Sie sind auf Seite 1von 19

Applications of Operational Research in Marketing

Author(s): Alan Mercer


Source: OR, Vol. 17, No. 3 (Sep., 1966), pp. 235-252
Published by: Operational Research Society
Stable URL: http://www.jstor.org/stable/3006556
Accessed: 08-12-2016 13:54 UTC
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/3006556?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Operational Research Society, Palgrave Macmillan Journals are collaborating with JSTOR to
digitize, preserve and extend access to OR

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Applications of Operational Research


in Marketing
ALAN MERCER
University of Lancaster

The pricing and selling of industrial products are considered in sections on decision
theory, bidding and personal selling. It is suggested that the marketing of consumer durables and semi-durables can be studied by means of models which use
aggregate data but that consumer panel data are needed for modelling consumer
expendable markets. Finally, some problems which might be suboptimized are
indicated.

INTRODUCTION

OPERATIONAL RESEARCH is concerned with creating models, in the form of

mathematical relationships, which will enable management to make more effective decisions. Consequently the models must be capable of forecasting the outcome of any decision and of comparing alternative decisions. Thus operational
research differs from market research as traditionally practised. The latter can
realistically be regarded as the application of statistical analyses to marketing

data to produce statistics, which can be used either by the marketing manager
in taking a decision or in the model to forecast and compare the results of decisions. This explains why there are few realistic operational research models in
marketing in spite of the enormous amount of data which has been collected

and analysed. Indeed much of what finds its way into the operational research
literature could be described more appropriately as sophisticated market research.
Obviously, in future the two cannot exist separately as each has much to offer to
the other and they must be regarded as the two halves of marketing research.
Marketing is a dynamic activity complicated by a high level of risk, which can
be attributed to two distinct causes. First, there is the uncertainty of how the

customers will react to any change. Model building can be very difficult when
customer behaviour depends on intangibles, such as product performance and

convenience, which are very difficult to quantify. Second, the effects of any
company's marketing policy will be influenced by the actions of its competitors.

Consequently any model must include the competitors' activities for which data
must necessarily be obtained and this is frequently far from easy. Indeed in an
environment where the objective of some of the participants is to ensure that
there will be no equilibrium, the data are changing continuously so that models

may be developed long after they could make a useful contribution to the deci-

sion taking. Any model must be either dynamic or capable of being used with
more recent data than those used in the model's construction.

17

235

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3


As far as the analyst is concerned, marketing may be broadly categorized as
industrial products, consumer durables and consumer expendables. For most
industrial products, the qualities of all competitors goods are comparable and
the purchases are made by well-qualified people who are unlikely to be greatly
swayed by advertising and promotion. Thus the main ingredients of the marketing mix are usually the product price and the selling effort of the salesman, who
is likely to stress the high level of his company's service. Advertising and in-store
promotions are important in the marketing of all consumer products but more
so for low-priced, frequently purchased expendables. Because of the higher cost
of a durable, the decision is less likely to be made on impulse and more thought
will be given to the purchase. Hence the pace of the marketing campaign for a
durable will be slower than that of an expendable and is likely to last for months
rather than weeks or even days. Thus the analyst must use data collected over a
longer period of time and this in turn affects the model building itself. Con-

sequently, models appropriate for one type of market may not be appropriate
in another. Whilst the overall plan for the following sections is to consider
industrial products first, followed by consumer durables and then expendables,
there is much in common and the appropriate markets are indicated where
necessary.

A decision on where to store products and how much to stock ought not to
be taken without considering what variety should be sold and this is clearly an

essential part of the marketing policy. In many organizations, the marketing


director is responsible for distribution and inventory control, although in other
companies the responsibility belongs to the transportation director, the commercial director or even the production director. However, this problem area is
deliberately excluded on the grounds that the published papers do not treat
distribution and inventory control as a directly competitive operation.

DECISION THEORY

In a rapidly changing market, decisions unlike any previously made must be


taken. A situation arises which demands that action be taken and a number of
alternatives are available. For example, suppose that a manufacturer is asked
by a chain of stores to produce a product with a private label for sale in their
stores. The decision which he makes might be any one of a number of alternatives which might include agreeing to the request, offering a price reduction on
the manufacturer's own brand label or ignoring the request. For each possible
decision there will be a number of alternative outcomes. These might include
the chain obtaining all its supplies or only the product from another manufacturer or the chain might accept the manufacturer's decision. If the former
happened, then the manufacturer to whom the request was originally made
might wish to take further action. Thus his second decision could be to retaliate
by attacking the second manufacturer's customers or he might decide not to
236

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


retaliate. If there were retaliation, then the outcomes of this decision might be

that the second manufacturer ceased supplying the chain of stores or there might
be a price war. The sequence of decisions and their resulting states are set out
in the form of a decision tree in Figure 1, from which it can be seen that decision

and outcome follow each other in an alternating sequence which must be terminated at some point in a fairly arbitrary manner. This is essential because a
payoff must be assigned to the end of each branch of the tree. In Figure 1, these

are denoted by Cil, i = 1, 2 and j = 1, ..., 6. For example, the payoff if ther

is a price war after the manufacturer has ignored the chain stores' request for a
private label might be minus one million pounds.

