Neural Networks for the

Analysis and Forecasting

of Advertising and
Promotion Impact
Hean-Lee Poh, Jingtao Yao* and Teo Jas̆ic
National University of Singapore, Singapore

Abstract Allocating advertising expenses and forecasting total sales levels are the key
issues in retailing, especially when many products are covered and significant
cross-effects among products are likely. Various statistical and econometric
methods could be applied for such analyses. We explore how well neural
networks can be used in analyzing the effects of advertising and promotion
on sales in this article. The results reveal that the predictive quality of neural
networks depends on the different frequency of data observed, i.e. daily or
weekly data models, and the specific learning algorithms used. The study also
shows that neural networks are capable of capturing the nonlinear aspects of
complex relationships in non-stationary data. By performing sensitivity analysis,
neural networks can potentially single out important input variables, thereby
making it useful for scenario development and practical use.  1998 John
Wiley & Sons, Ltd.
Keywords: neural networks; sensitivity analysis; sales prediction; advertisement allocaion.

INTRODUCTION them make informed choices. Standard econo-

metric techniques which deal with such a prob-
This study explores the use of neural networks lem have limited explanatory capabilities
for marketing decision support in the retail because they are based on linear models. Neu-
business. Its aim is to capture the complex ral network technology has seen many appli-
relationships between marketing factors, such cation areas in business especially when the
as advertising and promotion strategy, and the problem domain involves classification, recog-
total sales levels. The important influence fac- nition and predictions. With the capabilities of
tors could be found after neural network fore- neural networks, hidden trends and relations
casting models are built and sensitivity analysis among data which are previously unseen can
is then conducted. Faster and better decisions be deduced. In other words, this is obtaining
are an important way to assist the businesses information from information (White, 1990). A
to survive in the rapidly changing and competi- recent survey research conducted by Wong et
tive business environments. Marketing decision al. (1995) indicated that at least 127 neural
makers are increasingly drawn to computer network business application journal papers
based Decision Support Systems (DSS) to help had been published up to September 1994. Dif-
ferent neural network models have been
*
Correspondence to: Jingtao Yao, School of Comput- applied to solving hard real-world problems
ing, National University of Singapore, Singapore such as financial forecasting. With their success-
119260. E-mail: yaojt얀comp.nus.edu.sg ful application to business problems, intelligent

CCC 1055-615X/98/040253-16$17.50 Received August 1996

 1998 John Wiley & Sons, Ltd. Revised June 1998

International Journal of Intelligent Systems in Accounting, Finance & Management

Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)
decision support systems are seen as a way methodologies aim to capture the relationships
to overcome ill-structured business strategies between the targeted result, the sales volume,
problems. Examples of the neural networks and marketing strategies used by the retailers,
used to learn functional relationships from their and in particular, the marketing decision
input variables to predict results can be found related to advertising and promotion strategies.
in Dutta et al. (1994), Hill and Ramus (1994), Although there are several defined econometric
Poh (1991), Poh and Jas̆ik (1995), White (1990) models developed for capturing the relation-
and Yao et al. (1997). One of the key issues ship of these factors in marketing analysis,
concerned in retailing is deciding which pro- there are no general rules between strategic
ducts to be advertised and promoted in order promotion decisions and performance. Rules
to increase the overall store sales. Given the that have been developed are often a result of
competitive pressure in the retailing sector, the past experiences as well as an intuitive aware-
retailers of the future are likely to accept only ness of local market conditions, rather than
those promotions which lead to an increase in a consistently developed systematic approach.
category and overall store sales. This may be Given these circumstances, a neural network
achieved by designing promotional programs technology can shed some light on these com-
which increase category and overall store sales, plex relationships and potentially be used as
and by directing promotional efforts towards a complementary tool, offering some help to
categories which are more responsive to pro- managers facing important decisions.
motions (Raju, 1992). Cost decisions (i.e. budget This paper begins with a basic concepts of
allocations), copy decisions (i.e. what message neural networks used in this study. It is fol-
to use), and media decisions (i.e. what media lowed by a section on data summary and back-
to employ) are three major decision areas for ground information. Subsequently, a section
advertising (Aaker and Myers, 1982). which presents the training and forecasting
There are two aspects of the given problem: results, a section which discusses sensitivity
(1) planning and allocating advertising analysis and a section which gives major find-
expenses; and (2) forecasting total sales levels ings follow. The conclusion discusses areas for
across a wide product range where significant future research.
cross-effects among products are likely. For a
single brand, the optimum level of advertising
depends upon its margin and the advertising CONCEPTS OF NEURAL NETWORKS
elasticity of demand. However, for a retailer’s
product, two additional factors are important A neural network is a collection of intercon-
as stated by Doyle and Saunders (1990): the nected simple processing elements, or neurons.
cross-effects of the advertised products on other Neural networks are potentially useful for
items in the retailer’s assortment; and the studying the complex relationships between
impact on overall store traffic as the effective- inputs and outputs of a system (White, 1990).
ness of advertising is reflected in the overall The data analysis performed by neural net-
store performance rather than in the sales of works tolerates a considerable amount of
the promoted product. In addition to fore- imprecise and incomplete input data due to
casting the sales based on advertising, sensi- the distributed mode of information processing.
tivity analysis could be used to determine There are two neural network models investi-
which promoted categories have more impact gated in this research: backpropagation net-
on the sales volume. works and counterpropagation networks. There
The primary focus of this research is to are three major steps in the neural network-
explore the potential and to investigate the based forecasting proposed by this research:
efficacy of neural networks as an alternative to preprocessing, architecture, and postprocessing.
statistical methods such as multiple regression In preprocessing, information that could be used
to predict the variable of interest, and to con- as the inputs and outputs of neural networks
duct sensitivity analysis of their models. Both are collected. These data are first normalized
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

