VOL. 10, NO.

16, SEPTEMBER 2015 ISSN 1819-6608

ARPN Journal of Engineering and Applied Sciences
©2006-2015 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

DEMAND FORECASTING FOR PRODUCTION PLANNING IN

A FOOD COMPANY
N. de P. Barbosa, E.da S.Christo, and K. A. Costa
Department of Production Engineering, Fluminense Federal University, Volta Redonda, Brazil
E-Mail: elianechristo@id.uff.br

ABSTRACT
The food and beverage industry is one of the most important sectors of the Brazilian economy, with a significant
participation in GDP index. The Brazilian economy has been showing a relative stability in the last decades, which takes
the sales demand to be more predictable. Due to this scenario of economic stability, the companies has been worried about
investing in planning their operations, making use, mainly, of forecasting methods in order to become more competitive in
the market. In the case of food industry, the seasonal and the short perishability factors are a limitation to the maintenance
of stocks, requiring a forecast with a high accuracy level. The present work consists in applying methods to forecast the
demand for products of a food industry, which directs its sales to the food service market, in order to base the short to
medium term production planning. Posteriorly, the forecasts will be evaluated using the error measure MAPE and
compared to the demand currently considered by the company. The proposed methods feature a reduction of the error
approximately 5%.

Keywords: demand forecasting, production planning, exponential smoothing.

INTRODUCTION Exponential Smoothing Models

The supply chain activities planning and control The exponential smoothing models are based on
depends of accurate estimates of the volumes of products smoothing the past data of a time series to predict the
and services to be processed and the estimates come as future. Among the main advantages of this model are the
forecasts (Ballou, 2007). One of the biggest challenges of simplicity and the low cost of application. They are
food and beverage manufacturers is adjust the production recommended, mainly, when the time horizon is short.
and the stocks to minimize the loss of products due to its Exponential methods have also been applied in
short perishability. Time series analysis is very important other fields, as in the investigation of exponential random
in a wide range of applications, especially when it comes geometric graphs (RGGs) process models. The RGGs
to forecasting, and it encloses many different forecasting characterizes many randomly deployed networks, such as
models. However, it is necessary to determine which wireless sensor networks (Shang, 2009).
model best suits each situation (de Oliveira Silva et al. In the following, the main exponential smoothing
2014). methods will be presented: The simple exponential
There are countless models to develop forecasts smoothing method, the Holt’s method and the Holt-
presentedin the bibliography. From market research to the Winters method.
most complexmethods computational. The reference
(Ballou, 2007) presents the three main groups of these Simple Exponential Smoothing
methods, they are: qualitative, historicalprojections and According to (Krajewski et al. 2012), the simple
causal. Depending on the series and the desiredtime to be exponential smoothing is a sophisticated method of
forecasted it is possible to choose a technique thatbest fits. weighted moving average. In this method, each new
Besides choosing the best technique, the forecasting tobe forecast is gotten from the previous forecast, increased by
generated by the model chosen should be as close to real the error in the previous forecast which is corrected by a
aspossible (Junior and Filho, 2012). In other words, the smoothing coefficient.
errors of forecasting should beminimized, so the This method can be applied in forecasting stable
production managers plan the production inattention to the demand series, those who oscillate around a constant basis.
market and minimizing the costs. According to (Krajewski et al. 2012), the smoothing
The demand forecasting methods can be based in coefficient () balances the forecast sensitivity to the
mathematical models that use historical data or in demand changes and the forecast stability. This coefficient
qualitative methods, planned according to the has to be contained in the interval [0;1]. The greater is the
administrative experience and customers reviews. They value of, the faster the model will react to variations in
can also be based in a combination of both quantitative the real demand, because the adjustment will be more
and qualitative methods (Lewis, 1997). aggressive compared to the forecast error made in the
This study aims to analyze and forecast the sales previous period. Otherwise, the smaller is the value of ,
demand in order to improve the short to medium term the less aggressive the adjustment will be, in other words,
production planning. For minimize the group of analysis, the forecasts will be more smoothed by the previous
an ABC ranking was utilized to determine that products forecasts and the model will take more time to assume the
have bigger importance in demand and in sales. changes in the time series data pattern.

