Вы находитесь на странице: 1из 9

Expert Systems with Applications 38 (2011) 55445552

Contents lists available at ScienceDirect

Expert Systems with Applications


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

Identifying the underlying structure of demand during promotions: A structural


equation modelling approach
Usha Ramanathan a,, Luc Muyldermans b
a
Newcastle Business School, City Campus East, Northumbria University, Newcastle Upon Tyne NE1 8ST, UK
b
Nottingham University Business School, Jubilee Campus, Wollaton Road, Nottingham NG8 1BB, UK

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

Keywords: It is widely recognized that promotions enhance product sales. Other factors such as seasonality, special
Demand-factor model days, customer preferences and economic factors may also have an impact on demand. Yet, endeavours to
Structural equation modelling model the dependency of demand on various factors are scarce. We develop a conceptual demand-factor
Information exchange model to represent the dependency of demand on these factors. We examine and validate the conceptual
Supply chain collaboration
model through structural equation modelling, using sales data of a leading soft drink company involved
in supply chain collaboration with downstream retailers. We nd that promotional and seasonal factors
have a large and direct impact on sales, while special days like festivals and bank holidays have limited
inuence. Our approach is versatile and can assist demand planners in understanding the demand during
promotional events, and in the planning and execution of promotional deals.
2010 Elsevier Ltd. All rights reserved.

1. Introduction deals are in place. For example, the demand for soft drinks can be
much higher during promotions than at normal (non-promotional)
Supply chain collaboration and information exchange are key periods. In such cases, supply chain collaboration (and CPFR in par-
elements of success in many businesses such as Wal-Mart, Proctor ticular) may help to avoid excess inventory and stock-outs. In
& Gamble and West-Marines (Ireland & Crum, 2005). Collaborative CPFR, information from different supply chain parties (such as pro-
planning, forecasting and replenishment (CPFR) is one framework motional calendars, inventory levels and point of sales data) is
that integrates suppliers and buyers. It encourages collaborating shared so as to improve forecasting and to facilitate timely replen-
partners to exchange information for increased transparency ishment. Sharing of demand and inventory information with up-
throughout the supply chain (VICS, 2004). But, the level of informa- stream members helps in reducing supply chain costs (Cachon &
tion exchange required is highly dependent on the businesses Fisher, 2000; Chen, 1998; Gavirneni, Kapuscinski, & Tayur, 1999;
involved in collaboration. Some businesses require limited infor- Lee, So, & Tang, 2000; Raghunathan, 2001). This is also evident
mation such as inventory levels and historical sales data for plan- from Wal-Marts Retail-link, an exclusive network to share real
ning. Others may need additional details such as promotional time sales data with other partners in the collaboration (Ireland
calendars and pricing information at retailer outlets. The depen- & Crum, 2005).
dency on such information makes supply chain collaboration Although supply chain collaboration and advanced information
(SCC) more essential, and the level of SCC is directly reected in technology have given a certain amount of leverage to coordinate
the extent of joint planning activities required. and improve production and replenishment, demand forecasting
The structure of demand is one of the important characteristics is still a big issue in the area of supply chain management. Collab-
to decide on the level of collaboration and information exchange. orative forecasting, by combining relevant information from differ-
Therefore, it is essential to identify the reasons behind demand ent supply chain partners, is one of the potential avenues to
for a particular product at a particular time. Normally, functional improve forecast accuracy (Aviv, 2001). Various external and inter-
products (e.g., detergents, shampoo, soft drinks) will have low de- nal factors make forecasting of promotional sales a challenging
mand and supply uncertainties compared to innovative products task. Hence identifying the underlying structure of demand is
(e.g., computers, video games, mobile phones) (Fisher, 1997; Lee, important.
2002). But this is not always the case, especially when promotional In this study, we use the concept of supply chain collaboration in
the context of information exchange and promotional planning and
Corresponding author. Tel.: +44 7726732156. forecasting. We begin by developing a conceptual demand-factor
E-mail addresses: Usha.Ramanathan@northumbria.ac.uk (U. Ramanathan), Luc.- model for the underlying structure of demand. Then, we empirically
Muyldermans@nottingham.ac.uk (L. Muyldermans). examine whether the developed constructs really represent prod-

0957-4174/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.10.082
U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552 5545