254 H.-L. POH ET AL.

or scaled in order to reduce the fluctuation and den layer, which can be trained using backpro-
noise. In architecture, a variety of neural net- pagation. The shaded nodes in Figure 1 are
work models that could be used to capture the processing units. The arrows connecting input
relationships between the data of inputs and and hidden units and connecting hidden units
outputs are built. Different models and con- and the output unit represent weights.
figurations using different training, validation The backpropagation learning algorithm
and forecasting data sets are used for experi- (Rumelhart et al., 1986; Werbos, 1974) is formu-
ments. The best models are then selected for lated as a search in the space of the pattern
use in forecasting based on such measures as of weights, W, in order to find an optimal
out-of-sample hit rates. Sensitive analysis is configuration, W*, which minimizes an error or
then performed to find the most influential cost function, E(W). The pattern of weights will
variables fed to the neural network. Finally, in then determine how the network will respond
postprocessing, different trading strategies are to any arbitrary input. The error or cost func-
applied to the forecasting results to maximize tion is
the capability of the neural network prediction.
E=
1
2
冘 冘 (t
i p
ip − oip)2 (1)

Backpropagation Networks This function compares an output value oip to

A multilayer feedforward network with an a desired value tip over the set of p training
appropriate pattern of weights can be used to samples and i output units. The gradient
model some mapping between sets of input descent method is used to search for the mini-
and output variables. Figure 1 shows an mum of this error function through iterative
example of a simple feedforward network updates:
architecture, with one output unit and one hid- W(k+1) = W(k) − ␩ⵜE (2)

Figure 1 The architecture of a backpropagation network

 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 255

where ␩ is the learning rate, and ⵜE is an ⌬wij = wijnew − wold
ij = ␩(ti − oi)Vj (3)
estimate of the gradient of E with respect to W. where w , w new old
are the weights between
ij ij
The algorithm is recursive and consists of elements i and j, Vj is the output of the winning
two phases: forward-propagation and back- element (usually set to 1), ␩ is the learning rate,
ward-propagation. In the first phase, the input Oi is the output of the network and ti is the
set of values is presented and propagated for- desired output (target).
ward through the network to compute the out- The outputs of the other hidden elements are
put value for each unit. In the second phase, set to zero. It should be noted that an underly-
the total-squared error calculated in the first ing assumption of such a network is that input
phase is ‘backpropagated’, layer by layer, from vectors of similar Euclidean distances will give
the output units to the input units. During this similar outputs. This condition may cause prob-
process, the error signal is calculated recur- lems in situations where the output of the data
sively for each unit in the network and weight is sensitive to small changes in the input vector
adjustments are determined at each level. The space. Such situations are common in real-
two phases are executed in each iteration of world business and forecasting problems (e.g.
the backpropagation algorithm until the error time-related patterns of financial data series),
function converges. making the suitability of such networks depen-
dent on the type of data used. An appropriate
Counterpropagation Network data analysis is desirable before using this type
of network.
A counterpropagation network (Hecht-Nielsen,
1987) is a hybrid learning network. It combines
a so-called Kohonen layer of unsupervised
learning with another layer of supervised learn- VARIABLE SELECTION AND DATA
ing which uses the basic delta rule. It can speed SUMMARY
up learning considerably by training some lay-
ers in an unsupervised way (i.e. without a
Data Sample
teacher/specified output). This works well in a
situation where similar input vectors produce The data used in this study are collected from
similar outputs. Thus, the aim of Kohonen layer a retailer in Singapore who has a chain of
is to categorize the input data into clusters with six department stores at various locations. The
competitive learning, and then use only the company is centrally operated with inclination
category information for the supervised learn- toward a top-down product budgeting
ing. The classes of similar inputs are defined approach. Most of the promotion and advertis-
by a set of prototype vectors. The class of a ing decisions are made by the headquarters
given input is found by finding the nearest manager. The advertising and promotion cam-
prototype vector using the ordinary (Euclidean) paign data cover various product categories,
metric. The classes must be formed by the different media, expenditure levels and sales
network itself from the correlation of the data are collected for the period between April
input data. 1989 and March 1991. A campaign here can be
The architecture of the counterpropagation defined as the advertising expenditure aimed
network is shown in Figure 2. It consists of at either a specific merchandise or a general
three layers of neurons: the input layer where sales event. It is possible for campaigns to be
the input pattern is fed, the hidden layer where overlapping. The data for competitive positions,
competitive (Kohonen) learning takes place, i.e. market share, prices and costs relative to
and the output layer where supervised delta competition, advertising and promotion strat-
rule learning occurs. The winning element in egy of competitors, are not available. This set
the hidden layer sends the value 1 into the of data should capture important linkages
third layer. Then the output layer, which is between marketing strategies and affect sales
linear, is trained with the usual delta rule over the time and hence the overall profitability
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

