ARTICLE

International Journal of Engineering Business Management

Special Issue on Innovations in Fashion Industry

Fourier Analysis for Demand

Forecasting in a Fashion Company
Regular Paper

Andrea Fumi1, Arianna Pepe1, Laura Scarabotti1 and Massimiliano M. Schiraldi1,*

1 University of Rome “Tor Vergata” - Department of Enterprise Engineering, Roma, Italy
* Corresponding author E-mail: [email protected]

Received 1 June 2013; Accepted 15 July 2013

DOI: 10.5772/56839

© 2013 Fumi et al.; licensee InTech. This is an open access article distributed under the terms of the Creative
Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract In the fashion industry, demand forecasting is 1. Introduction

particularly complex: companies operate with a large
variety of short lifecycle products, deeply influenced by The role of demand forecasting has become increasingly
seasonal sales, promotional events, weather conditions, important within businesses that have the maximization
advertising and marketing campaigns, on top of of customer service and the optimization of capital
festivities and socio-economic factors. At the same time, investment operating costs as their main objectives [1,2,3].
shelf-out-of-stock phenomena must be avoided at all Demand forecasting is also considered key to effective
costs. Given the strong seasonal nature of the products supply chain management [4,5,6,7] [8,9] and it plays a
that characterize the fashion sector, this paper aims to crucial role, especially in the long-term, in identifying the
highlight how the Fourier method can represent an easy direction of the business strategy [10,11]. The forecast
and more effective forecasting method compared to other accuracy affects all levels of production systems, from the
widespread heuristics normally used. For this purpose, a generation of production plans to the calculation of
comparison between the fast Fourier transform algorithm material requirements [12,13,14,15] and, consequently, to
and another two techniques based on moving average supply chain management. An accurate forecast can lead
and exponential smoothing was carried out on a set of 4- to significant cost savings, reduced working capital in
year historical sales data of a €60+ million turnover safety stocks [16], the strengthening of customer
medium- to large-sized Italian fashion company, which relationships and increasing competitiveness [17].
operates in the women’s textiles apparel and clothing However, it is well-known that accurate forecasts are
sectors. The entire analysis was performed on a common extremely difficult to attain due to many factors, from
spreadsheet, in order to demonstrate that accurate results macro-economic changes and the unpredictability of
exploiting advanced numerical computation techniques markets to fashion effects [18]. The complexity of this
can be carried out without necessarily using expensive problem may push companies to purchase specific
software. software, such as forecasting decision support systems;
however, as these are extremely expensive, in most cases
Keywords Demand Forecasting, Fourier Analysis, companies choose to use simple statistical approaches,
Fast Fourier Transform, Fashion such as implementing moving average or exponential

www.intechopen.com Andrea Fumi, Arianna Pepe, Laura Scarabotti andbus.

