Int. J. Business Forecasting and Marketing Intelligence, Vol. 1, No.

2, 2009 139

Forecasting volumes for trade promotions in CPG

industry using market drivers

Balasubramanian Kanagasabapathi* and

K. Antony Arokia Durai Raj
Center for Knowledge Driven Information Systems,
Software Engineering and Technology (SET) Labs,
Infosys Technologies Limited,
44 Electronics City, Hosur Road, Bangalore – 560 100, India
E-mail: [email protected]
E-mail: [email protected]
*Corresponding author

B. Shoban Babu
Holistic Information Management Infrastructure (HIMI),
Software Engineering and Technology (SET) Labs,
Infosys Technologies Limited,
44 Electronics City, Hosur Road, Bangalore – 560 100, India
E-mail: [email protected]

Mitul Shah
Infosys Consulting,
44 Electronics City, Hosur Road, Bangalore – 560 100, India
E-mail: [email protected]

Abstract: Promotions are an integral part of the consumer packaged goods

(CPG) industry. Anywhere between 30–40% of the sales volumes are achieved
through various promotions. Promotions are instrumental in creating brand
visibility and awareness. In this study, we attempt to analyse the impact of
promotions with feature advertises, in-store display, temporary price discounts
etc. Five different multivariate regression models have been developed to
forecast the total sales of a product considering pricing and distribution
variables. The performance of these models has been analysed by using
syndicated data. Based on the results, it is found that the S-shaped (double-log)
model has shown superior performance over the other models considered in
this study.
Keywords: sales promotions; business forecasting; category management;
S-shaped models; CPG industry.
Reference to this paper should be made as follows: Kanagasabapathi, B.,
Antony Arokia Durai Raj, K., Shoban Babu, B. and Shah, M. (2009)
‘Forecasting volumes for trade promotions in CPG industry using market
drivers’, Int. J. Business Forecasting and Marketing Intelligence, Vol. 1,
No. 2, pp.139–152.

Copyright © 2009 Inderscience Enterprises Ltd.

140 B. Kanagasabapathi et al.

Biographical notes: B. Kanagasabapathi is working as a Research Associate at

the Center for Knowledge Driven Information Systems, SET Labs, Infosys
Technologies Limited, Bangalore, India. He received his PhD in Construction
Management from Indian Institute of Technology Madras, Chennai, India. His
research interests include scheduling, 4D CAD, forecasting, resource
management, statistical modelling and simulation modelling. He has published
research papers in journals such as International Journal of Operational
Research, International Journal of Industrial and Systems Engineering,
NICMAR Journal of Construction Management and Journal of Institution of
Engineers (India). He serves as a Referee for Construction Management and
Economics.

K. Antony Arokia Durai Raj is a Research Associate at the Center for

Knowledge Driven Information Systems, SET Labs, Infosys Technologies
Limited, Bangalore, India. He received his PhD in the area of Operation
Management from IIT Madras, Chennai, India. He has published research
papers in International Journal of Advanced Manufacturing Technology and
International Journal of Operational Research and in refereed international
and national conferences. His research interests are in the areas of forecasting,
scheduling, supply chain management, logistics and distribution management,
simulation of manufacturing systems and metaheuristics.

B. Shoban Babu is working as a Research Associate for SET Labs of Infosys

Technologies Limited, Bangalore, India. He has a joint research and
development experience in business forecasting, optimisation and
meta-heuristics with Intelligent Systems Research & Innovation Centre of BT
Innovate, British Telecom, Ipswich, UK. His research interests include
customer analytics, statistical modelling of retail, scheduling, logistics and
supply chain management and optimisation using meta-heuristics. He is a
Doctoral candidate of IIT Madras in the area of Production and Operations
Management. He has published and presented papers in refereed international
journals and conferences.

Mitul Shah is a Senior Associate at Retail, CPG & Logistics practice at Infosys
Consulting. He has a keen interest in forecasting and analytics and has
published several papers in international journals. He is a Key Contributor to
SCOR 9.0 and has been felicitated by Supply Chain Council in recognition of
outstanding contribution to Global Supply Chain Excellence in 2008. He has
spoken at numerous forecasting and supply chain conferences. He has patented
research in the area of risk assessment in forecasting.