256 H.-L. POH ET AL.

Figure 2 Structure of a counterpropagation network

and the growth of organization, sales pro- the macro level or seasonal effects, resulting in
motion, special events, marketing services and fewer training sets.
visual merchandising. There are four sections The graphs of daily sales and weekly sales
in the advertising and promotion department and advertising expenditures are shown in
of the retailer organization under study. The Figures 3 and 4, respectively. In Figure 3, the
sales promotion and special events section graph of the observed daily sales over the
make the core decision while the other two specified period is a series of discrete peaks,
play support roles. The manager’s decision in which represent sales during the weekends
advertising or promotion depends on the rec- (Saturday and Sunday). However, in some per-
ords of past advertising and promotion cam- iods, the graph allows the capture of continu-
paigns, the latest trend, competitors’ activities, ous trends, while giving rise to chaotic regimes
budgets, inventory data, sales figures, profit in others due to seasonal effects. Figure 4 shows
loss statement and other raw data. To establish data sets organized on a weekly basis. As this
budgets, a simple historical method, adjusted pattern represents an aggregation of daily
from previous-year budgets, is used. values, the seasonal effects are now more obvi-
The original form of data is the daily model, ous, although the trend and the cycle cannot
and the weekly model is easily derived from be detected a priori. It is interesting to observe
it. The daily data model takes into account a the dynamics of input variables, such as adver-
very short lag effect of advertising with obser- tising expenditure levels versus sales levels. At
vation of an immediate sales response. Sales times it constitutes a uniform policy, while at
results are taken one day after the associated other times, it is a pulsing policy. The high
advertisement placement. The weekly data peak in July 1989 is due to a massive promotion
model is expected to aggregate this effect at campaign for a special anniversary of the retail
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 257

Figure 3 Sales data of stores 1 and 4 on a daily basis for the period: April 1989 to March 1990 (The units
are in 1000 Singapore dollars, the solid line refers to sales in store 1 while the dotted line refers to store 4)

Figure 4 Sales data of store 1 and advertising expenditure on a weekly basis for the period: April 1989 to
March 1990 (units are in 1000 Singapore dollars)
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

258 H.-L. POH ET AL.

organization combined with the regular annual the number of advertisements appearing in that
summer sale. medium. As a few different advertisements can
be present in the same medium on the same
day, the value of the variables in this group is
Classification of Variables
sometimes greater than one. For the weekly
There are 22 input variables extracted for use data model, these values would be aggregated.
in this study. They follow the existing retailer’s Variable var14 (Newspaper 1) is the English
classification of advertising and promotion newspaper with the largest circulation, while
campaign practices. These input variables are Newspaper 2 (var15) is the local language
roughly divided into three major groups: (1) newspaper with major circulation. The other
variables related to the categories of products newspaper media, Newspaper 3 (var16) and
which have been advertised (Var1 to Var13); (2) Newspaper 4 (var17), do not feature advertise-
variables related to different media being used ments of the retailer so often.
for advertising and promotion (Var14 to Var20);
and (3) variables related to the cost of advertis-
ing and promotion activities with respect to the STATISTICAL MODELS FOR IMPACT OF
group of stores to which the budget is allocated ADVERTISING AND PROMOTION
(Var21, Var22). The classification of input and
output variables in the daily data model Two multiple linear regression models are built
together with their mean and standard devi- to evaluate the impact of advertising and pro-
ation is given in Table 1. motion. The models can be formulated as fol-
The effect of advertising and promotion is lows:
observed in an aggregate model comprising a
冘 ␤ var + ␮
n

different categories of products, and the SaleVolume = ␤0 + i i (4)

