
n
7
q1
Gij s2i
i1
j1
1
2
4
5
h1 i1
2
1 qr
Gij si P P
i1 j1
q X
m
X
s2i
2
1 Di si
r
h1 i1
i1
j1
Gij
qr
231
such that:
each group must contain at least one item m
X
Gij X1 for j 1; . . . ; n
MSEDGSIih
10
i1
Gij 1 for i 1; . . . ; m
11
j1
m
1 X
Sih
Sih
m i1
m2i
!2
s2i
m2i s21
m2 rm21
m1 X
m
m2i s2m
m2 X
1 1
2 2i
r sj sl
2
2
m rmm
m r j1 lj1 mj ml jl
12
Pm
2
m Sih
MSEWGSIih m2i Sih i1 i
s2i
mA
"
#
m1 X
m
X
m2i
2
2
2 s1 sm 2
rjl sj sl 13
rmA
j1 lj1
S
S
ih
ih
m i1
m2i
m2i
s21
s2m
m2 rm2m
m2 rm21
Pm
2
i1 mi Sih
Sih
mA
2
s
1
i2 2 s21 s2m
mi rmA
Pm
2
m Sih
s2 s2
i2 A2
Sih i1 i
mA
mi rmA
14
MSEWGSIi
m2i
15
Sih
Sjh q
m2 j1
m2
i1 i
h1 i1
2
q s1
s2m
16
mr m21
m2m
232
q X
m1
X
Di
h1 i1
m1
1 X
Sih
Sjh
m1 j1
!2
!2
m2
1 X
Di Sih
Sjh
m2 j1
h1 i1
2
m
X
s2i
q
s1
s2m
Di 2 2
q
Di 2
mm
mi mr
m1
i1
m
X
1 s2i
1 Di 1
r m2i
i1
q X
m2
X
7. Empirical investigation
7.1. Experimental structure
17
q X
m1
X
Pm1
Sih
h1 i1
q X
m2
X
j1
mA1
Pm2
j1
Sih
h1 i1
mq s2A1 s2A2
r m2A1 m2A2
mj Sih
!2
mj Sih
!2
mA2
m
X
s2
i
i1
m2i
18
A2
q X
m1
X
Pm1
Di Sih
h1 i1
q X
m2
X
Di Sih
h1 i1
m
X
j1
mj Sih
mA1
Pm2
j1
!2
mj Sih
!2
mA2
s2 1 s2A2
s2 mq
Di A
q
Di i2
r
mi
m2A1 m2A2
i1
m
X
1 s2i
1 Di 1
r m2i
i1
19
233
and the forecast produced for that period (or the average
of cumulative demand over the horizon and the cumulative
forecast for that time horizon) to form symmetric absolute
percentage errors (sAPEs). These errors are then summarised across time per series by taking their arithmetic
mean (sMAPE). An arithmetic mean is also used to
average the sMAPEs per series across all 218 series. The
advantages of the sMAPE over other relative error
measures have recently been discussed in a comprehensive
paper by Kolassa and Martin (2011).
In addition, and given that the theoretical results were
derived based on the MSE, we have decided to employ this
error metric as well. Squared errors are calculated per
period (or horizon) and summarised across time per series
by their arithmetic average (MSE). This particular measure
is known to be heavily scale dependent and unduly
inuenced by the volume of the series. As such, MSEs
are summarised across series by using a Relative Geometric
summarisation (RGMSE). That is, the MSEs per series per
method are summarised across all series with a geometric
summarisation (GMSE). The RGMSE then is the ratio of
the GMSEs related to any two methods. To standardise
the presentation of the results all methods are compared
against the ISI method (to be considered in the denominator). Values below 1 indicate performance in favour of
the method under concern; values above 1 indicate a
superior performance of the ISI method. Relative geometric summarisations have been shown (among others by
Fildes 1992; Syntetos and Boylan, 2005) to be very robust.
The theoretical properties and scale independent nature of
the RGMSE make it a natural measure to be considered
for the purposes of our experimentation. It allows the
linkage of our empirical results to the theoretical analysis
while a degree of fairness is ensured for the comparison
across series.