Int. j. eng. Massimiliano M. Schiraldi:
manag., 2013, Vol. 5, 1
Fourier Analysis
Specialfor Demand
Issue Forecasting
Innovations in a Fashion
in Fashion Industry,Company
30:2013
 smoothing heuristics [19] on a common spreadsheet. As a alternative. Other methods that have been used in the
consequence, these basic forecasts need to be reviewed in fashion industry - providing successful results in sales
order to take into account exceptional circumstances or prediction - include two-stage dynamic sales forecasting
events that may not emerge from historical data: as a models [41], fuzzy logic approaches [42], artificial neural
result, up to 80% of forecasts are adjusted [4], although networks [43] and extreme learning machines [44,8,45].
experimental evidence from some authors suggests that Despite their effectiveness, few commercial software
judgmental adjustments to statistical forecasts are usually solutions implement these methods for forecasting. The
unnecessary [20]. In the fashion industry, demand few that do are often too expensive for small- or medium-
forecasting is particularly complex [21,22,23,24]: sized companies. As a result, in practical cases, most
companies in this specific sector operate with a large companies use basic heuristics implemented on common
variety of short lifecycle products, deeply influenced by spreadsheets. However, common spreadsheets do not
seasonal sales [25,26,27], promotional events [28,29], support only basic techniques: for instance, the Fast Fourier
weather conditions, and advertising and marketing Transform - which has been widely used and applied in
campaigns [30], as well as festivities and economic and many fields ranging from physics, seismology, engineering
social factors. Moreover, these slow-moving expensive and economics and has been described as the “most
products usually face an intermittent or lumpy demand important numerical algorithm of our lifetime” [46] – is
[31], but at the same time – as they are usually high-margin easily available in Microsoft Excel. The use of Fourier
items - shelf-out-of-stock phenomena must be avoided at analysis in forecasting overcomes certain limitations that
all costs [32,33,8]. Thus, quick and effective supply chain other techniques have in capturing seasonality phenomena
management, with flexible production schedules and [47]. Therefore, this approach has been used for forecasting
appropriate inventory levels, becomes critical for each changes in electricity demand and/or prices [48,49], which
stock-keeping unit. For this reason, the demand forecasting are variables that are clearly related to light and
process needs to be timely and accurate. temperature cyclic variations. In forecasting consumers’
behaviour, it has been used for estimating the volumes of
Thanks to its capability in terms of decomposing a function incoming calls in a call centre [50] and has been integrated
into a sum of sinusoids of different frequencies, amplitude with a linear equation for forecasting motorcycle sales with
and phase, Fourier analysis can be used effectively for successful results [51], whilst it did not lead to significant
seasonal sales forecasting; this is because the Fourier improvements for forecasting automobile sales [52]. Apart
transform takes a time series and maps it into a frequency from these few examples, FFT seems to be applied only
spectrum in the frequency domain. The discrete version of rarely for forecasting consumer goods sales and – to the
the Fourier transform can be quickly calculated using fast authors’ knowledge – no contributions seem to be present
Fourier transform (FFT) algorithms. Given the strong in the literature for forecasting consumer goods sales,
seasonal nature of the products that characterize the specifically in the fashion industry. The fashion industry is
fashion sector and the simplicity of computing FFT on progressively drawing the attention of researchers due to
popular spreadsheets, such as Microsoft Excel, this paper its increasing importance in the worldwide economy and
aims to highlight how the Fourier method can represent an the peculiarities of its operations’ practices. For example,
easy and more effective forecasting method compared to recent studies focused on fashion firms in analysing brand
other widespread alternatives normally used. For this value [53], organizational innovation [54,55,56], production
purpose, a comparison between FFT and two other efficiency increase [57,58] and improvements in logistics
techniques based on moving average and exponential processes [59,24,60]. For an updated and complete review
smoothing was carried out on a set of 4-year historical sales of the forecasting techniques in the fashion industry, refer
data of a €60+ million turnover medium- to large-sized to Nenni et al. [61].
Italian fashion company operating in women’s textiles
apparel and clothing sectors which distributes over 2,500 The next section presents in detail both the methodology
products to more than 200 shops. and the step-by-step procedure used to apply FFT on
historical time series to predict sales. Next, the results of
2. Previous Research the validation of the proposed approach on the real data
of a fashion company are presented.
Forecasting approaches can be divided into qualitative
and quantitative methods [34]. Here, we focus on 3. Methodology
quantitative methods based on the study of historical
time series [35]. Among these, the most well-known are This section presents both the methodology and the step-
the moving average and the exponential smoothing by-step procedure used to apply FFT to historical time
methods - Holt-Winters’ method [36] and regressive series to forecast sales in detail. In the next section, results
methods [37]. However, in having to deal with lumpy proving the effectiveness of the suggested approach are
demand, Croston’s method [38] [39] and its evolution as shown by analysing a fashion company’s data. The whole
offered by Syntetos and Boylan [40] represent a better analysis was performed on a common spreadsheet in

2 Int. j. eng. bus. manag., 2013, Vol. 5, www.intechopen.com

Special Issue Innovations in Fashion Industry, 30:2013
order to prove that accurate analyses exploiting advanced imaginary part. Excel provides simple functions that can
numerical computation techniques can be carried out be easily used to calculate the amplitude and phase of
without needing to use expensive software. Microsoft each of the N/2 significant components, from each
Excel software is able to compute FFT on a data vector. complex number Ai:
The FFT algorithm significantly reduces computational
times compared to the standard Fourier transform [62], |�� | = IMABS (Ai) / (N/2) (1)
and specifically from O(n2) to O(n⋅log(n)) [63]. Although
Microsoft Excel is not primarily conceived to perform ϕi = IMARGUMENT (Ai) (2)
such mathematical analyses, it is easy and quick to
calculate the discrete Fourier transform (DFT) and its It is worth noting that the amplitude of the first
inverse function on a data range. The tool’s only component (the average value of the series,
constraint is that the number of input values must be a corresponding to the f0 frequency) is obtained by
power of 2, up to the value of 4,096; however, this should calculating the absolute value of the first complex number
not represent a limitation for these types of analyses of the array but dividing it by N and not by N/2, since
(4,096 values can describe a sales record over more than there is no conjugated value associated to it. The most
eleven years with daily samples, or over more than 78 significant components are those with the highest
years with weekly samples). amplitude. The sales forecast pattern can therefore be
attained by summing a specific number of components
In this analysis, the FFT was used to perform spectral (i.e., with the inverse Fourier transform) chosen among
analysis on sales patterns in order to attain a frequency the most significant of them. A critical issue is how to
spectrum (the representation of which is often referred to select the correct number of components to consider; this
as a “periodogram”). This is because the Fourier transform is explained in detail in the case study paragraph. Figure
can decompose a periodic series into a sum of sinusoidal 1 shows some periodogram examples resulting from the
functions (harmonic components [64], specifically cosine spectrum analysis of different signals.
functions in Microsoft Excel [65]). The FFT returns complex
numbers from which it is possible to extract information –
frequency (f), amplitude (A), and phase (ϕ) – of N/2
periodic waves that form the sales pattern, where N is the
total number of time periods of sales data, which is also the
size of the sample (for the limit of N/2 sine waves see
Nyquist –Shannon sampling criterion in signal theory [66]).
Thus, the result of the Fourier transform is:

���� � ∑������� ��� ���� � � �� ��� ���� �� (1)

However, the following simplification can be considered

[67]:

���� � ∑���� �� �������� � � �� � (2)

Where k = N/2.