uct sales. In our study, we use sales data of different soft drinks over relational competency in SCC. Although many of these articles
a time period of 36 months. We also use some publicly available are from the discipline of operations management, they do not ex-
information in our model. Our study makes an important contribu- plain the relationship between demand and its inuencing factors
tion to the literature through the development of an empirical de- in a great extent. This was also evident from the recent review pa-
mand model and identifying its underlying factors. This attempt per on supply chain coordination (Arshinder and Deshmukh,
is the rst of its kind in the literature. Because our approach reveals 2008).
the signicant variables and their inter-relationships, an interesting In the past two decades some analytical models assessed cus-
extension is to use these factors in causal forecasting methods. This tomer responses to various types of promotions, including price
in turn, may assist demand planners in producing more accurate discounts with advertisement, on-shelf discounts without adver-
forecasts for promotional events. tisement and coupon discounts (Cooper, Baron, Levy, Swisher, &
The remainder of the paper is organised as follows. Section 2 re- Gogos, 1999; Raju, 1995). A recent review article by Fildes, Niko-
views the literature on supply chain collaboration and promotional lopoulos, Crone, and Syntetos (2008) cited some interesting articles
forecasting. Section 3 describes our case study company a major from the Marketing literature, which relate demand and its under-
soft drink manufacturing company located in the UK. Its current lying causal factors in the context of forecasting. In an attempt to
supply chain collaboration practices for promotional events help forecast the effect of promotions, Cooper et al. (1999) identied
us to develop a conceptual model and also to propose four research that the display location of items in the stores had a positive im-
hypotheses. Section 4 describes the research approach, including pact on sales, whereas a longer duration of the promotional deal
data description, measurement model and structural equation (with major display) had mixed effects on sales, especially for slow
models. Section 5 discusses the results of our analysis, and con- moving items. Several attempts have been made to estimate the ef-
cludes with future research ideas. fect of promotions and price discounts on sales using store level
data (Abraham & Lodish, 1987; Blattberg & Levin, 1987; Naik,
Raman, & Winer, 2005) and market level data (Ali, Sayn, Van
2. Supply chain collaboration for improved forecasting Woensel, & Fransoo, 2009; Christen, Gupta, Porter, Staelin, & Witt-
ink, 1997; Dube, 2004). Dubes econometric model for consumers
Supply chain collaboration has a rich and wealthy literature purchase pattern related the time of buying and the time of
dealing with many aspects of different supply chain processes. consumption. Divakar, Ratchford, and Shankar (2005) developed
Information exchange in the supply chain focuses mainly on two CHAN4CAST, a sales forecasting decision support model incorpo-
areas. One is production and replenishment, and the other is fore- rating several variables (trend, promotional variables, seasonality,
casting and planning. Either cost reduction or inventory control are holidays and temperature). The authors considered channel level
the primary motive of these supply chain collaborations (Cachon & data, combining a variety of products. Recently demand forecasting
Fisher, 2000; Chen, 1998; Kulp, Lee, & Ofek, 2004). As our research in the presence of promotions was discussed by Ali et al. (2009).
aims to establish the relationship between sales and the underly- They supported detailed input data with advanced techniques to
ing factors, our literature study focuses on articles dealing with improve forecast accuracy in comparison with standard linear
supply chain collaboration and information exchange for the pur- regression models. Almost all of the above mentioned literature
pose of demand planning. concentrated more on forecasting rather than identifying the
Many researchers consider information exchange in the supply underlying demand variables and their relationship to sales.
chain as vital for improving performance (Byrne & Heavey, 2006; We build our research based on the previous work of Divakar
Cachon & Fisher, 2000; Lee et al., 2000). In this line, a special issue et al. (2005) by trying to understand how different demand factors
of the Management Science journal on Marketing and Operations inuence sales using structural equation modelling (SEM). SEM
Management Interfaces and Coordination debated various issues (see e.g., Hair, Black, Babin, Anderson, & Tatham, 2006) is a multi-
common to both the elds of operations management and market- variate statistical analysis technique, which can be used to estab-
ing, and having sales information as the main focus. While some lish and test the causal relationships between various latent
authors illustrated that point of sales (POS) data is important infor- constructs. Latent constructs are variables that cannot be mea-
mation across the supply chain, others doubted its role for demand sured directly, but which are estimated through other measurable
forecasting. For example, while Kulp et al. (2004) integrated differ- indicator variables. For example, one of the latent constructs in our
ent forms of information and knowledge to evaluate supply chain structural equation model is called promotional factors and this
performance, Steckel, Gupta, and Banerji (2004) questioned the construct is measured through variables such as price discount
importance of POS information. They argued that POS information and the duration of the promotional event. SEM is a relatively
may distract decision makers particularly if product demand was new approach in the eld of operations management and forecast-
highly uctuating. However, executing promotions at retail outlets ing, but some researchers have used this technique to nd relation-
usually requires intense collaboration between the manufacturer ships between different groups (suppliers and buyers) or
and the retailers (Ailawadi, Beauchamp, Donthu, Gauri, & Shankar, departments (marketing and operations) in analysing a rms per-
2009; Aviv, 2007) and different data may be needed in different formance (Mishra & Shah, 2009). Our proposed SEM was developed
phases of the planning/execution process: whereas the demand using weekly sales data of different products (2 l pet bottles dif-
during promotional events can be predicted based on historical ferent avours) of a leading soft drink company in the UK. The data
sales data and information about the promotional deals run by were not at the store level, but involved the aggregated sales vol-
the stores, the replenishment phase typically requires inventory umes among all stores of major retailers in the UK.
status and recent POS information in order to avoid stock-outs or
excess inventory.
Analytical models developed by Aviv (2001, 2007) showed sup- 3. Case study soft drink manufacturing company
port for sharing sales information and forecasts between retailers
and manufactures to improve forecast accuracy. It was also dem- 3.1. Case study research
onstrated that the overall supply chain performance was higher
with high quality centralized information (Forslund & Jonsson, We have chosen a case study approach to identify different fac-
2007). A recent article of Paulraj, Lado, and Chen (2008) empha- tors of demand. We conducted our study with a leading soft drink
sised the importance of inter-organizational communication for manufacturing company (SDM) in the UK. SDM practices CPFR
5546 U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552