234
Minimum
25th percentile
Median
75th percentile
Maximum
Demand
Multiplicative
seasonal indices
Mean
Max/min
scale changes
Min
Max
192.37
1195.70
3451.68
9940.31
563707.98
1.07
1.48
1.83
2.46
12.88
0.08
0.45
0.54
0.64
0.80
1.19
1.45
1.64
1.89
5.06
Table 2 Group sizes (and description) in descending order for the three grouping approaches
Group
#1
#2
#3
#4
#5
#6
#7
100
102
103
101
106
107
105
Mixed model
Description
Incandescent
Compact uorescent
Non-compact uorescent
Halogen general lighting
HQI power stars
HPS vialox
HQ/HW/SOX/UV
Total
Number
K-means
Average linkage
66
61
45
27
12
5
2
103
82
15
15
1
1
1
209
3
2
1
1
1
1
218
218
218
Table 3 Symmetric MAPE (sMAPE) results (%) for the mixed model
Method
Point forecasts
1-step
3-step
6-step
9-step
Cum3
Cum6
Cum9
50.43
51.67
54.07
54.66
39.91
36.60
34.66
49.29
51.21
55.44
56.70
39.25
36.20
34.27
48.98
50.85
54.83
56.19
39.48
36.50
34.57
49.15
51.17
55.72
57.39
39.43
36.54
34.64
48.51
49.59
51.72
51.84
38.24
35.04
33.40
48.32
49.54
51.41
51.01
37.93
34.92
33.47
235
VarAGGREGATE
m2
Table 4 Relative geometric mean squared error (RGMSE) results for the mixed model
Method
Point forecasts
1-step
3-step
6-step
9-step
Cum3
Cum6
Cum9
1.00
1.00
1.00
1.00
1.00
1.00
1.00
0.81
0.83
0.81
0.84
0.86
0.85
0.90
0.90
0.92
0.96
0.94
1.00
0.89
0.93
0.90
0.93
0.97
0.95
0.94
0.94
0.99
0.80
0.78
0.80
0.78
0.80
0.78
0.77
0.77
0.83
0.80
0.85
0.83
0.84
0.87
236
References
Bunn DW and Vassilopoulos AI (1993). Using group seasonal
indices in multi-item short-term forecasting. International Journal
of Forecasting 9(4): 517526.
Bunn DW and Vassilopoulos AI (1999). Comparison of seasonal
estimation methods in multi-item short-term forecasting. International Journal of Forecasting 15(4): 431443.
Chen H and Boylan JE (2007). Use of individual and group
seasonal indices in subaggregate demand forecasting. Journal of
the Operational Research Society 58(12): 16601671.
Chen H and Boylan JE (2008). Empirical evidence on individual,
group and shrinkage seasonal indices. International Journal of
Forecasting 24(3): 525534.
237
238
Appendix A
The additive model
The additive model is specied as:
Yith mi Sih eith
The forecast for item i, the hth season in year r 1 using ISI is:
FISIir1h mi ISIih
r
1X
Yith
r t1
2
1
E mi Sh eir1h mi Sh ei1h ei2h eirh
r
2
s
s2i i
r
Suppose we can nd m items with similar seasonal patterns. It might be better to estimate their seasonality from the
group.
The GSI estimator is given by:
q
m X
r
m X
r X
1 X
1 X
Yith
Yith
GSIh
mr i1 t1
mqr i1 t1 h1
^i
m
q
r X
1 X
Yith
qr t1 h1
Yith
Yith
Yith
qr t1 h1
mr i1 t1
mqr i1 t1 h1
The MSE of the forecast is as follows:
!2
q
q
r X
m X
r
m X
r X
1 X
1 X
1 X
MSEGSIi E Yir1h
Yith
Yith
Yith
qr t1 h1
mr i1 t1
mqr i1 t1 h1
0
12
q
r P
P
mi Sih eir1h mi qr1
eith
B
C
t1 h1
B
C
B
C
m
P
B mA 1
C
Sih
EB m m
C
B
C
i1
B
C
q
m P
r
m P
r P
@
A
P
P
mA
1
1
mr
eith m mqr
eith
i1 t1
i1 t1 h1
!
P
!2
2r s2i
rij si sj
m
jai
1 X
s2i
s2A
s2A
2
Sih
Sih si 2 2
m i1
qr m r m qr
mqr2
!
