Daniella
Frias
CEFET/RJ, Brazil
Email: daniellaffrias@gmail.com
Carolina
Cavour Siqueira Muniz
CEFET/RJ, Brazil
Email:
carolinacavour94@gmail.com
Pedro
Senna Vieira
CEFET/RJ, Brazil
Email: pedro.sennavieira@gmail.com
Dominique
Souza Sant’anna
CEFET/RJ, Brazil
Email: dominiquesouzasantanna@gmail.com
Augusto
da Cunha Reis
CEFET/RJ, Brazil
Email: augusto.reis@cefetrj.br
Submission: 2/28/2020
Revision: 3/2/2020
Accept: 3/11/2020
ABSTRACT
Demand forecasting has become a fundamental tool for companies' strategic planning. Represented by one of the highest growth rates in the country, the cosmetics industry faces numerous challenges in meeting the demand of consumers with a high level of service. Correctly identifying demand is critical to avoiding unnecessary extra costs for the business, such as stockout or stock over. The sales data of shampoo franchises are real values, covering the period from January 2013 to December 2018. After data organization, opentime and fixedtime time series techniques were analyzed in order to find the best forecasting technique for the type of product analyzed, i.e. the method with the smallest difference in absolute values between the actual demanded. and the estimated. The models were successfully applied, and we concluded that one of the analyzed methods could be applied in the company, because it presented smaller Mean Absolute Percentage Error.
Keywords: Risk; Mitigation; Supply; Demand; Forecast
1.
INTRODUCTION
The Supply Chain
analysis process makes it easy to make decisionmaking along the chain by
coordinating information and the flow of products and services. Among the
processes for supply chain management, one of the most important is demand management,
which seeks to align customers' needs with the company's ability to service
through a market perception. For Slack et al. (2003), without a demand
estimate, it is not possible to plan for future events, but only react to them.
According to Corrêa et al. (2010), this process is extremely important
for the company for identifying the factors that generates the initial
information of the production process. Several risks can lead to disruptions in
supply chains, and the consequences of economic crises are significant.
Therefore, it is necessary that companies are prepared for unstable situations
and, develop competitive advantages that involve supply chain resiliency, and
the ability to adapt to changes in the external environment.
One tool to protect
these companies, in view of the changes of recent years, is demand management.
As mentioned, demand forecasting activity is important to ensure the service of
the operating market and to avoid expenses due to overproduction and waste of
raw material or delay and lack of material when desired by the customer. From
this context, the present study analyzes methods for demand forecasting of a
cosmetics company.
In addition, this
paper also analyzes the demand forecasting techniques used, seeking to reduce
forecasting error and improve demand management, mitigating the threats mapped
in risk analysis. The study used the SCOR (Supply Chain Operations Reference)
framework that seeks to propose improvements, through the analysis of business
activities to manage the physical and temporal resources employed.
2.
LITERATURE REVIEW
Supply chain management
has been recognized as managing key business processes across the network of
organizations that compose the supply chain. In other words, it consists of
developing and seeking the continuous improvement of activities related to the flow
of transformation of products and services, from obtaining the raw material to
the arrival of the final product.
According to Juttner et al. (2007), Supply Chain Management focuses on
the efficiency of procurementrelated processes (demand fulfillment) and tends
to be costoriented, while marketing management seeks to generate revenue
(demand creation), identifying consumer value perceptions and translating into
product offerings. The implementation of structured processes within the
company has been stimulated over the last few years. It is now necessary to
integrate processes between the areas of the supply chain, reducing costs for
the supplier of the beginning of the chain, while adding value to the final
consumer meeting their needs.
Christopher and Peck
(2004) define resilience in supply chains as the ability to return to their
status quo or move to another more desirable after suffering disruption. For Ponomarov and Holcomb (2009), resilience in the supply
chain can be conceptualized as the ability to adapt the supply chain to
unexpected events, responding to interruptions, recovering and continuing
operations at the level Desired. One point that deserves to be highlighted in
demand management is the need for a contingency plan through internal or
external occasions that cause an imbalance in supply and demand.
The risk is related to
