You are on page 1of 10

A System of Promotional Models Author(s): Ambar G. Rao and Gary Lilien Source: Management Science, Vol. 19, No.

2, Application Series (Oct., 1972), pp. 152-160 Published by: INFORMS Stable URL: http://www.jstor.org/stable/2629329 Accessed: 14/12/2010 11:52
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=informs. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. 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.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.

http://www.jstor.org

we consider two kinds of promotions which were used extensively by oil companies in the sixties-competitive games and mass credit card mailouts. MV'onitoring sales and profit performance after the marketing plan has been implemented so that deviations from plan can be detected and appropriate action taken. on sales. 1972 Printed in U. The models were developed to be used with a computerized MIS for market planniilng anld sales forecasting and are validated usiilng actual sales data. 2. t New York University. we * Received August 1971. These particular types of promotions have declined in importance due to recent legislative changes. not stored. RAOt t AND GAARY LILIEN? A system of promotionial-effect models developed from certain behavioral assumptions about consumer btuyinig habits is described. 19. where the aggregated responses of millions of individuals to corporate strategies are measured. New York. New York. Such model building efforts have frequently accompanied the development of sales forecasting procedures which incorporate not only information gleaned from historic time series studies.SCIENCE MANAGEMENT Vol. Development of supporting plans once the marketing plan has been structured. Here the model building goal is to estimate the incremental effect of various marketing activities singly. 152 . The estimates could for instance be used as input for production scheduling and distribution models. A system of models that provides such estimates can be used for 1. Included are insights inito the cutstomer buying process that were revealed durinig the parameter estimationi and updating procedtures. an adaptive modelling capability is essential for long-term use of a system. ? Mobil Oil Corporation. but also the expected effects of planned corporate strategies and anticipated competitive actions. In as complex an area as marketing. 2. Development of good marketing plans.S. No. revised December 1971. October. also fulfills a vital model building requirement: it provides systematic feedback which can adjust model parameters ancdenrich the models themselves. In addition. and jointly. Parameter estimation procedures and methods for calcutlating the effect of simultaneous promotionis are discutssed. for example. A SYSTEM OF PROMOTIONAL MODELS* AMBAR G. but the model structures appear to be valid for most station-based promotions. at least. The value of such a data base without associated tools such as management science models to extract actionable information is questionable (see. M1anagementcan simulate and evaluate the effects of alternative strategies together with competitive counterstrategies and then select the "best" plan. This third function. This paper describes some of the efforts to model the effects of various kinds of promotions on sales. A. The authors were fortunate enough to participate in the development of a system for forecasting the gasoline sales of a major oil company. that of monitoring actual results and comparing them with plan. In particular. 3. t Work done while this author was at Mobil Oil Corporationi. Ackoff [l]). The same basic model struetture is shown to be applicable to several types of gasoline marketing promotionis anld to various nonigasoline promotioiis as well. Introduction Computerization of marketing information systems often leads to the retention of data that previously were not gathered or.

the approach appears to be applicable to products and services that are sold in franchised outlets. It was hypothesized that the incremental gain in sales due to a promotion depends on the following: . however. Acceptable fits to data were obtained. Thus the game was very similar to a lottery. although more precise estimates could. Sales gains could be achieved only by acquiring new customers-that is. The effects of games on sales were. not much a priori knowledge existed about customer heterogeneity. snack food chains.A SYSTEM OF PROMOTIONAL MODELS 153 indicate how the models can be adapted to reflect the effect on sales of price changes. reaching a peak in about four months. Preliminary analysis showed that the effects of a credit card mailout increased slowly. but the models were difficult to interpret and had poor predictive qualities. In essence. Periodically. Later on. Therefore. where each outlet carries the "products" made by a single corporation. M/lanv authors have used this approach to develop insightful models of customer behavior in different situations (see. The data that were maintained were monthly shipments to stations in various market areas. however. Since we were mainly interested in estimating expected changes in market share. we proceeded with an "expected value" type of approach. Model Concepts Initial efforts to explain sales changes due to promotions included a regression analysis approach with an assortment of lagged and transgenerated variables. in principle. after the game. we did not feel this was a significant omission. In suclh circumstanices. customers who normally purchased gasoliiie sold by other companies. however. This level was then maintained for a substantial length of time. although there were legal differences. Again. etc. ignoring differences anmongindividual customers. The hope was that customers would be drawn from competing stations that were not participating in game promotions and that. the promotions took on a more defensive aspect-they were instituted to avoid a loss of customers rather than to acquire new ones. Description of the Promotions and Data variants of the competitive game promotion were offered to customers during M\'Iany the sixties. "winners"of the game would be selected by a random drawing. be obtained for each market. tire stores. they were very similar: A customer would obtain a game form with a number or other identificationi from the promoting brand. In our case. some of these customers would be retained as regular purchasers. credit cards were sent to residents in certain neighborhoods which satisfied some demographic and socioeconomic criteria. and we decided to develop models from behavioral assumptions about individual consumer buying habits. models of individual response to marketing action can then be developed and aggregated to appropriately reflect such differences. [41). [31. It should be noted that unlike promotions for consumer-packaged goods. Hence the approach was dropped. for example. it was not expected that customers would buy more than their normal quantity. In the case of mailouts. Examples are banks. There is generally a lag of about one week between a shipment occurring and a sale. more immediate. it is often advanltageous to assume that the population is heterogeneous in terms of its behavior. of course. Shipment data were highly seasonal and in many markets exhibited a pronounced trend. a shift of customers to the mailout brand was the objective. Although these models were developed in the specific context of gasoline marketing.

