European Journal of Operational Research 220 (2012) 820–830

Contents lists available at SciVerse ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Innovative Applications of O.R.

Modeling a multi-attribute utility newsvendor with partial backlogging

Anderson J. Brito, Adiel T. de Almeida ⇑
Federal University of Pernambuco, Cx. Postal 7462, CEP 50630-970 Recife, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Traditional newsvendor models usually focus on single profit maximization or cost minimization
Received 24 October 2011 approaches. However, making a monetary estimate of the consequences of lost sales as a result of short-
Accepted 16 February 2012 ages is often a difficult task for many practitioners. Besides, there is still a lack of an explicit account of
Available online 24 February 2012
decision-making judgments on the multiple consequences of making decisions with regard to order
quantities. In order to deal with this problem, this paper presents a multi-attribute utility model for
Keywords: the newsvendor problem with regard to profit, the impacts of service level on corporate image and on
Multiple criteria analysis
customers’ goodwill, and the impact on the environment arising from the disposal of unsold products.
Multi-Attribute Utility Theory
Newsvendor problem
Demand is partially backlogged according to a decreasing exponential function of the waiting time.
Inventory control The fundamental principles and limitations related to the application of the model built are also
Partial backlogging discussed.
Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction Even though it has been widely studied, newsvendor models re-
main an important research topic and a useful theoretical tool for
Inventory management in complex environments is a process analyzing ordering policies as well as supply chain coordination
that generally presents direct impacts on the total costs and service initiatives (Zhou and Wang, 2009; Hau, 2010). The management
level of a company (and of the supply chain). This problem, and the of newsvendor-type products is a field of study that attracts
entire supply chain management under a more comprehensive researchers and practitioners, especially those with regard to
point-of-view, has been studied by means of Operations Research highly perishable items or environments in which technology
techniques for many years. In order to explore factors such as syn- develops quickly, in both of which life cycles get shorter and shorter.
ergy, conflicts of interest, and the effects of trade-off on inventory For these sorts of products, the shorter the life cycle, the more
holding costs, shortage costs and service level, simpler Economic similar to typical single-period items inventory behavior becomes
Order Quantity models (Hillier and Lieberman, 2005) have evolved (Khouja, 1999; Zhou and Wang, 2009).
into more complex models which address uncertainty, multiple Several models and extensions for the newsvendor problem are
objectives, coordination, incomplete and asymmetric information found in the literature. In a reference work, Khouja (1999) presents
and so on (Agrell, 1995; Benton and Park, 1996; Weng and McClurg, a taxonomy of a wide range of contributions, which include: exten-
2003; Zhou et al., 2004; Sucky, 2006; Jaber and Goyal, 2008). sions for objectives other than profit (such as maximizing the prob-
A reference problem within the context of inventory/ordering ability of achieving a target profit); extensions for different pricing
management is the classical analysis on how much to order/pro- and discounting policies; extensions for different utility functions;
duce for a single selling period, in order to minimize costs or to and extensions for more complex logistic frameworks, such as
maximize profit, by taking into consideration that demand is multi-product, multi-stage and multi-location models.
uncertain and that the stocks remaining after that period must One of the trends in the studies of newsvendor/inventory man-
be discarded or sold for a given bargain price (Kumar, 1988). This agement is to relax the premise of complete lost sales after a stock-
problem is widely known as the newsvendor problem or single-per- out, and to consider the possibility that part of, or even all the
iod problem (Hillier and Lieberman, 2005). As Keren and Pliskin demand, should be backlogged. As San José et al. (2006) point
(2006) stress, since its introduction by Whitin (1955), this problem out, many researchers have paid attention to demand backlogging
has been the subject of numerous research studies, which have on inventory management models. Zhou et al. (2004) present an
developed several extensions to its classical formulation. inventory model, where demand is a deterministic function of
time. The model assumes multiple and successive periods, and
the replenishment cost per order is considered as dependent on
⇑ Corresponding author. Tel.: +55 81 21268728; fax: +55 81 21268728x30. lot size. Demand is partially backlogged, and it is modeled as a
E-mail addresses: adiel.almeida@pq.cnpq.br, almeidaatd@gmail.com deterministic function of the waiting time until the next replenish-
(A.T. de Almeida). ment. Weng (2004) develops a newsvendor model to analyze the

0377-2217/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2012.02.027
 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830 821

coordination effects of decisions on order size taken between a were represented by fuzzy sets. The optimal order size for a fixed
manufacturer and a reseller in order to maximize their joint ex- period was set to minimize the total ‘‘possible’’ cost. Abad (2001)
pected profit. In his model, Weng addresses two scenarios of de- approached the newsvendor problem with backlogging in order
mand: in the first, all excess demand is backlogged and must be to determine an optimal price and an optimal order quantity, but
satisfied; in the second, it is assumed that all customers with backlogging costs and lost sales costs were ignored on that model
unsatisfied orders are willing to wait, but the reseller may choose because of the difficulty of estimating these parameters in practice.
not to meet any additional demand if the resulting profit is nega- The problem was set to maximize profit per period, and the back-
tive. For that situation, a threshold value q for the additional de- logged demand was a strictly decreasing function of the waiting
mand is set in order to start a second replenishment order. Zhou time.
and Wang (2009) extend the coordination analysis of Weng’s work Among most newsvendor models proposed hitherto, only single
to a partial backlogging situation. The fraction of the unfilled de- profit maximization or cost minimization approaches have been
mand that is willing to wait for the next replenishment is assumed considered. Besides the difficulties in estimating the costs of lost
to be known and fixed. sales as well as other monetary parameters, we still note a lack
San José et al. (2006) present an inventory model, where excess of an explicit account of decision-making assessments and prefer-
demand is partially backlogged according to a negative exponential ence judgments on multiple consequences that making decisions
function of the time that customers will wait until their orders are on one-period order quantities may bring. Current models do not
fulfilled. While Zhou et al. (2004) address the minimization of a to- directly explore decision-maker’s preferences and value judgments
tal cost function, the problem in San José et al. (2006) was set in on non-monetary factors such as customer goodwill, corporate im-
order to maximize the total profit per cycle. Total costs in their age under stock-out, pollution, and possible environmental im-
model comprise ordering costs, holding costs, backorder costs pacts due to the disposal of unsold products, for example. Also,
and costs of lost sales, which are assumed to be known parameters. they do not take into consideration value tradeoffs between these
Lodree (2007) presented a two-level supply chain within a factors and under-stock/over-stock costs. Instead, non-monetary
newsvendor framework, incorporating within shortages a combi- outcomes are often assumed to be implicitly considered within
nation of backorders and lost sales, and also allowing the buyer monetary parameters. As previously stated, these parameters are
to start emergency replenishments. The problem was addressed usually very hard to estimate, so such outcomes are converted into
with the objective of minimizing the total expected costs of the constraints or into a mono-attribute utility scale.
supply chain, within which a cost of ‘‘lost contract’’ between buyer This paper differs from the existing literature on newsvendor
and supplier, depending on the magnitude of shortage, was also topics by proposing a multi-attribute utility model for the newsven-
incorporated. Later, Lodree et al. (2008) presented a newsvendor dor problem based on the methodological background of Multi-
model with a similar objective function. However, they considered Attribute Utility Theory – MAUT (Keeney and Raiffa, 1976). A model
a complete backlogging situation for which backorder costs were was developed to explicitly represent the decision-maker’s prefer-
proportional to the lead-time for the emergency fulfillment, and ences and behavior regarding risk caused by the uncertainty of de-
the emergency replenishment process was characterized by multi- mand. An exponential backlogging rate is considered. Within a
ple shipments. wider decision-making framework, the newsvendor problem is ad-
Lee and Lodree (2010) studied the impatience behaviors of dif- dressed as a multiple criteria decision problem concerning other
ferent customers in newsvendor contexts under a condition of factors besides profit maximization or cost minimization. In this
time-dependent partial backlogging. The problem was formulated model, impacts of lost sales on corporate image and on customers’
in order to minimize total expected costs, and different backorder goodwill, as well as environmental impacts that may appear from
rate functions were used to classify customers as impatient, neu- scrapping unsold products, are also assessed in the decision-making
tral or patient regarding their willingness to wait for an emergency process of choosing an order quantity Q. This multi-attribute model
replenishment after a shortage. also helps to overcome difficulties involving monetary transforma-
Even if the backlogging phenomenon is not taken into consider- tions and assessments of costs due to the consequences of shortages
ation, and hence backorder costs are not calculated, several authors in complex contexts, where the use of economic analysis alone may
recognize that estimating cost parameters in newsvendor models induce a myopic representation of the aspects involved.
may be a difficult task for managers, especially for those costs re- This paper is organized as follows: In Section 2, some basic fea-
lated to shortages (Moon and Choi, 1994; Agrell, 1995; Petrovic tures of the newsvendor problem are presented; the backlogging
et al., 1996; Chen and Chuang, 2000; Abad, 2001). As Moon and condition is discussed; and a multi-attribute decision model is pro-
Choi (1994) stated, one basic reason is that stock-out costs are re- posed. An illustrative numerical experiment is presented in Section
lated to intangible components including impacts on customers’ 3, and a sensitivity analysis is discussed. Then, a discussion on re-
goodwill, and delays in other parts of the system. Agrell (1995) sults, in Section 4, highlights some aspects related to the back-
criticized the approach of minimizing total costs, and pointed out ground of the model and its application. Lastly, conclusions are
that a consistent monetary transformation for a mono-objective drawn and remarks made on this research theme in Section 5.
treatment of the inventory problem requires an adequate knowl-
edge of the costs involved in stock-outs, which is hard to find in
practice. 2. The multi-attribute newsvendor model
In order to deal with the absence of precise information on the
cost parameters of shortages, Moon and Choi (1994) presented a This section introduces a multi-attribute expected-utility model
distribution-free stochastic inventory model with a constraint on for the newsvendor problem. We consider a newsvendor decision-
the service level, while Chen and Chuang (2000) developed a news- maker (DM) who sells a short life-cycle product during a selling
vendor model with a shortage level constraint in order to deter- season. At the beginning of that season, the newsvendor faces
mine the optimal order quantity and optimal timing of an order. the choice of ordering a number of Q products for sale. The entire
Agrell (1995) proposed an inventory model formulated with three order quantity is received at the same time. Demand is stochastic,
objectives: minimization of total expected (regular) costs, minimi- and this choice must be made before demand realization x, which
zation of expected number of stock-outs, and minimization of ex- is known at the end of the period.
pected size of shortages. Petrovic et al. (1996) presented two fuzzy The DM must find a compromise solution between the potential
newsvendor models, where inventory costs, as well as demand, consequences of over-ordering and the potential consequences of
822 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830