In addition to payoffs, the probabilities of the different outcomes must be

estimated, and these are denoted by pi, where the probabilities of the outcom
of a decision must sum to unity. Thus the expected payoff of any decision can
be determined and the best decision is that which gives the greatest payoff, which

then becomes associated with the outcome which necessitated a decision. For
example, the expected payoff from attacking the customers of the alternative

supplier after ignoring the chain stores' request is P25 C25 +P26 C26, where
P26 = 1 -P25* If this is less than C24, then the better decision is not to retaliate
and C24 is the payoff associated with the chain stores' obtaining all their supplies

from elsewhere. Then P14 C14 +P5 C21 is the expected payoff of ignoring the
chain stores' request and if this is greater than both C11 and the expected payoff
when a price reduction is offered, the request should be ignored. Thus the best
strategy for the manufacturer is to ignore the chain stores' request and if the
stores obtain their supplies from another manufacturer, there should be no
attempt to attack the new supplier's other customers.

The reaction of many analysts to the non-repetitive problem is that they are
unable to assist the decision-taker. Being accustomed to probabilities based on
frequency theory and cost functions estimated from historical data by sophisti-

cated statistical methods, the subjective probabilities and payoffs of the Bayesian
decision theory approach are unfamiliar and there is scepticism about its value.
However, many of the problems facing the marketing manager fall into the

above category and cannot be ignored. First, the very act of formalizing the
decision process will ensure that it is faced realistically and that no alternative

decisions or outcomes will be neglected. The decision-taking sequence may be


regarded as the first stage in the construction of a mathematical model. It will
indicate which data must be used in the process and therefore what must be

collected. Frequently this operation will take the form of detailed discussions

with the organization's executives and it is important that no answers are

accepted without careful scrutiny. Opinions on the same subject should be

sought from more than one source. It may be argued that management finds
it extremely difficult to determine probabilities to the final outcomes resulting
from actions in any situation. This is true partly because the problem is one of
assigning relative values to a large number of possible final outcomes. Typically
237

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3

coo

P-4~~~~~~

Cl

0~~~~~~~~~~~~0 C

oto

238

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


these will average 0 01 or less but the decision tree method overcomes this
difficulty. The probabilities of the outcomes to a decision are estimated for each
decision so that they will be relatively few in number. However, the penalties

and their probabilities are seldom precise and become less so as the sequence of
decisions is lengthened so that the shorter it is, the more accurate the results are

likely to be. With a problem in which data might be in error by an order of

magnitude, it is essential to perform sensitivity analyses. Thus attention will be


focused on the vital decisions and outcomes whilst others can be discarded.
If a computer is available, then many marketing problems yield very readily
to this form of analysis. Its application to a typical industrial marketing problem

has been described by Green.' He considers a firm, selling in an oligopolistic


market, which wishes to improve its performance. The decisions which it takes
are the amounts and the timings of possible price reductions. The outcomes are
essentially the sales volumes and profits which it achieves. The sales are represented probabilistically depending on the price reductions and the reactions of
the competitors. A company horizon is fixed so that the decision process is time
structured and the company's cumulative, compounded net profits are calculated
over that period. The firm's decisions are influenced by fears of future price
weaknesses if immediate price reductions are not made in order to encourage an
increase in the total size of the market. Moreover, the firm believes that unless
it decides to reduce prices, new competitors might enter the market and this
would lead to an oversupply, with consequent drastic reductions in profits, at
some future time.

BIDDING

Green's example of the use of decision theory represents pricing in a market


where each firm fixes its price for a period of time and its sales to all customers
are then made at that price. In many markets, this mechanism does not operate.
For example, sales might be made by putting in tenders for the right to supply
products over a period of one year. Then a contract is awarded to the firm which
quoted the lowest price.

