International Journal of Forecasting 19 (2003) 95110 www.elsevier.

com / locate / ijforecast

Forecasts of market shares from VAR and BVAR models: a comparison of their accuracy
Francisco Fernando Ribeiro Ramos
Faculty of Economics, University of Porto, 4200 Porto, Portugal

Abstract This paper develops a Bayesian vector autoregressive model (BVAR) for the leader of the Portuguese car market to forecast the market share. The model includes ve marketing decision variables. The Bayesian prior is selected on the basis of the accuracy of the out-of-sample forecasts. We nd that BVAR models generally produce more accurate forecasts. The out-of-sample accuracy of the BVAR forecasts is also compared with that of forecasts from an unrestricted VAR model and of benchmark forecasts produced from three univariate models. Additionally, competitive dynamics are revealed through variance decompositions and impulse response analyses. 2002 International Institute of Forecasters. Published by Elsevier Science B.V. All rights reserved. Keywords: Automobile market; BVAR models; Forecast accuracy; Impulse response analysis; Marketing decision variables; Variance decomposition; VAR models

1. Introduction Multiple time series models have been proposed, for some time, as alternatives to structural econometric models in economic forecasting applications. One such class of multiple time series models, which has received much attention recently, is the class of Vector Autoregressive (VAR) models. VAR models constitute a special case of the more general class of Vector Autoregressive Moving Average (VARMA) models. Although VAR models have been used primarily for macroeconomic models, they offer an interesting alternative to either structural econometric, univariate (e.g., BoxJenkins /ARIMA or exponential smoothing) models, or multivariate (e.g., VARMA) models for problems in which simultaneous forecasts are required for a collection of
E-mail address: framos@fep.up.pt (F.F. Ribeiro Ramos).

related microeconomic variables, such as industry and rm sales forecasting. The use of VAR models for economic forecasting was proposed by Sims (1980), motivated partly by questions related to the validity of the way in which economic theory is used to provide a priori justication for the inclusion of a restricted subset of variables in the structural specication of each dependent variable.1 Sims (1980) questions the use of the so-called exclusionary and identication restrictions. Such time series models have the appealing property that, in order to forecast the endogenous variables in the system, the modeller is not required to provide forecasts of exogenous
1

In the Marketing eld these arguments are mutatis mutandis also valid. The lack of a generally accepted theory about aggregate market response and marketing mix competition means that there is little a priori reason to support or reject any of a number of plausible model specications.

0169-2070 / 02 / $ see front matter 2002 International Institute of Forecasters. Published by Elsevier Science B.V. All rights reserved. PII: S0169-2070( 01 )00125-X

96

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

explanatory variables; the explanatory variables in an econometric model are usually no less difcult to forecast than the dependent variables. In addition, the time series models are less costly to construct and to estimate. This does not imply, however, that VAR models necessarily offer a parsimonious representation for a multivariate process. While it is true that any stationary and invertible VARMA process has an equivalent representation as a VAR process of an eventual innite order (see, for example, Fuller, 1976), generally the VAR representation will not be as parsimonious as the corresponding VARMA representation, which includes lags on the error terms as well as on the variables themselves. Despite this lack of parsimony, and the additional uncertainty imposed by the use of a nite-order VAR model as an approximation to the innite-order VAR representation, VAR models are of interest for practical forecasting applications because of the relative simplicity of their model identication and parameter estimation procedures, and superior performance, compared with those associated with structural and VARMA models.2 Brodie and De Kluyver (1987) have re ported empirical results in which simple nave market share models (linear extrapolations of past market share values) have produced forecasts as accurate as those derived from structural econometric market share models. Furthermore, the same paper shows that using lagged market share often gives better results than an econometric model, which incorporates marketing mix variables. Indeed, Danaher and Brodie (1992) provided a criterion which determines whether it is advantageous to use marketing mix information for forecasting market shares. The number of parameters to be estimated may be very large in VAR models, particularly in relation to the amount of data that is typically available for business forecasting applications. This lack of par-

simony may present serious problems 3 when the model is used in a forecasting application. Thus, the use of VAR models often involves the choice of some method for imposing restrictions on the model parameters: the restrictions help to reduce the number of parameters and (or) improve their estimation. One such method, proposed by Litterman (1980), utilises the imposition of stochastic constraints, representing prior information, on the coefcients of the vector autoregression. The resulting models are known as Bayesian Vector Autoregressive (BVAR) models. In the Marketing literature, applications of multiple time series models (Transfer Functions, Intervention and VARMA models) include Aaker, Carman and Jacobson (1982), Adams and Moriarty (1981), Ashley, Granger and Schmalensee (1980), Bass and Pilon (1980), Bhattacharyya (1982), Dekimpe and Hanssens (1995), Franses (1991), Geurts and Whitlark (1992), Grubb (1992), Hanssens (1980a,b), Heyse and Wei (1985), Heuts and Bronckers (1988), Jacobson and Nicosia (1981), Krishnamurthi, Narayan and Raj (1986), Kumar, Leone, and Gaskins (1995), Leone (1983, 1987), Lui (1987), Moriarty and Salamon (1980), Moriarty (1985), Sturgess and Wheale (1985), Takada and Bass (1988), Umashankar and Ledolter (1983). In this paper, we develop a BVAR for the leader of the Portuguese car market for the period 1988:1 throughout 1993:6 using monthly data. The rationale for the choice of a multiple time series technique is two-fold. Structural models of market share, as surveyed in Cooper and Nakanishi (1988), tend to be based on a number of generalisations about the effectiveness and relative importance of advertising, price, and other elements of the mix, with little emphasis being placed on the correct determination of exogenous assumptions and on the appropriate dynamic model specication. Secondly, and pre-

Very few VARMA analyses of higher-dimensional time series (e.g., models with more than four series) are reported in the literature. The wider class of vector ARMA models were not considered because there was little evidence of moving average components and because both the identication and estimation of such models are relatively complicated. For a recent summary of the specication of VARMA models, see Tiao and Tsay (1989).

Apart from the multicollinearity between the different lagged variables leading to imprecise coefcient estimates, the large number of parameters leads to a good within-sample t but poor forecasting accuracy because, according to Litterman (1986a, p. 2), parameters t not only the systematic relationships . . . but also the random variation.