under-ordering (Ryu and Yücesan, 2010). Risks of under-ordering that b is a strictly decreasing function of s. Additionally, uncertainty
are about the consequences of ordering or producing fewer items over emergency replenishment lead time is not taken into account
than necessary to meet demand, leading to stock-out and lost sales. (Abad, 2001; San José et al., 2006; Lodree, 2007; Vijayan and Kuma-
Risks of over-ordering are related to ordering or producing an ran, 2008; Hu et al., 2009; Lee and Lodree, 2010). In order to model
excessive number of items, leading to disposal of extra stock after impatient behavior of customers regarding waiting time, we assume
the selling season (Hillier and Lieberman, 2005). In this model, a that the backlogging function of demand is represented as per the
compromise solution is found by taking into consideration the con- following exponential form (Lee and Lodree, 2010):
sequences on Profit, Service Level and Environmental Impacts. In or-
bðsÞ ¼ eas ð1Þ
der to detail the concepts and formulations of these attributes,
prior considerations on emergency options and on the behavior where a is a parameter that may be statistically obtained by the
of demand are presented. newsvendor based on historical correlated data.
For many cases, the replenishment waiting time s in (1) may be
2.1. The emergency replenishment options considered as a function of the backlogged orders (and, therefore, a
function of b). However, since b is already influenced by s, this
If a stock-out occurs, we consider the case, where the newsven- would incur in mutual dependence between b and s, leading to un-
dor may decide to start emergency replenishment in order to avoid friendly recursive calculations. Thus, for simplification reasons, this
the worst consequences from lost sales – on profit, on market paper follows the assumption made by previous studies such as
share, and on customers’ goodwill. Prior to the selling season, the Lodree et al. (2008) and Lee and Lodree (2010): in order to obtain
newsvendor must contract an order size Q and a contingency plan a generalized interpretation of b as a function of the size of the
si e S = {s1, s2, . . . , sn} from a finite set of n emergency replenishment shortage, the backlogging lead time for emergency replenishment
schemes. is set to be proportional to (x  Q), the number of units short (Lod-
The emergency supply schemes are contingency plans that in- ree, 2007). Thus, the replenishment waiting time is represented as
clude sets of emergency replenishment measures, which generally follows:
yield shorter procurement lead times when compared to the news-
si ¼ ki ðx  QÞ ð2Þ
vendor’s original orders. As Lodree (2007) states, a shorter lead time
is obtained at higher ordering costs via emergency procurements, where ki is the constant related to the emergency scheme si.
overtime, extra production runs, faster delivery services, etc. In Eq. (2), the emergency-order lead-time is assumed to be lin-
Although these backorder costs are sometimes high, they may be ear regarding (x  Q). However, a different form of lead-time could
less critical than the consequences of possible lost sales on custom- be useful for other contexts, since the lead-time can sometimes be
ers’ loyalty and even on the firm’s future competitiveness. independent of order size. For instance, a concave function could
Each scheme si is represented by a time coefficient ki e K = be considered. In that case, another model could be built for a dif-
{k1, k2, . . . , kn} which has a unit backorder cost bi e B = {b1, b2, . . . , bn} ferent situation than the one approached in this study.
associated with it. This cost may be regarded as a price per unit Based on (1) and (2), the fraction of unmet demand under back-
that the company is willing to pay in contingency plan si, which logging can be written as a function of the magnitude of the short-
may be implemented to reduce the emergency supply time by age (Lodree, 2007):
means of a coefficient ki, where 0 < ki < 1. Therefore, the shorter
bðx  Q Þ ¼ ea ki ðxQ Þ ð3Þ
the emergency lead time desired, the higher the unit backorder
cost bi that the newsvendor must be able to pay for an emergency
batch. 2.3. Decision attributes
Since backorder cost bi influences the newsvendor’s assessment
of the size of the ‘‘original’’ lot Q (as further demonstrated), the As presented at the start of Section 2, the decision-making pro-
choice of si affects the ordering decision. Therefore, the decision cess of the newsvendor problem is modeled according to three
alternatives for the newsvendor correspond to bi-dimensional vec- dimensions or decision criteria: Profit, Service Level, and Environ-
tors [si; Q], where si is discrete and Q is continuous or discrete. Even mental Impacts, which are detailed as follows.
for discrete products, newsvendor models usually assume continu-
ous order sizes, which are taken as good approximations for solv- 2.3.1. Profit
ing such stochastic single-period problems. For a given emergency replenishment scheme si, profit after the
selling season may be summarized as:
2.2. The backlogging behavior of demand 
ðp  cÞQ þ ðp  bi Þbðx  Q Þ  lð1  bÞðx  Q Þ if x P Q
pðsi ; Q Þ ¼ ð4Þ
px  cQ  hðQ  xÞ if x < Q
We consider a newsvendor who faces a partial backlogging con-
dition, where only part of demand will wait for an emergency sup- where the following notation is used:
ply during the stock-out interval. Following the notation
established in deterministic demand models (Montgomery et al., p – Newsvendor profit.
1973; San José et al., 2006; Pentico et al., 2009), the proportion p – Retail price of the products sold.
of shortages under backlogging is represented by b, which may also c – Regular unit ordering cost from supplier (or from the pro-
be interpreted as the probability that a single customer will be duction plant).
backlogged during the stock-out period, or the fraction of a cus- bi – Unit backorder cost related to the emergency replenish-
tomer’s unfulfilled demand to be backlogged (Lodree, 2007; Lee ment scheme si.
and Lodree, 2010). l – An opportunity cost of lost profit that may be financially
In several papers (see, for example, Montgomery et al., 1973; estimated.
Khouja, 1996; Vijayan and Kumaran, 2008; Hu et al., 2009; Leung, h – Unit disposal cost for unsold items less salvage value.
2009; Chang and Lo, 2009), the fraction b is assumed to be a fixed
and known parameter. However, an alternative approach addresses Maximizing expected profit or minimizing total expected cost is
b as a function of the replenishment waiting time s during the generally a single objective in most classical single period models
stock-out period (Abad, 2001; San José et al., 2006). We consider as well as in many research studies on this theme. There are also
 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830 823