Suppose that the name of the winning company and its bid are both known.
At the time that a firm made its own bid, it would have estimated the value of
the contract to itself. For example, if a company is selling plastic to firms of
boulders, the worth of a contract might be taken as the tonnage. Thus the

winning bids of all the companies can be plotted as a function of the analysing
company's estimate of the contract value. When this is done, it is found that the

average unit price for any given sized contract is a non-linear monotonically
decreasing function of the estimated worth. Moreover, the variation about the

average winning bid price decreases as the size of the contract increases. This
phenomenon is entirely as would be expected and is illustrated in Figure 2. The
average price decreases because selling organizations obviously prefer to make
large sales rather than small ones. Apart from ensuring a larger market share,
239

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3


there are also marketing, distribution and administrative economies of scale as
well as any production advantages which might accrue. The large purchasing
organizations clearly expect to receive a share of these economies and consequently exert pressure on the marketing companies to obtain lower prices.
The variation in the unit price of winning bids is equally explainable. If a contract

has a high value, then a marketing organization striving to win the contract will

not quote too high a price or it is certain to lose. On the other hand, if a
ridiculous price much below that of the competitors were quoted, then the

winning company would have taken an unnecessarily large decrease in total


profit. Hence the tight bidding and low variation in winning prices for large
contracts. For a small contract, a marketing organization is not going to become
unduly worried if its price is lower than necessary because the decrease in total

profit will only be small. Equally it is likely to adopt the attitude that if a high

price is quoted and wins, then the firm will make a good profit but if it loses,
then little total profit will have been missed. Hence loose prices, with the
consequent large unit price variation, are quoted for small contracts.

-IU

U )(

r
C~~~~ X
C

)X

.O

x
)<

0)%

x>KX
x X

X~- XXX_
Log (estimated worth)

FIG. 2. Winning bids.

Thus when a contract comes up for tender, the first step is to estimate the
worth. Then it may be argued that the distribution of winning bids for the

estimated worth should be determined. Essentially this would be done from the
data used in Figure 2 and the mean winning bid would be given by the curve
and the variance of the distribution would decrease as the mean decreased. If

the selling company wished to ensure that it has a 95 per cent chance of winning

the particular contract, the bid should be the value of the lower 5 per cent point
Alternatively, a company might wish to decide what prices to quote for a number

of different contracts for which it will be tendering and the company's objective
might be to maximize its profits subject to keeping its market share within
defined limits. Then since the probability gives the unit winning price as a
function of the worth, alias sales, the tendering prices may be obtained from the
240

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


solution of a non-linear mathematical programming computation. This is essen-

tially the approach used by Hanssmann and Rivett2 and may be reasonable if
a market is static or if the problem is the complementary one of purchasing in
a one-off situation, typified by the tendering for oil-drilling rights.
Unfortunately few markets are static and if one company were to reduce its
prices to the lower 5 per cent point of the winning price distribution, then other

companies would follow quickly and the lower 5 per cent point would rapidly
be transformed into the mean. Even if a company had computed the prices for
a sequence of bids in advance of the first tender in order to maximize its profits,
then the same result would occur if the calculations lead to the first few bids

being put in at low prices even if the intention were to put in the later bids at
higher prices. Thus the approach developed for a static market would have
disastrous results if used in a dynamic market, as it would rapidly take all the
profit out of the business.
The use of probabilities is often regarded as a form of sophistication and

many management decisions may be criticized because variability has been


completely ignored. Probabilities enable the irrational to be rationalized. Consider, for example, product use tests in which a customer might prefer product

A to B, B to C and C to A when presented with three paired comparison tests.


It may be said that the customer is irrational and the results meaningless but if

the choice process is regarded as probabilistic, then the results are compatible.
However, probabilities are used either because the process is not understood
completely or because it is too complicated to be explained simply. Essentially,
probabilities provide a way of making this unnecessary by integrating over the

complexities of the process so that it is only necessary to talk about outcomes


rather than their reasons. This is precisely what the use of the distribution of

the unit price of winning bids does and it means that the position of a particular

winning bid relative to the mean for all contracts of the same estimated worth is
determined purely by chance. It implies that the bidders are irrational but this
is obviously not the case. If a bid, whether successful or not, is made at a higher

than average price, then it is likely to be for some good reason, such as that
the customer's factory is a long way from the seller's factory with consequent
high distribution costs. If the bid is below average, it might be that the marketing
organization needs the sales in order to keep its factories in production. In other

words, there are usually very good reasons for whether the winning bid to a
contract is above or below average. Simply stated, these are the objectives of
the different companies and how successful they are in achieving them. Thus if

a company is selling in an oligopolistic market or if selling companies can be


grouped such that the market appears to be so, then the first essential in a

bidding study is to model each competitor's policy. This involves looking at the
winning bids of Figure 2 in terms of the companies who won the contracts, their
sales position at the time of the bid, their production capacities and locations

and the characteristics, including location, of the customers. For example, the
241