1 Introduction

Price promotions or trade promotions are an integral part of the marketing mix since the
mid 1960s (Taylor, 1965). Increasingly, they represent the major share of the marketing
budget for most consumer packaged goods (CPG). In 2007, CPG companies planned to
spend 14% of their revenue on trade promotions, which may grow by 4% in 2008 (Lora
and Fenella, 2007). Promotions account for 50% of the $70 billion marketing budget of
CPG manufacturers (Progressive Grocer, 1995).
Promotional efforts are recognised as a potent tool for managing brands, with in-store
displays, feature advertising and temporary price reductions are the key components of a
traditional promotional mix (Blattberg and Neslin, 1990). An extensive research has been
 Forecasting volumes for trade promotions in CPG industry 141

conducted by many researchers especially academicians that temporary price reductions

substantially increase short-term brand sales (Blattberg et al., 1995). Also, the slow
market growth and a lack of differentiation have resulted in a highly competitive
environment in many product categories. This, in turn, is forcing managers to use
promotions more frequently and engage in price wars (Wall Street Journal, 1991).
Therefore, managers are hard-pressed for immediate results hence they turn to
promotional tools. Because of frequent use of such tactics, some marketers are concerned
about promotional wear-out effects resulting in steadily declining redemption rates for
such promotional vehicles as coupons (Forbes, 1989). Two potential negative effects of
promotions that have been mentioned are an increase in price sensitivity and a decrease in
brand loyalty (Papatla and Krishnamurthi, 1996). The literature, however, does not
provide concrete empirical evidence regarding such effects. This is because most of the
empirical work investigates promotional effects for only a few, typically one or two,
occasions after a promotional purchase.
Causal models are widely used in retail marketing forecasts. Also, there are two basic
approaches to demand function modelling (Talluri and Van Ryzin, 2005). One way of
incorporating the effects of marketing variables on sales is through models of individual
consumer choice behaviour. Then in a bottom-up forecasting fashion, these individual
choices are aggregated to get total demand. Another approach – called aggregate
forecasting – is to model aggregate demand directly as a function of price and other
marketing variables. In this paper we focus on this latter approach, as it is prevalent in
marketing theory and practice (Talluri and Van Ryzin, 2005).
In this research study, we present different regression based forecasting models by
adding variables with respect to distribution channel and pricing for retail marketing
forecasts and compare their results. This paper is organised as follows: Section 2 presents
a review of the background literature on the price sensitivity and promotional
effectiveness. Section 3 describes about the data used in this study. Section 4 describes
the details of the forecasting models. In Section 5, we discuss the results and discussions
of the forecasting models considered in this study. And finally, Section 6 offers
managerial implications and the conclusions out of this study.

2 Background

Walters (1991) developed a conceptual framework for retail promotional effects that
included brand substitution effects, inter-store sales displacements and the effects of
promotions on complementary goods. Abraham and Lodish (1987) described a method to
measure the effectiveness of promotional events using store-level data. Because the
approach uses store level data, it does not decompose the promotional lift (i.e., the
volume of sales above baseline) into sales that are truly incremental versus those that are
borrowed. Their approach also does not incorporate dynamics in consumer response or
the effects of previous marketing activity on future promotional events. Papatla and
Krishnamurthi (1996) proposed a brand choice model that provides an estimate of the
dynamic effects of promotions on loyalty to the brand and customers’ sensitivity to the
price of the brand and measures whether promotional purchases reinforce or reduce
subsequent response to similar promotions. Mela et al. (1997) examined the long-term
effects of promotion and advertising on consumers’ brand choice behaviour. Their
142 B. Kanagasabapathi et al.