i=1
observed sales volume is the sum of sales of
all items. This focus is relevant from a retailer’s where n = 22 for daily model and n = 19 for
perspective because the retailer’s revenues are weekly model; ␤0, ␤1, ␤2, . . ., ␤n are regression
more closely linked to the overall category and coefficients; SaleVolume is the Sales Volume to
store sales than to the sales of any particular be forecast; and ␮ is the error term that cannot
brand or product category. For the first group be explained by independent variables.
of variables (categories of products), there can The number of input units in the weekly
be various advertisements related to different model is reduced to 18 by performing
categories of items appearing in different media regression analysis and excluding variables
on a daily basis. The variable for the product which are found not significant, such as the
category therefore assumes the value which is variable var11 (Stock Clearance Announcement).
equal to the number of advertisements on Furthermore, the variable var2 (Weekend
each day. Indicator) is excluded because in the weekly
Variables var8 (Gift Ideas), var10 (Storewide model only seasonal effects at the macro level
Advertising and Promotion Campaign), var11 are present. Two variables describing news-
(Stock Clearance Announcement), and var13 paper media in the daily model (var16/News-
(Fashion Shows) are not confined to any spe- paper 3 and var17/Newspaper 4) can be com-
cific category of products and are characterized bined together. Finally, two variables related to
as special events. These events can be grouped the cost of advertising (var21 and var22) in the
together as they do not coincide at the same daily model are also combined into one. There-
time and represent promotional sales of differ- fore, the definition of vari may differ from the
ent ranges of products across the stores. The daily model and thus a bracketed term is pro-
incorporation of these events is important vided after each variable.
because they can have a strong influence on The models aim to seek the least mean
the ‘normal’ pattern of sales. squared error between the actual and the fore-
For the second group (media), the value is cast sales volumes so as to determine the coef-
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 259

Table 1. Data summary of training data in the daily data model (var2 is a binary code to indicate the
weekend, var18 is the number of video clips or spots televised, var20 to var23 are in thousand Singapore
dollars and others are number of times)

Variable Meaning Mean Std. Dev. Max. value

var1 Man Fashion 0.05 0.2114 1.00

var2 Weekend Indicator – – –
var3 Ladies’ Fashion 0.1085 0.3123 1.00
var4 Ladies’ Accessories 0.0233 0.1513 1.00
var5 Toys/Stationery 0.0155 0.1240 1.00
var6 Children’s wear Products 0.0155 0.1240 1.00
var7 Household 0.0233 0.1513 1.00
var8 Gift Ideas 0.0310 0.1740 1.00
var9 Cosmetics 0.5814 0.5687 3.00
var10 Storewide Sales Advert. 0.4031 0.5379 2.00
var11 Stock Clearance Announcement 0.01 0.088 1.00
var12 Seasonal Fashion Statements 0.062 0.2421 1.00
var13 Fashion Shows 0.0775 0.2685 1.00
var14 Newspaper 1 1.2016 0.6541 3.00
var15 Newspaper 2 0.4264 0.4965 2.00
var16 Newspaper 3 0.0543 0.2274 1.00
var17 Newspaper 4 0.2868 1.7862 20.00
var18 TV (3 Channels) 2470.931 7665.473 66000.00
var19 Mailers 0.0465 0.2114 1.00
var20 Tourist Editions 0.0620 0.2421 1.00
var21 Cost (Group 1) 8023.465 232823.81 193320.00
var22 Cost (Group 2) 135665.40 25055.51 175000.00
var23 Sales Volume (Store 1) 69.28 50.38 409.00
var23 Sales Volume (Store 2) 57.34 37.18 317.00
var23 Sales Volume (Store 3) 72.78 48.50 385.00
var23 Sales Volume (Store 4) 144.37 80.36 654.00
var23 Sales Volume (Store 5) 129.88 90.40 809.00
var23 Sales Volume (Store 6) 149.57 107.21 967.00

ficient values. We used a t-test to test the sig- (where Sx and Sy are the standard deviations
nificance of each individual coefficient. For the of x and y respectively) is used to check on
daily model, the variables with significant or the collinearity of data set x and y. The results
nearly significant t-statistics are: var2 (Weekend have shown that there are seven variables that
Indicator), var4 (Ladies’ Accessories), var8 (Gift are correlated with dependent variables for the
Ideas), var9 (Cosmetics), var12 (Seasonal Fashion weekly model. They are var1 (Man Fashion,
Statements), var13 (Fashion Shows) and var15 0.335), var4 (Toys/Stationery, 0.204), var5
(Newspaper 2). For the weekly model, the vari- (Children’s wear Products, 0.296), var7 (Gift
ables with significant or nearly significant t- Ideas, 0.368), var8 (Cosmetics, −0.216), var10
statistics are: var4 (Toys/Stationery), var7 (Gift (Seasonal Fashion Statements, −0.202) and var13
Ideas), var8 (Cosmetics), var11 (Fashion Shows), (Newspaper 2, 0.267). More results on statistical
var13 (Newspaper 2) and var18 (Cost). These models may be found in Jas̆ic (1993) where the
variables are generally consistent for both mod- impact on seasonal indices is also discussed.
els except for Toys/Stationery and Statements.
Pearson’s correlation coefficient which is BUILDING THE NEURAL NETWORK
defined as MODELS

冘 (x − x) (y − y)/(N − 1)S S
N

r= (5) In this study there are 730 samples for the

i i x y
i=1 daily model and 104 for the weekly model in
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