Applying the Microsoft Excel Fourier analysis tool on a N-

Figure 1. Examples of periodograms generated from different
sized data range, it yields an N-sized range of complex signals.
numbers which contain information on N components.
However, as previously stated, only the first half should be
The procedure to be applied to a historical time series is
considered. This is because the second half of the series
summarized below. Clearly, before performing the
shows the conjugated values, symmetrical with respect to
analysis, the items/products to be analysed and the level
the Nyquist-Shannon frequency, which is fc/2 where fc is the
of detail of the analysis must be chosen. Specifically,
sampling frequency. The spectral resolution is:
when dealing with fashion products (and specifically
�� with the more expensive women’s clothing, purses or
�� � accessories) it is often impossible to drill down the
�
analysis to manage the volume of sales per product / per
The amplitudes of the i-th component can be calculated day / per shop due to the fact that, at such level of detail,
as the absolute value of the i-th complex number the average values can be close to zero or statistically
�� � �� � ��� , where a is the real part and b is the insignificant. Thus, data is usually analysed using weekly

www.intechopen.com Andrea Fumi, Arianna Pepe, Laura Scarabotti and Massimiliano M. Schiraldi: 3
Fourier Analysis for Demand Forecasting in a Fashion Company
 time buckets. On top of this, product sales are often Clearly, θk may easily differ from item to item. Despite
grouped by category or sub-category and/or shops are choosing a unique θ = θk ∀k, it may provide acceptable
grouped by market region. Thus, forecasts are first results through an easier and more straightforward
calculated at an aggregated level and then the volumes procedure, and a first tuning analysis is advisable,
are brought back to the most appropriate level of detail particularly when product categories drastically differ
through heuristic ratios. Hence, after having chosen the (e.g., purses, scarves and coats all belong to the general
most appropriate time bucket and product aggregation, a “women fashion products” category but may show very
10-step procedure to attain an accurate forecast through different sales patterns over the same time period). In the
Fourier analysis is employed, as follows: following section, the application of the procedure is
1. Extract the historical series of data related to the shown with two examples of the use of Fourier analysis
item to be analysed over a significant time interval to forecast the sales in the trolley category and the belt
(e.g., 4 years); category.
2. Divide the data into two subsets: a calibration data
set (e.g., the first 3 years) and a validation data set 4. Application to fashion products
(e.g., the 4th year), which allows the quality of the
results to be analysed; the calibration set should be This section presents the application of Fourier analysis to
composed of N values (e.g., in Excel, N needs to be a calculate the sales forecasts for a medium- to large-sized
power of two); Italian fashion company operating in the women’s
3. Calculate a linear trend of the series of the textiles, apparel and clothing sectors. Several product
calibration set and subtract it from the data series categories were analysed and, as a result, the application
[68] (e.g., in Excel, the TREND function may help); of the proposed method generally yielded more accurate
4. Calculate the FFT on the detrended data (e.g., in forecasts in comparison with two of the most commonly-
Excel, the Fourier analysis tool gives a range of used approaches based on moving average and
complex numbers); exponential smoothing. Two examples are presented
5. Create the frequency spectrum by calculating, for here: the trolley sub-category and the belt sub-category.
each of the first N/2 components, their amplitude The Fourier analysis applied to the historical series of the
and phase (e.g., in Excel, using (1) and (2) in the former returned a much more precise forecast, while with
complex numbers’ range); the latter the forecasting error was comparable with that
6. Except for f0, order the f1… fN/2 components in a obtained with the other two approaches. In all cases, the
decreasing order of amplitude; historical series included the calibration data set (sales
7. Perform the inverse Fourier transform N/2 times (or during 2007, 2008 and 2009) and the validation data set
sum the wave components), taking into account (sales during 2010). With the moving average and
each of the N/2 components progressively, starting exponential smoothing techniques, the traditional
from f0 up to the N/2-th; approaches [70] were used to calculate weekly ratios
8. Re-apply the eliminated trend, calculated in step 3, using three periods of historical data (2007, 2008 and
to each of the N/2 inverse Fourier transforms; 2009). The forecast (α) parameter in the exponential
9. Compare the validation data set with each of the smoothing was chosen in a different way in each case, by
N/2 inverse Fourier transforms, calculating the using the one that returned the best results. In the Fourier
forecast error with the preferred method [69] (e.g., analysis, since the input range must be a power of two,
mean absolute percentage error, MAPE). the calibration set was reduced to the period from
10. Choose the most appropriate number of wave 19/07/2007 to 31/12/2009 - i.e., N = 128 weeks - while the
components to consider in order to minimize the error. validation set was kept the same (from 01/01/2010 to
31/12/2010). As the sampling frequency is equal to fc = 1
ଵ
Note that this 10-step procedure needs to be performed week, the spectral resolution is equal to ο݂ ൌ ଵଶ଼. Despite
only the first time a given item k (product, category, that the Fourier analysis operated within a smaller
subcategory, etc.) is analysed. This is because, once the historical data range, the forecast turned out to be more
most appropriate number of wave components to be accurate, as may be seen in the following cases.
applied has been chosen (say, θk), to further improve the
forecast, steps 6, 7, 8, 9 and 10 can be substituted with the 4.1 The trolley sub-category
following: The calibration data set of the trolley sub-category sales is
6. Perform the inverse Fourier transform, taking into shown in Figure 2.
account f0 and the first θk components with
decreasing amplitude; Starting from f0, the other components in the spectrum
7. Re-apply the eliminated trend, calculated in step 3, were sorted by decreasing amplitude, according to step 6
to the inverse Fourier transform calculated in the in the mentioned procedure. The result is shown in
previous step 6. Figure 4.