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3

objective of a selling company might be to maximize its profit subject to selling


not less than a given tonnage per annum. Faced with this policy, a company
would be selective in its marketing, which implies putting in lower bids to

companies where its distribution costs, etc., would be least. This stage of the
study is far from easy. Even if the competitors' policies were known, much of
the data for the model would be unknown. For example, it is likely that a
competitor's capacity will have to be estimated from the bidding data rather
than being known from market intelligence. In a real situation, the competitors'
strategies are unknown so that it is necessary to estimate the parameters of the
models and to compare many different models. The mathematical modelling of
a policy can be very complex if the firm's bid for any contract relates to what
it expects to happen at future tenders as well as reflecting declared contracts. It
is desirable to formulate policies solely in terms of past events but even so the
analysis may have to resort to simulation.
If the unit price of the winning bid is unknown the approach to the problem
remains unchanged. However, for contracts which the analysing company failed

to win, only an upper bound for the successful price is known, i.e. the company's
own bid. A different approach is needed if the number of bidders is too large
for the policy of each competitor to be considered. In this case, one possibility
is to consider the price of the tender to be given by a Markov process, in which
both the price of the previous tender made by the company and its market share
are used.

Pricing in industrial markets has been the subject of this section and it was

used to illustrate the ideas of decision theory. In the bidding situation, prices
are submitted simultaneously and the lowest wins the contract. In the example,
there is a fixed price for all customers at any particular time. In many markets,
the price is fixed by a process which might be termed bidding with matching. A
competitor's salesman calls upon a firm and offers a price which is lower than
that which the firm is currently paying. Before accepting the offer, the firm tells
the present supplier's salesman, who has the opportunity to revise his price.

Thus he might decide to match the competitor's offer or he might take the
calculated risk of losing the business by changing his price to something higher
than the competitor's offer. This pricing problem is also one in which competitors' policies should be evaluated, where possible. The supplier's salesman's
reaction will take account of his estimate of the value of the quality of his
product, the service of his company and his own selling effort. The greater they
are, the less likely it is that the price offer will be fully matched. Thus the final
model must include these factors.
PERSONAL SELLING

Personal selling enters into the marketing of almost all products,


consumer fields there is a complex interaction with the non-person
such as advertising, and point-of-sale promotion. A company must
242

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


firms its salesmen are to call upon and how a salesman's effort should be divided
between existing and prospective new accounts. The optimal call frequency and
length of call must be determined for each type of customer and salesmen's
remuneration schemes must be designed. There have been many papers pub-

lished on these problems and Ackoff and Green3 give 117 references. For
example, the sales volume per customer in unit time was plotted against the
number of calls per unit time made by the salesmen of one large organization.

Although it had been anticipated that for a given class of customer, the curve
would have an S-shape, as shown in Figure 3, no correlation was found. Thus
it was concluded that the calling rate was too high and the number of calls was

0)

a) +7

CL

0)3

0)

Number -of cols per unit time


FIG. 3. Relation of sales to number of calls.

reduced without any adverse effects on the company's business. In another case,

the sales were actually found to decrease when the calling rate increased to a
high level suggesting that customers began to resent the attentions of the salesmen. In spite of the large number of published papers, Ackoff and Green concluded that there has been little serious research in the application of quantitative
techniques to personal selling. The work has lacked operational significance and
generalizable research into the productivity of personal selling and its relationship to optimal total marketing models is practically non-existent. Since personal
selling effort is often the only major factor apart from price, the situation in
industrial marketing is probably satisfactory in that the relatively simple models
can be used. However, there remains much to be done in the more complex,

dynamic consumer market, where personal selling forms a smaller part.


ADVERTISING MODELS USING AGGREGATE DATA

In some marketing companies, the advertising budget is set by appropriating a

fixed proportion of the gross profit and the gross profit is determined as a given
proportion of the cost. Hence the advertising budget is a proportion of sales,
so that there is a linear relationship between the two. However, since it is the
sales which determine the advertising budget, any attempt to use the line to
243

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3


forecast sales for a particular advertising expenditure will be completely
fallacious.

Kuehn and Rohloff4 have discussed a proposed model, in which the ratio of

the sales for a region in one year to those in the previous year were plotted for
different sales regions of America as a function of the ratio of advertising
expenditure in the region for the year to the sales in the previous year. A line

fitted the data extremely well so that it might be concluded that the line could
be used to allocate advertising expenditure to the sales regions for the following
year. However, the sales regions were of different population sizes and the
marketing management had related the advertising budgets to such measures

when making the appropriations. But since sales are related to the size of the
population, the advertising budgets had been implicitly related to the sales so
that the cause-effect relationship was reversed. Moreover, each pair of variables
had a common denominator in the form of the previous year's sales for a region
and this produced a spuriously high correlation.
The two previous examples of fallacious arguments relate to static models.