with most of its retail customers for promotional events. Every From discussion with SDM managers, it was also clear that the
year SDM produces and distributes about 240 million cases of soft collaboration with the retailers was much more intense during
drinks. The product portfolio of the company includes pet bottles promotional events. Under normal circumstances (no promotions)
of 2 l, 1.5 l, 500 ml and 330 ml cans. For our study we contacted the sales of soft drinks appeared to be rather stable (apart from
the Customer Logistics Ofce, which looks after the largest 400 for some products seasonal uctuations during summer). Auto-
customers, including main grocery retailers, wholesale customers, matic replenishment systems can handle these situations very
national retail outlets and main brewers. well. But, promotions may heavily distort the normal demand
The case-study approach included company visits and semi- prole. For example, bogof and half price promotions could easily
structured interviews with managers and demand planners uplift the sales of 6-packs of 330 ml cans of a popular soft drink
responsible for sales, forecasting and promotional plans. We left by a factor 30 or more. SDMs customer demand analysts forecast
the questions for the interviews as provisional and updated on the sales for each promotional event using the promotional calen-
completion of each stage of the study as our research was explor- dar and the historical performance of similar deals. This was (and
atory in nature (Voss, Tsikriktsis, & Frohlich, 2002). In fact, this ap- still is) a non-automatic, time-consuming process and no specic
proach was very versatile to capture the interesting points on forecasting models (e.g. causal models) were used. During a pro-
promotional planning and forecasting, and information exchange motional event, the actual sales (POS) and inventory levels are
between SDM and the retailers. Our goal during the interviews monitored closely to make sure that the right amount of products
was to understand SDMs promotional planning and forecasting are deployed in the distribution network in order to prevent stock-
approach as well as to identify various factors that may impact outs.
on the sales during promotional events. SDMs promotional plan-
ning and forecasting processes are briey described below. 3.3. Conceptual demand-factor model

3.2. Promotional planning and forecasting at SDM Apart from the promotional elements discussed in Section 3.2,
there are other factors such as weather and temperature, Christ-
The sales of soft drinks (and other functional products) in the mas and holiday periods, which may impact on sales. These factors
UK are heavily dependent on the promotional deals run by the are external to the supply chain and hence difcult or impossible
retailers. Approximately 60% of SDMs sales volume is achieved to control by the supply chain members. The external factors are
during promotional events. In the past ve years, SDM had en- represented by three constructs. The rst, customer preferences
gaged in collaborative planning, forecasting and replenishment (CP) includes preferences of customers for certain products (fast
(CPFR) with the largest retailers. SDM discusses promotional plans or slow moving products) and life cycle aspects (i.e., whether the
with their customers and develops a promotional calendar for each product is an established product or a new product). The second,
retailer. The promotional calendar shows the products (or product special days (SD) includes holiday periods, bank holidays and ma-
families) on promotion, the timing and duration of the promotions, jor festivals (Christmas and Easter); and the third, seasonal factors
the types of promotions, the price discounts, the promotional dis- (SF) includes temperature and time in the year. These four factors
plays and possibly other promotional features. may act independently or inter-dependently in realizing the actual
SDM normally offers the same type of promotion at the same sales. This idea of relating promotional factors and non-promo-
time to all different avours within the same product family. For tional factors with the actual sales performance is represented con-
example, if a 2-for-2.50 promotion is planned for 2 l cherry ceptually in Fig. 1.
drinks, then this promotion is also offered for 2 l strawberry drinks. The four factors may or may not be required to explain the de-
The timing of a promotional deal is specied by xing the starting mand of all soft drinks. The objective of this research is to uncover
week and the number of weeks the deal runs in the stores. Typi- the relationships between the different factors and the overall
cally, the company offers promotions for 25 times a year, while sales performance. It is also important to mention that there may
each deal lasts for 15 weeks. The types of promotion differ widely or may not be a relationship between these four constructs. At this
across various retailers, and include: half-price, buy-one-get-one- stage, we hypothesize the following four relationships:
free (bogof), buy-2-get-2-free, 2-for-x, 3-for-x, and save x. Each
of these deals can be converted into a percentage discount on the Hypothesis 1. Promotions have a direct positive impact on sales.
normal (non-promotional) sales price. Promotions such as half- Hypothesis 2. Customers preferences for product are directly
price and bogof have the highest price discount (50%) and result related to sales.
in steep increases in the sales volume. The display of the products
on promotion can also have an impact on the sales. Products can be Hypothesis 3. Special days have a positive impact on sales.
displayed at gondola end, end aisle, side aisle, plinth, stack, etc. The Hypothesis 4. Seasonal factors have an impact on the sales of soft
gondola end display is at the end of a major aisle in the store and drinks.
with a prominent sign board. This display captures maximum These four hypotheses are very much in line with the general
attention from potential buyers. Other promotional features in- belief and common sense. There is strong support in the literature
clude in-store advertisements (sign boards, sports features with a for some of the hypotheses, and in particular for Hypotheses 1 and
product logo, etc.) and media advertisements outside the stores 4. Many causal (regression) models do incorporate explanatory
(on the radio, television and in magazines). All the above promo- variables to model different promotional aspects (type, price dis-
tional elements may result in an increase of sales over the normal count, duration and type of display) and usually, stronger promo-
base sales. The increase in sales can be captured using a measure tional mechanisms result in a larger sales uplift (see e.g., Cooper
called sales uplift, dened as the ratio of the average sales volume et al., 1999). Regarding Hypothesis 4, there is evidence in the liter-
during a promotional event and the average normal sales level ature (Ailawadi et al., 2009; Sun, 2005) that even similar products
when no promotion is offered. Information on promotions can be are valued differently at different points in time. In addition, there
termed as internal to SDMs supply chain or under control of is strong support that the sales of soft drinks are higher when the
the supply chain partnership. All these promotional elements will weather is hot (see e.g., Cooper et al., 1999; Dube, 2004). Hypothe-
be grouped in a single latent variable called promotional factors sis 2 is based on the fact that some products have a higher demand
(PF) in our conceptual demand-factor model and structural equa- than others (fast vs. slow moving items) and that customers may
tion model (see Section 4). respond differently to newly introduced products compared to
U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552 5547