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

97

sumably in part because of this, the forecasting performance of such models has been poor compared to that of time series models: for such evidence, see Brodie and De Kluyver (1987), Danaher and Brodie (1992), and Brodie and Bonfrer (1994). Out-ofsample one-through 12-months-ahead forecasts are computed for the leaders market share and their accuracy is evaluated relative to that of forecasts from an unrestricted VAR model and from three best-tting univariate models.4 The paper is organised as follows. The rst section briey describes the VAR and the BVAR modelling methodologies. The second section describes the database used and the rationale behind the choice of the variables. The third section presents the selected models, the main empirical results, and illustrates the use of impulse response analysis and variance decomposition as a marketing tool in providing information about the competitive dynamics of the market. We conclude with a section on the limitations of our research and possible extensions.

processes with positive denite contemporaneous covariance matrix and zero covariance matrices at all other lags, and the Bk s are (n 3 n) coefcient matrices with elements b ijk . This approximation assumption holds, in fact, if Yt is a covariancestationary linearly regular process. Eq. (1) can be used to generate the forecast ft , h at time t of Yt 1 h , with subsequent forecast error e t , h 5 Yt 1 h 2 ft , h and error variancecovariance matrix Vh 5 E(e t , h ? e 9 t , h ). Granger and Newbold (1986, chapter 7) show that the optimal (in terms of minimising the quadratic form associated with Vh ) h period ahead forecast ft , h of Yt 1 h made at time t is ft , h 5

k 51

OB f

k t,h 2k

(2)

where ft , h 2 k 5 Yt 2 (k 2 h ) for k 5 h,h 1 1, . . . , p, and Bk s are the coefcient matrices in Eq. (1).

2.1. The unrestricted VAR

2. VAR and BVAR modelling The theory underlying VAR models has its foundation in the analysis of the covariance stationary linearly regular stochastic time series. We assume here that Yt is (n 3 1) in dimension, i.e. Y t9 5 (Y1 t , . . . ,Ynt ). By Wolds decomposition theorem, Yt possesses a unique one-sided vector moving-average representation which, assuming invertibility, gives rise to an innite-ordered VAR. In empirical work it is assumed that Yt can be approximated arbitrarily well by the nite p th-ordered VAR: Yt 5 In a VAR with n variables there is an individual equation for each variable. For the unrestricted case there are p lags for each variable in each equation. For example, the equation for the i th variable is Yit 5

k 51

Ob

i 1 k Y1, t 2 k 1 ? ? ? 1

k 51

Ob

ink n , t 2 k

1 e it .

(3)

k 51

OB Y

k t 2k

1 et

(1)

where e t is a zero-mean vector of white noise

The basic metric of forecasting comparisons is the calculation of MAPE, RMSE and Theils U statistics across the range of forecasting horizons (out-of-sample forecasts) for the several models.

As in the problem of seemingly unrelated regressions, when the right-hand-side variables are the same in all equations, the applications of the OLS equation by equation is justied. The coefcient estimates are maximum likelihood estimates (MLE) only if the e s are normally distributed, otherwise quasi-MLE. The unrestricted VAR has been used extensively by Sims (1980), and in the initial stages of model building by Caines, Keng and Sethi (1981), Tiao and Box (1981), and Tiao and Tsay (1983). The main problem with the unrestricted VAR is the large number of free parameters that must be estimated. Since the number of parameters increases quadratically with the number of variables, even moderately sized systems can become highly overparameterised relative to the number of data points.

98

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

This over-parameterisation 5 results in multicollinearity and loss of degrees of freedom that can lead to inefcient estimates and large out-of-sample forecasting errors. While estimation of such a highly parameterised system will provide a high degree of data tting, the out-of-sample forecasts can be very poor in terms of mean square error. Because of these problems, researchers have suggested imposing various types of parameter restrictions on VAR models. Several types of these restrictions are described in the literature. One solution is to exclude insignicant variables and (or) lags based on statistical tests. An alternative approach to overcome over-parameterisation is to use a BVAR model as described in Litterman (1980), Doan, Litterman and Sims (1984), Todd (1984), Litterman (1986b), and Spencer (1993).

similar in many respects to the ridge and Stein estimators. Since there is a ridge regression analogy to the BVAR, it is not surprising that BVAR solved the multicollinearity problem. As is well known, from a Bayesian standpoint, shrinkage estimators can be generated as the posterior means associated with certain prior distributions. While Littermans estimator can be justied as a posterior mean, the economic content of the prior information is not strong. To demonstrate Littermans procedure, consider the i th equation of the VAR model (3):

Yit 5 d it 1

k 51

Ob

i 1 k 1, t 2 k

1???1

k 51

Ob

ink n , t 2 k

1 e it (4)

2.2. The Bayesian VAR

The Bayesian approach starts with the assumption that the given data set does not contain information in every dimension. This means that by tting an over-parameterised system some coefcients turn out to be non-zero by pure chance. Since the inuence of the corresponding variables is just accidental and does not correspond to a stable relationship inherent to the data, the out-of-sample forecasting performance of such models deteriorates quickly. The role of the Bayesian prior can therefore be described as prohibiting coefcients to be non-zero too easily. Only if the data really provides information, will the barrier raised by the prior be broken. In an attempt to reduce the dimensionality of VARs, Litterman (1980) applied Bayesian techniques directly to the estimation of the VAR coefcients. His procedure generates a shrinkage type of estimator
5

where d it is the deterministic component of Yit and can include the constant, trend, and dummies. Littermans prior is based on the belief that a reasonable approximation of the behaviour of an economic variable is a random walk around an unknown, deterministic component. For the i th equation the distribution is centred on the specication Yit 5 d it 1 Yi , t 2 1 1 e it . (5)

An alternative to a VAR is a simultaneous equations structural model. However, there are limitations to using structural models for forecasting since projected values of the exogenous variables are needed for this purpose. Further, Zellner (1979), and Zellner and Palm (1974) show that any linear structural model can be expressed as a VARMA model, the coefcients of the VARMA model being combinations of the structural coefcients. Under certain conditions, a VARMA model can be expressed as a VAR model and a VMA model. A VAR model can therefore be interpreted as an approximation to the reduced form of a structural model.