260 H.-L. POH ET AL.

the period between April 1989 and March 1991. ber of training sets), a cross-validation pro-
The first 360 daily samples and 52 weekly cedure may be needed due to the large number
samples data are used for training, while 92 of parameters provided which can cause the
daily samples (from January 1991 to March oversized network to fit a noise (Weigend et al.,
1991) and 26 weekly samples (from October 1990). Alternatively, certain pruning algorithms
1990 to March 1991) are used to observe how may be pursued.
well the neural network performs. The remain- In general, some experiments are carried out
ing data can be used for validation purposes to find the configuration of the neural network
to avoid over fitting in backpropagation. for a particular problem. For the number of
The data size requirement depends on the hidden-layer units, a rule of thumb is used that
number of input and output units of the neural defines the number of hidden units as half of
network. Data are normalized to values within the sum of the input and output units. Thus,
the interval [0,1] for use by the continuous- the network configuration used in this study
valued units in the neural network. The nor- for the daily model are 22–10–1 (i.e. 22 input
malizing factor can be taken as the maximum units, one hidden layer with 10 units, and 1
absolute value of each variable in the observed output unit), and 18–9-1 for the weekly data
period of 2 years. An important issue here is model.
the treatment of outliers and seasonal effects. For a network with two hidden layers, there
In order to enhance the predictive and are no specific rules of thumb with respect to
explanatory power of the model, further infor- the number of units in the first and second
mation can be provided for the neural network. layer or their ratio. In this study, the 22–15–5-
This can be achieved by adding auxiliary units 1 and 22–10–3-1 configurations are chosen for
to the neural network model. An additional the daily model, and 18–9-3–1 for the weekly
indicator for the long weekend (Saturday and model. The training and forecasting errors are
Sunday) is provided in the daily model because measured in terms of a goodness of fit R2
the sales pattern shows peaks at weekend. (where R2 is calculated as the average fore-
casting error divided by the variance of the
forecast data). The results are shown in Table 2.
Backpropagation Results for Daily and
Figure 5 shows the prediction performance of
Weekly Data Models
a 22–10–1 network, using a data set correspond-
The configuration of the neural network is one ing to a time period of 3 months. The network
of the most important issues when using the is trained with a learning rate of 0.05 and
backpropagation method. There are a number 3000 iterations. The actual output (solid line)
of theoretical observations which show that a is compared with the neural network output
backpropagation neural network is an efficient (dotted line).
function approximator, given sufficiently many For the 18–9-1 network which is chosen for
hidden nodes with sigmoid output functions the weekly data, the choice of a cut-off point
(Funahashi, 1989; Hornik et al., 1989; Lippmann, of training is not straightforward. In contrast
1987). However, these theorems do not specify to the daily data, the forecasting error for the
the optimal or minimal number of hidden units weekly data is not greater than two times the
needed to reach a desired performance. training error for a relatively large number of
With regard to the number of units and training cycles (e.g. 1400 cycles for a 18–9-1
weights, Baum and Haussler (1989) suggested network). Table 3 shows the best generalization
that the number of weights for a satisfactory results for the 18–9-1 network which is trained
network performance should be less than one up to 10,000 iterations at a learning rate of 0.05,
tenth of the number of training patterns. An using the data from different stores.
obvious shortcoming of this rule is that it is
very restrictive for limited data sets and large Counterpropagation Results for Daily and
networks. For an oversized network (i.e. the Weekly Data Models
number of weights is of the order of the num- For the counterpropagation network, the suit-
ability of its use depends on the relationships
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 261

Table 2. The best forecasting results of a network (22-10-1) using daily data for different stores

Stores 1 2 3 4 5 6

Testing error 0.018188 0.015776 0.010636 0.016885 0.013949 0.013312

Training error 0.010754 0.004215 0.013811 0.004370 0.004502 0.004400
Iteration 200 575 150 350 500 475

Figure 5 Three months’ daily prediction for store 1 using a 22-10-1 network

Table 3. The best forecasting results of a network (18-9-1) using weekly data for different stores

Stores 1 2 3 4 5 6

Testing error 0.009827 0.011415 0.007105 0.009750 0.012436 0.010038

Training error 0.008958 0.006300 0.009849 0.005192 0.005790 0.006409
Iteration 425 500 350 500 575 550

between the inputs and outputs. If the points then the network will perform well, as the
in the input space that are close in Euclidean initial clustering in the Kohonen layer will be
distance have widely varying outputs, the net- easier.
work will not perform well. In that case, a For both the daily and weekly data models,
network may require a large number of the counterpropagation network is trained with
Kohonen nodes in order to adequately rep- 100, 500, and 1000 iterations using the following
resent the problem. On the other hand, if the parameters: 1 to 60 nodes in the Kohonen layer,
input data are from a homogeneous source, the learning rate of the Kohonen layer is set to
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

262 H.-L. POH ET AL.

Figure 6 Forecasting error for the different counterpropagation daily models (data of store 1)

0.05, and the learning rate of the delta layer is

set to 0.1. Figures 6 and 7 present the fore-
casting results in terms of the average error.