Fig. 1. Conceptual demand-factor model.

well-established avours. Finally, Hypothesis 3 tests whether spe- the year (SF2). The average (historical) temperature was collected
cial days such as Christmas, Easter or other holidays have an im- from the MET ofce and is measured on a continuous scale. The
pact on sales. week in the year is represented by an integer number between
1 (rst week) and 52 (last week).
4. Model construction We have initiated the data analysis using the sales of different
soft drinks (2 l bottles) sold through one of the retail chains. The
4.1. Data description and analysis data matrix was checked for coding errors. In case of any mismatch
of sales gures and promotional events (as specied on the promo-
SDM observes electronic point of sales data (EPOS) from its cus- tional calendar), it was corrected through discussion with the key
tomers. We consider three years of sales data (20052007) of 10 informant at SDM (see Voss et al., 2002). The data used in our anal-
different soft drinks (different avours) in 2 l pet bottles, sold ysis can be considered as error free and complete (weekly sales
through three major retail chains in the UK. The data are used data for the years 20052007) with ten variables. Table 1 reports
for studying the impact of demand factors on the actual sales vol- the descriptive analysis results namely, mean, standard deviation
ume but not for studying the changing interests of buyers. Sales (SD) and correlation matrix of the data.
volume is used as the endogenous variable to measure sales perfor- The four promotional variables are signicantly correlated to
mance in our structural equation model. All the demand factors as each other at p < 0.01 level. The correlation values vary between
specied below (promotional factors, special days, seasonal factors 0.26 and 0.80. This shows that each of the variables under the con-
and customer preference) are exogenous variables. struct promotional factors is representing a different characteristic
We measure the latent construct promotional factors (PF) of the promotional event. The non-promotional construct cus-
through four indicator variables: type of promotion (PF1), size tomer preference has two correlated variables: product rank and
and/or display of promotion (PF2), duration of the promotional new/established products. It is important to note that these two
event (PF3) and price discount (PF4). The different types of promo- variables are not correlated to any other variables of the other
tions (PF1) are scored against a scale of 1-to-5, with the strongest three constructs. The next non-promotional construct special
mechanism having the highest score: save x% (score 1), 3-for-x days has two variables: festivals and other holidays. These two
(score 2), buy 2-for-3 (score 3), 2-for-x (score 4), and bogof variables are signicantly correlated at p < 0.01 level with correla-
(score 5). In a similar way, the size and/or display of a promotion tion value 0.39. The festivals and the holidays are signicantly cor-
(PF2) is scored against a 1-to-5 scale, with the highest score cor- related with the duration of the promotional event (PF3), with a
responding to the most prominent display location. The size of a value of 0.13 and 0.10, respectively at p < 0.01. The last non-pro-
promotional event relates to the number and/or type of retail motional construct seasonal factors has two variables tempera-
stores that run the promotional deal. For example, the display in ture and week-in-year. Both these variables are correlated with
smaller stores can be on the shelf (score 1), end aisle (score 2) or value 0.18. Temperature has a high signicant correlation with size
gondola end (score 3), whereas the display on shelf and gondola of promotion and price discount. But, temperature has a negative
end at mega-stores is scored as 4 and 5 respectively. The duration correlation with festivals and holidays. The negative correlation
of a promotional event (PF3) ranges from one to ve weeks and is shows that the main festivals in the UK are happening during sea-
also measured on a 1-to-5 scale. The price discount (PF4) is mea- sons with lower temperature.
sured on a continuous 0-1 scale (with 0 corresponding to no price
discount and 1 corresponding to a 100% price discount). 4.2. Structural equation modelling
We measure the second latent construct customer preferences
through two variables: product rank (CP1) and new or established We use structural equation modelling (SEM) for exploring the
product (CP2). Product rank is measured on a 1-to-5 scale: the dif- relationships between the different factors and their impact on
ferent avours are ranked according to their annual sales volumes, sales performance (see Fig. 1). SEM follows a two steps approach
and the top selling avours receive a score of 5, whereas the slow- (Anderson & Gerbing, 1988) constructing the measurement model
est moving avours receive a score of 1. CP2 is measured on a and constructing and testing the structural model. Measurement
dichotomous 0-1 scale, with 1 for established products and 0 and structural models are usually represented by diagrams (see
for new products. Figs. 24). The measurement model deals with the relationships
We measure the third latent construct special days (SD) between measured input variables and the latent constructs (PF,
through two variables: major festivals such as Easter and Christ- SD, SF and CP). Principal component analysis was conducted to
mas (SD1) and other special days including bank holidays, back identify and conrm the different factors under each construct in
to school, and mothers day (SD2). Both these variables are mea- our conceptual model. We used SPSS 15 for exploratory factor
sured on a dichotomous 0-1 scale. analysis (EFA) and for conrmatory factor analysis (CFA). In stage
We measure the fourth latent construct seasonal factors (SF) two, we test the relationships between these constructs and sales
through two variables: average temperature (SF1) and week in performance, which is measured through sales volume. We used
5548 U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552