research concluded that the consumers have become more price and promotion sensitive
over time. Kopalle et al. (1999) developed a dynamic, descriptive model of brand sales,
and integrated it with a normative model of retailer and manufacturer pricing decisions.
They have also analysed the trade-off between the effect of a promotion in the period it
offered and also the potentially negative effect it may have in subsequent periods.
Silva-Risso et al. (1999) developed a decision support system to determine the
manufacturer’s sales promotion calendar, which eventually improves routine promotional
decisions and potentially enhances the efficiency of the marketing function. Srinivasan et
al. (2004) investigated the manufacturer revenue, the retailer revenue and the retailer
margin effects of price promotions and reported that promotional planning is more
tactical than strategic. Also concluded that each promotion should be evaluated based on
its own financial impact during the dust-settling period. Hanssens et al. (2005) identified
and addressed the key issues related to present and future implementation of marketing
models in practice. Their research have reviewed existing market response models and
suggested areas of potential growth in model use; discussed the key characteristics of
successful implementation, standardisation and empirical generalisation of the market
response models. Lee et al. (2007) investigated the effectiveness of providing various
levels of support for the use of analogies in sales forecasting. Their analysis suggested
that a simple, easily implemented form of adaptation support could significantly improve
forecast accuracy under some conditions. Slotegraaf and Pauwels (2008) examined the
long-term effectiveness of marketing efforts by focusing on display, feature advertising
and price promotions. Their research concluded that both permanent and cumulative sales
effects from marketing promotions are greater for brands with higher equity and more
product introductions.
From the literature review it is evident that there is scope to develop regression based
forecasting models to predict total sales more accurately. Hence in this paper, an attempt
is made to develop linear and non-linear regression models to forecast sales volume. The
sales distribution variables along with pricing related variables are identified and
considered to forecast the total sales in this study. Earlier studies considered the variables
like actual price of the product, percentage of stores participating in in-store display and
percentage of stores participating in feature promotions to forecast the total sales
(Walters, 1991; Boatwright et al., 1999). We reviewed and implemented the models
reported in the literature in CPG context using price promotion, feature advertising,
in-store display and promotion reach through retail outlets.

3 Data for modelling

Most CPG manufacturers run various promotion activities that let retailers reduce the
price of the product, advertise the product on newspapers or display it on the floor. This
promotion, however, mostly occurs in a specific ‘key’ account or product level due to
budget limit. In this study, in order to model the current scenario, it is important to
consider the relationship between retail account strategy, competition and promotion
sensitivities; also we need sales and promotional data across different retailers. An
‘account’ is a particular market-retailer combination. Most researchers have used market
or chain-level data due to availability, estimation convenience and the fact that managers
usually do not have access to data at lower levels of aggregation (Dekimpe et al., 2005).
The use of such data may have the potential problem of aggregation bias (Christen et al.,
 Forecasting volumes for trade promotions in CPG industry 143

1997). Nijs et al. (2001) and Srinivasan et al. (2005) found that this bias to have a limited
impact. So, this study uses weekly data for chosen category to estimate our models. The
information in the syndicated data from the sample of stores is aggregated to the account
level. The data contains information about the regional market, product manufacturer,
brand description and Universal Product Code (UPC) for a consumer nondurable
category in one market area for the last three years, from July 2005 to July 2008. The
chosen category is ideal because there is sufficient discounting to calibrate the model and
the product is stock pilable.
This research uses weekly data for modelling which may contain ‘extreme’
observations in sales and/or the marketing-mix variables. If, however, these unusual data
points are numerous and not the main focus of this paper, they may be labelled as outliers
(e.g., caused by data errors, competitive promotions on which information is not
available, etc.). If not properly accounted for, such data points can produce sizeable
biases in the estimation of long-run marketing effects. In order to deal with this data
problem in this study, we neglect those data points for modelling. For example, if the
total sales or average product price values are zero then we classify that as an outlier and
eventually drop those data points.
Out of the 156 weeks of data, the first 104 weeks of data are used to develop the
model and the rest of the weeks data are used for validation. Since consumer promotions
are planned on the level of the account (Boatwright et al., 1999; Boatwright et al., 2004),
we assume correlation across accounts within a chain to be negligible, although this and
other types of correlations would be feasible to model. The 104 weeks of actual sales data
for three different UPCs in the same category is shown in Figure 1.