The probability that a randomly chosen customer does not normally buy the promoting brand or brands. Assuming that G is constant for a given market at a given time. and could be allocated as an approximation proportional to the shares of market of the nonpromoting brands. The more attractive the promotion is. 'as Therefore. then we can make some observations about the exponents a and A.m)/ma. since it will mean less of a detour from his normal driving patterns. the required probability is an increasing function of (1 . (b) For a given promotion.B 1. This factor is termed the STRENGTH (S) of a promotion. the number of stations operated by the promoting brand(s) increases and the likelihood of finding one of these stations within a given radius of a fixed point will also increase. By the time a brand (or brands in the case of simultaneous promotions) has a 20-25 % share of the market. and S = K (a constant). that P = (1 . RAO AND GARY LILIEN 1. then VG = KG(1. for simplicity. the higher the likelihood that the customer will respond to it. In equation (1).154 AMBAR G. 1. Thus &R/&inshould be positive when m is close to zero. If we assume that the promoting brand(s) have a joint share of market equal to m. Consider a nonpromoting brand with share in0. Consequently. the loss by the nonpromoting brands is also V. and let VL be the loss to this brand. As m increases. the larger d can be and still have the customer respond to it. Then. This probability depends on two factors: (a) Consider a customer who has to travel an additional distance d to take advantage of a particular promotion. Since the sales increment due to a gasoline promotion arises out of purchases by new customers. The probability that a randomly chosen customer who does not buy the promoting brand(s) normally takes advantage of the promotion. V was defined as the volume gained by the promoting brand (s). If there are C customers in the market and Cg is denoted by G. the distance d will decrease. 0 < a < 1.1. the probability defined above can be regarded as the POTENTIAL (P) of a promotion. In particular. and should gradually decline to a value -y close to zero but positive when m is close to 1. Then VL = KGma(l -_r)O(mo/1 1 - in) (2 ) = KG MP MO(- A m)o . rather than by increased purchase quantities by existing customers. Thus the probability of finding a station with a promotion within a given distance is an increasing function of m. the smaller d is. where VGis the volume (sales) gain due to a game promotion. . we assert 0 < a < 1 and A -. the probability that a randomly chosen customer will respond to a promotion is given by P X R X S. aMal > O at in = 0+ and aml -y m -. the total gasoline sold in the market. This is because as m increases. almost everybody should be within reach of a station which is conducting a promotion.1. where g is the average quantity purchased by a randomly chosen customer during the promotional period. If this is indeed a reasonable form of the model. Formulation for a Game Promotion The next step is to assume mathematical forms for the factors P and R. Hence. R = ma. We shall call this probability the REACH (R) of a promotion. then the incremental sales gain to the promoting brands is given by (1) Mathematical V = P X R X S X G.m). the expected gain per customer in the market from a promotion is given by V* = P X R X S X g. Assume.m)o. 0 < a < 1. 2. 2.