Table 1
Descriptive statistics of rst order observed variables.

Variables Mean SD PF1 PF2 PF3 PF4 CP1 CP2 SD1 SD2 SF1 SF2
PF1 promotion type 1.47 1.68 1
PF2 size of promotion 0.33 0.75 0.54* 1
PF3 duration of promotion 1.28 1.76 0.72* 0.26* 1
PF4 price discount 0.14 0.18 0.80* 0.74* 0.59* 1
CP1 product rank 3.73 1.67 0.03 0.01 0.04 0.03 1
CP2 New/established 0.57 0.5 0.01 0.03 0.03 0.01 0.87* 1
SD1 festivals 0.02 0.14 0.06 0.06 0.13* 0.02 0.07 0.01 1
SD2 other holidays 0.12 0.33 0.04 0.07 0.10* 0.02 0.03 0.01 0.39* 1
SF1 temperature 13.27 4.92 0.02 0.30* 0.10* 0.22* 0.01 0.02 0.17* 0.09** 1
SF2 week-in-year 26.81 15.11 0.05 0.01 0.08** 0.06 0.02 0.02 0.23* 0.07** 0.18* 1

n = 826 observations.
*
Correlation is signicant at the 0.01 level
**
Correlation is signicant at the 0.05 level.

(a) Conceptual measurement model (b) Inital model

PF1 PF1
0.915*

0.748*
PF2 PF2
Promotional Promotional
factors - PF 0.761* factors - PF
PF3 PF3
0.936*

PF4 PF4

CP1 Customer 0.07***


preferences -
CP
CP2

SD1 SD1 0.833*


-0.21***
Special days - Special days -
SD 0.841* SD
SD2 SD2
-0.24**

SF1 SF1 0.769*


Seasonal Seasonal
factors - SF 0.780* factors - SF
SF2 SF2

*** - significant at p < 0.10


** - significant at p < 0.05
* - significant at p < 0.01

Fig. 2. (a) Conceptual measurement model and (b) initial model.

the software Amos 16 for developing the structural equation mod- Conrmatory factor analysis (CFA) was conducted to conrm
els and to assess the overall t using the maximum likelihood the factors under each latent variable and an initial model was
method. Three different data sets corresponding to different retail developed (Fig. 2b). The CFA results are shown in Table 3.
chains were used. Three latent constructs (promotional factors, special days and
seasonal factors) are highly signicant at p < 0.01 level, whereas
4.2.1. Stage 1: Measurement model construction and validation the construct customer preferences fails the Bartlett test of sphe-
We have formed four latent constructs from 10 observed vari- ricity (not signicant) although KaiserMeyerOlkins (KMO) mea-
ables: promotional factors (PF), customer preferences (CP), special sure of sampling adequacy has a value of 0.5 (Hair et al., 2006).
days (SD), and seasonal factors (SF). The conceptual measurement Hence, we considered dropping this factor from our analysis. (Cus-
model is shown in Fig. 2a. tomer preference (CP1) has a signicant high correlation with sales
Principal component analysis was conducted to identify the volume, and we will re-introduce this variable in our later struc-
number of groups formed in these variables. To validate the tural models (see Figs. 3 and 4)).
groups, the varimax rotation method with Kaiser normalization CFA analysis conrmed the signicance of the entire path be-
was applied. An eigenvalue of one or more identies the factors. tween observed variables and rst order latent variables with
Any element with a factor loading smaller than 0.4 was not consid- p < 0.01. The construct validity of our model is explained through
ered for further analysis as it will not measure a specic construct the percentage of variance extracted (Fornell and Larcker, 1981).
(Hair et al., 2006). The results of the explanatory factor analysis The total variance explained by each construct is in the range of
(EFA) are shown in Table 2. The total variance explained in EFA is 5271% (see Table 3). The reliability of the constructs was veried
68.28%. This represents the amount of stability of the constructs. through the Cronbach alpha a = 0.63. Here, a Cronbach alpha
U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552 5549

-0.09**

0.004

0.43*
-0.23*

Fig. 3. Test structural equation model.