Figure 7 Forecasting error for the different counterpropagation daily models (data of store 1)
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 263

Results for the Daily Data Model period of 6 months. The network is trained
Given the results obtained for the counter- with 100 iterations, using 20 nodes in the
propagation network using the daily data sets, Kohonen layer. Compared with the actual tar-
the major indications are that there is no sig- get (solid line), it can be seen that the neural
nificant difference in the forecasting results for network output(dotted line) closely follows the
a different number of iterations. In addition, actual target values.
the following observations are made.
The best forecasting result is obtained using
1000 iterations and 40 nodes in the Kohonen SENSITIVITY ANALYSIS
layer (data of store 1), with a generalization
error being 0.019806. When compared with the A good and understandable data model can be
results obtained using backpropagation, the provided by reducing the number of variables
forecasting results are not any better. This may through a weight sensitivity analysis. The
be attributed to the fact that changing a single impact of each independent variables on the
bit in the weekend indicator (var2), causes the dependent variables can also be found from
output to change significantly, and this violates the sensitivity analysis. The analysis of weights
the assumption for using counterpropagation. can be accomplished using the following
three methods.
Results for the Weekly Data Model The first method is the Equation Method. For
For the counterpropagation network using the a feedforward neural network with one hidden
weekly data sets, the forecasting results for layer and one output unit, the influence of each
data of store 1 are shown in Figure 7. As in input variable on the output can be calculated
the daily data model, these results are not any from the equation:
better than the forecasting results obtained
using backpropagation and there is no signifi- 冘 O(1 − O)w v (1 − v ) w
k
2 2
k1 k
2
k
1
ik ∀i (6)
cant difference in the forecasting results
obtained using different number of iterations. where O = the value of the output node;*
In addition, the following points should be w2k1 = the outgoing weight of the kth node in
emphasized. the hidden (second) layer
The best result is obtained using 20 nodes in v2k = the output value of the kth node in the
the Kohonen layer for data of store 1 after 500 hidden (second) layer
iterations, with the generalization error being w1ik = the connection weights between the ith
0.009042. Increasing the number of the Kohonen node of the input (first) layer and the kth node
nodes over 30 will generate very poor gen- in the hidden layer.
eralization results. This is contrary to the Using the Equation Method, there will be n
expected behavior as more Kohonen nodes readings for n input variables for each input
should give better performance. The reason row into the network. If there are r input rows,
may be that some information is filtered out in there will be r readings for each of the n input
the Kohonen layer due to the limited number variables. All the r readings for each input
of forecasting data and training data, while the variable are subsequently plotted to obtain its
properties of forecasting data are different from mean influence, In, on the output. These Ins
those in the training data. A counterpropa- indicate the relative influences each input vari-
gation network is more suitable for the weekly able could have on the output variable: the
data model because the fluctuations of the greater the value, the higher the influence.
dependent variables at the micro level (daily The second method is the Weight Magnitude
data model) are not reflected in the weekly Analysis Method. The connecting weights
data model.
Figure 8 shows the prediction performance of *The notation wabc here can be read as the weight
a counterpropagation network, using the from the bth node in the ath layer to the cth node
weekly forecasting set corresponding to a time in the next layer.
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

264 H.-L. POH ET AL.

Figure 8 Six months’ weekly prediction for store 1 using a counterpropagation network

between the input and the hidden nodes are

observed. The rationale for this method is that
Ii = 冘 max w (w )
k
0
ik

All i,k
0
ik
(7)

variables with higher connecting weights

between the input and output nodes will have where the notation is as same as in equation (6).
a greater influence on the output node results. The third method is the Variable Perturbation
For each input node, the sum of its output Method and tests the influence of the inputs on
weight magnitudes from each of the hidden the output. The method adjusts the input
layer nodes is the relative influence of that values of one variable while keeping all the
hidden node on the output. It is done for other variables untouched in this approach.
all input nodes. To find the sum of weight These changes could take the form of In ⇒ In
magnitudes from each input node, weight mag- + ␦ or In ⇒ In * ␦, where I is the input variable
nitudes of each of the input nodes are first to be adjusted and ␦ is the change introduced
divided by the largest connecting weight mag- into I. The corresponding responses of the out-
nitude between the input and the hidden layer. put against each change in the input variable
This is called normalization. The normalization are noted. The node whose changes affect the
process is a necessary step whereby the weights output most is the one considered most influ-
are adjusted in terms of the largest weight ential relative to the rest.
magnitude. The weight magnitudes from each Sensitivity analysis of a suitably trained
input node to the nodes in the hidden layer neural network can determine a set of input
are subsequently summed and ranked in a variables which have greater influence on the
descending order. The rank is an indication of output variable. In the experiment, sensitivity
the influence an input node has on the output analysis for different variables in the weekly
node relative to the rest. The rank formula is model is performed. These variables are per-
as follows: turbed for different ranges of their values in
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 265