Figure 1 Actual total sales units for three data sets for 104 weeks (see online version for colours)

4 Model development

The total sales of a product can be written as a function of marketing mix variables that
include price, feature, display and interactions of feature and display promotions with any
price reductions.
144 B. Kanagasabapathi et al.

4.1 Variables in the model

The data source providers’ offers data for three basic types of variables. They are:
1 Sales in units and sales in dollars of each SKU.
2 Average retail price, base price (data source providers).
3 Estimates of the extent of promotional activities measured in percentages of all
commodity volume. Percentage of all commodity volume (%ACV) is a measure of
the extent of promotional activity over a group of stores in a retail account.
Among the variables available from data sources, the following variables are identified in
this study to develop the model. The earlier studies used pricing related variables to
predict the total sales (e.g., Boatwright et al., 1999; Blattberg and Levin, 1987). In this
study the variables with respect to distribution channel like percentage of stores
participating in promotional activities along with pricing related variables are considered.
The variables considered in this study are discussed in Sections 4.1.1 and 4.1.2.

4.1.1 Dependent variable

Total sales of a product (TS)
Total sales volume is defined as the number of units of a product sold during a particular
time period. In this study, total sales have been considered as a dependent variable.

4.1.2 Independent variables

Average price (AP)
The AP of the product is introduced as an independent variable in the model. The AP is
the average of promotional price (price at which a product is sold while the promotion is
run on a store during a time period) and non-promotional price (price at which a product
is sold while the promotion is not run on a store during a time period). Boatwright et al.
(1999) have used actual or shelf price in their model as the pricing related variable, but
the change in sales volume could be due to promotions run by the CPG industry. The AP
is considered in this study to capture the effect of promotions as the AP is a function of
promotional price and non-promotional price.

No. of stores participating in only in-store display advertise (DPA)

This variable is an estimate of the extent of display activity measured in percentage of all
commodity volume (%ACV) units. ACV is a measure of the extent of display activity
over a group of stores, thus bringing in important distribution related information in the
model. It is the sum of total value of all the commodities sold in a particular market.
ACV is typically used in the context of measuring the reach of a promotional activity.
The display promotion is carried out by displaying the product at a prominent place in a
retail store. This variable has been included as we use the account level data for
modelling. Moreover, it is intuitively appealing that more the number of stores
participates in display promotion, the more the consumer gets influenced and hence the
change in sales is imminent.
 Forecasting volumes for trade promotions in CPG industry 145

No. of stores participating in only feature advertise (FPA)

The promotion through features is run at retail stores, by circulating the information
about the product and promotional offer at a retail store, in a printed hard copy. The
number of stores participating in feature promotion will have direct impact on the sales
volume of a product. Hence, ACV of stores involved in such feature advertisements is
considered as an independent variable.

No. of stores participating in both feature and display advertise (DFPA)

If both display promotion and feature promotions are run simultaneously in a given time
period at retail stores, then the sales data are recorded under the head ‘feature and
display’. Displays and features are similar types of promotions, because both bring
product to the consumer’s attention. Furthermore, similar, if not identical, price
promotional tactics are used with displays and features. Nevertheless, it is not possible to
figure out the effect of feature or display promotion on total sales, when both feature and
display promotions are run simultaneously in a store. Hence, the number of stores
participating in both feature and display promotions is considered to develop the
regression models.

Percentage discount (PD)

The discounted price is computed based on AP and non-promotional price of the product
and is mathematically given as
⎧ No Promo Price − Average Price ⎫
PD = ⎨ ⎬ × 100.
⎩ No Promo Price ⎭
The price reduction influences the change in sales volume. Hence, the percentage
discount over the price of the product is introduced to capture the effect of price
reduction on total sales volume.
In this study, we have considered all the above mentioned variables for ‘own’ and
‘competitive’ products. The β’s have the interpretation of elasticity.