An example of a dynamic model is expressed by the dynamic difference formula,


which states that a brand's market share will increase in direct proportion to

the difference between its market share and its share of the total market advertising. This is expressed mathematically by the first-order Markov model:

St-St-, = K(at-St-,),
where St = market share in tth period

and at:= brand's share of total market advertising expenditure in tth period.
Now if at were constant, then the model gives an exponential growth or decline
of sales and indeed the starting point of many models is to assume an exponen-

tial change. Thus if it is reasonable to assume this exponential change, then the
dynamic difference formula is likely to fit well but the advertising share adds
little to the prediction of sales. A further objection to this formula is that it
implies that K will be the same constant for all brands.

The inference to be drawn from the preceding three examples of this section
is that it is ridiculously easy to confuse cause and effect when constructing
marketing models with historical, aggregate data. Because of this, it has been

argued that marketing models should only use data from experiments or data
relating to individual purchasers or households. Carrying out experiments

removes much of the ill-conditioning inherent in historical, aggregate data but


they are expensive when carried out at the market level. Certainly the ill-

conditioning makes the linear multiple regression approach, of which the three

preceding examples are all simple cases, of doubtful value. Even if there is no
ambiguity in the interpretation of the resulting relationships, the predictions
are usually insensitive to large changes in the set of coefficients. However, these
difficulties can be overcome by constructing models, which are capable of explicit
244

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

A fan Mercer - Operational Research in Marketing


interpretation rather than fitting a response surface to large quantities of data.
Clearly in a rapidly changing market, typified by many of the low cost, frequently
purchased consumer expendables, a realistic model will be complicated and need

structuring so finely that historical, aggregate data would both be too coarse
and average over too long a period. Thus in such a market, it is necessary to use
consumer panel data.

Marketing is essentially a two-stage process and the first stage is to get th


product into the shops. Thus the dependent variable will be a measure of the
distribution and the most important independent variables will be the personal
selling effort of the manufacturer's representatives, the inducements such as
large profit margins to the retailers and the product sales in previous periods.
Past sales are important because if a line is selling well, then more outlets will

wish to retail it and those who already sell the line are going to carry larger
stocks and probably therefore give the product better display. The second stage
is to persuade the consumer to buy the product from the retailer. In very recent
research not yet published, Emshoff and Mercer have considered the sales of a
particular product to the consumer to be primarily dependent on price, distribution and the advertising of all companies in the market. The advertising expenditure in a given period of time was assumed to still have some, although reduced
effect in later time periods and a non-linear transformation was made. The form
of this was based on market research studies although the parameters were
estimated from the model itself. The reason for this transformation is that
increasing the advertising effort by a certain proportion does not mean that its
effectiveness will increase by the same proportion. Many alternative but plausible
market mechanisms were modelled and the best was also the most realistic in
that advertising of all competitors contributed to the customers' awareness. The
conversion into sales was allowed to differ between companies and the advertising to sales conversion ratios were estimated from the data. In some of the
models, press and television advertising were considered separately. The value
of tied outlets was considered and the increased sales of a new product were
assumed to be made to customers who heard about it either from advertisements
or from retailers or from friends who had already bought the product. In the
final analysis, the models for the two stages must be brought together to formulate operational rules. Then it might be thought that one composite model
should be created by substituting the distribution function of the first model in
the second model. If this is done, then the sales in one period are a function of
those in the previous period and, of course, many models start with this assumption and ignore distribution entirely. However, whilst such a model might give
a good forecast of future sales for different strategies, it fails to convey the

feeling of why the market reacts in a particular way. Indeed distribution is a


marketing statistic which the decision taker can understand and measure and
should therefore not be eliminated from the model. The two-stage model clearly
illustrates the dynamic nature of the marketing process and the answers to such
245

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3


questions as to whether sales are dependent on distribution or vice versa are
immediately answerable.

Generally speaking, the published practical papers relating to model building


with historical, aggregate data have considered the effect on sales of varying only
one of the constituents of the promotional mix and the models have been static.

Thus at best they may be regarded as giving some indication of the likely
functional form for part of the total marketing model.

CONSUMER BUYING MODELS USING PANEL DATA


For marketing models of low cost, frequently purchased items, the dynamic

nature of most markets dictates that the data should be closely tied to the time

and point of the sale. One source of such data is the consumer panel, which is
usually a sample of housewives who are demographically and geographically

representative of the total population. These housewives record each purchase


of the relevant products by brand, size, variety (colour, flavour, etc.), price, date
of purchase, place of purchase, together with any details of price reductions or
other promotional offers.