the forecasting set in order to obtain the corre- nonlinear characteristic of the problem which
sponding changes in the output variable, i.e. is captured by the neural network.
var19 (Sales Volume). The changes of the output The results of var8 are shown in Table 5.
variable can then be observed for the direction From here we can find that var8 is negatively
of change and the degree of reaction, and then correlated to var19 (Sales Volume) and has a
compared to changes obtained by the linear significant negative impact on dynamics of total
regression model. It should be noted that this sales. A sensitivity analysis of var12 finds that
is a weak sensitivity measure because of the it also has a less significant impact on sales.
possibility of nonlinear interactions. Changes to In the second experiment, two pairs of vari-
two or more inputs in pair can have a different ables (var8 and var12, and var8 and var13) in the
effect from that of change to one input alone. weekly model are perturbed simultaneously to
However, sensitivity analysis can also be done observe the sales volume change. In other
for a combination of input variables to examine words, the interaction among these variables is
the interactions among the variables. tested. For both pairs of variables, the direction
For this experiment, the 18–9-1 network is of change is the same as in the case of a
first trained with weekly data sets of store 1 sensitivity analysis of variable var8, which
using backpropagation, with a learning rate set dominates the change in the output variable.
to 0.05. The network with 900 training cycles
is chosen.
Sensitivity analysis of variables var8 (Number FINDINGS AND DISCUSSION
of Advertisements for Cosmetics Items), var12
(Newspaper 1), var13 (Newspaper 2), and var18
Technical Observation
(Cost of Advertising) in the weekly model is
conducted in our first experiment. The results In this study, the forecasting problem is limited
of var13 are shown in Table 4. The values of to an explicative approach because the auto-
var13 are discrete, and they are either 0.00, 0.25, regressive method is not observed. However,
0.50, 0.75 or 1.00. These values are then per- the primary concern is the use of a neural
turbed by 0.25 or −0.25 except for the extreme network to capture the mapping between the
values (0 and 1) which are perturbed only in dynamics of input variables and the fluctu-
one direction. In the linear model, the coef- ations of the output variable. Such a model is
ficient for the variable var13 is 0.172368 with a able to yield a good out-of-sample performance
standard error 0.104700 (the t-test value is given the short-time prediction horizon.
1.646). Therefore, the sales volume change is The backpropagation network is suitable for
0.043092 if the change of the variable var13 prediction in both the daily and the weekly
is 0.25. data models. Training for the daily model is a
The results show that the neural network more difficult task but the prediction perform-
reacts similarly to the linear model with respect ance in a short-time period is acceptable. A
to the direction of change and the degree of counterpropagation network is not suitable for
reaction. In addition, the change of sales vol- the daily model as explained above but for a
ume is sensitive to different values or ranges weekly model it gives good predictions (see
of the perturbed input variable. This is the Figure 8). This may point to the existence of

Table 4 Sensitivity analysis of var13 at values 0.00, 0.25, and 0.50 for a perturbation value of ±0.25

var13 0.00 0.25 0.50

⌬ Input +0.25 −0.25 +0.25 −0.25 +0.25 −0.25

⌬ Sales +0.0425 N.A. 0.0464 −0.0436 0.0485 −0.045

 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

266 H.-L. POH ET AL.

Table 5 Sensitivity analysis of var8 at values 0.00, 0.25, and 0.50 for a perturbation value of ±0.25

var8 0.00 0.25 0.50

⌬Input +0.25 −0.25 +0.25 −0.25 +0.25 −0.25

⌬Sales −0.0664 N.A. −0.0577 +0.0628 −0.0474 +0.0552

higher degrees of correlations of input variables ent product categories and different periods
in the weekly model which is captured by of time.
the counterpropagation network. However, this
type of neural network cannot be used for
sensitivity analysis as opposed to the back- CONCLUSION AND FUTURE RESEARCH
propagation network.
Another point which should be noted is that The results from this study indicate that a neu-
the validity of neural network prediction per- ral network is able to capture the nonlinear
formance assumes that the dynamics of input relationships in a causal model even though
variables is similar to that of the recent past, there is no explicit structures of the domain
especially if insufficient data sets are available field. The model developed thus can be used
for training. in assisting in the short-term forecasting of a
variable of interest. In addition, it can be used
to conduct a sensitivity analysis, which is a
Marketing Implications
useful procedure for analyzing possible stra-
By conducting a sensitivity analysis on a suit- tegies with respect to the observed variables.
ably trained neural network model, the follow- The results indicate that backpropagation is
ing findings of this study can be translated into an efficient method for studying the relation-
practical applications in marketing manage- ship between input and output variables. It
ment. First, the results indicate that continuous could be a useful tool for planning and allocat-
and heavy promotion of certain items (e.g. ing advertising expenses. However, the fore-
cosmetics) is either not correlated or may have casting accuracy of the dependent variables
negative effects on the overall sales. Second, depends on different data models and size of
special events (shows, statements) and the data sets. In the case of the daily model, a
intermittent advertising and promotion of some suitable network architecture is not easy to
product categories are positively correlated and identify. For the weekly data, the choice of a
have a significant positive impact on the overall suitable network architecture is easier and can
sales. Third, the results indicate that advertising be determined by looking at the best fore-
in the leading local language (Chinese) news- casting result.
paper has a more significant effect on the total Speed is an advantage when using a counter-
sales than through other channels. On the other propagation network as compared to backpro-
hand, featuring advertisements in the English pagation. However, for the daily data model,
newspaper that has the largest circulation, counterpropagation does not yield better results
which carried the majority of advertisements, than backpropagation. A possible explanation
is found to have no significant effect on the is that some points in the input space that are
overall sales. close in Euclidean distance have widely varying
These results have been verified by the esti- outputs, causing neural networks to perform
mation of different regression methods. How- poorly. In particular, changing one bit in the
ever, it is the intention of this study to gen- input corresponding to a weekend indicator in
eralize the results. It is quite possible that the daily model induces significant change in
different advertising retention exists for differ- the output. This will cause unsatisfactory per-
 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