Fig. 4. Complete structural equation model.

value P 0.60 is valid in our exploratory case. Hence, we claim that relationships between the three latent constructs, is shown in
both the construct validity and the construct reliability are satis- Fig. 2b.
factory. We have further tested our constructs for discriminant
validity to check how each construct is distinct from the others 4.2.2. Stage 2: Construction and testing of structural equation models
(Anderson & Gerbing, 1988). Since none of the inter-correlations Structural theory is a conceptual representation of relation-
among the three constructs contains 1.0, we conclude that each ships between constructs. In Section 4.2.1 we constructed a
construct is distinct. The initial model, including the covariance measurement model with four latent constructs to identify the
5550 U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552

Table 2 chi-square (v2/df), comparative t index (CFI), goodness of t in-


Exploratory factor analysis. dex (GFI), incremental t index (IFI), Trucker Lewis index (TLI)
Constructs Variables Factor loadings and root mean square error of approximation (RMSEA) at 90%
Promotional factors PF1 0.913 condence interval (CI). Table 4 gives a brief description of these
PF2 0.727 t indices (Hair et al., 2006).
PF3 0.776 We rst evaluated the measurement model (Fig. 2), which con-
PF4 0.932 siders rst order latent constructs and observed variables. In this
Customer preferences CP1 0.737 model (Fig. 2b) three rst order constructs correlate with each
CP2 0.699 other. The models t indices are acceptable which is reected in
Special days SD1 0.835 the signicant factor loadings. The models t has GFI = 0.988,
SD2 0.724 CFI = 0.989, v2 = 45.42 with df = 18, normed v2 = 2.52, IFI = 0.990,
Seasonal factors SF1 0.792 TLI = 0.979 and RMSEA = 0.043 at 90% CI. This model performed ex-
SF2 0.629 tremely well in terms of the suggested level of t in empirical re-
search (Shah & Goldstein, 2006).
Then, we tested the measurement model linked with sales
Table 3 performance the latter being measured through sales volume.
Conrmatory factor analysis (factor loading). In this model, all rst order constructs except customer prefer-
Constructs Variables Factor loadings Total variance
ence were found signicant. In customer preferences, CP1 (prod-
explained (%) uct rank) was signicant but not CP2 (new or established
product). The model was then re-tested after dropping the ob-
Promotional PF1 0.915 71.32
factors PF2 0.748 served variable CP2. The resulting model t (Fig. 3) has normed
PF3 0.761 v2 = 3.379, GFI = 0.981, CFI = 0.983, RMSEA of 0.052 at 90% con-
PF4 0.936 dence interval (0.041, 0.067), IFI = 0.983, and TLI = 0.968. This t
Customer CP1 0.723 52.31 index is well under the suggested t (Kline, 1998). In this model,
preferences the covariance relationships between latent constructs PFSD
Special days SD1 0.833 70.05 and SDSF are signicant at p < 0.05 and p < 0.01, respectively.
SD2 0.841 The negative correlation of 0.09 between the constructs promo-
Seasonal SF1 0.769 59.99 tional factors and special days shows that these two aspects are
factors SF2 0.780 not planned or occurring at the same time. In other words, not
n = 826 observations; p-value < 0.01 for all factor loadings. many promotions are executed during holidays. Similarly, the
correlation between SD and SF is 0.23. This shows that good
weather conditions or high temperature and important festivals
Table 4
(Christmas and Easter) are not occurring at the same time in
Description of t indices. the UK.
After removing the (non-signicant) covariance relationships
Fit Description Suggested
index cut-off
between PF and SF, the model t has not changed much. The
resulting model t has normed v2 = 3.379, GFI = 0.981,
v2/df Normed chi-square: chi-square divided by degree of <4
CFI = 0.983, RMSEA of 0.054 at 90% condence interval (0.041,
freedom
0.067), IFI = 0.983, and TLI = 0.968.
CFI Comparative t index: compares the model t with a >0.90
baseline model
Next (see Fig. 4) we also remove the covariance relationships
between PFSD and SDSF (though signicant with small path
GFI Goodness of t index: indicates how well the model >0.90
reproduces the covariance matrix among the
coefcients). We note that the missing covariance relationships
observed variables PFSD and SDSF do not affect the t apart from a small change
IFI Incremental t index: group of goodness of t indices >0.90
in the path coefcients between SD and sales performance, and be-
that assesses how well a specied model ts relative tween SF and sales performance. Hence, it is clear that the inter-
to some alternative baseline model relationships between the rst order latent constructs are not
TLI Trucker Lewis index >0.90 highly inuencing the sales performance of 2 l bottles.
RMSEA Root mean square error of approximation <0.07
The model t has normed v2 = 3.741, GFI = 0.976, CFI = 0.978,
RMSEA of 0.058, IFI = 0.978, and TLI = 0.963. This t index is
again well under the suggested t. This model is considered as
a complete structural model for 2 l bottles. The R2-value of this
interrelationship between promotional and non-promotional fac- t is 0.748. In the structural model (Fig. 4) the promotional fac-
tors. In this section, we investigate the relationships between the tors have direct positive impact on the sales with path coef-
rst order constructs and sales performance, which is measured cient of 0.79 at a signicance level p < 0.01. Also, SD and SF
through sales volume. Figs. 3 and 4 show the initial and nal have a positive impact on the sales performance. The customer
structural equation model. We discuss these models and how preference factor CP1 has a positive path coefcient of 0.43 at
they were derived below. The models t is evaluated by normed signicance level p < 0.01.