4.2 Regression models considered in this study

We have modelled the problem under consideration using five different regression
models and they are described as follows:

4.2.1 Multivariate linear regression (MLR) model

This research has used the variables presented in Section 4.1 which describes price,
feature, display and price discounts to specify the following function for the total sales.
TS = β 0 + β1 × ( APown ) + β 2 × ( DPAown ) + β 3 × ( FPAown ) + β 4 × ( DFPAown ) + β 5
× ( PDown ) + β 6 × ( APcomp ) + β 7 × ( DPAcomp ) + β8 × ( FPAcomp ) + β 9
× ( DFPAcomp ) + β10 × ( PDcomp ) + ξ.
146 B. Kanagasabapathi et al.

4.2.2 Multiplicative or power model

The generic multiplicative model is given as follows:

ln Z = β 0 × y1β 1 × y2β 2 × y3β 3 × ... ... ... ... ... ... × yM

βΜ
+ ξ.

where Z is a dependent variable (total sales) and y1, y2, y3,… are causal effects
(i.e., regular price, discount, display). Taking double-log transformation on variables
allows us to run linear regression for modelling. The transformed model is written as,
ln(ln TS ) = ln( β 0 ) + β1ln( APown ) + β 2ln( DPAown ) + β 3ln( FPAown )
+ β 4 ln( DFPAown ) + β 5ln( PDown ) + β 6ln( APcomp ) + β 7 ln( DPAcomp )
+ β8ln( FPAcomp ) + β 9ln( DFPAcomp ) + β10 ln( PDcomp ) + ξ.

4.2.3 Working-Leser (semi-log) model

Using Working-Leser model, the total sales can be computed as:
TS = β 0 + β1ln( APown ) + β 2 ln( DPAown ) + β 3ln( FPAown ) + β 4 ln( DFPAown )
+ β 5ln( PDown ) + β 6 ln( APcomp ) + β 7 ln( DPAcomp ) + β 8ln( FPAcomp )
+ β 9ln( DFPAcomp ) + β10ln( PDcomp ) + ξ.

4.2.4 Cobb-Douglas (double-log) model

The total sales can be computed as:
ln TS = ln( β 0 ) + β1ln( APown ) + β 2ln( DPAown ) + β 3ln( FPAown ) + β 4 ln( DFPAown )
+ β 5ln( PDown ) + β 6 ln( APcomp ) + β 7ln( DPAcomp ) + β8ln( FPAcomp )
+ β 9ln( DFPAcomp ) + β10 ln( PDcomp ) + ξ.

4.2.5 S-shaped (double-log) model

The generic S-shaped double-log model is given as follows:

⎛ Z − Z min ⎞
ln ⎜ ⎟ = ln β 0 + β1 ln y1 + β 2 ln y2 + β 3 ln y3 + ... ... ... ... ... ... + β M ln yM + ξ.
⎝ Z max − Z ⎠
The S-shaped double-log model with respect to the problem under consideration is

⎛ TS − TSmin ⎞
ln ⎜ ⎟ = ln( β 0 ) + β1 ln( APown ) + β 2 ln( DPAown ) + β 3 ln( FPAown )
⎝ TSmax − TS ⎠
+ β 4 ln( DFPAown ) + β 5 ln( PDown ) + β 6 ln( APcomp ) + β 7 ln( DPAcomp )
+ β 8 ln( FPAcomp ) + β 9 ln( DFPAcomp ) + β10 ln( PDcomp ) + ξ.
 Forecasting volumes for trade promotions in CPG industry 147

4.3 Relevant notations

TS total sales of the product
TSmin minimum of total sales during the time period considered for modelling
TSmax maximum of total sales during the time period considered for modelling
APown average price of the own product
DPAown no. of stores participating in only in-store display advertise on the own
product
FPAown no. of stores participating in only feature advertise on the own product
DFPAown no. of stores participating in feature and display advertise on the own product
PDown price discount on own product
APcomp average price of the competitor’s product
DPAcomp no. of stores participating in only in-store display advertise on the competitive
product
FPAcomp no. of stores participating in only feature advertise on the competitive product
DFPAcomp no. of stores participating in feature and display advertise on the competitive
product
PDcomp price discount on competitor’s product.