for sake of generality. HM(t) = actual major price at t._ (3-1)ma(1-m) 2] o 0 which implies . . The strength of the draw of the price difference should be related to the magnitude of the change in that difference. and the gain is allocated in proportion to shares. Thus. and (5) PL = Kmt. a price differential change can be viewed as a promotion staged by the independent brands against the majors. the changes which will be considered here. A market-usual price difference between majors and independents can be established for most markets from historical price data. If one of the promoting brands has a share mo.(6). Note also that limm1m PL K. 0 < a < 1. the effect of an additional brand having a game on proportional loss is also small (loss rate levels out). Let IIM prevailing or market-usual major price. and their total market is giv~enby m. HI* -prevailing or market-usual independent price. K[Ma-l1 )M1a. It will be a reasonably tight bound if mois small. Price Changes The structure developed above can be modified to consider price changes. In most gasoline markets. changes within the group of major brands and within the group of independent brands are small relative to between-group price changes. represent the market share of independent brands (as a group). Thus. Then let = ( M(t) - 111(t)) - (M M HII) where 6 thus represents the change in price difference at time t in a given market. To use reasoning analogous to that for games. We have assumed in the above. so that the proportional loss is (3) PL = = VL/mnoG Km(n(1 It is reasonable here to assume that aPLiamM-* 0 as in -I 1. 2 1. Thus. so K can be interpreted as an upper bound of the proportional loss that a company can incur by not participating in the promotion. In other words. equation (1) can be rewritten as (4) V = KGm'(1 - In). if the -l number of nongame playing brands is small (mn 1). Reach MI Strength . when In is near 1. the proportional gain by a brand with share mo when a total share of m is being promoted is (6) PG = KGma (1 m) (nojmn)(1/moG) = Kml1(1I ). HI(t) = actual independent price at t. that there are several brands simultaneously using similar promotions. we have: Potential = (1 -m).A SYSTEM OF PROMOTIONAL MODELS 155 The expected sales volume of the brand when there are no promotions in the market is Gmo. letting m.

If another brand. and suppose a proportion cj of its sales are ePedit sales. The basic problem is that the potential of a particular customer (and also the strength of the promotion to him) debe pends on whether he already purchases gasoline on credit or not. Interaction Effects Thus far. whether the independents increase or decrease their price relative to the majors. is given by (7) VI = f (8)I(1 -I)- If f (6) can be approximated. EN_] Now let the strength of the promotion be K1 for a customer of the first and K2 for a customer of the second of the tw-otypes discussed above. RAO AND GARY LILIEN wYhere (8) is some unknow. Then the volume gained by the promoting brand. the probability given purchase occasion and purchases it on credit is 111jCj.cj). then (7) simplifies to (8) VI = GKo&iiI'(1 - ). If more than one promotion occurs.Cj-)]. and obtain a formulation quite similar to the previous ones. We therefore face the problem of modelling interaction effects. Mathematical Formulation for Mailouts MA4odelling effect of a credit card mailout is more complex than the two cases the considered previously.jcj. where p is the car population in the market. (10) VC = G (/I + /p)miio"[K1 E jFo m{1jcj K2 EjFO (. Let n be the number of cards mailed out. the joint effect of the two promotions would not. however. is given by (9) Vc = m17o0f(n)G[Ki Ejxo mjcj + K2 EjZO Mi (1 - Ci)].. has a simultaneous mailout of L cards. The probability that a randomly chosen customer buys Brand j on a Similarly. We assume that there are N brands in the market.j and makes a cash purchase is m1(1 . w-heref(n) is some unknown function of n. and the promoting brand is Brand 0. we have indicated how game. the mailout size. We can modify the concept of potential here.156 AMBAR G. due to the price change only. recalling that the reach of the brand is in0. f Then the volume gain to the independent brands (as a group). be expected to be simply the sum of the individual effects. If we assume that f(n) is approximately a linear function of n given by n/p. in general. say Brand K. due to the price and the proportional loss in sales to a major brand with share mll0 change by the independents is KAO8ni. that he buys Brand. Let miii the market share of the jth brand in a given market. for small 6. . above development is The for the case when Brand 0 is the only one conducting a mailout promotion. lllj and the proportional gain to the company is P = VcVmiioG.n function of price change. that is. Note that this loss can be positive or negative depending on the sign of 6. Thus the probability that a randomly chosen customer buys any brand but Brand 0 on credit is given by Vl. by KO. credit card and pricing effects are modelled under the assumption that only one effect is present in a market at a time. and the probability that the purchase is on cash is N=i mlj (1 .cj). then the gain to Brand 0 is V/c = Vc -1G(L/p)1PK"[K1ooco -K2))?0(1-CO)] where the second term on the right is the gain Brand K would obtain from Brand 0.