NEURAL NETWORKS FOR ADVERTISING AND PROMOTION IMPACT 267

formance of Kohonen layer, or require a large Funahashi, K., ‘On the approximate realization of
number of Kohonen nodes, which is imprac- continuous mappings by neural networks’, Neural
Networks, 2, 1989, 183–192.
tical. For the weekly model, the counterpropa- Hecht-Nielsen, R., ‘Counterpropagation networks’,
gation network produces better results, but IEEE Proc. of the International Conference on Neural
again, not better than the corresponding results Networks, Vol. 2, 19–32, 1987.
of the backpropagation method. Hill, T. and Ramus, W., ‘Neural network models for
Practitioners may deal with similar problems intelligent support of managerial decision making’,
in retail markets using market findings in this Decision Support Systems, 11, 1994.
Hornik, K., Stinchcombe, M. and White, H., ‘Multi-
study. A decision support system could be built layer feedforward networks are universal approxi-
using a neural network as a black box to assist mators’, Neural Networks, 2, 1989, 359–366.
in human decision making. Combining with Jas̆ic, T., A Neural Network Approach to Retail Market-
statistic models and linking to a variety of data ing Analysis and its Implications for Decision Support,
sources, an expert system can also be built. Masters thesis, National University of Singapore,
1993.
Finally, further research should lay the
Lippmann, R.P., ‘An introduction to computing with
ground for the incorporation of the neural net- neural nets’, IEEE ASSP Magazine, April, 1987,
work models into a broader framework of a 4–22.
marketing decision support system (MDSS). Poh, H.L., A Neural Network Approach for Marketing
The neural network model in such a system Strategies Research and Decision Support, PhD thesis,
should be able to deal with different problems Stanford University, March 1991.
Poh, H.L. and Jas̆ic, T., ‘Forecasting and analysis of
(e.g. sales analysis, market share evaluation, marketing data using neural networks: a case of
assessment of the impact of different marketing advertising and promotion impact’, Proceedings of
strategies etc.) using extrapolative (time series) the 11th Conference on Artificial Intelligence for Appli-
as well as explicative (causal) analyses, or a cations, Los Angeles, 1995, pp. 224–230.
combination of both. Raju, J.S., ‘The effect of price promotions on varia-
bility in product category sales’, Marketing Science,
11, 3, 1992, 207–220.
Acknowledgments Rumelhart, D.E., Hinton, G.E. and Williams, R.J.,
‘Learning internal representations by error propa-
The authors are grateful to the anonymous ref- gation’, in Rumelhart, D.E. and McClelland, J.L.
erees whose insightful comments enabled us (eds), Parallel Distributed Processing (PDP): Explo-
to make significant improvements. We thank rations in the microstructure of cognition, Vol. 1, MIT
Press, Cambridge, MA, 1986.
Professor Chew Lim Tan for his invaluable
Weigend, A.S., Rumelhart, D.E. and Huberman, B.A.,
advice and helpful comments. Thanks also go ‘Predicting the future: a connectionist approach’,
to Robyn E. Wilson for her precious comments International Journal of Neural Systems, 1, 3, 1990,
on the technical writing. 193–209.
Werbos, P.J., Beyond Regression: New Tools for Predic-
tion and Analysis in the Behavioral Sciences, PhD
References Thesis, Harvard University, November, 1974.
Aaker, D.A. and Myers, J.G., Advertising Management, White, H., ‘Connectionist nonparametric regression:
Prentice Hall, Englewood Cliffs, NJ, 1982. multilayer feedforward neural networks can learn
Baum, E.B. and Haussler, D., ‘What size net gives arbitrary mappings’, Neural Networks, 3, 1990.
valid generalization?’ Neural Computation, 1, 1989, Wong, B.K, Bodnovich, T.A. and Selvi, Y., ‘A bibli-
151. ography of neural metwork business applications
Dutta, S., Shekhar, S. and Wong, W.Y., ‘Decision research: 1988-September 1994’, Expert Systems,
support in non-conservative domains: generalis- August 1995.
ation with neural networks’, Decision Support Sys- Yao, J.T., Li, Y.L. and Tan, C.L., ‘Forecasting the
tems, 11, 1994. exchange rates of CHF vs USD using neural net-
Doyle, P. and Saunders, J., ‘Multiproduct advertising works’, Journal of Computational Intelligence in Fin-
budgeting’, Marketing Science, 9, 2, 1990, 97–113. ance, 5, No. 2, 1997, 7–13.

 1998 John Wiley & Sons, Ltd. Int. J. Intell. Sys. Acc. Fin. Mgmt. 7, 253–268, (1998)

268 H.-L. POH ET AL.