5 Results and discussions

In this section we present the results of the models for the category considered in this
study and discussions on our model validation analysis. The models presented in
Section 4.2 are programmed and implemented in SAS Enterprise Guide 4.1 and
Weka 3.4. In the forecasting literature, the Mean Absolute Percentage Error (MAPE) has
been used for measuring the forecasting accuracy. The MAPE has certain advantages:
a it can be averaged across years or months or weeks
b it is easy to interpret.
The limitation of using MAPE is that if it is not free of outliers, the value can be skewed.
Later researchers attempted to use the median rather than the mean and some researchers
used a cutoff or threshold value to eliminate outliers (e.g., Brown and Niederhoffer,
1968). Unlike the MAPE, Median Absolute Percentage Error (MdAPE) is not distorted
by extreme percentage errors. In this study, the forecasting accuracy is measured by
considering MAPE, MdAPE and Standard Deviation of Absolute Percentage Error
(SdAPE). The SdAPE is reported in order to understand how far the Absolute Percentage
Error (APE) is scattered around the MAPE. The MAPE, MdAPE and SdAPE for each of
148 B. Kanagasabapathi et al.

the models across three data sets are tabulated in Tables 1 to 3. The mean of different
data sets with respect to each of the models have also been computed and presented.
Table 1 The MAPE for the data sets across the different models

Multiplicative Working-Leser Cobb-Douglas S-shaped

MLR model
model model model model
Data 1 24.83% 14.11% 37.66% 13.09% 13.00%
Data 2 27.31% 18.97% 57.33% 18.67% 17.09%
Data 3 20.57% 18.45% 18.92% 18.43% 16.64%
Mean 24.24% 17.18% 37.97% 16.73% 15.58%

Table 2 The MdAPE for the data sets across the different models

Multiplicative Working-Leser Cobb-Douglas S-shaped

MLR model
model model model model
Data 1 18.87% 13.86% 31.44% 12.76% 13.13%
Data 2 16.94% 16.71% 45.03% 15.25% 13.65%
Data 3 14.49% 12.57% 12.50% 12.33% 13.22%
Mean 16.77% 14.38% 29.66% 13.45% 13.33%

Table 3 The SdAPE for the data sets across the different models

Multiplicative Working-Leser Cobb-Douglas S-shaped

MLR model
model model model model
Data 1 31.83% 9.76% 38.03% 9.53% 7.98%
Data 2 27.29% 13.31% 53.85% 13.30% 10.52%
Data 3 17.41% 16.57% 16.42% 16.56% 16.64%
Mean 25.51% 13.21% 36.10% 13.13% 11.71%

Figure 2 MAPE obtained by the models considered in this study on three different data sets
(see online version for colours)

The computation result shows that the S-shaped model predicts better than the other
models with a forecasting accuracy (considering MAPE presented in Table 1) as high as
 Forecasting volumes for trade promotions in CPG industry 149

82.91% to 87% (approximately 83% to 87% respectively). Also, the S-shaped model
performs consistently better than the other models while considering other performance
measures such as MdAPE and SdAPE. Figures 2, 3 and 4 represents the MAPE, MdAPE
and SdAPE obtained by the different models on three data sets respectively. Based on the
results presented in Tables 1 to 3, the double-log models like S-shaped model,
Cobb-Douglas model and multiplicative or power model have shown superior
performance over the Working-Leser model (semi-log model) and MLR model. It is
interesting to see that the MLR model performs better than the Working-Leser model
(refer to Figure 2). The superior performance of double-log models is due to the
normalisation of values by taking logarithmic values of both dependent variable and
independent variables.

Figure 3 MdAPE obtained by the models considered in this study on three different data sets
(see online version for colours)

Figure 4 SdAPE obtained by the models considered in this study on three different data sets
(see online version for colours)
150 B. Kanagasabapathi et al.