7137
 VOL. 10, NO. 16, SEPTEMBER 2015 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2015 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

Holt’s Method - Exponential Smoothing with Trend (8)

Adjustment
The exponential smoothing with trend adjustment Dt = demand for the period t;
sometimes referred as double exponential smoothing or Mt = forecast of the level for the period t;
Holt’s Method, is a variation of the simple exponential Tt= forecast of the trend for the period t;
smoothing method and is used to treat seasonal demand St = seasonal índex for the period t;
with trend (Alexandrov et al. 2012). = smoothing coefficient for mean (0 <  < 1);
This method uses two smoothing coefficients,  = smoothing coefficient for trend (0 <  < 1);
and (with values between 0 and 1) and consists in making  = smoothing coefficient for seasonality (0 <  < 1);
the forecast based in two factors: the forecast of the s = a complete seasonal period (Ex: s = 12 when having
average using the exponential smoothing and a trend monthly data and annual seasonality);
exponential estimative. = demand forecasting for the next k periods ahead.
Holt-Winters Method –Exponential Smoothing The index values for level, trend and seasonality
with Trend and Seasonal Adjustment can be calculated through Eq. 9 to 11:
According to (Giacon and Mesquita, 2011), the Holt-
Winters method incorporates not only trend, but also a (9)
seasonal component. The seasonal demand data are
characterized by the occurrence of cyclic patterns of


variation that repeat in constant time intervals (Tratar, (10)

2015).
The Holt-Winters method is based in three
(11)
smoothing equations that are associated to each one of the
series pattern components: level, trend and seasonality
(Christo et al. 2013). ABC Analysis
For series where the range of the seasonal cycle The ABC analysis is based on the Pareto’s
keeps constant over time (additive seasonality), the Principle, defined by the Italian economist Vilfredo
forecast can be calculated through Eq. 1 to 4 (Hyndman et Pareto, who says that the majoritarian values (80% of its
al. 2005): value) of a particular group comes from a relatively short
portion of its components (20 % of its quantity).
The most applied method to aggregate products is
(1)
the ABC ranking that determines the importance of the
(2) product, relating demand and sales (Krajewski et al.
2012).
In this case, the products can be grouped into
(3) three categories:
 A Items – Represents 80% of the company sales and
about 20% of the products sold. The demand forecast
(4)
is made individually for each product from this
category, however, it maybe manager interest the
Dt = demand for the period t;
stratification of the time series according to the
Mt = forecast of the level for the period t;
region, costumer or sealer;
Tt= forecast of the trend for the period t;
 B Items - Represents 15% of the company sales and
St = seasonal índex for the period t;
about 30% of products sold. As for the A category,
= smoothing coefficient for mean (0 <  < 1); the forecasting is made individually for each product
= smoothing coefficient for trend (0 <  < 1); from this category, however, it does not need
 = smoothing coefficient for seasonality (0 <  < 1); stratification. If any stratification is made over the
s = a complete seasonal period (Ex: s = 12 when having series correspondent to the A category, the statistical
monthly data and annual seasonality); treatment of the series are the same.
= demand forecasting for the next k periods ahead.  C Items– Represents 5% of the company sales and
For series where the range of the seasonal cycle about 50% of the products sold. The demand forecast
varies over time (multiplicative seasonality), the forecast is made in an aggregated way for the products of this
can be calculated through Eq. 5 to 8: category.

(5)

(6)

(7)

7138
 VOL. 10, NO. 16, SEPTEMBER 2015 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2015 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

The most applied method to aggregate products is

the ABC ranking that determines the importance of the
product, in demand and in sales (Eksoz et al. 2014).
In this case, the products can be grouped into
three categories as illustrated in the Figure-1.

Figure-2. Monthly Sales of Product 1 (Source: Author).

Figure-1. ABC curve (Source: Author).