several models the objective function of which is to maximize the 2.3.3. Environmental Impacts
expected utility (see Eeckhoudt et al., 1995; Khouja, 1999; Keren Concurrently with profit-maximization behavior and service le-
and Pliskin, 2006; Wang et al., 2009, for examples). However, they vel-maximization behavior, the DM also seeks to reduce the effects
usually have only considered a given economic attribute, such as of pollution generated by scrapping unsold products. Sometimes,
profit, cost or wealth. scrapping involves the disposal of hazardous material and toxic
Although they are single or heavy-weighted factors for ordering substances on landfill sites, the use of incineration facilities, etc.
decisions in inventory control, cost/profit analyses are often not For example, products such as mobile telephones and computer
able to translate all the complexity in a business environment. equipment have parts that contain harmful metals such as lead,
They also may be insufficient for newsvendor-type companies, nickel and cadmium, which require appropriate treatment on dis-
especially for those acting in highly competitive markets, for those posal in order to avoid harming humans and the environment. Be-
selling products with an important social role, and also for prod- cause of its potentially hazardous nature, the scrapping of
ucts closely associated with the corporate image. In these contexts, electronic devices is considered to be a priority waste stream in
managerial subjective assessments of non-monetary factors are Europe (Clift and Wright, 2000). Minimizing excess stock, there-
also taken into account. fore, may reduce the risk of environmental impacts related to the
In order to tackle the newsvendor problem in a more compre- final disposal of unsold items.
hensive way, a new approach that goes beyond this profit criterion We consider such a newsvendor, who is also concerned about
is needed to address the multiple dimensions of outcomes related reducing the Environmental Impacts of solid waste emissions due
to choosing an order size Q. These dimensions can be regarded as to the disposal of unsold items. This is more common for highly
multiple attributes or multiple objectives to be achieved. There- perishable products or those with a short life-cycle. It is considered
fore, the multi-attribute newsvendor evaluates impacts and that these products cannot be stored for future purposes, and the
trade-offs concerning these other attributes in the decision-making remaining stock after the selling season must be dumped.
process, and thus Eq. (4) alone cannot represent the decision-ma- In a simple context of reverse logistics and recycling networks
ker’s objectives. being absent, we assume that environmental impacts are propor-
tional to the stock of unsold items to be discarded. The conse-
2.3.2. Service level quences on this attribute may be represented as follows:
In this paper, the consequences for Service Level are associated 
0 if x P Q
with the proportion of lost sales over total demand x faced by EIðQ Þ ¼ ð6Þ
the newsvendor, as shown in: wðQ  xÞ if x < Q
(
ð1bÞðxQ Þ where w is a constant associated with the environmentally unde-
x
if x < Q
LSðsi ; Q Þ ¼ ð5Þ sired physical or chemical properties per unit scrapped. This de-
0 if x P Q
pends on the particularities of the product and on the final
For generalization purposes, this is assumed to be the fraction of disposal destination, for example, grams of lead per unit in landfill,
the newsvendor’s market with orders that have not been filled. For mass of toxic gas emitted per unit incinerated, etc.
the case in which each customer orders only one single unit, the Reducing environmental impacts is an important objective for
proportion of lost sales represents the fraction of customers who companies with corporate policies on social responsibility (Cruz,
balked. 2009), and for which the adoption of commitments with respect
The service level provided by an order size Q may affect vari- to the environment are strategic issues which address customers’
ables other than profit, inventory holding costs, opportunity costs environmental concerns, tighter governmental regulations and
and business financial indicators. In an under-ordering situation, the risk that investors perceive in investing in companies which
a service level associated with the size of stock-out, the number themselves may pose risks to the environment (Rao and Holt,
of stock-outs or the backlogging lead-time may critically affect 2005; Cruz, 2009). These companies are concerned about minimiz-
relationship with customers, may erode customers’ loyalty and ing emissions of gases or effluents, the disposal of waste, and so
jeopardize the company’s market-share and long-term competitive forth. These concerns can take the form of non-monetary objec-
position. The more relevant these factors are, the harder it is to set tives to be considered in production/purchasing ordering deci-
values for shortages and the cost of lost sales. In this situation, esti- sions. As Cruz (2009) points out, a stance of commitment to the
mating the impacts of shortages merely as the opportunity costs of environment helps to enhance corporate image, and also mini-
lost revenues may be unrealistic, and thus weaken the decision- mizes the possibility of stressing the relationship with regulators,
maker’s confidence in the recommendations from the model local communities and environmental groups. Most of all, mini-
(Agrell, 1995). mizing the generation of waste reduces the chance of being re-
Despite the difficulties in estimating in monetary terms the garded as negligent by public opinion in situations where there
consequences of service levels for certain newsvendor situations, occur real incidents related to soil contamination, water pollution
it is possible to explore the decision-maker’s preferences and and other forms of damage to the environment.
trade-off assessments between economic results and those intan- Due to the complexity of environmental consequences from
gible consequences yielded by an order [si, Q]. For example, a news- waste generation, sometimes these consequences cannot be di-
vendor may be willing to accept a reduction of 5% in expected rectly estimated from the mass of waste generated (Clift and
profit in order to increase the expected service level by 20%, which Wright, 2000). However, in order to set and illustrate the environ-
may be obtained by reducing lost sales through a higher lot size Q mental attribute for newsvendor decision-making, we take the
and a faster contingency scheme si. With adequate support from assumption previously highlighted that the environmental impacts
company staff in order to evaluate the consequences generated are proportional to the stock of remaining units to be discarded. As
by different proportions of lost sales, this attribute is thus used for the Service Level attribute, the support of staff from Product
by the DM to analyze the impacts of different options for [si; Q] Development, Health and Safety, Environmental Management
on customer’s goodwill, on the corporate image and on the com- and related areas is needed in order to help the newsvendor officer,
pany’s competitive position. After gaining these insights, the news- as a decision-maker, to identify the nature of parameter w to be
vendor position regarding service level is represented on a utility tackled and to evaluate the environmental risks for different op-
preference scale, as further detailed. tions for Q.
824 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830

From (6), it can be observed that Environmental Impacts is the The expected performance of [si; Q] in the Service Level attribute,
only attribute which does not depend on the chosen value for ki. with regard to the uncertainties about x, is obtained from (3) and
In an under-ordering scenario, where the emergency replenish- (5) as follows:
ment scheme si is started, there will be no more stocks after meet- Z Z  
Q
ing the backlogged demand. So, this attribute is influenced only by
1
ð1  eaki ðxQ Þ Þðx  QÞ
U SL ðsi ; Q Þ ¼ 1  f ðxÞdx þ uSL f ðxÞdx ð10Þ
the second component of the decision vector [si; Q]. 0 Q x