the occurrence of uncertainties. According to Heckmann
et al. (2015), the risk in the supply chain lies in the possibility of loss of efficiency
and effectiveness motivated by some event likely to occur. An example of a risk
to the supply chain would be the mismatch between supply and demand that would
result in disruption. For Juttner et al. (2007), this
incompatibility can be caused by lack of information, shortage of raw materials
and flow of unbalanced products.
Risk management
involves planning processes, identification, quantitative and qualitative
analysis, response planning and risk monitoring and control. Bradley (2014)
suggests that the supply chain risk management process follows the steps (a)
risk identification; (b) risk measurement; (c) prioritization of risk for
mitigation; (d) evaluation of risk mitigation tactics; (e) implementation of
mitigation tactics. Padoveze (2010) conceptualizes
risk as follows: opportunity (the higher the risk, the greater the potential
for return); hazard or threat (potentially negative events such as: financial
losses, fraud, reputational damage or theft, death or injury, system failures
or legal claims) and uncertainty (related to the distribution of all possible
outcomes, whether positive or negative).
In this context, Felea et al. (2013) states that the goal of having good and
structured risk management is to manage events related to any kind of
uncertainty in the face of an increasingly unstable markets. From this
perspective, Kleindorfer (2005) suggests that
effective risk mitigation can only be achieved through close collaboration
between supply chain partners. Thus, risk management also translates into the
resilience of the supply chain. The ability to adapt quickly to changes in the
competitive environment is one of the main skills that an organization can
develop to minimize the risks and impact of unmapped events.
The SCOR model that was
established in 1996 and is regularly updated to be used as a tool to align and
adapt to changes. With the main objective of integrating techniques, metrics
and best practices to ensure more effective communication of the various poles
between suppliers, products and services offered to the end customer, the
method seeks to improve the management of activities related to the improvement
of the supply chain, implement systems that support members, and prepare the
organization to better adapt to changes.
SCOR was developed to
describe all business activities associated with customer demand service and
can be used from the simplest to the most complex chains. The model is based on
the six main management processes: Plan, Source, Make, Deliver, Return and
Enable and can successfully be used to describe and provide a basis for chain
improvement involving specific and global projects.
The model describes
activities of an organization, i.e., focuses on the activity developed by the
organization and not on the person or element responsible for the activity. In
other words, it does not determine how the company should conduct its business.
It is necessary that the company understand which business model is inserted
and how to behave in its market of operation.
The SCOR structure
consists of 4 main sections:
(a) Performance: For this section it is
essential to define metrics to describe process performance.
(b) Processes: Description of the
management processes used and the relationship between these processes.
(c) Practices: Management practices that
produce significantly better process performance.
(d) People: Defining the skills needed
to execute supply chain processes.
Open model time series
techniques analyze series identifying their components and creating a unique
model that designs such components. A minimum of 48 periods of historical data
is recommended. Classic decomposition is a method used for demand forecasts
from time series in general made annually or monthly. According to Corrar and Theóphilo (2004) data
from a time series may be influenced by some macroeconomic, technological
factors, variations in nature and unpredictable phenomena. These and other
factors determine the components of the time series that need to be decomposed.
(a) Cycle (C): represents the undulating
motion or cycle of a time series. Consists in the oscillations or deviations
around the trend line.
(b) Seasonality (S): Series fluctuation
over a year. It refers to similar patterns that a series can offer and acts as
a multiplier index. May be influenced by the weather or commemorative dates
such as the sale of ice cream in summer.
(c) Residual fluctuations (U): It is the
component that represents the random fluctuations of the series, which are
hardly predicted.
The method consists of
decomposing the sales series in the components mentioned. The first step is to
evaluate the duration of the seasonal period. Then moving average of the
seasonal pattern of n duration periods is calculated, according to the
equation (1):