The A items represents 80% of the company sales

and about 20% of the products sold. The demand forecast
is made individually for each product from this category,
however, it may be manager interest the stratification of
the time series according to the region, costumer or sealer.
The B items represents 15% of the company sales and
about 30% of products sold. As for the A category, the
forecasting is made individually for each product from this
category, however, it does not need stratification. If any
stratification is made over the series correspondent to the Figure-3. Monthly Sales of Product 2 (Source: Author).
A category, the statistical treatment of the series are the
same. The C items represents 5% of the company sales During this period, the product 1 has shown
and about 50% of the products sold. The demand forecast seasonal pattern, despite its aleatory pattern in the first
is made in an aggregated way for the products of this months. Most of time, cycles of 3 months can be
category (Schmit and Kaiser, 2006). observed, although this pattern is more marked in the last
10 months. We can also notice that the cycle range varies
Case Study over time. In the last 7 months is possible to notice a sales
The company to be analyzed in this case study is increasing trend.
a food industry that directs its sales to the food service During the analyzed period, an increasing trend
market. The midsize industry is located in Volta Redonda, can be clearly observed in the sales of this product 2. A
Rio de Janeiro and focuses its production in pasta and seasonal pattern with cycles from 3 to 4 months also can
sausage, composed of 6 different production lines that be observed, although this pattern is more marked in the
produces 49 different types of products and supply two last 10 months. The seasonal cycle range keeps constant as
restaurants chains and a pizzeria chain located all over the time goes by.
Brazil. From the analysis of the demand pattern through
For The ABC analysis, were considered the sales the graph (Figure-2), the most suitable method, for this
of all products from January 2012 to January 2014. This product, is Multiplicative Holt-Winters Method, because
plot allows not only identify how much each product this time series presents a marked trend in the last 7
represents over the total sales, but also, determine the three months and also a seasonal pattern, which amplitude
product categories. Among all products from A category, varies over time.
the first two products of the ranking, that represent From the analysis of the demand pattern through
together 27% of the company sales in the considered the graph (Figure-3), the most suitable method, for this
period were chosen to be analyzed. The monthly sales of product, is the Additive Holt-Winters Method, because
the selected products from January 2012 to January 2014 this time series presents not only trend, but also a seasonal
were plotted using the software Minitab®. The plots are pattern. The seasonal cycle range remains constant over
shown in the following figures (Figure-2 and Figure-3). time.
The respective Holt-Winters Methods were
applied using MS-Excel®. The index values of M0, T0, e
S0 were settled through the Eq. 5 to 7 and a seasonal cycle

7139
 VOL. 10, NO. 16, SEPTEMBER 2015 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2015 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

of 3 periods were considered in both cases. To determine volumes of products and services to be processed to satisfy
the optimal values for the α, β and γ coefficients the Solver customer’s needs.
tool of the software Excel® was used (Hyndman et al. It can be concluded that the Holt-Winters method,
2008). which was applied in the time series analyzed in this work,
The nonlinear programming problem solved by showed its effectiveness for forecasting demand of
Solver aims to determine the values of α, β and γ that products that presents trend and seasonality patterns in
minimize the value of MAPE (Mean Absolute Percent sales history. The use of the tool Solver of Excel® made it
Error), with the restrictions that α, β and γ must be possible to obtain smoothing coefficients in a simple way,
contained in the interval [0,1] (de Oliveira Silva et al. working as an effective alternative for obtaining them
2014). even though when the access to robust computational
After setting the Solver parameters, and ask for packages is not possible.
solving the problem, the tool provides the optimal values The method applied in this work showed its
for the smoothing coefficients. simplicity and accessibility due to the low cost and
The optimal values of ,  and  found for easiness of application. By having these characteristics,
product 1 were: this method can be used by small and medium-sized
companies, where is not possible to make huge
investments in planning their operations.
The optimal values of ,  and  found for The food products have a factor that limits the
product 2 were: maintenance of stocks, the short perishability. These
products have a period in which they keep their
characteristics and should be consumed before being
considered unsuitable for consuming. Thus, it is suggested
Considering the smoothing coefficients found and for future works that the short perishability of products
a seasonal cycle of 3 periods, the Minitab® software was must be taken into account when evaluating the results
used to forecast the demand of the product for the months obtained by the quantitative methods. To make possible
of February, March and April 2014 using the not only plan the production to satisfy the forecasted
Multiplicative Holt-Winters Method for product 1 and the demand, but also contribute to minimize the loss of
Additive Holt-Winters Method for product 2, with a products due to its short perishability and consequently,
confidence level of 95%. improving the profitability of the company.
Considering the forecasts obtained for the months
of February, March and April 2014 and the real sales in REFERENCES
this period, a Mean Absolute Percent Error (MAPE) of
1,38% was obtained for product 1 and of 3,14% for the [1] Alexandrov T., Bianconcini S., Dagum E. B., Maass
product 2. P. and Mcelroy T. S. 2012 'A review of some modern
To verify the performance of the obtained approaches to the problem of trend extraction',
forecast and the effectiveness of the proposed methods Econometric Reviews, Vol. 31, pp. 593-624.
over the method currently used by the company to forecast
de demand, a comparison was made using the value of [2] Ballou R. H. 2007. 'The evolution and future of
MAPE (Table-1). logistics and supply chain management', European
business review, Vol. 19, pp. 332-48.
Table-1. Comparison of the results.
[3] Christo E., Ferreira M. and Alonso K. 2013. 'Use of
Statistical Control for Improved Demand Forecasting',
Computational Intelligence and 11th Brazilian
Congress on Computational Intelligence (BRICS-CCI
& CBIC), 2013 BRICS Congress on. IEEE.