According to (10), the utility of [si; Q] in the Service Level attri-

2.4. Multi-attribute decision-making bute is assessed as the expected utility of the proportion of lost
sales given that [si; Q] is chosen. For the cases where x < Q (first
In this paper, we seek to maximize the decision-maker’s utility integration), the proportion of lost sales is zero and the utility on
function regarding the multiple concerns presented in Section 2.3. Service Level – on a [0; 1] scale – is 1.
The use of von Neumann and Morgenstern utility functions over Unlike the previous attribute, the consequences of [si; Q] in Envi-
the consequences instead of the consequence values themselves ronmental Impacts are actually related only to Q, by means of the
allows the newsvendor model to incorporate the decision-maker’s expected utility of the impacts generated by scrapping the unsold
preferences and behavior regarding risk, due to uncertainties that items. They also depend on the uncertainties of demand (6). This is
inherently exists in this problem (Keeney and Raiffa, 1976). represented by means of the following integration (11):
The next step of this model is to obtain a general utility value Z Q Z 1
for every option [si; Q] based on its performance on Profit, Service U EI ðsi ; Q Þ ¼ uEI ½wðQ  xÞf ðxÞdx þ 1  f ðxÞdx ð11Þ
0 Q
Level, and Environmental Impacts. This is brought about by eliciting
the utility functions on the set of three-dimensional attributes for In the cases, where x P Q (second integration), waste disposal is
the newsvendor decision problem, presented in Section 2.3. This zero and the utility on Environmental Impacts is 1.
process involves obtaining subjective data from the DM in accor- By taking into consideration the uncertainty of demand, the
dance with some structured elicitation protocols, which are ap- decision problem involves selecting a pair [si; Q] that maximizes
plied in order to fit a utility function that is representative of the the expected value of newsvendor’s multi-attribute utility func-
DM’s preferences (see Raiffa, 1970; Keeney and Raiffa, 1976). tion in (8). Since expected value calculations present additive
The newsvendor multi-attribute utility function may be sum- property, this is obtained by applying to (8) the integrations from
marized as in: (9)–(11), which results in the following function to be maximized
(12):
Uðsi ; Q Þ ¼ U½pðsi ; QÞ; LSðsi ; Q Þ; EIðsi ; Q Þ ð7Þ 0R 1
Q
In this model, a decision-maker’s compensatory rationality is 0
up ½ðh þ pÞx  ðc þ hÞQ f ðxÞ dx
B R1 C
assumed, and the three-dimensional expression in (7) is assumed Max kp B
@ þ Q up ½ðp  cÞQ þ ðp  bi Þðx  Q Þea ki ðxQ Þ C
A
si; Q
to be additive. This is a typical multi-attribute utility function lðx  Q Þð1  ea ki ðxQ Þ Þf ðxÞdx
which is often found in practice (Vincke, 1992), for example, for Z Z 1 "  # !
problems in logistics and in maintenance (Almeida and Bohoris,
Q
1  eaki ðxQ Þ ðx  Q Þ
þ kSL 1  f ðxÞdx þ uSL f ðxÞ dx
1996; Almeida, 2005; Almeida, 2007; Ferreira et al., 2009). How- 0 Q x
ever, there are situations where this function may not be applied, Z Q Z 1 
as discussed in Brito et al. (2010). The use of additive utility func- þ kEI uEI ½wðQ  xÞf ðxÞdx þ 1  f ðxÞdx ð12Þ
0 Q
tions implies a decision-maker’s independence of preferences
among Profit (p), Service Level (SL) and Environmental Impacts Besides choosing Q, selecting a pair [si; Q] involves assessing the
(EI) (for more details, see Keeney and Raiffa, 1976). Thus, the mul- options from the set S = {s1, s2, . . . , sn}, which is available to the
ti-attribute utility function in (7) can be obtained based on the newsvendor’s procurement division. Every si will affect the Profit
one-dimensional utility functions up(si; Q), uSL(si; Q), and uEI(si; Q), performance through the related backorder cost bi and backlogging
as follows: parameter ki (9), while the Service Level performance is also influ-
enced by ki (10). Thus, for every emergency replenishment scheme,
Uðsi ; Q Þ ¼ kp up ðsi ; Q Þ þ kSL uSL ðsi ; Q Þ þ kEI uEI ðsi ; Q Þ ð8Þ the newsvendor decision-maker chooses a value for Q that maxi-
The following property holds: kp + kSL + kEI = 1. The parameters mizes (12). The pair [si; Q] that offers the highest value for (12),
kp, kSL and kEI are scaling constants estimated by means of elicita- as per the preferences expressed by the decision-maker, is the
tion procedures based on comparisons of payoff lotteries (Keeney solution recommended for the multi-attribute newsvendor
and Raiffa, 1976). problem.
When the inherent uncertainties of the newsvendor model are As an examination of expressions (9)–(11) shows, the choice of
taken into consideration within a Decision Analysis framework [si; Q] is influenced by the conflicting nature of the decision attri-
(Berger, 1985), the utility of each option [si; Q] in each attribute butes. While lower values for Q (order sizes) may increase or re-
in (8) is assessed as the expected utility of [si; Q] for deterministic duce expected profit (depending on the offset of the expected
consequences on those attributes. For the Profit attribute, the util- costs of under-ordering and over-ordering), they reduce the
ity for every pair [si; Q] may be derived from (3) and (4) as follows: expected service level offered by the newsvendor to his market.
However, they imply the risks of dumping unsold items are lower,
8R
> Q and therefore, they reduce the possible adverse environmental im-
< 0 Rup ½ðh þ pÞx  ðc þ hÞQ f ðxÞdx
>
pacts of the newsvendor activity. On the other hand, higher values
U p ðsi ; Q Þ ¼ þ Q1 up ½ðp  cÞQ þ ðp  bi Þðx  Q Þea ki ðxQ Þ ð9Þ
>
> of Q increase the service level by reducing the expected proportion
:
lðx  QÞð1  eaki ðxQÞ Þf ðxÞdx of lost sales, although they incur higher environmental risks (and
therefore a lower environmental utility) related to discarding some
where f(x) represents the probability density function of demand; excess stocks. Furthermore, for that case, higher order sizes do not
up denotes the mono-attribute utility function for profit; and Up de- necessarily yield higher profit values.
notes the expected value of this function by taking into consider- Conflicts are also present when choosing si. Lower values for ki
ation the uncertainties about x. (faster replenishment schemes) are possible by means of higher
 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830 825

backorder costs bi. However, such values provide shorter backlog- within the company context. Preliminary meetings are scheduled
ging time for customers and reduce the proportion of lost sales with this group in order to present the main concepts of Multi-
(thus increasing service level). Therefore, trade-offs and preference Attribute Utility Theory (Keeney and Raiffa, 1976) and Decision
judgments regarding si must be made on the multiple objectives in Analysis (Raiffa, 1970; Berger, 1985). The steps of this model are
order to identify the best compromise solution for the multi- also explained. Furthermore, an adequate identification of the deci-
attribute newsvendor. sion-maker is stressed as a fundamental issue for this process be-
cause of the judgments that need to be made on the monetary,
service level and environmental consequences. The decision ana-
3. Numerical experiment and sensitivity analysis
lyst also emphasizes that the DM should have enough power to
solve conflicts of interest between the company’s departments,
In this section, we present a numerical study in order to illus-
and to make trade-offs among the attributes in order to set an or-
trate this multi-attribute approach for the newsvendor problem.
der size Q representing the best ‘‘compromise’’ among the conflict-
It is generally not feasible to obtain a closed-form expression for
ing decision attributes.
the compromise solution of this model; therefore, a computer tool
As to applying this model, the Operations Executive is identified
has been used to find the results.
as the decision-maker. A multi-attribute utility function U[p(si, Q),
Consider a retail company which is going to trade a computer
SL(si, Q), EI(si, Q)] is then elicited according to his expressed prefer-
product with a short-life-cycle, during a selling season, in a news-
ences. As previously stated, the multi-attribute utility function in
vendor context. After this period, it is expected that this product
(7) is assumed to be additive for illustration purposes. This implies
will be replaced by an upgraded version which is currently under
that the decision-maker’s expressed preferences present an inde-
development. This product is sold in a highly competitive market
pendent nature among the attributes p, SL and EI. Therefore, utility
and belongs to a series of products which has been historically
functions are obtained for each attribute.
associated with the company’s trade mark and corporate image.
For the case of the Profit attribute, the analyst follows an elici-
In this numerical study, it is assumed that demand is normally
tation protocol as summarized on the following lines. First, the
distributed (Khouja, 1995; Chen and Chuang, 2000; Chung and
Executive is invited to express a profit result p⁄ that is reasonable,
Flynn, 2001; Mostard and Teunter, 2006; Su and Pearn, 2011),
but sufficiently high, for the Executive (DM) to feel confident in
although this model is not restricted for this particular case. The
associating a utility value of 1 for this consequence. Latter, in the
parameters of demand are l = 240,000 and r = 80,000. Ordered
opposite direction, the analyst asks the DM to identify p0, a profit
items are sold at a retail price p = 250, and are commonly pur-
result (loss) that is sufficiently low for the DM to feel confident in
chased at a cost c = 150 per unit. The net unit disposal cost less
associating a utility of 0. Then, for some points p0 within the range
residual value is h = 60. A lost sale cost l = 90 may be accounted
[p0; p⁄], the analyst presents some hypothetical lotteries involving
for, based on a low-risk alternative investment of lost profit due
p0 , p0 and p⁄, for the DM to compare and to express his preference
to a unit of demand unmet. The backlogging behavior of demand
judgments. Based on the theoretical framework of Decision Analy-
is described as per (1), and analysis of the historical data from re-
sis (Raiffa, 1970; Berger, 1985; Keeney and Raiffa, 1976), the an-
lated products indicates a = 5  106 as a good estimate for the
swers are used to calculate utility values up (p0 ) for these points.
parameter of b function.
Next, the consequence values p0 and the associated utility values
In order to cope with the uncertainty on demand, the company’s
up (p0 ) are plotted onto a dispersion graph. A regression method
procurement division asks suppliers for options for emergency
is then used in order to find the function that is best adjusted to
replenishment schemes. Based on different configurations for pro-
these points.
duction schedules and delivery services, seven emergency options
For the Profit attribute, the following logistic-type utility func-
are obtained. These schemes are generally represented, for illustra-
tion was obtained based on plotted points of profit-utility values
tion purposes, by the following set of emergency replenishment
elicited from the newsvendor officer:
alternatives and their associated backorder costs (Table 1):
For every si in Table 1, an ‘‘optimal’’ regular order size Q⁄ is cal- 1
culated, in a parallel analysis, following the traditional model for up ðsi ; Q Þ ¼ ð13Þ
1 þ ed½pðsi ;Q Þn
maximization of expected profit. However, the company’s opera-
tions officer decided to run this multi-attribute model in order to where d = 6.5  107 and n = 11  106 (R2 = 0.90) . This function was
set Q under a more comprehensive approach regarding profit, ser- able to translate the decision-maker’s risk-prone behavior for
vice level and environmental impacts. This study was undertaken losses, relative risk neutrality for small monetary gains, but a risk-
from a managerial point of view, in order to assess preferences, averse attitude vis-à-vis high levels of profit on hand.
to translate behavior regarding risk, and to evaluate the influence After that, consistency checks are performed when, for new
of trade-offs among economic, service level and environmental consequences, utility values calculated from the curve equation
goals on the order size Q. are compared to utility values obtained through elicitation lotter-
A group of company staff is initially selected from the Logistics, ies analyzed by the decision-maker. This process is carried out in
Trading, Marketing and Environment departments to help a deci- order to ensure that the utility function obtained in (13) consis-
sion analyst consultant to carry out the application of this model tently represented the decision-maker’s preferences.
The same procedure is applied for Service Level consequences
and latter for Environmental Impacts outcomes in order to find each
Table 1 utility function. When analyzing the Service Level attribute, which
Set of available emergency replenishment schemes.
was modeled by means of the proportion of lost sales over total
Option ki bi demand, the utility values expressed by the decision-maker
s1 0.95 160 showed a decreasing exponential tendency as the proportion of
s2 0.70 205 lost sales increased. The following exponential curve was obtained:
s3 0.60 250
s4 0.50 290 uSL ðsi ; QÞ ¼ euLS% ðsi ;Q Þ ð14Þ
s5 0.40 350
s6 0.30 425
where u = 4. This negative exponential form provided a good trans-
s7 0.25 475
lation of the decision-maker’s preferences (R2 = 0.93), for which
826 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830