4 Int. j. eng. bus. manag., 2013, Vol. 5, www.intechopen.com

Special Issue Innovations in Fashion Industry, 30:2013
 Component Amplitude Frequency Phase
Trolley sub-category sales
f0 63.8 - -
1000
f5 67.7 0.04 -1.04
800 f3 40.5 0.02 -0.73
f4 40.0 0.03 +2.72
600
f8 33.9 0.06 -1.77
400 f1 30.5 0.01 +2.31
f10 27.2 0.08 -2.47
200
Table 1. Amplitude, frequency and phase of the chosen components.
0
15
22
29
36
43
50
57
64
71
78
85
92
99
106
113
120
127
1
8

Weeks y(t) = 63.8 + 67,7 ⋅ cos(2π ⋅ 0.04 ⋅ t – 1.04) +

+ 40.5 ⋅ cos(2π ⋅ 0.02 ⋅ t – 0.73) +
Figure 2. Calibration data set of the trolley sub-category sales (in
+ 40.0 ⋅ cos(2π ⋅ 0.03 ⋅ t + 2.72) +
units).
+ 33.9 ⋅ cos(2π ⋅ 0.06 ⋅ t – 1.77) +
+ 30.5 ⋅ cos(2π ⋅ 0.01 ⋅ t + 2.31) +
The FFT yielded the frequency spectrum shown in Figure 3. + 27.2 ⋅ cos(2π ⋅ 0.08 ⋅ t – 2.47)

MAPE for N/2 different forecasts

Trolley sub-category sales spectrum
140%
80
120%
70
100%
60
50 80%

40 60%
30 40%
20 20%
10
0%
0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64
Fo
0.01
0.02
0.02
0.03
0.04
0.05
0.05
0.06
0.07
0.08
0.09
0.09
0.10
0.11
0.12
0.13
0.13
0.14
0.15
0.16
0.16
0.17
0.18
0.19
0.20
0.20
0.21
0.22
0.23
0.23
0.24
0.25
0.26
0.27
0.27
0.28
0.29
0.30
0.30
0.31
0.32
0.33
0.34
0.34
0.35
0.36
0.37
0.38
0.38
0.39
0.40
0.41
0.41
0.42
0.43
0.44
0.45
0.45
0.46
0.47
0.48
0.48
0.49

Frequencies Number of considered components

Figure 3. Frequency spectrum computed on the calibration data set. Figure 5. MAPE for all of the 64 possible forecasts.

Figure 6 shows the comparison between the forecasts

Components ordered by decreasing amplitude
obtained using Fourier analysis, exponential smoothing and
80
moving average. Obviously, the Fourier forecast succeeded in
70
following the validation data pattern more effectively, while
60
exponential smoothing and moving average techniques
50
yielded more or less the same results. Specifically, it is clear
40
that the huge peak that both the exponential smoothing and
30
moving average techniques forecast for weeks 162-165
20
results from a similar peak in weeks 85-88 of the calibration
10
0
data set shown in Figure 2. This is because neither of these
classical techniques can ignore the high values recorded in
0.5
Fo
0.04
0.02
0.03
0.06
0.01
0.08
0.13
0.17
0.16
0.20
0.10
0.07
0.09
0.34
0.16
0.48
0.32
0.38
0.13
0.23
0.19
0.35
0.05
0.38
0.28
0.41
0.24
0.49
0.20
0.14
0.48
0.30
0.45
0.31
0.05
0.39
0.27
0.45
0.29
0.33
0.22
0.36
0.02
0.21
0.18
0.27
0.23
0.42
0.26
0.34
0.37
0.41
0.25
0.30
0.12
0.44
0.46
0.09
0.15
0.11
0.43
0.47
0.40

Frequencies
the same period of the previous season; on the contrary, the
Figure 4. Components ordered by decreasing amplitude Fourier analysis forecast tends to confirm only those
(the first is f0). increments that are recorded cyclically.