(1) 
This calculation purges
the seasonal effect and most of the random variations, maintaining the cyclic
trend and movement as in the equation (2).

(2) 
Thus, to identify the
trend of a series it is necessary to determine the curve that best adjusts to
the series of moving measures. Next, cyclic movements are calculated,
expressing them as a trend percentage, through the equation (3):

(3) 
As the original series
contains the trend, cycle, seasonality and residual fluctuations components and
the series of moving averages contains trend and cyclical variations, we must
divide the first by the second to have a series containing fluctuations and
seasonal effects (4):

(4) 
Then the seasonal
factors are calculated for each historical period. As random fluctuations are
being considered, it is expected that the seasonal factor of a period will be
different from the calculated factor of n periods ago. To work around
this question, seasonal indices are calculated, i.e., the average seasonal
factors of that period. The equation (5) presents the calculation of the
seasonal factor:

(5) 
If the historical
series is long enough, seasonal indices can be calculated from the adjusted
average, that is, purging the largest and lowest seasonal factor of each month.
By calculating seasonal indices, these can be used to measure unexplained
variations (random or residual). The forecast is performed by individually
projecting each of these components (trend, seasonality and cycle), and then
combine projections. The trend (T) is the component that shows growth or
decline over time and can be calculated through the equation (6):

(6) 
In equation (6), α
is the slope of the centered moving average and β the. For the initial
periods in which the moving average was not calculated, the first value of the
calculated cycle is repeated. For the final periods, the last calculated value
is repeated.
Seasonality (S) is the
fluctuation of the series over the course of a year. It refers to similar
patterns that a series can offer and acts as a multiplier index. It may be
influenced by the weather or commemorative dates such as the sale of ice cream
in summer. If the classical decomposition is calculated for an annual
seasonality time series, seasonality will be the average seasonal indices of
the same months of the year. Residual fluctuation (U) is the component that
represents random fluctuations in the series. Classical decomposition can
generate predictions through additive or multiplier models. For the additive
model, the only change would be in the final of the calculated components. The
prediction formula in the Classical Decomposition method  multiplicative and
additive method, respectively represented by equations (8) and (9).

(8) 

(9) 
Time Series Techniques of Fixed Models stand out mainly because they are of
simple implementation and do not require very large historical series. In this
way, most of these techniques quickly adjust to changes in sales behavior and
are thus appropriate for short and mediumterm forecasts. It is a commonly
used method and has as main advantages the operational ease and smoothing of
short fluctuations, since it considers the average of a certain number of
periods. Its application is indicated only for series without trend and
seasonality. The use of this method in time series with such behavior can lead
to unsatisfactory results because it considers that the forecast for a later
period involves the addition of new data and the disregard of the previous
ones. For these cases, a better choice would be the Dual Moving Average method.
The longer the calculation period of the simple moving average, the longer the
time limit of the monitored trend, according to equation (10):

(10) 
Where:
· = Forecast for the next period.
· = Forecast for the next period.
· = Actual value observed in the
period t.
· = Number of periods considered in
the moving average.
The Dual Moving Average
method is one of the methods for predicting trended series. The calculation can
be basically summarized in five steps:
(a) Calculation of simple moving average
(equation (11)):
(11) 
(b) Calculation of moving average based
on previously calculated moving average series (equation (12)):
(12) 
(c) Add to the simple moving average and
the difference between the two series of moving averages (α_{t})
according to the equation (13):
(13) 
(d) To consider the trend, the
additional adjustment factor (β_{t}), similar to an angular
coefficient (equation (14)):
(14) 

(e) Finally, the forecast for future
periods is calculated (equation (15)):
(15) 
Where:
· = Actual value observed in period t.
· = Number of periods considered in
the moving average.
· = Number of future periods to be
forecast.
The forecast for the period t is the amount sold in the period t1
(equation (16)):

(16) 
Where:
· = Forecast for the period t.
· = Real value observed in the t1
period.
The weight of demand
decreases in time in a geometric progression. In the case of the exponential
moving average, a new forecast with bigger weight is generated at the most
recent value, based on the previous forecast and from the calculation of an
error, this error is corrected by an alpha coefficient (equation (17)):

(17) 
Where:
· = Forecast for the period t.
· Actual value observed in the t1
period.
· = Weighting coefficient.
· Demand in the t1 period.
Simple
exponential smoothing’s main advantage is the fact that it is nonparametric
(not associated with a given probability distribution). In exponential moving
average all elements of the historical series have weighted importance
differently: the elements closest to time n have greater weight, while
the farthest ones have lower weight. On the other hand, as a negative point,
simple exponential smoothing does not consider possible trends of growth,
growth or seasonality. The forecast is given by the equation (18):