higher percentage values for lost sales were much more undesirable As can be seen throughout the third row of Table 2, [s2; 184,
than lower ones. LS(si, Q) was calculated as per (5). 643] provided the DM with the most satisfactory compromise be-
The company is also concerned about minimizing the release of tween the three dimension of impacts involved. Therefore, the
hazardous materials into the soil as a consequence of potential ex- emergency replenishment option s2 (ki = 0.70, bi = 205) and a regu-
cess stock. In the context of this company, unsold products are lar order size Qc = 184,643 form the vector recommended to the
scrapped and several components from these products are buried newsvendor as the solution for this decision-making problem, for
in landfill sites. The Environmental Manager highlighted the nega- they provided the most preferred trade-off and thus the highest
tive impacts of any groundwater pollution or soil contamination on expected utility among the whole set of decision options.
the corporate image, mainly with regard to the company’s stated When [s2; 184,643] is compared to the ‘‘classical’’ newsvendor
position on its responsibility to the environment. A value for solution [s2; 207,956], for example, the compromise lot size (Qc =
parameter w in (6), equal to 0.9 kg/unit, was set by the Environ- 184,643) causes a loss of 2.86% in expected profit E[p] and an in-
mental Management department as the amount of heavy metals crease of 2.3 percentage points in the expected proportion of lost
present in each scrapped unit. In a similar way to the Service Level sales, E[LS]. However, due to the DM’s behavior regarding risk on
attribute, utilities on the Environmental Impacts attribute presented profit, the lower order size of Qc provides to this newsvendor a
a decreasing exponential behavior. Using a curve fitting procedure higher expected utility value Up[si; Q] in the economic dimension.
over some utility points that had been elicited, the following expo- Additionally, the higher performance of [si; Q⁄] on the service level
nential expression was obtained: dimension was offset by a better performance of [si; Qc] on the
environmental attribute, where Qc provides a reduction of more
uEI ðQ Þ ¼ ewEIðQÞ ð15Þ than 6 tonnes (37.55%) on the expected amount of harmful metals
released into the environment within scrapped units. It means that
where w = 3  104 (R2 = 0.89) Also for this attribute, higher conse- this DM accepts giving up some advantage on expected results of
quences of environmental impacts have been taken as much more service level yielded by [s2; Q⁄] in order to obtain a gain provided
undesirable than lower outcomes. by [s2; Qc] on the company’s environmental performance.
In a similar way to the elicitation of utility functions for each But improvements on profit and on environmental results do
attribute considered by the newsvendor, an elicitation procedure not always compensate, for the newsvendor, an increase on the
(Raiffa, 1970) is also used to obtain the scaling parameters kp, kSL proportion of lost sales. It depends on the balance of benefits
and kEI. The DM is invited to face a set of probabilistic choices of among the three attributes. As a comparison of [s1; Qc] and
lotteries involving three-dimensional payoffs. These payoffs are [s2; Qc] in Table 2 points out, [s1; 183,761] enables the expected
triples of Profit, Service Level and Environmental Impacts outcomes, profit to increase by 6% and to reduce the expected environmental
with compositions of the reasonably best and worst single- impacts by 2% when compared to [s2; 184,643]. However, these re-
attribute values which the newsvendor pointed out (for details sults could not offset, for the DM, the better performance of
on this procedure, see Keeney and Raiffa, 1976). As the result, [s2; 184,643] regarding service level, which yields a reduction of
the following multi-attribute scaling constants are obtained: 2% points in the proportion of lost sales (22% of reduction on the
kp = 0.55, kSL = 0.30 e kEI = 0.15. 9.1% obtained by [s1; 183,761]).
For every contingency option si, an associated lot size Qc is cal- In order to analyze the effect of changes of parameters on the
culated based on (12) in order to maximize the DM’s multi- results of this numerical experiment, a sensitivity analysis was per-
attribute utility function. The result is 7 pairs of the type [si; Qc] formed on the scaling constants kp, kSL and kEI from the DM’s multi-
as decision alternatives. At the same time, lot size values Q⁄ are ob- attribute utility function. A Monte Carlo method was used in order
tained from the traditional newsvendor method, which provides 7 to verify how random variations or elicitation errors on the scaling
pairs of solutions of the type [si; Q⁄]. In order to compare these two constants affect the reliability of the results obtained. Every scaling
sorts of results, the utility values as well as the expected values of constant was set to follow a uniform distribution, with parameters
the consequences for every option are listed in Table 2: (1  20%)k and (1 + 20%)k respectively, where k was the value
As shown in Table 2, when preferences regarding service level previously obtained for the scaling coefficient by the elicitation
and environmental impacts are taken into consideration in the procedure. After being simultaneously and randomly generated,
newsvendor decision-making process, the compromise order sizes the scaling constants were normalized by their sum such that
Qc provide more desirable balances of performance than the profit- kp + kSL + kEI = 1. For this study, 100,000 cases were simulated.
maximizing solutions for all options of emergency replenishment In a first analysis focused on disturbances on the compromise
schemes. This may be observed by higher (or at least, equal) values solution [s2; 184,643] recommended by the model, the emergency
of U[si; Qc] than U[si; Q⁄] for every si e S. replenishment option s2 stayed robustly as the contingency

Table 2
Results of the multi-attribute and traditional newsvendor models.

Alternative Lot size (Q) U[si; Q] Up[si; Q] UNS[si; Q] UEI[si; Q] E[p] E[LS] (%) E[EI]
c
[s1; Q ] 183,761 0.790 0.820 0.745 0.773 15,712,488 9.1 9984
[s1; Q⁄] 204,361 0.780 0.789 0.811 0.688 16,063,724 6.4 15,215
[s2; Qc] 184,643 0.800 0.816 0.787 0.770 14,815,806 7.1 10,176
[s2; Q⁄] 207,956 0.786 0.781 0.851 0.672 15,251,768 4.8 16,297
[s3; Qc] 198,195 0.786 0.774 0.842 0.715 13,647,948 5.0 13,478
[s3; Q⁄] 216,683 0.774 0.753 0.883 0.632 13,937,508 3.6 19,152
[s4; Qc] 212,462 0.763 0.725 0.889 0.652 12,640,157 3.3 17,731
[s4; Q⁄] 224,763 0.757 0.718 0.911 0.593 12,777,051 2.6 22,086
[s5; Qc] 233,578 0.714 0.638 0.936 0.550 11,044,380 1.8 25,615
[s5; Q⁄] 237,129 0.714 0.640 0.940 0.533 11,057,145 1.7 27,135
[s6; Qc] 250,765 0.660 0.547 0.964 0.465 9,110,904 0.97 33,476
[s6; Q⁄] 251,055 0.659 0.547 0.964 0.463 9,111,029 0.96 33,640
[s7; Qc] 257,410 0.633 0.503 0.973 0.432 7,955,956 0.70 36,896
[s7; Q⁄] 259,137 0.633 0.504 0.974 0.424 7,959,643 0.67 37,811
 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830 827