The parameters are all assumed to have means of zero except for the coefcient on the rst lag of the dependent variable, which has a prior mean of one. All equations in the VAR system are given the same form of prior distribution. In addition to the priors on the means, the parameters are assumed to be uncorrelated with each other (the covariances are set equal to zero) and to have standard deviations which decrease the further back they are in the lag distributions. The standard deviations of the prior distribution on the lag coefcients of the dependent variable are allowed to be larger than for the lag coefcients of the other variables in the system. Also, since little is known about the distribution of the deterministic components, a at or uninformative prior giving equal weight to all possible parameter values is used. In equation form the standard deviation of the prior distribution for the coefcient on lag k of variable j in equation i is

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

99

g ] kd s ijk 5 g ? w ? s i ]]] d k ? sj

if i 5 j (6) if i j.

suggest minimising the log determinant of the sample covariance matrix of the one-step-ahead forecast errors for all the equations of the BVAR.

j is the estimated standard error of In Eq. (6), s residuals from an unrestricted univariate autoregression on variables j. Since the standard deviations of lag coefcients on variables other than the dependent variables are not scale invariant, the scaling factor i /s j is used. This ratio scales the variables to s account for differences in units of measurement and thus enables specication of the prior without consideration of the magnitudes of the variables. The term g the overall tightness of the prior is the prior distribution standard deviation of the rst lag of the dependent variable. A tighter prior can be produced by decreasing the value of g. The term d the decay parameter is a coefcient that causes the prior standard deviations to decline in a harmonic manner. The prior can be tightened on increasing lags by using a larger value for d. The parameter w the relative tightness is a tightness coefcient for variables other than the dependent variables. Reducing its value, i.e. decreasing the interaction among the different variables, tightens the prior. Note that the prior distribution is symmetric. The same prior means and standard deviations are used for each independent variable in each equation and across equations, and the same priors are used for each dependent variable across equations. Doan et al. (1984) have also considered another type of prior, known as general. In a general prior the interaction among the variables leads to the specication for the weighing matrix, f (i, j ), given by 1 if i 5 j f (i, j ) 5 f if i j (0 , f , 1).. ij ij The BVAR model is estimated using Theils (1971) mixed-estimation technique which involves supplementing data with prior information on the distribution of the coefcients. To apply Littermans procedure one must search over the parameters g, d, and w until some predetermined objective function is optimised. The objective function can be the out-ofsample mean-squared forecast error, or some other measure of forecast accuracy. Doan et al. (1984)

3. Data The database used for this study is a monthly time series sample of market shares, and marketing mix variables, for the period 1988:1 to 1994:6 in the car market in Portugal. The marketing mix variables included retail prices, advertising expenditures by media (TV, radio, and newspapers), and an age variable for the brand leading the car market. The Portuguese car market consists of 25 imported car brands, but the top seven account, on average, for 82.3% of the total market, with a standard deviation of 4.75%. The leader is a general brand, present in all segments of the market and represents, on average, 16.8% of the total market, with a standard deviation of 3.56%. The time series variables used in this study are dened as follows: MS 1 t is the market share of the leader brand; A1 t is the relative age of the leader brand; P 1 t is the relative price of the leader brand; TVS 1 t is the TV advertising expenditures in shares of the leader brand; RS 1 t is the Radio advertising expenditures in shares of the leader brand; and PS 1 t is Press (newspapers and magazines) advertising expenditures in shares of the leader brand. The data on MS 1 t is calculated from the monthly new automobile registrations. This data is published by the Portuguese General Directorate of Transports. As can be seen, Fig. 1 shows that our brand is not successful in increasing its market share. This series seems to be stationary and does not present seasonal uctuations.

Fig. 1. Market share of the Portuguese market car leader.

100

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

The marketing decision variable A1 t measures the age (in months) of the different brand car models after their introduction in the market. This variable represents the models life cycle of the brand, and can be seen as the product decision variable. It was obtained as follows: for the leader, we measure the age (in months) after the launch of the most representative model of each market segment, i.e. the model with the highest segment share. We apply a pseudosegmentation method based on the horsepower, and followed by the Portuguese Trade Automobile Association. This segmentation creates four segments: S1 (lower), S2 (lower-middle), S3 (upper-middle), and S4 (upper); for the competing brands, we calculate the simple average age of the most representative model of each segment; to obtain the brand average age of its competitors, we calculate the weighted average age for the models chosen on each segment. The weights are given by the relative importance of each segment on total demand (S1 1 S2 1 S3 1 S4); the relative age, called A1, is then calculated as the ratio between the weighted average age of the brand and the weighted average age of its competitors. The variable P 1 t is obtained following the steps just described for A1 t . The weights are the same, and the price for each model is the consumer price (all taxes included) of the most representative model of each segment. The price data are published on a monthly basis and are recorded in the Guia do Automovel (The Portuguese Car Magazine). Fig. 2

Fig. 3. Advertising expenditures in share by media.

plots the relative price and the relative age of our leading brand. The overall trend of A1 t is upward, and seems negatively related to the market share series. The price variable is stationary around the value of 1, indicating that our brand price is approximately equal to the competitors brand prices. The data on TVS 1 t , RS 1 t , and PS 1 t are expressed as shares of total advertising expenditures by media and are obtained from Sabatina. This Portuguese rm records on a monthly basis the advertising expenditures by media and brand. These advertising expenditures represent only ofcial or contractual prices, and we know in the industry that prices are frequently lower. As indicated in Fig. 3, the shares of the major media advertising expenditures of our brand uctuated widely over the observed time period. All variables are measured in logs to help reduce the problem of heteroscedasticity. The data set is included in Appendix A.

4. Empirical application

4.1. Models selected

Three classes of models are included in our empirical comparisons, each class being represented by one or more specic models. The classes are univariate, unrestricted VAR, and BVAR. The usual criteria, for example stationary, autocorrelation, and partial autocorrelation functions, signicance of coefcients, and the Akaike Information Criterion, are used to select the best models. In all computations we have used the RATS program (RATS 386, version 4.02).

Fig. 2. The brands age and price evolution.