Table 5
Fit indices for three different data sets.

v2 (df) Normed v2 GFI CFI RMSEA (90% CI) IFI TLI


Data set 1 101.01 (27) 3.741 0.976 0.978 0.05 (0.04, 0.07) 0.978 0.963
Data set 2 121.36 (27) 4.495 0.978 9.978 0.05 (0.04, 0.06) 0.978 0.964
Data set 3 157.09 (25) 5.818 0.964 0.959 0.07 (0.06, 0.08) 0.959 0.932
U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552 5551

Finally, to validate our model, we performed tests using differ- tors such as cannibalisation/competition, economic factors (ina-
ent samples from the same population. We used sales and promo- tion or recession), and whether the retail stores execute the
tional details from three different retail chains. The model ts were promotions properly. Another interesting extension is to mea-
acceptable using these three different sets and the t indices are sure sales performance by other or different metrics (e.g., sales
shown in Table 5. The good model ts provide good support for revenue and prot).
our conceptual demand-factor model.

5. Discussion and conclusion References

Abraham, M., & Lodish, L. (1987). PROMOTER: An automated promotion evaluation


The measurement model (Fig. 2) was just a representation of system. Marketing Science, 6(2), 101123.
covariance relationships between the latent constructs. Hence, it Ailawadi, K. L., Beauchamp, J. P., Donthu, N., Gauri, D. K., & Shankar, V. (2009).
Communication and promotion decisions in retailing: A review and directions
could not be considered as a complete model (Mishra & Shah,
for future research. Journal of Retailing, 85(1), 4255.
2009). In our test structural equation model (Fig. 3), we linked Ali, O. G., Sayn, S., Van Woensel, T., & Fransoo, J. (2009). SKU demand forecasting in
the measurement model with sales. However, we rejected this the presence of promotions. Expert Systems with Applications, 36(10),
model because there was no signicant correlation between pro- 1234012348.
Anderson, J. C., & Gerbing, D. W. (1988). Structural equation modelling in practice: A
motional factors and seasonal factors. Hence, there is no need to review and recommended two-step approach. Psychological Bulletin, 103(3),
model this link. Both promotional factors and seasonal factors do 411423.
have a positive impact on sales (as indicated by the positive path Arshinder, K. A., & Deshmukh, S. G. (2008). Supply chain coordination: Perspectives,
empirical studies and research directions. International Journal of Production
coefcients between PF-sales performance and SF-sales perfor- Economics, 115(2), 316335.
mance in Figs. 3 and 4). In the test model (Fig. 3), there are Aviv, Y. (2001). The effect of collaborative forecasting on supply chain performance.
small (but signicant) negative path coefcients between PF Management Science, 47(10), 13261343.
Aviv, Y. (2007). On the benets of collaborative forecasting partnerships between
SD and SDSF. The negative correlation ( 0.23) between special retailers and manufacturers. Management Science, 53(5), 777794.
days and seasonal factors may be justiable in the UK context as Blattberg, R. C., & Levin, A. (1987). Modeling the effectiveness and protability of
both Christmas and Easter do not match with periods of high trade promotions. Marketing Science, 6(2), 124146.
Byrne, P. J., & Heavey, C. (2006). The impact of information sharing and forecasting
temperature. The very small negative correlation ( 0.09 signi- in capacitated industrial supply chains: A case study. International Journal of
cant at p < 0.05) between special days and promotional factors Production Economics, 103(1), 420437.
only shows that promotions not necessarily occur on special Cachon, G. P., & Fisher, M. (2000). Supply chain inventory management and the
value of shared information. Management Science, 46(8), 10321048.
days. However, the positive path coefcient between SD-sales
Chen, F. (1998). Echelon reorder points, installation re-order points, and the value of
performance does indicate that the sales of soft drinks is higher centralized demand information. Management Science, 44(12), 221234.
during holidays and festivals. This is in line with previous re- Christen, M., Gupta, S., Porter, J., Staelin, R., & Wittink, D. (1997). Using market level
search ndings (Cooper et al., 1999; Divakar et al., 2005; Dube, data to understand promotion effects in a nonlinear model. Journal of Marketing
Research, 34(3), 322334.
2004). We eliminated the covariance relationships among three Cooper, L. G., Baron, P., Levy, W., Swisher, M., & Gogos, P. (1999). PromoCast: A
rst order latent constructs in the nal model (Fig. 4). Still we new forecasting method for promotional planning. Marketing Science, 18(3),
nd a direct signicant impact of PF, SD and SF on sales. Hence 301316.
Divakar, S., Ratchford, B. T., & Shankar, V. (2005). CHAN4CAST: A multichannel,
we can conclude that Hypotheses 1, 3 and 4 are supported. Fur- multi-region sales forecasting model and decision support system for consumer
thermore, customer preference (here represented by product packaged goods. Marketing Science, 24(3), 334350.
rank CP1) has a signicant positive factor loading of 0.43. This Dube, J. (2004). Multiple discreteness and product differentiation: Demand for