(18) 
Being
equivalent to the equation (19):

(19) 
· = Forecast for the next period.
· =
smoothing coefficient (0 ≤ α ≤ 1)
· = Actual value observed in the period t and.
· = Forecast for period t.
These models stand out
mainly because they are of simple implementation and use and do not require
very large historical series. Therefore, most of these techniques quickly
adjust to changes in sales behavior and are thus, appropriate for short and
mediumterm forecasts. When using DES, it is necessary to pay attention to the
initial values A_{0} and A'_{0} (first and second smoothing).
The use of the first observation for these values implies underestimating the
existing trend in a series and therefore they should be calculated according to
equations (20) and (21):

(20) 

(21) 
Where:
· = Linear regression coefficient of series
values.
· : = Angular coefficient of regression
of series values.
Holt's model is widely
used for forecast when the series presents randomness and a linear trend of
growth but does not present seasonality. In addition, it can be used when the
components of the series can be scorned. As it has a gradual and longterm
trend, the ideal is to assume that the behavior between demand and time is
linear. Α is the linear
coefficient, which will be the initial estimate of the mean, and β is the
angular coefficient, specific to adjust the trend estimate. The prediction in
the application of the Holt method is in the selection of these two
coefficients.
In this method, three
equations are used:
· Level: Equation that results from the addition
of the term related to trend estimation to the formulation of Simple
Exponential smoothing (equation (22)).
(22) 
· Trend: Equation used to adjust the
trend estimate by pondering the previous and the most recent from the β
coefficient (equation (23)) coefficient of regression of series values.

(23) 
Prediction: Equation that results in
forecast for p periods (equation (24)):

(24) 
Where:
· = Component level;
· = Trend component.
· = Smoothing coefficient (0 ≤ ≤ 1).
· = smoothing coefficient for trend estimation
(0 ≤ β ≤1).
· = Real value observed in period t.
· = Number of future periods to be
foreseen and.
· = Forecast for the period t + p.
Triple Exponential
smoothing is usually used when the time series has level, trend and seasonality
components. At first, the correct thing is to take the seasonality of the
series and calculate the level and trend to obtain demandbased factors after
extracting seasonality.
In the TES method there
are three equations – level N component, seasonal component S_{t} and T
trend component – and in addition we have three smoothing coefficients for
seasonality estimation – α, β, and all vary between 0 and 1. Below,
below, below seasonal adjustment equation for each period (equation (25)):

(25) 
In the equation, the
term (R_{t}/N_{t}) represents the seasonal adjustment for each
period t. The term S_{t } c refers to the seasonal adjustment
calculated c periods ago. Therefore, c = 12 is considered. The coefficient is
used to weigh the two plots. Equation for calculating trend component T, is
given by equation (26):

(26) 
The level is
calculated, considering the seasonal adjustment, through the equation (27):

(27) 
The forecast
is found through the equation (28):