Through the MAPE values obtained with the [4] De Oliveira Silva R., Da Silva Christo E. and Alonso
proposed methods, it can be concluded that the applying of Costa K. 2014. 'Analysis of Residual Autocorrelation
these methods in the respective time series offer in Forecasting Energy Consumption through a Java
satisfactory results due to the significant reduction of the Program', Advanced Materials Research. Trans Tech
error approximately 5%when compared to the current Publ.
method.
[5] Eksoz C., Mansouri S. A. and Bourlakis M. 2014
CONCLUSIONS 'Collaborative forecasting in the food supply chain: A
Demand forecasts is, with no doubt, the basis for conceptual framework', International journal of
developing an efficient supply chain. The supply chain production economics, Vol. 158, pp. 120-35.
planning and control depends of accurate estimates of the

7140
 VOL. 10, NO. 16, SEPTEMBER 2015 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2015 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

[6] Giacon E. and Mesquita M. a. D. 2011. “A survey on

detailed production scheduling in manufacturing
plants in São Paulo, Brazil', Gestão & Produção, Vol.
18, pp. 487-98.

[7] Hyndman R. J., Akram M. and Archibald B. C. 2008.

“The admissible parameter space for exponential
smoothing models', Annals of the Institute of
Statistical Mathematics, Vol. 60, pp. 407-26.

[8] Hyndman R. J., Koehler A. B., Ord J. K. and Snyder

R. D. 2005. 'Prediction intervals for exponential
smoothing using two new classes of state space
models', Journal of Forecasting, Vol. 24, pp. 17-37.

[9] Junior M. L. and Filho M. G. 2012. 'Production

planning and control for remanufacturing: literature
review and analysis', Production Planning & Control,
Vol. 23, pp. 419-35.

[10] Krajewski L. J., Ritzman L. P. and Malhotra M. K.

2012. Operations management, Pearson Education
Limited.

[11] Lewis C. 1997. “Demand forecasting and inventory

control”, Control-Coventry-Institute Of Operations
Management, Vol. 23, pp. 20-23.

[12] Schmit T. M. and Kaiser H. M. 2006. “Forecasting

fluid milk and cheese demands for the next decade',
Journal of dairy science, Vol. 89, pp. 4924-36.

[13] Shang Y. 2009. 'Exponential random geometric graph

process models for mobile wireless networks', Cyber-
Enabled Distributed Computing and Knowledge
Discovery, 2009. CyberC'09. International
Conference on. IEEE.

[14] Tratar, L. F. (2015) 'Forecasting method for noisy

demand', International Journal of Production
Economics, 161, 64-73.

7141