Then, N/2 inverse Fourier transforms can be calculated These results are confirmed by the numerical comparison
using different numbers of components, as mentioned in shown in Table 2, where the Fourier analysis is shown to
step 7 of the procedure described. Next, the trend is yield a much smaller error, both in terms of MAPE and
applied to each of them and the results are compared MAD (the exponential smoothing α parameter was set to
with the validation set. Figure 5 shows the MAPE index its best value, α = 0.96).
for each of the N/2 possible forecasts: the smallest error is
Forecasting technique MAD MAPE
recorded using 6 components, thus θtrolley = 6.
Moving average 38.0 104.0%
Trolley

Exponential smoothing 39.7 108.4%

The following Table 1 shows the characteristics of the 6
Fourier analysis 27.0 74.0%
chosen components – starting from f0 – followed by the
mathematical expression of the forecast function y(t). Table 2. Comparison of forecast errors.

www.intechopen.com Andrea Fumi, Arianna Pepe, Laura Scarabotti and Massimiliano M. Schiraldi: 5
Fourier Analysis for Demand Forecasting in a Fashion Company
 Forecasting results comparison
Components ordered by decreasing amplitude
Fourier
400 Validation data set 1200
350 Mov average
1000
Exp smoothing
300
800
250
units

200 600
150
400
100
50 200
0 0
129 132 135 138 141 144 147 150 153 156 159 162 165 168 171 174 177 180

0.04
0.08
0.13
0.06
0.09
0.09
0.15
0.30
0.10
0.13
0.03
0.02
0.18
0.19
0.34
0.22
0.26
0.14
0.40
0.07
0.27
0.44
0.12
0.38
0.05
0.48
0.23
0.17
0.11
0.16
0.36
0.05
0.30
0.34
0.48
0.20
0.47
0.37
0.02
0.21
0.41
0.33
0.46
0.41
0.45
0.32
0.25
0.38
0.45
0.42
0.24
0.16
0.31
0.29
0.43
0.28
0.49
0.01
0.27
0.35
0.39
0.20
0.23
Fo
Weeks Frequencies

Figure 6. Forecasting results comparison. Figure 9. Components ordered by decreasing amplitude

(the first is f0).
4.2 The belt sub-category
MAPE for N/2 different forecasts
The calibration data set of the belt sub-category sales is 140%
shown in Figure 7.
120%
Belt sub-category sales 100%
10000 80%

8000 60%
40%
6000
20%
4000 0%

2000 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64
Number of considered components
0 Figure 10. MAPE index for all 64 possible frequencies.
15
22
29
36
43
50
57
64
71
78
85
92
99
106
113
120
127
1
8

Weeks
It is clear that, in this example, the forecasts gave more
Figure 7. calibration data set of belt sub-category sales (in units)
accurate results compared to the trolley case depicted in
The Fast Fourier Transform yielded the frequency Figure 5: the smallest error is recorded using 17
spectrum shown in Figure 9.
components, thus θbelt = 17. Figure 11 shows the
Belt sub-category sales spectrum comparison between the forecasts attained through
Fourier analysis, exponential smoothing and moving
1200
average. In this case, despite the fact that the Fourier
1000
analysis followed the validation data set more accurately,
800 the difference between the suggested method and the
600 results attained with the moving average and the
exponential smoothing techniques is not as clear as in the
400
trolley case. In terms of MAD and MAPE, the numerical
200
values are shown in Table 3 (the exponential smoothing α
0 parameter was set to its best value, α = 0.98).
0.01
0.02
0.02
0.03
0.04
0.05
0.05
0.06
0.07
0.08
0.09
0.09
0.10
0.11
0.12
0.13
0.13
0.14
0.15
0.16
0.16
0.17
0.18
0.19
0.20
0.20
0.21
0.22
0.23
0.23
0.24
0.25
0.26
0.27
0.27
0.28
0.29
0.30
0.30
0.31
0.32
0.33
0.34
0.34
0.35
0.36
0.37
0.38
0.38
0.39
0.40
0.41
0.41
0.42
0.43
0.44
0.45
0.45
0.46
0.47
0.48
0.48
0.49
Fo

Frequencies
The differences between the results obtained in the trolley
Figure 8. frequency spectrum calculated on calibration data set
and belt cases originate from the specific patterns of their
historical time series. As may be seen comparing Figure 2
Starting from f0, the other components in the spectrum with Figure 7, the belt sub-category historical sales show
were sorted by decreasing amplitude, according to step 6 a much clearer cyclic pattern, with periodic peaks
in the mentioned procedure. The results are shown in repeating over time and gradually decreasing in value. In
Figure 9. the belt sub-category, both the moving average and the
exponential smoothing methods were able to capture the
The results for all 64 possible frequencies, expressed by pattern’s cyclicality, despite the decreasing trend in the
the index MAPE, are represented in Figure 10. peak values not being perfectly forecast, as would be
expected when using these original techniques [19]. On
the other hand, the trolley sub-category historical sales
pattern was more irregular, with a sudden high peak of

6 Int. j. eng. bus. manag., 2013, Vol. 5, www.intechopen.com

Special Issue Innovations in Fashion Industry, 30:2013
sales in the last season. This recent high peak significantly distributes over 2,500 products to more than 200 shops.
influenced the moving average and exponential The results show how Fourier analysis represents a valid
smoothing techniques, and both methods forecasted a alternative forecasting technique among those that can be
higher sales volume compared to what really occurred. easily implemented on a common spreadsheet. Clearly,
On the contrary, in both cases the Fourier analysis its effectiveness varies with the sales pattern
yielded much more accurate results. characteristics: for certain sets of historical data, the
suggested approach can attain much more accurate
Forecasting results comparison results compared to other simple heuristics, such as
Validation data set moving average or exponential smoothing methods. In
Fourier other patterns, the accuracy can be comparable. In this
7000 Mov average
6000 Exp smoothing
paper, the weekly sales of two product categories were
5000
analysed: for both categories, Fourier analysis yielded a
4000 smaller error both in terms of MAPE and MAD compared
units