carbonated soft drinks. Marketing Science, 23(1), 6681.
provides support for the second hypothesis. Fildes, R., Nikolopoulos, K., Crone, S. F., & Syntetos, A. A. (2008). Forecasting and
The purpose of this paper was to identify the underlying operational research: A review. Journal of the Operational Research Society, 59(9),
structure of demand during promotional events by identifying 11501172.
Fisher, M. L. (1997). What is the right supply chain for your product? Harvard
the signicant explanatory factors. We applied structural equa- Business Review, 75(2), 105116.
tion modelling for this purpose. SEM is a multivariate statistical Fornell, C., & Larcker, D. (1981). Structural equation models with unobservable
analysis technique, which looks for relationships between vari- variables and measurement error. Journal of Marketing Research, 18(1),
3950.
ous constructs. Our SEM model links sales performance with four
Forslund, H., & Jonsson, P. (2007). The impact of forecast information quality on
latent constructs (promotional factors, seasonal factors, special supply chain performance. International Journal of Operations & Production
days and customer preferences), each measured through a num- Management, 27(1), 90107.
ber of indicator variables. Sales performance was measured by Gavirneni, S., Kapuscinski, R., & Tayur, S. (1999). Value of information in capacitated
supply chains. Management Science, 45(1), 1624.
sales volume (exogenous variable). We have tested the SEM Hair, J. F., Black, W. C., Babin, B. J., Anderson, R. E., & Tatham, R. L. (2006).
model for one family of products (2 l bottles) sold through three Multivariate data analysis (6th ed.). New Jersey: Prentice-Hall.
different retailers, using the data from a leading soft drink man- Ireland, R. K., & Crum, C. (2005). Supply chain collaboration: How to implement CPFR
and other best collaborative practices. Florida: J. Ross Publishing, Inc..
ufacturer in the UK. We believe that decomposing the promo- Kline, R. B. (1998). Principles and practice of structural equation modeling. New York,
tional sales in its underlying explanatory factors through SEM, NY: Guilford Press.
is an interesting new approach to gain more insight in the struc- Kulp, S. C., Lee, H. L., & Ofek, E. (2004). Manufacturer benets from information
integration with retail customers. Management Science, 50(4), 431444.
ture of demand and to assist managers in promotional planning. Lee, H. L. (2002). Aligning supply chain strategies with product uncertainties.
SEM also reveals the signicant variables and their inter- California Management Review, 44(3), 105119.
relationships. Hence, an interesting extension is to use these fac- Lee, H. L., So, K. C., & Tang, C. S. (2000). The value of information sharing in a two-
level supply chain. Management Science, 46(5), 626643.
tors in causal forecasting methods. Another interesting extension Mishra, A. A., & Shah, R. (2009). In union lies strength: Collaborative competence in
is to test the SEM models for different product families (e.g. 6- new product development and its performance effects. Journal of Operations
packs of 330 ml cans and 500 ml bottles). This could reveal dif- Management, 27(4), 324338.
Naik, P. A., Raman, K., & Winer, R. S. (2005). Planning marketing-mix strategies in
ferent demand structures for different product families (e.g.,
the presence of interaction effects. Marketing Science, 24(1), 2534.
500 ml bottles are impulse buy products and their sales are Paulraj, A., Lado, A. A., & Chen, I. J. (2008). Inter-organizational communication as a
much more impacted by the local weather). Our SEM models relational competency: Antecedents and performance outcomes in
can be modied to study the moderating and mediating effect collaborative buyersupplier relationships. Journal of Operations Management,
26(1), 4564.
of certain factors (e.g., are promotions more effective during Raghunathan, S. (2001). Information sharing in a supply chain: A note on its value
periods of high temperature?), or extended by incorporating fac- when demand is non stationary. Management Science, 47(4), 605610.
5552 U. Ramanathan, L. Muyldermans / Expert Systems with Applications 38 (2011) 55445552

Raju, J. S. (1995). Theoretical models of sales promotions: Contributions, limitations, Sun, B. (2005). Promotion effect on endogenous consumption. Marketing Science,
and a future research agenda. European Journal of Operational Research, 85(1), 24(3), 430443.
117. VICS (2004). CPFR guidelines, voluntary inter-industry commerce standards.
Shah, R., & Goldstein, S. M. (2006). Use of structural equation modeling in Available at: <http://www.cpfr.org>.
operations management research: Looking back and forward. Journal of Voss, C., Tsikriktsis, N., & Frohlich, M. (2002). Case research in operations
Operations Management, 24(1), 148169. management. International Journal of Operations and Production Management,
Steckel, J. H., Gupta, S., & Banerji, A. (2004). Supply chain decision making: Will 22(2), 195219.
shorter cycle times and shared point of sale information necessarily help?
Management Science, 50(4), 458464.

Вам также может понравиться