Consumer panel data are obviously a rich source of information and they
have been subjected to much difficult analysis in order to try to express pur-

chases in mathematical terms. The distribution of the total quantity of a brand


bought in a given time period and the chance that a customer's next purchase
will be of a particular brand are two examples of the type of study to which a
lot of effort has been devoted.

The work of Herniter5 is typical of much in this field and examples of the
kind of results are given in Table 1. The variable 0 in the purchase sequence

signifies that some other brand was purchased and the variable 1 indicates that
the brand in question was bought. Thus the sequence 011 means that the brand
being studied was bought on the last two occasions and some other brand was
bought on the previous occasion and then the probability that the brand was
bought at the next purchase after this sequence was observed to be 0-64. The
binomial model assumes that each customer has a given set of probabilities of

buying the different brands and that there is a prior distribution of this probability set over the population of customers.

Each purchase is independent of previous purchases and the probabilities do

not change with time. Thus the same probability of 0 37 is associated with the
three sequences 001, 010 and 100, reflecting the fact that predictions depend only
on the number of l's in the sequence rather than their position and this is

clearly incorrect. In the Markov models, the probability of purchase depends

only on whether or not the brand was purchased on the previous occasion and
purchase transition matrix does not change with time. In the simple case, this
transition matrix is assumed to be the same for all customers, whereas in the

Bayesian case a symmetrical 2 x 2 matrix is assumed and the transition probability has a distribution over the population. The quantitative fits to the data

246

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


are not good and the simple Markov model gives poor qualitative results because
all the probabilities are 0-81 if the sequence ends with a 1 and 0-21 if it finishes

with a 0. In both Bayesian models, the prior distribution was assumed to be a


beta distribution, and the good fits when only one purchase in the sequence is
defined reflect the fact that these were used for the parameter estimation.
TABLE 1. PREDICTIONS OF NEXT PURCHASE PROBABILITIES

Basis of next purchase probability after given sequence


Historical

purchase Binomial Simple Bayesian Linear

sequence Empirical model Markov Markov learning


model model model

054

054

053

050

050

0 021 021 021 0-20 0-23


1 081 082 081 080 077
00 015 0 13 0-21
01 056 051 081
10 043 051 021
11 086 089 081

0-14 015
056 051
044 049
086 085

000 011 010 021 011 012


001 0 52 0-37 0 81 0 66 0 40
010 036 037 021 057 037
011 064 065 081 066 065
100 0 38 0 37 0 21 0-34 0-35
101 064 065 081 043 063
110 049 065 021 034 060
111 089 092 081 089 088

Obviously the mathematic


reflect reality and the reas
realistic, then important f

promotional activities suc


can be improved at the e
enormously

but

even

then

of all marketing manage


products, so that it is nai

can ultimately serve any u


of these models are causal. Management cannot use them to take decisions on
the size of its advertising budget or the number-of salesmen to employ. Clearly

the objective of the authors of papers on consumer models must have been to
first fit models to the purchase data and then to fit the parameters of them to
the causal factors. In other words, the models so far have been nothing more

than a means of summarizing panel data and must therefore be regarded as a


247

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3

sophisticated tool of market research. However, it seems that the best way of
using panel data is to simulate market behaviour in different, alternative environ
ments. This avoids much of the controversial data analysis, the conceptual
difficulties and the complex mathematics. Such a simulation would include consumer's attitudes, purchases and experiences of competing brands as they are
influenced by the products' characteristics, distribution, packaging, price, instore promotion and advertising. The market research which has been carried
out has been very useful in highlighting characteristics which must be explained
and included in the simulation models. More needs to be done, but the objective
of providing an operational tool for decision making must not be overlooked as
it appears to have been in the past.

One Markovian model using panel data which has proved useful in analysing
data for decision taking is the linear learning model which was first proposed by
Kuehn and Day.6 In this, the probability of buying the brand at the previous
purchase is the state variable rather than the brand bought at the last purchase.
Thus if the probability of purchasing the brand at the nth purchase is p(n) and
the brand is actually purchased on that occasion, then:

p(n+ 1) = (l-a)p(n)+bs+ a-b,


where a, b and s are constants. This expression is known as the gain operator
and if some other brand was purchased on the nth occasion, then the loss
operator

p(n?+1) (l-a)p(n)+bs

is used to estimate p(n + 1). This process is shown diagrammatically in Figure 4


It will be seen that the probability of buying a brand at the next purchase can
never exceed 1 - {b(l - s)/a} or be less than bs/a, which implies that no outcome
is ever certain. Whilst this assumption might be unrealistic for an individual