The S-shaped model has found to be performing well in forecasting retail markets for
different data sets. The reason behind the superior performance of S-shaped model is
attributed to the consideration of normalised dependent variable (total sales) with respect
to the minimum and maximum total sales units of a product, in the modelling equation
(as opposed to the other models considered in this study). The S-shaped model has been
found to be best suited for forecasting the total sales with promotional effects in CPG
industry. One of the potential explanations of superior performance of S-shaped models
is that they closely reflect the consumer buying behaviour with respect to variables like
price. Price impacts consumer buying behaviour in a certain range, where price elasticity
may have a linear or log-linear relationship with the sales. Once price breaches through
this threshold point, we see a sudden change in elasticity till the time price vs. volume
reaches to the next level. Till the next threshold price point is crossed, it reflects steady
relationship between price and volume. This is also reflective of other marketing
variables considered in this study. S-shape very closely imitates this behaviour and hence,
we believe, it’s the best suited model for this business scenario. Even though the data
sources supplies data for more than 20 variables, this study identifies and considers the
factors or variables that have direct influence on the total sales. This study is unique in
the sense that it considers important distribution related variables and pricing related
variables to develop the statistical models to forecast sales volume, whereas earlier
studies have omitted pricing related variables and distribution related variables that are
considered in this study.

6 Summary and conclusions

Accurate sales promotion forecasting is a key for the managers to plan and effectively
manage various promotions to increase the sales and return on investment (ROI).
Equally, it is important to understand the isolated impact of each of the market driver on
the sales volume. In this paper, we have investigated the effects of different promotional
activities on total sales of the product for a selected category over 152 weeks of sample
data. The breadth of the sample allows us to derive empirical generalisations on
price-promotion effectiveness and drivers. The impact of promotions on total sales has
been presented by considering important factors like average price, percentage of stores
that run features promotion, percentage of stores that run display promotion, percentage
of stores that run both feature and display promotion and percentage discount offered on
a particular product. This study also indicates that both manufacturers and retailers can
benefit from the use of account-level data. The parameters estimated in this research
would allow managers to assess the impact of variable prices on sales of their brand or
category.
The study also presented the performance evaluation of linear and non-linear
regression models using real-life data sets in the context of CPG industry. This study
considers MAPE, MdAPE and SdAPE for comparing the performance of all models. It is
evident from the computational results and comparisons that the S-shaped models can
provide superior forecasts over the other models considered in this study. We are
researching techniques like Bayesian shrinkage along with multivariate models to arrive
at more accurate coefficients. Bayesian shrinkage method can also be used with grid
search algorithms to simulate coefficient values. This line of research will allow us to
 Forecasting volumes for trade promotions in CPG industry 151

obtain estimates that will more closely resemble the true underlying distribution of
marketing variable by shrinking the estimates towards a common mean.

Acknowledgements

The authors are thankful to the referee and the editor for their valuable suggestions to
improve the earlier version of this paper. The authors extend their heartfelt thanks to
CATSTAT team for their support.