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

101

In the class of univariate models we have considered four models (an ARIMA, a NAIVE, a REGRESSION and an exponential smoothing, specically a HoltWinters, model). The best-tting ARIMA model for MS1 is as follows: (1 2 0.916B )MS 1 t 5 21.898 1(1 1 0.814B )e t a
(0.03) a ( 0.06 ) (0.07) a

Standard errors are in parentheses and superscript a indicates signicance at the 0.01 level. As Montgomery and Weatherby (1980, p. 306) note: The Box Jenkins approach uses inefcient estimates of impulse response weights which are matched against a set of anticipated patterns, implying certain choices of parameters . . . the analyst skills and experience often play a major role in the success of the model building effort. For other criticisms of ARIMA, see Chateld and Prothero (1973), Hillmer and Tiao (1982), and Prothero and Wallis (1976). The NAIVE (MSt 5 a 1 b MSt 2 1 ), the REGRESSION (MSt 5 a 1 b MSt 2 1 1 marketing mix cov ariates) and the HoltWinters models are compared in terms of forecasting accuracy with the other four models in Table 1. The nal specication of these models is not presented here due to space
Table 1 Accuracy of out-of-sample forecasts (1993:11994:6) Month ahead 1 Accuracy statistic U MAPE RMSE U MAPE RMSE U MAPE RMSE U MAPE RMSE U MAPE RMSE N 18 NAIVE 0.965 0.025 0.184 1.102 0.039 0.228 0.985 0.035 0.257 1.39 0.049 0.256 1.110 0.037 0.231 Holt Winters 0.901 0.021 0.170 0.841 0.025 0.173 0.685 0.020 0.179 1.15 0.040 0.216 0.895 0.027 0.185

limitations. However, these can be obtained from the author upon request. To determine the optimal lag length of the unrestricted VAR in levels [VAR(U)] we have employed the likelihood ratio test statistic (LR), suggested by Sims (1980).6 Given the number of observations, we have considered a maximum lag of nine and then tested downwards. The LR test supports the choice of six lags. The model is then estimated in levels (so that it is comparable to the BVAR model) with 37 parameters (including the constant) in each equation. In the class of BVAR models, the variables are specied in levels because, as pointed out by Sims, Stock and Watson (1990, p. 360), . . . the Bayesian
6

If AR(m) is the unrestricted VAR and AR(l ) the restricted VAR, where m and l are the respective lags, then the LR statistic for testing AR(l ) against AR(m) is given by LR 5 (T 2 c)(lnuVl u 2 lnuVm u ) where T is the number of observations, c is the correction factor which is equal to the number of regressors in each equation in AR (m), and V is the covariance matrix of residuals of AR(l ) and AR(m), respectively. The statistic LR is asymptotically distributed as chi-squared with k 2 (m 2 l ) degrees of freedom, where k is the number of regressions.

REG. 0.904 0.021 0.172 1.013 0.038 0.210 0.919 0.033 0.240 1.30 0.045 0.239 1.034 0.034 0.215

ARIMA 0.979 0.029 0.187 0.846 0.025 0.175 0.665 0.017 0.174 1.13 0.039 0.207 0.905 0.028 0.186

VAR(U) 0.999 0.031 0.191 0.894 0.029 0.185 0.596 0.014 0.156 1.06 0.032 0.195 0.887 0.027 0.182

BVAR(S) 0.813 0.108 0.155 0.761 0.017 0.158 0.65 0.016 0.169 1.09 0.034 0.20 0.828 0.021 0.17

BVAR(G) 0.831 0.019 0.159 0.747 0.016 0.149 0.633 0.015 0.162 0.04 0.031 0.191 0.812 0.020 0.165

16

13

12

Average

Note: N is the number of observations. The RMSEs, the MPEs and the U statistics are reported for log MSI. Average is the average of the 1-, 3-, 6- and 12-months-ahead RMSEs and the U statistics.

102

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110 MS 1 1 0.5 0.5 0.5 0.5 0.5 A1 0.75 1 0.75 0.5 0.5 0.5 P1 0.75 0.75 1 0.5 0.5 0.5 TVS 1 0.5 0.75 0.5 1 0.75 0.75 RS 1 0.5 0.75 0.5 0.75 1 0.75 PS 1 0.5 0.75 0.5 0.75 0.75 1

approach is entirely based on the likelihood function, which has the same Gaussian shape regardless of the presence of nonstationarity, [hence] Bayesian inference needs to take no special account of nonstationarity (see also Sims, 1988, for a discussion on Bayesian scepticism on unit root econometrics).7 The models are estimated with six lags of each variable. Longer lags (up to nine) were also tried, but the substantial results remained unchanged. The optimal Bayesian prior is selected by examining the Theil U and the RMSEs values for the out-of-sample forecasts. In a rst step we assume a symmetric prior, i.e. f (i, j ) 5 w, i j [BVAR(S)], then we relax this assumption to take into account a more general interaction between the variables [BVAR(G)]. In our search for the symmetric prior we have considered three values for w : 0.25, 0.5, 0.75. For the parameter g we have assumed a relatively loose value of 0.3 and a tight value of 0.1. We set the harmonic lag decay, d, to 1 as recommended by Doan et al. (1984). This has given us six alternative specications. The best values according to our criterion function 8 were obtained for g 5 0.3, w 5 0.5 and d 5 1. To specify the general prior we must dene g, d, and the interseries tightness parameters, f (i, j ). Using the information provided by impulse response analysis and variance decompositions, and after some initial search, the best values were obtained for d 5 1, g 5 0.15 and

f (i, j ) 5

MS 1 A1 P1 TVS 1 RS 1 PS 1

4.2. Evaluation of accuracy

The accuracy of the forecasts for 1993:1 to 1994:6 is measured by the MAPE, the RMSE and the Theil U statistics for 1- to 12-months-ahead forecasts. If A t denotes the actual value of a variable, and Ft the forecast made in period t, then the MAPE, the RMSE and the Theil statistic are dened as follows: 1 K u A t 1 j 1 k 2 Ft 1 j 1 k u MAPE 5 ] ]]]]] N j 51 A t 1j 1k
k 2 t 1 j 1 k 2 Ft 1 j 1 k ) / N