consumer, it will be reasonable if the model is considered to be an average for


a group of consumers. The predictions of the linear learning model based on the
steady-state distribution, which assumes that the process has been functioning
for sufficiently long for this to be reached, are given in Table 1 for the parameter
values a = 0-20, b = 0-02 and s = 050. Although these values were selected as
an example rather than to fit the data, the linear learning model is better than
the other three. However, the linear learning model is most useful in analysing
the effects of market deals such as price reductions, gift packs, etc., which are
dynamic but frequently short-term forms of promotion. To determine the effect
of a deal, it is necessary to know what would have happened if there had been
no deal. This can be estimated as the difference between what actually happened

and the predictions made by the learning model, aggregated for the whole
market. For such short-range predictions, the assumed linearity of learning may

well be unimportant. However, the model must not be used for comparing
promotions of different products but rather for comparing promotions of the
248

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


same product. The reason for this is apparent when it is seen that gain and loss
operators imply that the expected probabilities are given by:

E{p(n + l)} = (1 -b) E{p(n)} + bs,


so that s is the equilibrium brand share. Clearly the learning effect depends on
the time interval between purchases. Thus for heavy buyers, both a and b are

small and both the operators approach the diagonal. For light buyers, a and b
tend to one and the operators approach the horizontal. Obviously in using the

I / t ~~~~~~~~b (1-s)
Gain

operator

a-b

pin+I)

i j operator
bs{_

(n)

FIG. 4. The learning model purchase probabilities.

learning model, these differences in habit must be taken into account. Either the
model must be used for individual customers and the prediction for each used
to derive a market estimate or the customers must be grouped and group predictions used to estimate the overall market effect. In both cases, the marketing
organization must decide whether or not an average weighted with offtake is
used for the total market. Clearly there are also difficulties in determining operational measures of the success of a deal because some of the customers who were
attracted by the deal might not be converted to the brand and revert to a
competitor's product. Thus different conclusions might be reached if the sales
of the product over a 3-month period after the deal were used rather than the
sales during the deal. It is obvious that the parameters of the model should be
estimated from data for short periods before the deal. However, only experience
with particular products can determine how the probability at a purchase should
be estimated from the previous probability if the deal were taking place then
but the assumption was necessarily one of no dealing. One possible solution
would be to proceed as though there were no deal and another would be to use
the equation involving the expected probabilities. In all this, it is necessary to
watch for competitor's activities which might give rise to a misleading interpretation. The estimates of the effects of deals enable marketing management to
realistically compare alternative deals and should be invaluable in the construction of models containing all the major marketing factors. Collecting data for
249

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3


marketing models is very expensive and the use of the learning model, amongst
others, may enable companies to obtain short-term rewards whilst awaiting the
more long-term results.

SUBOPTIMIZATION PROBLEMS IN MARKETING

In "Personal Selling", it was suggested that whilst personal selling could be


studied in semi-isolation for industrial products, it is undesirable to do so for

consumer products because of the interactions with other marketing factors.

However, this should not imply that a study of personal selling alone in the
consumer field would not be rewarding if it were carried out carefully. One
reason for this is because it seems that the sales force cannot be reorganized
with the same frequency with which new promotional campaigns can be

mounted. Hence the personal selling cannot be truly optimal all the time.
Because it appears to be more difficult for a company to monitor the personal
selling of its competitors, this aspect seems to be. less competitive and hence
there does not seem to be the constant pressure for change. Suboptimizing the

personal selling of an organization might be reasonable if the solution were fairly


insensitive to other marketing changes.

The number of product lines and items per line must also influence the success
and profitability of any marketing organization, but it is seldom studied by the
analyst as part of the total marketing model or even on its own. The efficiency

of personal selling decreases if the product range is too large because the salesman is reduced to a passive role. However, if it is suggested that some lines or
items of a line should be dropped, which in most cases would also increase
production efficiency, the salesman will object because he fears that competitors
will acquire the business. Moreover, he believes that competitors would then be
in a strong position to successfully attack his sales of the continuing lines so

that he would eventually lose all his business. This highlights one of the two
difficulties in that any study of the product range must either predict the longterm effects of any changes in addition to the short-term results or the study

must be adaptive over a long time. In some markets, the reaction to any change
might come directly from the marketing organization's customers, but in others
the reaction will only reflect the behaviour of the ultimate consumers. Hence it
is necessary to know how this consumer will react to being unable to obtain a

particular size of a line previously produced by a given manufacturer. This

behaviour might vary at successive attempted purchases and will depend on the
consumers' size, product and brand loyalties. Thus it is obvious that the second
difficulty is that of obtaining adequate data even in slow-moving markets and
the best solution would appear to be to carry out market experiments.
One of the reasons that it is necessary to cut out old lines is that new products

are being introduced continuously. Indeed the survival of most companies


depends on their ability to introduce new products successfully. During the
time that a new product is being launched, the promotional expenditure is far
250