References
Abraham, M.M. and Lodish, L.M. (1987) ‘Promoter: an automated promotion evaluation system’,
Marketing Science, Vol. 6, No. 2, pp.101–123.
Blattberg, R. and Levin, A. (1987) ‘Modelling the effectiveness and profitability of trade
promotions’, Marketing Science, Vol. 6, No. 2, pp.124–146.
Blattberg, R.C. and Neslin, S.A. (1990) Sales Promotion, Prentice Hall, Englewood Cliffs, NJ.
Blattberg, R., Briesch, R. and Fox, E. (1995) ‘How promotions work’, Marketing Science, Vol. 14,
No. 3, pp.G122–G132.
Boatwright, P., Dhar, S. and Rossi, P.E. (2004) ‘The role of retail competition, demographics and
account retail strategy as drivers of promotional sensitivity’, Quantitative Marketing and
Economics, Vol. 2, No. 2, pp.169–190.
Boatwright, P., McCulloch, R. and Rossi, P.E. (1999) ‘Account level modelling for trade
promotion: an application of a constrained parameter hierarchical model’, Journal of
American Statistical Association, Vol. 94, No. 448, pp.1063–1073.
Brown, P. and Niederhoffer, V. (1968) ‘The predictive content of quarterly earnings’, Journal of
Business, Vol. 41, No. 4, pp.488–497.
Christen, M., Gupta, S., Porter, J.C., Staelin, R. and Wittink, D.R. (1997) ‘Using market-level data
to understand promotion effects in a nonlinear model’, Journal of Marketing Research,
Vol. 34, No. 3, pp.322–334.
Dekimpe, M.G., Hanssens, D.M., Nijs, V.R. and Steenkamp, J.E.M. (2005) ‘Measuring short- and
long-run promotional effectiveness on scanner data using persistence modelling’, Applied
Stochastic Models in Business and Industry, Vol. 21, No. 4–5, pp.409–416.
Forbes (1989) ‘Stealing the right shoppers’, 10 May, pp.104–105.
Hanssens, D.M., Leeflang, P.S.H. and Wittink, D.R. (2005) ‘Market response models and
marketing practice’, Applied Stochastic Models in Business and Industry, Vol. 21, No. 4–5,
pp.423–434.
Kopalle, P.K., Mela, C.F. and Marsh, L. (1999) ‘The dynamic effect of discounting on sales:
empirical analysis and normative pricing implications’, Marketing Science, Vol. 18, No. 3,
pp.317–332.
Lee, W.Y., Goodwin, P., Fildes, R., Nikolopoulos, K. and Lawrence, M. (2007) ‘Providing support
for the use of analogies in demand forecasting tasks’, International Journal of Forecasting,
Vol. 23, No. 3, pp.377–390.
Lora, C. and Fenella, S. (2007) ‘What is an effective trade promotion?’, AMR Research Report, 9
November 2007.
Mela, C.F., Gupta, S. and Lehmann, D.R. (1997) ‘The long-term impact of promotion and
advertising on consumer brand choice’, Journal of Marketing Research, Vol. 34, No. 2,
pp.248–261.
 152 B. Kanagasabapathi et al.

Nijs, V.R., Dekimpe, M.G., Steenkamp, J-B.E.M. and Hanssens, D.M. (2001) ‘The category
demand effects of price promotions’, Marketing Science, Vol. 20, No. 1, pp.1–22.
Papatla, P. and Krishnamurthi, L. (1996) ‘Measuring the dynamic effects of promotions on brand
choice’, Journal of Marketing Research, Vol. 33, No. 1, pp.20–35.
Progressive Grocer (1995) ‘The prisoner’s dilemma’, May, Vol. 74.
Silva-Risso, J.M., Bucklin, R.E. and Morrison, D.G. (1999) ‘A decision support system for
planning manufacturers’ sales promotion calendars’, Marketing Science, Vol. 18, No. 3,
pp.274–300.
Slotegraaf, R.J. and Pauwels, K. (2008) ‘The impact of brand equity and innovation on the
long-term effectiveness of promotions’, Journal of Marketing Research, Vol. 45, No. 3,
pp.293–306.
Srinivasan, S., Pauwels, K. and Nijs, V.R. (2005) ‘Dynamic drivers of retail prices’, Working
paper, University of California, Riverside.
Srinivasan, S., Pauwels, K., Hanssens, D.M. and Dekimpe, M.G. (2004) ‘Do promotions benefit
manufacturers, retailers, or both?’, Management Science, Vol. 50, No. 5, pp.617–629.
Talluri, K.T. and Van Ryzin, G.J. (2005) The Theory and Practice of Revenue Management,
Springer Science+Business Media, Inc., New York, USA.
Taylor, J.W. (1965) ‘Two requirements for measuring the effectiveness of promotion’, Journal of
Marketing, Vol. 29, No. 2, pp.43–45.
Wall Street Journal (1991) ‘Grocery price wars squeeze marketers’, 7 November, B1.
Walters, R.G. (1991) ‘Assessing the impact of retail price promotions on product substitution,
complementary purchase, and interstore sales displacement’, Journal of Marketing, Vol. 5,
No. 2, pp.17–28.

The author has requested enhancement of the downloaded file. All in-text references underlined in blue are linked to publications on ResearchGate.