O RMSE 5 HO (A
j 51

0.5

U 5 RMSE(model) / RMSE(random walk) where k 5 1,2, . . . ,12 denotes the forecast step and N is the total number of forecasts in the prediction period. The U statistic is the ratio of the RMSE for the estimated model to the RMSE of the simple random walk model which predicts that the forecast simply equals the most recent information. Hence, if U , 1, the model performs better than the random walk model without drift; if U . 1, the random walk outperforms the model. The U statistic is therefore a relative measure of accuracy and is unit-free. The forecasted value used in the computation of the MAPE, the RMSE and the U statistics is the level (in logarithms) of the market share, so these statistics can be compared across the different models. The accuracy measures are generated using the Kalman lter algorithm in RATS. The models are estimated for the initial period 1988:1 to 1992:12. Forecasts for up to 12 months ahead are computed. One more observation is added to the sample and forecasts up to 12 months ahead are again generated, and so on. Based on the out-of-sample forecasts, MAPEs, RMSEs and the Theil U statistics are computed for 1- to 12-months-ahead forecasts. The three accuracy measures for MS 1 for seven models are reported in Table 1. The table also

This kind of discussion about classical vs. Bayesian analysis of time series was the subject of a special issue of the Journal of Applied Econometrics (OctoberDecember 1991). One of the most interesting contributions seems to be the article of Phillips which is an answer to Sims (1988). The criticism of Phillips agrees with the view that the tool used by Sims to criticise the legitimacy of unit roots tests is based on the mechanical use of at priors in a Bayesian analysis of time series models. Phillips demonstrates that at priors are not uninformative but unwittingly introduce a tendency towards stationary models. 8 Instead of preselecting some values we could select the Bayesian hyperparameters by minimising the following function: Min U ( g,w,d ) 5

i 5 1h 5 1

OO u
n H

ih

where u ih is the Theil U for time-series i h -forecast steps ahead. Because the functional relationship is highly nonlinear, numerical methods must be used to minimise U. In particular, we could use a grid search over the arguments h g,w,d j.

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

103

reports the average of these statistics for the 1-, 3-, 6and 12-month-ahead forecasts. The conclusions from Table 1 are as follows: 1. MAPEs versus Theil U statistics: as both MAPE and Theils U indicate, the forecasting performance of all models deteriorates with larger forecasting horizons (6 and 12 months). The RMSE do not follow a consistent pattern with an increase in the forecast horizon. 2. BVAR versus univariate models: BVAR models produce more accurate forecasts (for all horizons) than the corresponding univariate models (NAIVE, REG., HoltWinters, and ARIMA). 3. The HoltWinters model, on average, performed the same as the ARIMA, but is clearly superior to the NAIVE and REG. models. We did not conrm in this study one of the main conclusions of the Big Mac paper (Makridakis et al., 1982) that . . . exponential smoothing was more accurate than ARIMA models on average . . . . 4. BVAR versus the unrestricted VAR models: in all cases, except for forecast horizon 6, BVAR(G) outperforms the unrestricted VAR model. The comparison between BVAR(S) and VAR(U), however, yields mixed results. VAR(U) seems to be better for longer horizons (6 and 12). 5. BVAR(G) versus BVAR(S): BVAR(G) always provides the most accurate forecasts. This is not surprising since the prior for the model was selected on the basis of minimisation of the average of 1- to 12-month-ahead RMSEs and the U statistics. 6. The Theils U for 12-month-ahead forecasts are larger than one for all the models, which indicates that the random walk model is an improvement on all the time series models. The results, in general, show that there are gains from using a BVAR approach to forecasting. On average, the BVAR models produce more accurate forecasts than the alternative forecasts. Finally, in the class of BVAR models, BVAR(G) always produces the most accurate forecasts except for 1-month-ahead forecasts. Finally, like other authors (Hafer and Sheehan, 1989) we found that the accuracy of the forecasts is sensitive to the specication of the priors. If the prior

is not well specied, an alternate model such as an unrestricted VAR or an ARIMA model may have a better performance.

4.3. Performance of alternative models

While the BVAR models, in general, produce the most accurate forecasts, another way to evaluate the performance of alternative models is to examine their ability in predicting turning points.9 We focus on the performance of the BVAR models relative to that of the unrestricted VAR and the univariate ARIMA models. As can be seen in Fig. 4, it seems that the ARIMA model is predicting a near immediate return to the mean, the BVAR models are predicting a time trend (differencing and a constant) and the unrestricted VAR gives the best visual forecasts. The unrestricted VAR is better at picking turning points than is the BVAR model. In fact, the BVAR model appears to do a very poor job at forecasting turning points in the market share data. Besides generating excellent baseline forecasts, BVAR models can also be used to study the effects of movements in one variable on movements in others using impulse response analysis and variance decompositions.

4.4. Impulse response analysis

Impulse responses are the time paths of one or more variables as a function of a one-time shock to a given variable or set of variables. Impulse responses are the dynamic equivalent of elasticities. For example, in a static multiplicative interaction model of the form ms 5 a ? p b , price elasticity [ 5 (dms / dp)( p / ms) 5 b ] is constant. However, in a dynamic system, changes in market share in 1 month are a function of changes in price over several months. The net effect must be represented as a convolution sum. The graphic representation of this sum provides a com-

This topic is of vital importance, as statistical methods can perform extremely well in terms of forecasting overall levels, and yet still perform poorly in the prediction of turning points. Unfortunately, the question of turning points does not easily lend itself to quantitative analysis due to difculties in, rstly, dening turning points, and, secondly, knowing when a given method has adequately predicted a turn.