3000 to the other two classical techniques; however, for one

2000 category, which displayed quite an irregular historical
1000 pattern, Fourier analysis reduced the error by 30% on
0 average between the two indexes; on the contrary, for the
129 132 135 138 141 144 147 150 153 156 159 162 165 168 171 174 177 180 second category, which displayed a more regular
Weeks
historical pattern, the average error reduction was 22%. A
Figure 11. Comparison of the forecast results. 10-step simple and straightforward procedure is described
and no complex data processing procedure is required.
Forecasting technique MAD MAPE Clearly, further improvements can originate from the
Moving average 532.0 55.1% application of those techniques that were conceived of in
signal theory to refine the Fourier transform (e.g.,
Belt

Exponential smoothing 533.1 55.2%

Fourier analysis 414.0 43.0% reducing spectrum leakage through an appropriate signal
windowing procedure). However, in this study the
Table 3. Comparison of forecast errors.
approach was intentionally described in its simplest form,
in order to provide a solid foundation both for
5. Conclusion researchers and practitioners.

Accurate forecasts are extremely difficult to attain. As a 7. References

result, large-sized companies may be pushed to purchase
expensive forecasting software, whilst small- and [1] Rexhausena D, Pibernik R, Kaiser G. (2012)
medium-sized enterprises generally choose to use simple Customer-facing supply chain practices—The impact
statistical approaches on common spreadsheets. of demand and distribution management on supply
However, common spreadsheets do not support only chain success. Journal of Operations Management.
basic techniques: for instance, the fast Fourier transform 30(4): 269-281.
(FFT) - which has been widely used and applied in many [2] Syntetos AA, Nikolopoulos K, Boylan JE. (2010)
fields ranging from physics, seismology, engineering and Judging the judges through accuracy-implication
economics - allows certain limitations present in other metrics: The case of inventory forecasting.
techniques in capturing seasonality phenomena to be International Journal of Forecasting. 26(1): 134–143.
overcome. Therefore, thanks to its capability in terms of [3] Graves SK. (1998) A dynamic model for requirements
decomposing a function into a sum of sinusoids of planning with application to supply chain
different frequencies, amplitude and phase, Fourier optimization. Operations Research. 46(3): 35-49.
analysis has been used for forecasting changes in [4] Fildes R, Goodwin P, Lawrence M, Nikolopoulos K.
electricity demand and/or prices. However, it seems to be (2009) Effective forecasting and judgmental
only rarely applied in consumer goods sales forecasting, adjustments: an empirical evaluation and strategies
and no contributions seem to be present in the literature for improvement in supply-chain planning.
for forecasting consumer goods sales, specifically in the International Journal of Forecasting. 25(1): 3-23.
fashion industry. In this paper, a comparison between the [5] Acara Y, Gardner ES. (2012) Forecasting method
FFT analysis in Microsoft Excel and another two selection in a global supply chain. International
techniques based on moving average and exponential Journal of Forecasting. 28(4): 842-848.
smoothing methods was carried out on a set of 4-year [6] Turrado García F, García Villalba LJ, Portela J. (2012)
historical sales data (3-years calibration data set, 1-year Intelligent system for time series classication using
validation data set) of a €60+ million turnover medium- to support vector machines applied to supply chain.
large-sized Italian fashion company operating in the Expert Systems with Applications. 39(12): 10590-
women’s textiles, apparel and clothing sectors, and which 10599.