Vc/G = proportion of volume attracted to credit card promotion. if these populations can be considered independent random samples (with replacement) from F. M'-akingthe connection with the associated volumes. where VG iS game-promotion volume. members of the population attracted by the credit card promotion. for example) then fc n fG = min {fc. Vc) > VG n Vc > VC VC/G. w e can now state that min (VG. Little for this suggestion. fG = members of the population attracted by the game. . since these populations are hardly likely to be independent. M\Aultiplyingby F to convert this proportion to an absolute population size implies: fG n fc = fGjfC fanf F2 F = fG fc F This should understate the interaction. fG}. It is then necessary to calculate Va u Vc = VG + VC - VGan Vc. then VG/G = proportion of total market volume attracted to the game promotion.(fc/F). 0< X 1. sales/Forecast sales for the period. Let F = total population size. C. An intuitively appealing approach to modelling VGn IVcis to formulate this intersection as a convex combination of these boundss:' (11) VGfn Vc = Xmin {VG. On the other hand. If we let G = total market volume. These populations are probably at least partially distinct. Vc is credit card promotion volume. Estimation of Parameters Consider a market where some promotion or combination of promotions is being conducted by a company.A SYSTEM OF PROMOTIONAL MODELS 157 Suppose there is both a game and a credit card promotion in a market at a certain time. Assume the following relationship: (12) where Y -Actual e= e = Y - + Proportional change in sales due to promotions + E. random error (E (ti) = 0. Since our model gives us expressions for the second term on the right of (12). An estimation procedure for X along with the other model parameters is developed in the next section. so this should overstate the interaction. we can obtain estimates for 1 The authors are indebted to John D. Ve} + (1 - X)V Vc/G. We now assume that these volumes will have a one-to-one relationship with respective customer sets. Note that A here represents 1 + proportional forecasting bias. fc If either fc or fG is a proper subset of the other (all those who are attracted bv a credit card promotion would have been attracted by a game. E (Y I no Promotions). then the proportion of the population we would expect to have duplicated is the product (f0/F). fE4j uncorrelated over time).

Letting Yit refer to Actual Sales/Expected Sales in market i at time t.10 0.10 0.10 0.30 0.35 0.u. Then.35 0. No other company has a credit card mailout.05 0.20 0. TABLE 1 Model-Estimated Changes in Market Share* Market Share in Games Company Has Game? Change in Price Diff.10 0.15 0.15 0.30 0.35 0. Yt- j + Git + Et. we can measure the constants for the game effect as follows: Suppose a set of markets {I } is isolated for which.25 0.20 0.15 0.158 AMBAR G.30 0.15 0.20 0.10 0.15 0.30 0.45 0.30 0.10 0.25 0.20 0.15 0.45 0.15 0.10 0.05 0.10 0.05 0.05 - -12% -2% +8% +17% + 10% +11% +18% +21% -3% +6% +9% + 19% - +4% +10% +12% -4% +5% + 10% +17% +18% +21% +16% +21% -5% -5% -16% 50% of market buys using cash.25 0. and the {Kil by a least squares procedure if we know .45 * Assumptioins: 2 10 -l1 2? 20 10 10 - Y Y Y Y N N N N Y Y Y Y N N N N Y Y Y Y N N N -10 -10 . where Git.15 0.15 0. and a time series analysis of historical sales was indirectly used to obtain estimates for .uwe reason as follows: Although the oil industry had promoted gasoline in the past using station-based promotions.05 0. Market Share of Independents Part of Population in Mailout Percent Change in Market Share 0 0 0 0 0.20 0.30 0.20 20 20 10 10 -10 -10 -20 -26 -20 -10 -10 - 10 20 0.10 0.25 0. the effect of games (referring to (5) and (6)).30 0.. The company has an 8% expected market share.10 0.25 0.10 0. PL if the brand does not have a game promotion in i at time t.15 0.20 0.20 0. ic {I}.10 0.25 0.30 0.30 0. at t.25 0.20 0.25 0.To estimate . RAO AND GARY LILIEN a. the only promotional disturbance is a game (i.10 0.10 0.05 0. PG if the brand has a game promotion in market i at time t.10 0.15 0. Thus the effect of historical promotions on sales could be treated as a component of random error.15 0.uin various markets. prices are stable. no credit card mailout).15 0. .10 0.e.35 0. these promotions never had as large an effect as games and credit card mailouts.05 0. X.15 0.