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Alan Mercer - Operational Research in Marketing


greater than the gross profit and a company only begins to show a profit on a
product when it is successfully established in the market. Thus it is important to

know at the earliest possible moment whether or not a launch is going to be

successful and what market share is going to be obtained. Two typical curves
for new product sales with time are given in Figure 5, where the sales for the
first four time periods are plotted. In one case, many of those who are attracted
to the product continue to buy, whereas in the other there is a smaller conversion
rate. This might make the difference between success and failure but if only the

sales were plotted as a function of time, no distinction could be made after four

0 1 High conversion
rote

x
x/ \ Low conversion

Time

FIG. 5. Sales of a new product.

time periods. Consequently it is desirable to consider individual consumers in


order to measure the penetration and the repeat purchasing rate. Even then the
repeat purchasing rate will fall from the peak achieved during the first few weeks
of the launch and it is clearly essential to know when this has reached its final
level. Lipstein7 equated the transition matrix at a purchase with that at the next
purchase by post multiplying it by a matrix and eigenvalue analyses were then
performed on the resulting set of matrices. If the maximum eigenvalue at a
purchase was close, to 1, then it was concluded that the market had become
stable and the brand share achieved at that time could be taken as the final
market share. The size of the maximum eigenvalue is a measure of the amount

of turbulence in the market and may also be used for deciding whether a
particular moment in time is opportune for introducing a new product. None
of this work on new product introductions is causal so that whilst marketing
management may decide whether or not to continue a particular promotion,
the question of how much money should be devoted to a launch and how it
should be allocated still remains to be answered.

In marketing models, advertising is usually measured by the size of the


appropriation but a given amount of money may be spent in many different
ways. This is the media scheduling problem where questions of the size,
frequency and advertising vehicle for the insertions must be resolved. It is
18

251

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Operational Research Quarterly Vol. 17 No. 3


interesting that whereas the analyst previously attempted to obtain an optimum
schedule using techniques of mathematical programming, the present approach
is to compare possible schedules by using simulation methods which incorporate
the subjective judgements of the media manager. The basic weakness in all this
work is that it is very difficult to measure the response function of a consumer
to advertising so that most schedules must be prepared without there being any

data on this aspect. Since the objective of all advertising must be to eventually
affect the product sales, this might be very serious and there is a need to demonstrate that extensive calculations using data which are expensive to acquire are
not worthless because of this weakness. Media scheduling is certainly an area
in which it is very reasonable to suboptimize in the sense of getting the best

value for a predetermined cost. However, the results will also depend on the
quality of the advertising material, so that it might be worth while spending
more or less on the number of truly alternative advertisements which are

generated and pretested. This is a problem which ought to be studied at the


same time as that of media scheduling, with a view to obtaining a better allocation between these two essential aspects of advertising.
ACKNOWLEDGEMENTS

This paper was read at the First Research Conference of the Department of Operational
Research of the University of Lancaster and I am grateful to those who contributed to
the discussion. I am also indebted to all those with whom I have talked about marketing
problems during the years in which my present views have developed.
REFERENCES

p. E. GREEN (1963) Bayesian decision theory in pricing strategy. J.


2 F. HANSSMANN and B. H. P. RIVETT (1959) Competitive bidding. O
3 R. L. ACKOFF and P. E. GREEN (1965) Status report: The role of personal selling in the
promotional mix. Report of Management Science Centre, Wharton School of Finance and
Commerce, University of Pennsylvania.
4 A. A. KUEHN and A. C. ROHLOFF (1964) Fitting models to aggregate data to measure
advertising and promotional effectiveness. Paper presented to the American Statistical
Association Meeting.
5 J. D. HERNITER (1965) Stochastic market models and the analysis of consumer panel data.

Paper presented to the Twenty-seventh National Meeting of the Operations Research


Society of America.

6 A. A. KUEHN and R. L. DAY (1 964) Probabilistic models of consumer buying behaviour.


J. Marketing, 28, 27-31.
7 B. LIPSTEIN (1965) Test marketing-A perturbation in the market place. Paper presented
to the Twenty-seventh National Meeting of the Operations Research Society of America.

252

This content downloaded from 14.139.246.18 on Thu, 08 Dec 2016 13:54:13 UTC
All use subject to http://about.jstor.org/terms

Das könnte Ihnen auch gefallen