104

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

Fig. 4. Market share forecasts for 1993:71994:6 (made in 1993:6).

plete description of the dynamic structure of the model (it is incorrect to directly interpret estimated coefcients for the variables of a BVAR model). The innovations in the BVAR models have been orthogonalised via a Choleski decomposition, and unless the error terms of each equation are contemporaneously uncorrelated it may not make sense to assume that a shock occurs only in one variable at a time.10 The variables are ordered in the following sequence: MS 1 t , A1 t , P 1 t , TVS 1 t , RS 1 t , PS 1 t . This ordering is based partly on our prior belief that changes in MS 1 t precede those in A1 t , P 1 t , TVS 1 t ,

RS 1 t , PS 1 t and partly on the timing of the availability of data. For instance, information on prices and market shares is released prior to the advertising expenditures. To illustrate, Fig. 5 shows the impulse responses of MS1 to a shock for each one of the six variables of the system and Fig. 6 shows the cumulative responses over 12 months. Both gures illustrate that: 1. The response of MS1 to a positive shock in MS1 decreases rapidly and is cancelled after 4 months. The cumulative effect after 12 months is signicant. 2. Over the sample period an unexpected increase in A1 produces a longer decrease (over 6 months) of MS1, which is not totally recuperated after 12 months, i.e. A1 has a negative permanent impact on MS1. 3. Surprisingly, the response of the MS1 to a price shock does not appear to be signicant. A plausible explanation for this behaviour is that our brand does not compete on a price basis and even if it changes its prices, these increases will be associated with the launch of new models (versions). 4. The MS1 responses to shocks on advertising expenditures (TV, Radio, and Press) are positive,

10

The correlation matrix of the residuals shows that the correlation among the off-diagonal elements seems to be small, thus changes in the order of the variables are likely to have minor effects on the impulse response results: MS 1 A1 P1 MS 1 1 2 0.21 2 0.23 A1 1 2 0.19 P1 1 TVS 1 RS 1 PS 1 TVS 1 0.07 2 0.22 90.24 1 RS 1 0.30 2 0.18 2 0.25 0.17 1 PS 1 2 0.23 0.08 0.15 . 0.26 0.04 1

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

105

Fig. 5. Responses of MS1 to shocks on all variables.

Fig. 6. Cumulative responses of MS1 to shocks on all variables.

lagged and varying in magnitude. TV advertising effects begin after 3 months, but seem to rest for a period longer than that of Radio and Press. Our ndings are consistent with the theory of the cumulative advertising effects on sales (e.g., Palda, 1964; Clarke, 1976), and even the measurement and duration of these effects are easy to calculate.

4.5. Variance decomposition

Variance decompositions give the proportion of

the h -periods-ahead forecast error variance of a variable that can be attributed to another variable. The pattern of the variance decomposition also indicates the nature of Granger causality among the variables in the system, and, as such, can be very valuable in making at least a limited transition from forecasting to understanding. If innovations in A1 t result in unexpected uctuations in MS 1 t , then information on A1 t would be useful in predicting MS 1 t . In interpreting these variance decompositions, one should bear in mind Runkles (1987) criticism that the implicit condence intervals attached to both

106

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

variance decompositions and impulse response functions are often so large as to render precise inferences impossible. The Choleski decomposition is used for the BVAR(G) model. The variables are ordered in the following sequence MS 1 t , A1 t , P 1 t , TVS 1 t , RS 1 t , PS 1 t . This ordering is based partly on our prior belief that changes in MS 1 t precede those in A1 t , P 1 t , TVS 1 t , RS 1 t , PS 1 t and partly on the timing of the availability of data. For instance, information on prices and market shares is released before that on advertising expenditures. The variance decompositions for our six-variable model for the period 1988:1 to 1993:6 are reported in Table 2. For each variable in the left-hand column, the percentage of the forecast error variance for 1, 6 and 12 months ahead that can be attributed to shocks in each of the variables in the remaining columns is reported. Each row sums to 100% (ignoring rounding errors) since all the forecast error variance in a variable must be explained by the variables in the model. If a variable is exogenous in the Granger

sense, i.e. if other variables in the model are not useful in predicting it, a large proportion of that variables error variance should be explained by its own innovations. How large is large? According to Doan (1992), in a six-variable model such as ours, 50% is quite high. If another variable is useful in explaining a left-column variable, that useful variable will explain a positive percentage of the prediction error variance. In practice, it is difcult to distinguish between a variable that has no predictive value and one that has little predictive value. Some conclusions, however, can be derived by comparing the magnitudes. Table 2 shows that, at a forecast horizon of 12 months, only 35.6% of the forecast error variance in the MS1 is explained by its own innovations, supporting the assumption that MS1 is not exogenous, and that other variables such as A1, TVS1, and PS1 can be equally useful in forecasting MS1. The market share is extremely stable from 1 month to the next none of the other variables gure at all in its 1-month-ahead forecast. Moreover, longer-term fore-

Table 2 Variance decompositions Variable MS1 Step 1 6 12 1 6 12 1 6 12 1 6 12 1 6 12 1 6 12 MS1 100 51.01 35.6 5.87 3.95 4.08 23.86 16.25 12.69 0.54 9.34 9.7 27.64 21.79 18.15 10.47 10.3 13.98 A1 0 29.03 31.9 0 8.09 19.85 65.76 52.67 40.44 5.94 4.56 7.75 0.32 1.18 9.49 0 5.42 8.56 P1 0 3.87 3.91 0 8.09 19.85 65.76 52.67 40.44 5.94 4.56 7.75 0.32 1.18 9.49 0 5.42 8.56 TVS1 0 3.61 15.8 0 3.74 6.06 0 9.55 15.06 83.88 56.87 50.79 1.22 14.52 18.23 12.16 19.01 17.17 RS1 0 6.61 4.97 0 2.45 1.34 0 8.96 13.88 0 3.9 4.79 68.3 36.05 29.01 4.4 10.14 8.91 PS1 0 6.43 7.8 0 2.94 3.45 0 4.96 5.21 0 3.68 5.9 0 8.99 9.19 72.96 39.75 34

A1

P1

TVS1

RS1

PS1

Note: entries in each row are the percentages of the variance of the forecast error for each variable indicated in the rows that can be attributed to each of the variables indicated in the column headings. Decompositions are reported for 1-, 6- and 12-month horizons.

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110

107

casts of market shares are heavily inuenced by age, TV advertising and not much at all by price. The exogenous behaviour of A1 seems to be reected in the 65.2% error variance explained by its own innovations. The results for the price variable vary within the forecasting horizon. For instance, the market share variable seems more important at shorter horizons (1 to 6 months), while the age and the advertising (TV and Radio) variables become the largest contributors for longer horizons (12 months). There are some interesting media differences in the advertising variables. TV advertising seems much more exogenous than the other media advertising (50.79% vs. 29.01% and 34%). Long-term forecasts of TV advertising are more explained by innovations in A1 than in MS1. However, for RS1 and PS1 the innovations in MS1, A1, and TVS1 help explain most of the forecast error variance attributable to other innovations.