component for all cases generated. Regarding initial lot size, the differences between the original ranking and each ranking gen-
changes in the value of Qc have been represented in the histogram erated by the Monte Carlo simulation were statistically significant.
presented in Fig. 1. Basically, Kendall’s s coefficient is a rank-based correlation
The values found for Qc had a mean l = 184,751 and a standard measure between variables. It may be interpreted as the difference
deviation r = 1993. Despite the relative symmetry and unimodality between the probability that both rankings are in the same order
of data, the normality hypothesis was rejected by the Kolmogorov– versus the probability that the rankings are in different orders, gi-
Smirnov and Lilliefors tests at significance level p < 0.01. Since the ven the observed data (Sheskin, 2007). Based on the s statistic,
sample size was sufficiently large (100,000 cases), the following Kendall’s Test evaluates the null hypothesis that there is no corre-
95% confidence interval for the mean of Qc was obtained: lation between the two rankings.
[184,738; 184,763]. However, the low coefficient of variation For 16% of the rankings randomly simulated, the value obtained
(1.08%) was reflected in a narrow confidence interval, which does for the s statistic was 0.905 (high correlation). The other 84% of re-
not contain Qc = 184,643, the compromise solution found. sults yielded a s equals to 1 (perfect correlation). For both cases, the
Therefore, the sensitivity analysis indicates a statistically signif- null hypothesis was rejected at p 6 0.001. Thus, it was possible to
icant influence of the scaling constant values on the compromise conclude that there is a very strong correlation between the original
lot size obtained. Values on the right tail in Fig. 1, for example, ranking (from the elicitation procedure), and the rankings gener-
are yielded in cases where kSL is close to the upper bound of the ated via the Monte Carlo simulation. This result indicates that the
uniform distribution simulated for the Service Level scaling con- original order among emergency replenishment options remained
stant. In those cases, stronger concerns for a customer’s goodwill within the [20%; +20%] range simulated for kp, kSL and kEI, which
influence the multi-attribute utility function in order to maximize is an indication of the robustness regarding the DM’s rank of prefer-
lot size and thus minimize the expected proportion of lost sales. On ences among the emergency replenishment schemes si.
the other hand, Qc values on the left side of Fig. 1 are associated
with cases, where kEI is close to its simulated upper bound. In such
cases, the newsvendor’s preferences for lower environmental im- 4. Discussion of model and results
pacts are more influential in the multi-attribute decision-making
to reduce the lot size in order to minimize the expected number This section presents a discussion of fundamental principles re-
of units to be discarded into the environment. lated to the model built, connected to its application. Also, some
Since the lot size Qc was shown to be statistically sensitive to limitations are discussed pointing out alternatives of implementation.
variations of the scaling parameters, the analyst and the DM revis- A successful use of this multi-attribute newsvendor model re-
ited the procedure for eliciting the scaling constants, confirming lies on the assumption that the distribution of demand is known,
the values previously obtained. Additionally, a more comprehen- and that parameter a from the backlogging function may be ob-
sive and rigorous study on the effect of variations of the scaling tained. Regarding the value of a, for newsvendor-type products
parameters over the first component of the decision alternatives, with previous sale seasons, historical records of sales may be avail-
si, was undertaken. This analysis aimed to verify if the random dis- able to be statistically handled in order to provide an estimate for
turbances on the scaling constants significantly changed the initial a. For ‘‘completely new’’ newsvendor-type products such as hard-
ranking of contingency options si regarding the compromise solu- ware or IT equipment, for instance, a proper sales data base is often
tions [si; Qc] presented in Table 2. not available. However, as cited in the numerical study, the sales of
Following a decreasing order of utility values, the original rank- products from the same family or from related items may provide
ing was s2, s1, s3, s4, s5, s6, s7. For each one of the 100,000 cases sim- data to guide the estimation of a, or to support the analysis of busi-
ulated, the ranking of contingency options was registered. Thus, ness experts (Berger, 1985). These experts sometimes are highly
Kendall’s Tau non-parametric statistic test was used to assess if experienced in trading, in customer relationships, in analysis of

Fig. 1. Distribution of observed lot size values in Monte Carlo simulation.

828 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830

the market’s behavior, etc. The knowledge accumulated over many decision-making process. For this context, a ‘‘forward’’ approach
years of experience by these business professionals may be may be adequate. As Doumpos and Zopounidis (2011) point out,
exploited in cases where recorded data are lacking, thus providing this approach is based on the active participation of the DM, and
a way to overcome inertia and to tackle the uncertainty using a also on the ability of the analyst in guiding an interactive elicita-
quantitative approach. tion session in order to elicit the decision-maker’s preference infor-
Most newsvendor approaches address uncertainty by means of mation (weights, trade-offs, etc.). In this context, a multi-criteria
probability distributions (see, for example, Pasternack, 1985; aggregation method such as MAUT (Keeney and Raiffa, 1976), for
Khouja, 1995; Eeckhoudt et al., 1995; Khouja, 1996; Hill, 1997; obtaining and aggregating single-criteria preferences may be use-
Khouja, 1999; Gurnani et al., 2000; Lariviere and Porteus, 2001; ful. This paper is concerned with this latter case.
Weng and McClurg, 2003; Hillier and Lieberman, 2005; Keren Although the use of MAUT often requires adequate support
and Pliskin, 2006; Arcelus et al., 2007; Yang et al., 2007; Lodree, from a decision analyst in order to clarify concepts and to help
2007; Lee and Lodree, 2010; Qin et al., 2011; Xu et al., 2010). How- the DM’s judgments to be consistent, it provides a solid theoretical
ever, for situations where uncertainty is taken into account, but the background to translate a newsvendor’s preferences and behavior
decision-maker feels uncomfortable about using parametric distri- regarding risks involving multiple criteria. Grounded on MAUT
butions, several different approaches are proposed in the literature and Decision Analysis frameworks, this also allows the conflicting
in order to deal with uncertainty, such as fuzzy methods (Zadeh, nature of different decision criteria to be incorporated into an addi-
1971; Zimmermann, 1985; Petrovic et al., 1996; Dutta et al., tive multi-attribute utility function used to solve the newsvendor
2005) and Rough Sets Theory (Pawlak and Slowinski, 1994; Greco problem.
et al., 2001, 2006, 2008). Another approach is to use distribution- As to the use of additive utility functions, the utility values in
free procedures (Alfares and Elmorra, 2005; Mostard et al., 2005; Table 2 cannot be regarded as a simple weighted sum of one-
Yue et al.,2007), but in this last case the conclusions are less precise dimension utility values. Actually, this is a special case of prefer-
and some information about probability distributions such as ences that, when addressed under the axiomatic background of
mean and standard deviation are often required. MAUT, must obey some conditions which imply a DM’s indepen-
Even though the above approaches are often applied for several dence of preferences among p, SL and EI (for more details, see
practical situations, there are also many cases where a parametric Keeney and Raiffa, 1976). For simplification purposes, this multi-
approach is useful. It is important to highlight that the use of prob- attribute approach for the newsvendor problem has taken into
abilistic techniques may not be difficult for several newsvendor consideration the particular case of additive utility functions. How-
decision-makers, since they are used to tackling uncertainty and ever, future developments may address the use of more complex
making probabilistic choices in decision-making situations regard- utility functions, such as those with multi-linear and multiplicative
ing classical inventory management. The model proposed in our forms for Eq. (7).
work was developed for the particular case of decision-makers As exemplified in the numerical experiment, the trade-off anal-
who feel able to apply a probabilistic thinking and to use a para- yses performed by the DM among decision criteria are a distinctive
metric approach in newsvendor decisions. Otherwise, another ap- feature of this multi-attribute approach in relation to the mono-
proach such as Rough Sets Theory should be applied, which is a attribute profit-maximizing models proposed hitherto. It allows
mathematical tool to analyze inconsistent or ambiguous descrip- exploring compensations among conflicting criteria of different
tions of actions in decision-making problems (Greco et al., 2001). natures, for instance, service level versus environmental concern-
Yet another alternative to a probabilistic approach is to tackle ments. For a newsvendor who presents such non-monetary goals
the problem from a deterministic point-of-view, which is question- along with economic objectives, the results of Table 2 show that
able since uncertainty is an essential element of the Newsvendor the profit-maximizing solution is not necessarily the most pre-
Problem. ferred solution.
This model also assumes that the newsvendor’s DM feels confi- Although the utility numbers in Table 2 are apparently close,
dent to provide preference information and to use the conceptual this should not be misinterpreted as an equivalent or indifferent
framework of Multi-Attribute Utility Theory (Keeney and Raiffa, performance by the DM. A higher divergence between the perfor-
1976) and Decision Analysis (Raiffa, 1970; Berger, 1985). Indeed, mance of the decision options is reduced by the fact that, for all op-
many DMs may feel uncomfortable about providing preference tions in Table 2, utility values regarding service level are equal to 1
information for single criteria, although they may feel able to ex- for all scenarios where x < Q Eq. (10). Similarly, utility values
press global judgments on alternatives by taking into consideration regarding environmental impacts are equal to 1 for all scenarios
multiple attributes. For this case, a group of increasingly popular where x > Q Eq. (11). Additionally, it is important to point out that,
methods in the MCDA literature, deemed ‘‘disaggregation meth- since the utility values are elicited into an interval scale (Berger,
ods’’ (Jacquet-Lagrèze and Siskos, 2001; Doumpos and Zopounidis, 1985), there is no sense in assessing an ‘‘intensity of preferences’’
2011) may be used. These methods are used for preference elicita- for option [s1; Qc] in relation to [s1; Q⁄], by saying that ‘‘[s1; Qc] is
tion by developing decision models based on reference alterna- y times more preferable than [s1; Q⁄]’’, for example. Actually, this
tives, which are usually examples of decisions taken by a DM type of numerical scale allows us to compare the intensity in dif-
(Doumpos and Zopounidis, 2011). By means of regression tech- ference of preferences among pairs of decision options. For exam-
niques, global evaluations of alternatives are used to build a func- ple: the difference between the utility values of [s1; Qc] and [s1; Q⁄]
tion to represent the DḾs system of preferences. This may be may seem quite limited, as may the difference between [s5; Qc] and
observed in compensatory methods such as UTA (Jacquet-Lagrèze [s5; Q⁄]. However, we are able to state that the preference for
and Siskos, 1982), UTADIS (Zopounidis and Doumpos, 2001), and [s1; Qc] in relation to [s5; Qc] is 14% higher than the preference of
also in outranking methods such as in ELECTRE TRI applications [s1; Q⁄] in relation to [s5; Q⁄]. Therefore, the performances of
for sorting problems (Mousseau and Slowinski, 1998; Mousseau [s1; Qc] and [s1; Q⁄] are not equivalent, not are the performances
et al., 2001). Also, the concept of representative value function of [s5; Qc] and [s5; Q⁄], which highlights that analyses of utility val-
may be considered in robust multi-criteria decision problems ues must be carefully performed by taking into consideration the
(Kadzinski et al., 2012). interval scale.
However, there are also cases where the DM does not feel able In fact, the influence of risk-aversion on lot sizing has been al-
to express global judgments, but he is confident about providing ready explored in the newsvendor problem (Eeckhoudt et al.,
preference information regarding the single criteria used in the 1995; Keren and Pliskin, 2006). However, as the results in Section
 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830 829