!
m1
m
P
P
P
2 qrs2i qr rij si sj
2r s21 s22 s2m 2
rjl sj sl
jai
mq2 r2
j1 lj1
m2 qr2
m
1 X
Sih
Sih
m i1
239
!2
s2i
s2
s2
2A 2A
qr m r m qr
!2
m
1 X
1 2 q 1s2A
s
Sih 1
Sih
m i1
qr i
m2 qr
s2i
Appendix B
The mixed model
The mixed model is specied as:
Yith mi Sih eith
In order to derive the rules for the mixed model, it is assumed that:
^1
^i pi m
m
for i 1; 2; . . . ; m
^A p1 p2 pm ^
m1
m
mi pi m1
for i 1; 2; . . . ; m and p1 1
DGSIh
mr
mrp2
240
2
mi i1 Sih pi e11h e12h e1rh
pi em1h em2h emrh
E mi Sih eir1h
mr
mrpm
m
!2
m
m
1 X
m
X
1 X
p2 rs2
p2 rs2
pi pi
2
mi Sih
Sih s2i i 2 21 2i 2 m2 2
rr sj sl
mr
m r pm
mrpj mrpl jl
m i1
j1 lj1
m2i
m
1 X
Sih
Sih
m i1
m2i
m
1 X
Sih
Sih
m i1
m2i
m
1 X
Sih
Sih
m i1
!2
s2i
m
XX
p2i rs21
p2i rs2m
p2i m1
1 1
2
r sj sl
2
2
2
2
2
2
mr
m r pm
m r j1 lj1 pj pl jl
s2i
m1 X
m
p2i s21
p2i s2m
p2i X
1 1
2
r sj sl
m2 r
m2 rp2m
m2 r j1 lj1 pj pl jl
s2i
m1 X
m
m2i s21
m2 s2
m2 X
1 1
2i m2 2 2i
r sj sl
2
2
m rmm
m r j1 lj1 mj ml jl
m rm1
!2
!2
m2i
m
1 X
Sih
Sih
m i1
m2i
m
1 X
Sih
Sih
m i1
m
1 X
Sih
Sih
m i1
!2
s2i
!2
o
!2
1 X
m
X
m2i s21
m2i s2m
m2i m
1 1
s2i
2
2
r
s
s
os
j
l
i
m2 rm2m
m2 r j1 lj1 mj ml jl
r
m2 rm21
2 2
m1 X
m
s2i
mi s1
m2i s2m
m2i X
1 1
2
r sj sl
r
m2 rm2m
m2 r j1 lj1 mj ml jl
m2 rm21
"
!#
m
m
1 X
m
X
1 s2i
1 X
s2i
1 1
o
2
r sj sl
r m2i m2 i1 m2i
m m jl
j1 lj1 j l
WGSIh
r
Y11h ::: Y1rh ::: Ym1h ::: Ymrh
rp1 p2 ::: pm ^
m1
The forecast of hth season in year r 1 for the ith item using WGSI is:
^1 WGSIh
FWGSIir1h pi m
241
2
m
m Sih pi e11h e1rh em1h emrh
E mi Sih eir1h i i1 i
mA
rp1 p2 pm
Pm
2
m Sih
p2i
m2i Sih i1 i
s2i
E e11h e1rh em1h emrh 2
mA
r2 p1 p2 pm 2
"
#
Pm
2
m1 X
m
X
p2i
2
2
2
2
i1 mi Sih
mi Sih
si
rs1 rsm 2
rrjl sj sl
mA
r2 p1 p2 pm 2
j1 lj1
"
#
Pm
2
m1
m
XX
m2i
2
2
2
2
i1 mi Sih
si 2 s1 sm 2
rjl sj sl
mi Sih
mA
rmA
j1 lj1
MSEWGSIio MSEISIi if and only if:
m2i
m2i
Pm
i1
Sih
mi Sih
mA
Pm
2
i1 mi Sih
mA
Sih
2
s2i
"
#
m1
m
XX
m2i
s2
2
2
2 s1 sm 2
rjl sj sl os2i i
r
rmA
j1 lj1
"
#
m1 X
m
X
s2i
m2i
2
2
o 2 s1 sm 2
rjl sj sl
r
rmA
j1 lj1
0
Pm
Sih
i1
mi Sih
mA
2
2
1B
Bs
o B i2
r @m i
s21
s2m
m1
P
m
P
j1 lj1
m2A
rjl sj sl C
C
C
A