www.intechopen.com Andrea Fumi, Arianna Pepe, Laura Scarabotti and Massimiliano M. Schiraldi: 7
Fourier Analysis for Demand Forecasting in a Fashion Company
 [7] Fisher ML, Hammond JH, Obermeyer WR, Ramen A. [21] Yu Y, Hui CL, Choi TM. (2012) An empirical study of
(1994) Making Supply Meet Demand in an Uncertain intelligent expert systems on forecasting of fashion
World. Harvard Business Review. May-June 1994: color trend. Expert Systems with Applications. 39(4):
83-93. 4383–4389.
[8] Xia M, Zhang Y, L. W, Ye X. (2012) Fashion retailing [22] Au KF, Choi TM, Yu Y. (2008) Fashion retail
forecasting based on extreme learning machine with forecasting by evolutionary neural networks.
adaptive metrics of inputs. Knowledge-Based International Journal of Production Economics.
Systems. 36: 253–259. 114(2): 615-630.
[9] Guo ZX, Wong WK, Li M. (2013) A multivariate [23] Hammond JH. (1991) Quick response in the apparel
intelligent decision-making model for retail sales Industries. Harvard Business School Background
forecasting. Decision Support Systems. 55(1): 217-255. Note 690-038, April 1991..
[10] Smith CD, Mentzer JT. (2010) Forecasting task- [24] Battista C, Schiraldi MM. (2013) The Logistic
technology t: The inuence of individuals,systems Maturity Model: application to a fashion firm.
and procedures on forecast performance. International Journal of Engineering Business
International Journal of Forecasting. 26(1): 144–161. Management. in press.
[11] Haq AN, Kannan G. (2006) Effect of forecasting on [25] Mostard J, Teunter R, de Koster R. (2011) Forecasting
the multi-echelon distribution inventory supply demand for single-period products: A case study in
chain cost using neural networks, genetic algorithm the apparel industry. European Journal of
and particle swarm optimization. Internation Journal Operational Research. 211(1): 139–147.
of Services Operations. 1(1/2): 1-22. [26] Wong WK, Guo V. (2010) A hybrid intelligent model
[12] Bregni A, D’Avino M, De Simone V, Schiraldi MM. for medium-term sales forecasting in fashion retail
(2013) Formulas of Revised MRP. International supply chains using extreme learning machine and
Journal of Engineering Business Management. 5(10): harmony search algorithm. International Journal of
1-10. Production Economics. 128(2): 614–624.
[13] D’Avino M, De Simone V, Schiraldi MM. (2013) [27] Sen A. (2008) The US fashion industry: A supply
Revised MRP for reducing inventory level and chain review. International Journal of Production
smoothing order releases: a case in manufacturing Economics. 114(2): 571–593.
industry. Production Planning & Control. in press - [28] Thomassey S. (2010) Sales forecasts in clothing
DOI:10.1080/09537287.2013.764579. industry: The key success factor of the supply chain
[14] D’Avino M, Bregni A, Schiraldi MM. (2013) A management. International Journal of Production
revised and improved version of the MRP algorithm: Economics. 128(2): 470-483.
Rev MRP. Applied Mechanics and Materials. 328: [29] Chen FL, Ou TY. (2011) Sales forecasting system
276-280. based on Gray extreme learning machine with
[15] D'Avino M, Macry Correale M, Schiraldi MM. (2013) Taguchi method in retail industry. Expert Systems
No news, good news: positive impacts of delayed with Applications. 38(3): 1336–1345.
information in MRP. International Journal of [30] De Toni A. (2000) The production planning process
Management and Decision Making. 12 (3), in press. for a network of firms in the textile–apparel industry.
[16] Nenni ME, Schiraldi MM. (2013) Validating Virtual International Journal of Production Economics. 65(1):
Safety Stock Effectiveness through Simulation. 17-32.
International Journal of Engineering Business [31] Chatfield DC, Hayya JC. (2007) All-zero forecasts for
Management. in press. lumpy demand: a factorial study. International
[17] Moon MA, Mentzer JT, Smith CD. (2003) Conducting Journal of Production Research. 45(4): 935–950.
a sales forecasting audit. International Journal of [32] Sanuwar R (2013) The Role of Quick Response for
Forecasting. 19(1): 19-25. Demand Driven Globalized Apparel Supply Chain
[18] Fildes R, Beard C. (1992) Forecasting systems for Management London: Springer.
production and inventory control. International [33] Martin C, Lowson R, Peck H. (2004) Creating agile
Journal of Production and Operations Management. supply chains in the fashion industry. International
12(5): 4-27. Journal of Retail and Distribution Management.
[19] Holt CC. (2004) Forecasting seasonals and trends by 32(8): 367-376.
exponentially weighted moving averages. [34] Caniato F, Kalchschmidt M, Ronchi S. (2011)
International Journal of Forecasting. 20(1): 5-10. Integrating quantitative and qualitative forecasting
[20] Lawrence M, Goodwin P, O’Connor M, Onkal D. approaches: organizational learning in an action
(2006) Judgmental forecasting: A review of progress research case. Journal of the Operational Research
over the last 25 years. International Journal of Society. 62: 413–424.
Forecasting. 22(3): 493-518. [35] Abraham B, J Ledolter J (2009) Statistical Methods for
Forecasting New Jersey: Wiley Interscience.