3 point out, the risk-averse/risk-prone behavior to economic con- We also suggest the development of multi-attribute newsvendor
sequences is not the only factor to reduce/increase the order size models which encompass the entire supply chain, by aggregating
Q in relation to the traditional newsvendor model. Non-monetary different actors and stages into the decision-making process.
concerns on service level and environmental impacts may also sig-
nificantly influence a DM’s subjective assessment of Q along with
associated contingency options. Therefore, a multi-criteria ap- Acknowledgments
proach for supporting decisions in newsvendor problems should
be applied when preferences regarding issues other than profit The authors’ research is partially supported by the Brazilian Re-
must be taken into consideration. search Council (CNPq).
The sensitivity analysis enabled a study to be made of how a The authors would like to acknowledge the Editor and the
‘‘noise’’ effect on the elicitation of the scaling constants could im- anonymous reviewers for considering the potential of a previous
pact the results of applying the model. Modeled as a continuous version of this paper and for encouraging its improvement, with
variable (such as in most newsvendor models), the compromise the insightful critique and valuable suggestions.
lot size Qc was shown to be significantly sensitive to changes in
the scaling constants. This highlights the condition that, to imple-
References
ment this multi-attribute newsvendor approach successfully, pref-
erence elicitation procedures must be carefully conducted in order Abad, P.L., 2001. Optimal price and order size for a reseller under partial
to obtain results that consistently reflect the newsvendor’s prefer- backordering. Computers & Operations Research 28, 53–65.
ences and trade-off judgments. On the other hand, the sensitivity Agrell, P.J., 1995. A multicriteria framework for inventory control. International
Journal of Production Economics 41, 59–70.
analysis also indicated that the first component of the solution vec- Alfares, H.K., Elmorra, H.H., 2005. The distribution-free newsboy problem:
tor [si; Qc] is a much stronger recommendation for the DM when extension to the shortages penalty case. International Journal of Production
compared to Qc. This provides newsvendors with the confidence Economics 93–94, 465–477.
Almeida, A.T., Bohoris, G.A., 1996. Decision theory in the maintenance strategy of a
they need to contract an emergency replenishment option based standby system with gamma distribution repair time. IEEE Transactions on
on which they may carefully plan the initial order size Qc, reflect Reliability 45 (2), 216–219.
on their preferences, and make any refining adjustment before Almeida, A.T., 2005. Multicriteria modelling of repair contract based on utility and
ELECTRE I method with dependability and service quality criteria. Annals of
the selling season begins. Operations Research 138, 113–126.
Almeida, A.T., 2007. Multicriteria decision model for outsourcing contracts selection
based on utility function and ELECTRE method. Computers & Operations
Research 34, 3569–3574.
5. Conclusions
Arcelus, F.J., Kumar, S., Srinivasan, G., 2007. Pricing and rebate policies for the
newsvendor problem in the presence of a stochastic redemption rate.
A multi-attribute utility approach for the newsvendor problem International Journal of Production Economics 107, 467–482.
or single-period problem has been presented in this paper. We dis- Benton, W.C., Park, S., 1996. A classification of literature on determining the lot size
under quantity discounts. European Journal of Operational Research 92, 219–
cussed that traditional newsvendor models are mostly focused on 238.
profit maximization or cost minimization objectives, and therefore Berger, J.O., 1985. Statistical Decision Theory and Bayesian Analysis. Springer, New
they face two main limitations: first, shortage costs due to lost York.
Brito, A.J., de Almeida, A.T., Mota, C.M.M., 2010. A multicriteria model for risk
sales and due to loss of customers’ goodwill are difficult for man- sorting of natural gas pipelines based on ELECTRE TRI integrating utility theory.
agers to estimate in practical situations. Secondly, single-objective European Journal of Operational Research 200 (3), 812–821.
models do not adequately address decision-making judgments on Chang, C., Lo, T., 2009. On the inventory model with continuous and discrete lead
time, backorders and lost sales. Applied Mathematical Modelling 33, 2196–
non-monetary consequences such as corporate image, long-term 2206.
competitiveness, and the environmental impacts of newsvendor Chen, M.S., Chuang, C.C., 2000. An extended newsboy problem with shortage-level
activities. constraints. International Journal of Production Economics 67, 269–277.
Chung, C., Flynn, J., 2001. A newsboy problem with reactive production. Computers
This multi-attribute newsvendor model contributes to over- & Operations Research 28 (8), 751–765.
coming these limitations and explores the newsvendor problem Clift, R., Wright, L., 2000. Relationships between environmental impacts and added
from a wider decision-making point-of-view. The selection of an value along the supply chain. Technological Forecasting and Social Change 65,
281–295.
order size and an emergency replenishment scheme was modeled
Cruz, J.M., 2009. The impact of corporate social responsibility in supply chain
with regard to profit, service level and by considering the environ- management: multicriteria decision-making approach. Decision Support
mental impacts that may result from discarding unsold products. Systems 48, 224–236.
These attributes represented three different dimensions that deci- Doumpos, M., Zopounidis, C., 2011. Preference disaggregation and statistical
learning for multicriteria decision support: a review. European Journal of
sions on order quantities may bring in a newsvendor context, and Operational Research 209, 203–214.
they were dealt with by using concepts from Multi-Attribute Util- Dutta, P., Chakraborty, D., Roy, A.R., 2005. A single-period inventory model with
ity Theory – MAUT (Keeney and Raiffa, 1976). fuzzy random variable demand. Mathematical and Computer Modelling 41,
915–922.
In this model, a partial backlogging condition of demand was ta- Eeckhoudt, L., Gollier, C., Schlesinger, H., 1995. The risk averse (and prudent)
ken into consideration, and a negative exponential form for the newsboy. Management Science 41 (5), 786–794.
backlogged demand was assumed. Results obtained by a numerical Ferreira, R.J.P., de Almeida, A.T., Cavalcante, C.A.V., 2009. A multi-criteria decision
model to determine inspection intervals of condition monitoring based on delay
experiment indicate that the compromise solutions afforded more time analysis. Reliability Engineering & System Safety 94, 905–912.
preferable performances (higher utility) than profit-maximizing Greco, S., Inuiguchi, M., Slowinski, R., 2006. Fuzzy rough sets and multiple-premise
solutions. From a multi-attribute point-of-view for the newsven- gradual decision rules. International Journal of Approximate Reasoning 41,
179–211.
dor problem, these compromise solutions represent a better bal- Greco, S., Matarazzo, B., Slowinski, R., 2001. Rough sets theory for multicriteria
ance of the decision maker’s concerns about profit, service level decision analysis. European Journal of Operational Research 129, 1–47.
and environmental issues. Greco, S., Matarazzo, B., Slowinski, R., 2008. Parameterized rough set model using
rough membership and Bayesian confirmation measures. International Journal
For future studies on this theme, we suggest tackling multi-
of Approximate Reasoning 49, 285–300.
attribute newsvendor models incorporating other non-monetary Gurnani, H., Akella, R., Lehoczky, J., 2000. Supply management in assembly systems
attributes, different assumptions on backlogging rate, as well as with random yield and random demand. IIE Transactions 32 (8), 701–714.
other forms for the multi-attribute utility function. Order size Hau, A., 2010. Comparative statics of changes in risk on monotonically and partially
responsive payoffs. European Journal of Operational Research 201, 267–276.
and time may be considered as discrete, and different premises Hill, R.M., 1997. Applying Bayesian methodology with a uniform prior to the single
for the emergency replenishment lead-time could be explored. period inventory model. European Journal of Operational Research 98, 555–562.
830 A.J. Brito, A.T. de Almeida / European Journal of Operational Research 220 (2012) 820–830