a 10 % sales boost can be expected. i E {IG}I Yit = Ai Kmnt + Eit. was approximately half its original value one year after games were introduced. in a market that was used for validation. including the company of interest. when the market usual price difference has been increased by 20 and independents have 10 % of the market (with no credit card mailout in effect). The game-strength parameter. However. Referring to equations (5) and (6) again: (13a) (13b) Yit = Ait + Kmit' - (1 -n2. Figure 1 compares. the results of the time series forecasts alone (the alternative system and the one in use at the time) with the results after the addition of the promotions models. J F M A M J 1 967 J A S O N D FIGURE I . the variance of the error between twelve monthly actual and predicted sales figures was reduced by over 50 %. ING} where {IG} is that subset of the markets where the company has a game. Its effective 'half-life' (time until half the peak incremental sales were proCOMPARISON FORECASTING FzGURE 1 st OECASALE O <t ~~~~~~~~~~~~~~~PLUS TIME SERIES FORECAST PROMOTIONS MODEL Credit Cord Compony Leove Compony E..A SYSTEM OF PROMOTIONAL MODELS 159 Let I = {IG. For example. the fifth line indicates that when 15 %of the market. K. is participating in a game promotion. Some interesting insights into consumer behavior were revealed from updating parameter estimates... It should be noted that our goal at this stage of the procedure is to obtain working parameter estimates. In a test of 19 markets. Estimation procedures for the other parameters are similar. the model system was incorporated into the existing time-series forecasting package and was used to predict sales in various markets where promotions were being conducted. Least squares estimates of K and a can now be made from (13) using a nonlinear optimization technique. i E {ING}. + Eit. Since we derived no closed-form results for the estimates. Model Characteristics and Results Once parameters had been obtained. Table 1 shows a sample of the output of the models as estimated in June 1968 for a variety of situations. we do review their predictive qualities in the next section. The lag in pickup of the July-August 1967 sales spurt was attributed to an advertising campaign preceding the mailout. we can make no claim for their statistical properties. and advertising promotions were not treated in the model structure.. The strength of a credit card mailout was found to peak about four months after the original mailing and then slowly decline. {ING} where the company does not.

R. 16. Cambridge. this system could be used in several ways: With the addition of competitive information and competitive response assumptions." 2. D. ACKOFF. StochasticModels of Buying Behaviot. References Management Science. the system could be used more as an "aid-to-judgment" to simulate the effects of promotional strategies in different competitive situations.160 AMBAR G. The MIT Press. 3. G. W. 4 (December 1967). Because of the amount of additional competitive data and model-building effort needed for promotional expense allocations.. Johni Wiley & SoIis. Little [2] discusses this method of using models. MIanageinent Science. An on-line interactive program was developed to make the system of models accessible to marketing management and to increase the probability that they be properly used for planning and control. J.. we could not complete the cycle of development-experimentation-model refinement as we had planned at the beginning of the project. 1. the "aid-to-judgment" approach was chosen initially. G. Implementation In addition to its use in a sales forecasting procedure. A manager could access the models through a console and experiment with promotional assumptions and timings. No. C. Howevei. B466-485. Price parameters showed the largest sample-tosample variation indicating either highly erratic consumer price consciousness or possible weakness in the price-model structure. Quantitative Theories in Advertising. No. AND MORRISON. Even more importantly. RAO AND GARY LILIEN duced) was about one and a half years.. B147-156. LITTLE. F. "Models aiid Maniagers: The Conicept of a Decision Calceltis. New York. MONTGOMERY. MASSY. 1970. Vol. R1AO. pp.. Vol. Alternately. a framework for so doing had been developed and could be used in the future for other station-based promotions. 4. D. pp. Mass. D. 14. the effects of assumed or ainticipatedcompetitive actions and reactions could be examined. B. 1970.. optimal allocations of promotional expenses within and across markets could be developed." . 8 (April 1970). Due to the termiination of the specific promotions discussed in this paper. "Managemenit Misiniformationi Systems.. A. L.