8 Int. j. eng. bus. manag., 2013, Vol. 5, www.intechopen.com

Special Issue Innovations in Fashion Industry, 30:2013
[36] Winters P. (1960) Forecasting sales by exponentially [52] Hülsmann M, Borsheid D, Friedrich CM, Reith D.
weighted moving averages. Management Science. (2012) General Sales Forecast Model for Automobile
6(3): 324-342. Markets and their Analysis. Transactions on Machine
[37] Mills TC (2013) A Very British Affair: Six Britons and learning and Data Mining. 5(2): 65-86.
the Development of Time Series Analysis During the [53] Battistoni E, Fronzetti Colladon A, Mercorelli G.
20th Century UK: Palgrave McMillan. (2013) Prominent determinants of consumer based
[38] Croston J. (1972) Forecasting and stock control for brand equity. International Journal of Engineering
intermittend demands. Operational Research Business Management. in press.
Quarterly. 23(3): 289-304. [54] De Felice F, Petrillo A, Autorino C. (2013) Key
[39] Paliwal M, Usha A. Kumar UA. (2009) Neural success factors for organizational innovation in the
networks and statistical techniques: A review of fashion industry. International Journal of
applications. Expert Systems with Applications. Engineering Business Management. in press.
39(1): 2-17. [55] Pirolo L, Giustiniano L, Nenni ME. (2013) The Italian
[40] Syntetos AB. (2005) The accuracy of intermittent footwear industry: an empirical analysis.
demand estimates. Internationl Journal of International Journal of Engineering Business
Forecasting. 21(2): 303-314. Management. in press.
[41] Ni Y, Fan F. (2011) A two-stage dynamic sales [56] D'Amico S, Giustiniano L, Nenni ME, Pirolo L. (2013)
forecasting model for the fashion retail. Expert Product Lifecycle Management as a tool to create
Systems with Applications. 38(3): 1529-1536. value in the fashion system. International Journal of
[42] Thommassey S. (2005) A global forecasting support Engineering Business Management. in press.
system adapted to textile distribution. International [57] De Carlo F, Borgia O, Tucci M. (2013) Bucket
Journal of Production Economics. 96(1): 81-95. brigades to increase productivity in a luxury
[43] Gutierrez RS, Solis AO, Bendore NR. (2004) Lumpy assembly line. International Journal of Engineering
demand characterization and forecasting Business Management. in press.
performance: an exploratory case study. In [58] De Carlo F, Arleo MA, Borgia O, Tucci M. (2013)
Proceedings of the WDIS 2004 Conference Layout design for a low capacity manufacturing line:
[44] Tsan-Ming C. (2012) Supply Chain Management in a case study. International Journal of Engineering
Textiles and Apparel. Journal Textile Science & Business Management. in press.
Engineering. 2(2): 1-2. [59] Iannone R, Ingenito A, Martino G, Miranda S, Pepe
[45] Yu Y, Choi TM, Hui CL. (2011) An intelligent fast C, Riemma S. (2013) Merchandise and replenishment
sales forecasting model for fashion products. Expert planning optimization for fashion retail.
Systems with Applications. 38(6): 7373–7379. International Journal of Engineering Business
[46] Strang G. (1994) Wavelets. American Scientist. 82(3): Management. in press.
250-255. [60] Costantino F, Di Gravio G, Shaban A, Tronci M. (2013)
[47] Lye KW, Yuan M, Cai TX. (2009) A spectrum Exploring bullwhip effect and inventory stability in a
comparison method for demand forecasting. seasonal supply chain. International Journal of
SIMTech technical reports. 10(1): 32-35. Engineering Business Management. in press.
[48] McLoughlin F, Duffy A, Conlonb M. (2013) [61] Nenni ME, Giustiniano L, Pirolo L. (2013) Demand
Evaluation of time series techniques to characterise forecasting in the fashion industry: a review.
domestic electricity demand. Energy. 50(1): 120-130. International Journal of Engineering Business
[49] Pedregal DJ, Trapero JR. (2007) Electricity prices Management. in press.
forecasting by automatic dynamic harmonic [62] Stremler F (1990) Introduction to Communications
regression models. Energy Conversion and Systems Third edition USA: Addison-Wesley.
Management. 48(5): 1710-1719. [63] Gonzalez O. (1990) Pitfalls of signal analysis.
[50] Lewis BG, Herbert RD, Bell RD. (2003) The Mechanical Engineering. February 1990: 72-74.
Application of Fourier Analysis to Forecasting the [64] Marhavilas PK, Koulouriotis DE, Spartalis SH. (2013)
Inbound Call Time Series of a Call Centre. In Harmonic analysis of occupational-accident time
Proceedings of the International Congress on series as a part of the quantified risk evaluation in
Modeling and Simulation MODSIM03); Townsville, worksite: Application on electic power industry and
Australia p. 1281-1286. construction sector. Reliability Engineering and
[51] Oladebeye DH, Ejiko OS. (2012) Developmento of System Safety. 112(April 2013): 8-25.
Fourier series forecasting model for predicting [65] Kerr DA. (2009) The Fourier Analysis Tool in
selected company sales volume. Science Engineering Microsoft Excel (working paper)..
Environmental Management Research & [66] Proakis J, Manolakis D (1992) Digital Signal
Development Journal. 1(1): 42-54. Processing : Principles, Algorithms, and Applications
New York: Macmillan Publishing Company.

www.intechopen.com Andrea Fumi, Arianna Pepe, Laura Scarabotti and Massimiliano M. Schiraldi: 9
Fourier Analysis for Demand Forecasting in a Fashion Company
 [67] Bartafai I. (2002) An illustration of harmonic [69] Armstrong JS, Collopy F. (1992) Error measures for
regression based on the results of the Fast Fourier generalizing about forecasting methods: Empirical
transformation. Yugoslav Journal of Operations comparisons. International Journal of Forecasting.
Research. 12(2): 185-201. 8(1): 69-80.
[68] Garcia-Ferrer A, Del Hoyo J. (1992) On trend [70] Wei WWS (1994) Time Series Analysis Boston:
extraction models: Interpretation, empirical evidence Pearson / Addison Wesley.
and forecasting performance. Journal of Forecasting.
11(8): 645-655.

10 Int. j. eng. bus. manag., 2013, Vol. 5, www.intechopen.com

Special Issue Innovations in Fashion Industry, 30:2013