Hillier, F.S., Lieberman, G.J., 2005. Introduction to Operations Research, eighth ed. Pentico, D.W., Drake, M.J., Toews, C., 2009. The deterministic EPQ with partial
McGraw Hill, New York. backordering: a new approach. Omega 37, 624–636.
Hu, W., Kim, S., Barnejee, A., 2009. An inventory model with partial backordering Petrovic, D., Petrovic, R., Vujosevic, M., 1996. Fuzzy models for the newsboy
and unit backorder cost linearly increasing with the waiting time. European problem. International Journal of Production Economics 45, 435–441.
Journal of Operational Research 197, 581–587. Qin, Y., Wang, R., Vakharia, A.J., Chen, Y., Seref, M.M., 2011. The newsvendor
Jaber, M.Y., Goyal, S.K., 2008. Coordinating a three-level supply chain with multiple problem: review and directions for future research. European Journal of
suppliers, a vendor and multiple buyers. International Journal of Production Operational Research 213, 361–374.
Economics 116, 95–103. Raiffa, H., 1970. Decision Analysis. Addison-Wesley, Reading, MA.
Jacquet-Lagrèze, E., Siskos, Y., 1982. Assessing a set of additive utility functions for Rao, P., Holt, D., 2005. Do green supply chains lead to competitiveness and economic
multicriteria decision making: the UTA method. European Journal of performance? International Journal of Operations and Production Management
Operational Research 10, 151–164. 25 (9), 898–916.
Jacquet-Lagrèze, E., Siskos, Y., 2001. Preference disaggregation: 20 years of MCDA Ryu, K., Yücesan, E., 2010. A fuzzy newsvendor approach to supply chain
experience. European Journal of Operational Research 130 (2), 233–245. coordination. European Journal of Operational Research 200, 421–438.
Kadzinski, M., Greco, S., Słowinski, R., 2012. Selection of a representative value San José, L.A., Sicilia, J., García-Laguna, J., 2006. Analysis of an inventory system with
function in robust multiple criteria ranking and choice. European Journal of exponential partial backordering. International Journal of Production
Operational Research 217, 541–553. Economics 100, 76–86.
Keeney, R.L., Raiffa, H., 1976. Decision with Multiple Objectives: Preferences and Sheskin, D.J., 2007. Handbook of Parametric and Nonparametric Statistical
Value Trade-Offs. John Wiley & Sons, New York. Procedures. Chapman & Hall/CRC, Boca Raton, USA.
Keren, B., Pliskin, J.S., 2006. A benchmark solution for the risk-averse newsvendor Su, R.H., Pearn, W.L., 2011. Product selection for newsboy-type products with
problem. European Journal of Operational Research 174, 1643–1650. normal demands and unequal costs. International Journal of Production
Khouja, M., 1995. The newsboy problem under progressive multiple discounts. Economics 132, 214–222.
European Journal of Operational Research 84, 458–466. Sucky, E., 2006. A bargaining model with asymmetric information for single
Khouja, M., 1996. A note on the newsboy problem with an emergency supply supplier – single buyer problem. European Journal of Operational Research 171,
option. Journal of Operational Research Society 47, 1530–1534. 516–535.
Khouja, M., 1999. The single-period (newsvendor) problem: literature review and Vijayan, T., Kumaran, M., 2008. Inventory models with a mixture of backorders and
suggestions for future research. Omega 27 (2), 537–553. lost sales under fuzzy cost. European Journal of Operational Research 189, 105–
Kumar, G.S., 1988. Newsboy problem revisited. Production Engineer 67, 22–23. 119.
Lariviere, M., Porteus, E., 2001. Selling to the newsvendor: an analysis of price-only Vincke, P., 1992. Multicriteria Decision-Aid. Wiley, Bruxelles.
contracts. Manufacturing and Service Operations Management 3 (4), 293–305. Wang, C.X., Webster, S., Suresh, N.C., 2009. Would a risk-averse newsvendor order
Lee, H., Lodree, E.J., 2010. Modeling customer impatience in a newsboy problem less at a higher selling price? European Journal of Operational Research 196,
with time-sensitive shortages. European Journal of Operational Research 205, 544–553.
595–603. Weng, Z.K., McClurg, T., 2003. Coordinated ordering decisions for short life cycle
Leung, K.F., 2009. A generalization of sensitivity of the inventor model with partial products with uncertainty in delivery time and demand. European Journal of
backorders. European Journal of Operational Research 196, 554–562. Operational Research 151, 12–24.
Lodree, E.J., 2007. Advanced supply chain planning with mixtures of backorders, lost Weng, Z.K., 2004. Coordinating order quantities between the manufacturer and the
sales and lost contract. European Journal of Operational Research 181, 168–183. buyer: a generalized newsvendor model. European Journal of Operational
Lodree, E.J., Kim, Y., Jang, W., 2008. Time and quantity dependent waiting costs in a Research 156, 148–161.
newsvendor problem with backlogged shortages. Mathematical and Computer Whitin, T.M., 1955. Inventory control and price theory. Management Science 2, 61–
Modelling 47, 60–71. 68.
Montgomery, D.C., Bazaraa, M.S., Keswani, A.K., 1973. Inventory models with a Xu, M., Chen, Y., Xu, X., 2010. The effect of demand uncertainty in a price-setting
mixture of backorders and lost sales. Naval Research Logistics Quarterly 20, newsvendor model. European Journal of Operational Research 207, 946–957.
255–263. Yang, S., Yang, J., Abdel-Malek, L., 2007. Sourcing with random yields and stochastic
Moon, I., Choi, S., 1994. The distribution free continuous review inventory system demand: a newsvendor approach. Computers & Operations Research 34, 3682–
with a service level constraint. Computers & Industrial Engineering 27 (1–4), 3690.
209–212. Yue, J., Wang, M., Chen, B., 2007. Mean-range based distribution-free procedures to
Mostard, J., Koster, R., Teunter, R., 2005. The distribution free newsboy problem minimize ‘‘overage’’ and ‘‘underage’’ costs. European Journal of Operational
with resalable returns. International Journal of Production Economics 97, 329– Research 176, 1103–1116.
342. Zadeh, L.A., 1971. Quantitative fuzzy-semantics. Information Science 3 (2), 159–
Mostard, J., Teunter, R., 2006. The newsboy problem with resalable returns: a single 176.
period model and case study. European Journal of Operational Research 169 (1), Zimmermann, H.J., 1985. Fuzzy Set Theory and Its Applications. Kluwer Nijhoff
81–96. Publishing, Boston.
Mousseau, V., Slowinski, R., 1998. Inferring an ELECTRE TRI model from assignment Zhou, Y., Lau, H., Yang, S., 2004. A finite horizon lot-sizing problem with time-
examples. Journal of Global Optimization 12, 157–174. varying deterministic demand and waiting-time-dependent partial
Mousseau, V., Figueira, J., Naux, J., 2001. Using assignment examples to infer backlogging. International Journal of Production Economics 91, 109–119.
weights for ELECTRE TRI method: some experimental results. European Journal Zhou, Y., Wang, S., 2009. Manufacturer–buyer coordination for newsvendor-type-
of Operational Research 130, 263–275. products with two ordering opportunities and partial backorders. European
Pasternack, B.A., 1985. Optimal pricing and return policies for perishable Journal of Operational Research 198, 958–974.
commodities. Marketing Science 4 (2), 166–176. Zopounidis, C., Doumpos, M., 2001. Multicriteria preference disaggregation for
Pawlak, Z., Slowinski, R., 1994. Decision analysis using rough sets. International classification problems with an application to global investing risk. Decision
Transactions in Operational Research 1 (1), 107–114. Sciences 32 (2), 333–385.