5. Limitations and extensions In this paper, we demonstrate the utility of VAR and BVAR methodologies as a marketing tool that fullls two requirements: it forecasts market shares, and it provides insights about the competitive dynamics of the marketplace. We compared the forecasting accuracy of BVAR with several traditional approaches. Using data, we establish that BVAR is a superior forecasting tool compared to univariate ARIMA and VAR models. Because BVAR uses few degrees of freedom and is easy to identify, it satises the practical requirements as a marketing forecasting tool. Finally, using impulse response functions and variance decompositions, we illustrate that BVAR provides important insights for marketing managers. Although BVAR is a promising and reliable forecasting tool, certain limitations should be pointed out. First, BVAR models are highly reduced forms. Structural interpretations based on the signs and magnitudes of estimated parameters should be avoided. Hypotheses about effects should be tested using impulse response analysis. Second, the accuracy of the forecasts is sensitive to the specication of the prior. If the prior is not well specied, an alternate

model such as an unrestricted VAR or an ARIMA model may perform better. Third, the prior that is selected on the basis of some objective function (e.g. the Theils U ) for the out-of-sample forecasts may not be optimal for beyond the period for which it was selected. This model, like all time series models, is best suited for stable environments (e.g. wide-sense stationary processes) where sufcient numbers of observations are available. Thus, BVAR is not a new product model and its forecasts may be unreliable in markets characterised by frequent new entries or dropouts. We propose several extensions of this initial application of BVAR to forecast brand market shares. First, the Bayesian approach can be improved by putting more structure based on marketing theory into the prior, thereby abandoning the symmetric treatment of all variables. This would make the approach more Bayesian in spirit, since the prior can now reect better the a priori beliefs of the investigator. On the other hand, the greater exibility makes it more difcult to nd the optimal forecasting model. Second, the inclusion of the contemporaneous values of some variables in some equations (using, for example, a Wold causal ordering) may result in improved forecasting accuracy due to a simpler model specication. Third, we used one specication of each model to forecast over the entire testing period. More frequent specications (e.g., a timevarying parameter BVAR model) would undoubtedly improve accuracy. An important problem deals with how often a model should be re-specied. Finally, we used single equation procedures to estimate all models. Forecasting accuracy may improve by estimating all equations in each model simultaneously and exploiting the information in the cross-equation residual covariance matrix.

Acknowledgements C. Ribeiro, the The author is grateful to Jose associated editor and two anonymous referees for helpful and constructive comments. Responsibility for any error is solely mine.

108

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110 06-92 07-92 08-92 09-92 10-92 11-92 12-92 01-93 02-93 03-93 04-93 05-93 06-93 07-93 08-93 09-93 10-93 11-93 12-93 01-94 02-94 03-94 04-94 05-94 06-94 0.183 0.151 0.143 0.143 0.160 0.164 0.161 0.175 0.147 0.164 0.159 0.134 0.121 0.147 0.125 0.128 0.121 0.145 0.184 0.126 0.175 0.132 0.147 0.113 0.119 1.23 1.57 1.60 1.84 2.28 2.18 2.25 2.37 2.19 2.24 2.27 2.54 2.57 0.91 1.05 1.05 1.05 0.95 1.04 0.95 0.94 1.02 0.95 0.91 0.97 0.98 0.209 0.241 0.443 0.211 0.135 0.166 0.145 0.149 0.159 0.091 0.170 0.097 0.069 0.072 0.114 0.071 0.197 0.130 0.143 0.111 0.123 0.061 0.061 0.089 0.021 0.025 0.171 0.161 0.134 0.196 0.164 0.162 0.171 0.212 0.180 0.173 0.166 0.079 0.089

Appendix A
Period 01-88 02-88 03-88 04-88 05-88 06-88 07-88 08-88 09-88 10-88 11-88 12-88 01-89 02-89 03-89 04-89 05-89 06-89 07-89 08-89 09-89 10-89 11-89 12-89 01-90 02-90 03-90 04-90 05-90 06-90 07-90 08-90 09-90 10-90 11-90 12-90 01-91 02-91 03-91 04-91 05-91 06-91 07-91 08-91 09-91 10-91 11-91 12-91 01-92 02-92 03-92 04-92 05-92 MS1 0.273 0.229 0.217 0.234 0.234 0.245 0.169 0.178 0.126 0.192 0.220 0.180 0.215 0.223 0.181 0.149 0.203 0.223 0.181 0.149 0.142 0.202 0.209 0.120 0.151 0.173 0.161 0.141 0.164 0.146 0.150 0.109 0.185 0.171 0.130 0.177 0.181 0.188 0.217 0.186 0.170 0.161 0.138 0.132 0.143 0.169 0.166 0.216 0.115 0.175 0.142 0.149 0.165 A1 0.97 1.06 1.14 1.39 1.35 0.64 0.86 0.40 0.77 0.97 0.73 0.62 0.84 0.74 0.56 1.00 0.62 0.52 0.74 0.88 0.54 1.05 0.91 1.44 1.27 1.12 1.14 1.38 1.13 1.30 1.02 0.54 0.26 0.40 0.34 0.46 0.52 0.57 0.59 0.70 0.76 0.76 0.80 0.82 1.66 1.88 1.93 1.43 1.72 1.12 1.11 1.23 1.36 P1 0.81 0.90 0.90 0.83 0.86 1.09 0.98 0.85 0.96 0.99 0.99 1.05 1.00 1.05 1.02 0.94 1.11 1.04 1.03 1.08 1.17 1.15 1.13 1.20 1.14 1.06 1.15 1.17 1.15 1.13 1.09 1.17 1.06 1.08 1.01 1.08 1.05 1.07 0.96 0.93 0.98 0.96 1.16 1.11 1.02 1.02 1.08 1.10 1.05 0.91 0.99 0.96 1.07 TVS1 0.325 0.372 0.246 0.189 0.257 0.338 0.389 0.375 0.265 0.326 0.285 0.233 0.193 0.120 0.083 0.147 0.170 0.205 0.260 0.244 0.198 0.280 0.216 0.185 0.171 0.165 0.102 0.142 0.092 0.103 0.084 0.297 0.327 0.184 0.143 0.186 0.170 0.277 0.121 0.084 0.128 0.192 0.246 0.264 0.134 0.216 0.175 0.102 0.160 0.090 0.121 0.181 0.142 RS1 0.176 0.223 0.132 0.122 0.001 0.117 0.068 0.176 0.214 0.168 0.208 0.157 0.253 0.196 0.143 0.145 0.064 0.224 0.154 0.133 0.148 0.196 0.184 0.164 0.001 0.026 0.047 0.067 0.001 0.035 0.001 0.054 0.157 0.058 0.035 0.083 0.001 0.001 0.001 0.100 0.097 0.185 0.001 0.001 0.089 0.071 0.214 0.116 0.205 0.102 0.169 0.092 0.001 PS1 0.385 0.333 0.258 0.214 0.169 0.193 0.315 0.189 0.190 0.128 0.200 0.320 0.376 0.256 0.185 0.169 0.124 0.075 0.100 0.124 0.201 0.143 0.097 0.151 0.113 0.085 0.150 0.233 0.190 0.141 0.092 0.115 0.151 0.183 0.197 0.216 0.299 0.277 0.110 0.132 0.130 0.147 0.090 0.186 0.169 0.123 0.153 0.063 0.057 0.098 0.163 0.265 0.188