(28) 
· = Seasonal component.
· = Component level;
· = Trend component.
· = smoothing coefficient (0 ≤ ≤ 1).
· = smoothing coefficient for trend estimation
(0 ≤ β ≤1).
· = Actual value observed in period t.
· = Number of future periods to be
foreseen and.
· = Forecast for the period t + p
and.
·
Ɣ = smoothing coefficient for
estimating seasonality (0 ≤ Ɣ ≤ 1).
As in the Holt Method,
the selection of smoothing coefficients is essential to obtain a satisfactory
degree of accuracy. The prediction through the exponential model presents
better trend and seasonality results when compared to the classic decomposition
model that uses simple moving averages.
3.
METHODOLOGY
The data used in the
present study were provided by employees in the Supply Chain department of a
French multinational company based in Brazil. The company shared data about the
demand and sales information and explained which forecast methods are used. After collecting sales data from the
products covered at this paper, analyses are performed from the application of
time series techniques for demand forecasting, both from the fixed model and
the open model. All analysis will follow the premise of adopting the model with
the lowest MAPE – Mean Absolute Percentage Error – which means the average
percentage of the error estimated in the forecast.
The scope of this paper
involves the processes of demand forecasting of a multinational company in the
cosmetics field, taking into account the fluctuations in market demand, high
competition due to low barriers of entry and strategy of production of large
quantities of inventory so that the product is always distributed and
available. Although the company operates in several countries, the study
centralizes its analysis in the forecast model used in the subsidiary installed
in Brazil, with national management and production and focuses on the
political, economic and social scenario of the country considering the risks of
production and supply chain throughout the process.
4.
CASE STUDY
The organization
studied is a multinational present in 130 countries and in 2018 was considered
the third most valuable French brand. However, despite the expertise gained
after years of experience in the beauty sector, the company needs to have a
detailed planning of the strategies and forms of action in the different
markets it operates, each with its specificity and consumption profile. The
economic crisis culminated in the closure of the Rio de Janeiro plant. This
fact indicates that despite high investment and planning, there are points that
can be improved. One of the reasons raised, in addition to the high cost to
maintain the factory, was the idle capacity of employees, a factor that should
be more analyzed mainly when there are situations of rupture of some items sold
by the company.
The current sales
forecast of the analyzed company is made through “future master”, a tool that
consolidates information on production capacity and availability of supplies
aligned with the demand and supply of factories. The method used in the
software is Double Exponential smoothing and MAPE is the error indicator used.
Nowadays, the MAPE obtained by the company is about 32%. The equation (29) is
used for calculating MAPE. The calculation of the percentage errors set out
below reflects the difference in absolute values between the actual value
demanded in a period t and the estimated value for this same period.

(29) 

(30) 
5.
ANALYSIS OF RESULTS AND DISCUSSION
Two datasets were
collected for analysis and testing of prediction methods, both from January
2013 to December 2018. The first dataset contains the sales of a onebrand
item: Shampoo 200ml of the ABC xyz brand franchise. In a second moment, sales
data of all Shampoos in the 200ml format of the xyz brand were collected, that
is, involving all franchises and resulting in about 35 products.
For the calculation of
both data series, techniques of open model time series  Classic Decomposition
with some variations in calculation  and fixed model  Simple Moving Average
(SMA) of two to ten periods were used to calculate the least error, Double
Moving Average (DMA) of nine and ten periods (periods with the lowest mms
error), Naive Method, Simple Exponential Smoothing (SES), Double Exponential
smoothing (DES) applying the Brown and Holt methods and Triple Exponential
Smoothing (TES) by applying the Winter method. Figure 1 and Figure 2 represent
the comparison between the forecasting methods for a specific shampoo item 200
ml for the ABC franchise and for franchises of all 200 ml shampoos,
respectively:
Table 1: Methods comparison summary for ABC 200 ml
shampoo
Table 2: Methods comparison summary of the 200 ml
shampoo family
After the application
and analysis of the methods and their variations it is perceived that there are
several variables that can determine the service of a company and influence its
positioning in the market. The level of service, for example, is a decision
that directly influences the customer's perception because it is the
measurement of processes according to the expectation and quality.
Making decisions to
increase the level of service means better service to customers, receiving the
products in the requested period, making ontime deliveries, producing quality
products and having a good aftersales service. Reviewing methods to improve
the level of service should be a constant study for highsized companies to
mitigate risks and position themselves better versus their market competitors.
Reducing the percentage
of forecast error means better alignment with market demands, leveraging
resources and labor more efficiently and making the company more prepared to
react to external factors. A low level of service enables it to increase but
remain at a level that ensures the minimum level of service required by
consumers. It is necessary to seek the balance between these two indicators.
In the highly
competitive beauty market and with the consumer's buying profile with low
loyalty, delaying the delivery of a family of shampoos for example in a
shopkeeper means that the customer will buy from the competition to keep the
shelves stocked. This fact presents several negative factors for the company
such as loss of space of its products on the shelves, decreased market share in
the customer, the public accustomed to buying in that retail environment
migrates to substitute products from others brands, in addition there can be
damage to the company's commercial reputation with the shopkeeper.
Studies and investment
to better empower teams can bring not only longterm financial gains, but also
operational for the company in question and to the Supply Chain area. With a
routine analysis of the forecast methods every two years, comparing the
company's strategy with market reality, it is possible to serve internal and
external customers with a higher level of service and reduce the average
percentage error of calculation. It is extremely important to keep a constant
study of identification and mitigation of risk variables always reviewed and
updated. Identifying the balance between the level of service and inventory
management is one of the biggest challenges that organizations face in demand
management. After all, the interpretation of the inventory level says a lot
about the demand forecasting process: Inventory surplus raises the financial
costs and the lack of inventory translates into loss of sales. In both, there
is the opportunity cost behind the misguided demand planning.
6.
CONCLUSION
The main objective of
this paper was to identify the current demand forecasting method in the company
analyzed and to evaluate the importance of generating improvements to the
existing process. Thus, other methods were analyzed that could return smaller
statistical errors. From the entire
theoretical framework studied, we observed quantitative methods of demand
management and apply them on top of the demand history of different product
franchises of the mentioned company. Quantitative models offer projected values
in statistical calculations, thus providing a better dimension of the expected
margin of error and, therefore, a greater accuracy in the information.
In the case of
qualitative models, the margin of error, as well as the forecast itself, is
based on subjective criteria, thus being able to influence from the
subjectivity of those who are analyzing the data. Therefore, the use of
qualitativeonly models may not be the best choice, and the influence of
quantitative easing scans is fundamental to bring unbiased results. Analyzing the case study in the company's
data in the cosmetics field, it was clear the need to review methods used every
two years mainly when it comes to a very competitive market such as beauty and
in a country like Brazil , with its cultural, political and environmental
complexities.
Looking for ways to
find the lowest percentage error of calculation for the forecast, linked with a
higher level of service for consumers and a study to minimize the costs of the
company is an objective that all teams should keep in mind and, with the
present study , it was clear the importance of having a team completely aligned
with the management of the supply chain and use of resources, since even a
company with 110 years of experience in the market and presence in 130
countries, presents opportunities for improvement in the effectiveness of its
processes.
REFERENCES
BRADLEY, J. R. (2014) An
improved method for managing catastrophic supply chain disruptions. Business
Horizons, v. 57, n. 4, p. 483495.
BURGO, R. N. S. (2005) Supply Chain Management: Uma
Introdução à um Modelo de Gestão da Cadeia de Suprimentos para Obtenção de
Diferencial Competitivo. Revista Científica Eletrônica de Administração, v. 5, n.
9, p. 17
CHRISTOPHER, M.; PECK, H.
(2004) Building the resilient supply chain. International Journal of
Logistics Management, v. 15, p. 113.
CORRAR, L.; THEÓPHILO, C. (2008) Pesquisa operacional
para decisão em contabilidade e administração, Ed. Atlas, São Paulo.
CORRÊA, H. L.; CORRÊA, C. A. (2010) Administração de
produção e operações, Ed. Atlas, São Paulo
FELEA, M.; ALVASTROIU, I.
(2013) Managing Supply Chain Risks. Journal Supply Chain Management
Journal, v. 4, n. 2.
JUTTNER, U.; CHRISTOPHER,
M.; BAKER, S. (2007) Demand chain managementintegrating marketing and supply
chain management. Industrial Marketing Management, v. 36 n. 3, p.
377392.
KLEINDORFER, P. R.; SAAD, G.
H. (2005) Managing disruption risks in supply chains. Production &
Operations, v. 14, p. 5368.
LAMBERT, D. M.; COOPER, M.
C. (2000) Issues in Supply Chain Management. Industrial Marketing Management, v. 29, p. 6583.
PADOVEZE, C. (2010) Contabilidade Gerencial: Um enfoque em sistema de informação
contábil, ed. Atlas, São Paulo.
PONOMAROV, S. Y.; HOLCOMB,
M. C. (2009) Understanding the concept of supply chain resilience. The
International Journal of Logistics Management, v. 20, p. 124143.
SLACK, N.; LEWIS, M. (2003) Operations Strategy, ed. Pearson
Education, New Jersey.