References
Aaker, D., Carman, J., & Jacobson, R. (1982). Modeling advertisingsales relationships involving feedback: a time series analysis of six cereal brands. Journal of Marketing Research, 19 (February), 116125. Adams, A., & Moriarty, M. (1981). The advertisingsales relationship: insights from transfer-function modeling. Journal of Advertising Research, 21 (June), 4146. Ashley, R., Granger, C., & Schmalensee, R. (1980). Advertising and aggregate consumption: an analysis of causality. Econometrica, 48 (July), 11491167. Bass, F., & Pilon, T. (1980). A stochastic brand choice framework for econometric modeling of time series market share behavior. Journal of Marketing Research, 17 (November), 486497. Bhattacharyya, M. (1982). Lydia Pinkham data remodelled. Journal of Time Series Analysis, 3 (2), 81102. Brodie, R., & De Kluyver, C. (1987). A comparison of the short extrapolaterm forecasting accuracy of econometric and nave tion models of market share. International Journal of Forecasting, 3, 423437. Brodie, R., & Bonfrer, A. (1994). Conditions when market share models are useful for forecasting: further empirical results. International Journal of Forecasting, 10, 277286. Caines, P., Keng, C., & Sethi, S. (1981). Causality analysis and multivariate autoregressive modelling with an application to supermarket sales analysis. Journal of Economic Dynamics and Control, 3, 267298.

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110 Chateld, C., & Prothero, D. (1973). BoxJenkins seasonal forecasting: problems in a case-study (Series A: General) (Part 3). Journal of the Royal Statistical Society, 136, 295315. Clarke, D. (1976). Econometric measurement of the duration of advertising effect on sales. Journal of Marketing Research, 13 (November), 345357. Cooper, L., & Nakanishi, M. (1988). Market-share analysis: evaluating competitive marketing effectiveness. Boston: Kluwer Academic. Danaher, P., & Brodie, R. (1992). Predictive accuracy of simple versus complex econometric market share models. International Journal of Forecasting, 8, 613626. Dekimpe, M., & Hanssens, D. (1995). The persistence of marketing effects on sales. Marketing Science, 14 (1), 121. Doan, T. (1992). RATS. Users manual (version 4). Estima, IL. Doan, T., Litterman, R., & Sims, C. (1984). Forecasting and conditional projection using realistic prior distributions. Econometric Reviews, 3, 1100. Franses, P. (1991). Primary demand for beer in the Netherlands: an application of ARMAX model specication. Journal of Marketing Research, 28 (May), 240245. Fuller, W. (1976). Introduction to statistical time series. New York: Wiley. Geurts, M., & Whitlark, D. (1992). Forecasting market share. Journal of Business Forecasting, 11 (4), 1722. Granger, C., & Newbold, P. (1986). Forecasting economic time series. New York: Academic Press. Grubb, H. (1992). A multivariate time series analysis of some our price data. Applied Statistics, 41 (1), 95107. Hanssens, D. (1980a). Bivariate time series analysis of the relationship between advertising and sales. Applied Economics, 12 (September), 329340. Hanssens, D. (1980b). Market response, competitive behavior, and time series analysis. Journal of Marketing Research, 17 (November), 470485. Hafer, R., & Sheehan, R. (1989). The sensitivity of VAR forecasts to alternative lag structures. International Journal of Forecasting, 5, 399408. Heuts, R., & Bronckers, J. (1988). Forecasting the Dutch heavy market: a multivariate approach. International Journal of Forecasting, 4, 5779. Heyse, J., & Wei, W. (1985). Modelling the advertisingsales relationship through use of multiple time series techniques. Journal of Forecasting, 4, 165181. Hillmer, S., & Tiao, G. (1982). An ARIMA-model-based approach to seasonal adjustment. Journal of the American Statistical Association, 77 (March), 6370. Jacobson, R., & Nicosia, F. (1981). Advertising and public policy: the macroeconomic effects of advertising. Journal of Marketing Research, 18 (February), 2938. Krishnamurthi, L., Narayan, J., & Raj, S. (1986). Intervention analysis of a eld experiment to assess the build-up effect of advertising. Journal of Marketing Research, 23 (November), 337345. Kumar, V., Leone, R., & Gaskins, J. (1995). Aggregate and disaggregate sector forecasting using consumer condence measures. International Journal of Forecasting, 11, 361378.

109

Leone, R. (1983). Modeling salesadvertising relationships: an integrated time series econometric approach. Journal of Marketing Research, 20 (August), 291295. Leone, R. (1987). Forecasting the effect of an environmental change on market performance. International Journal of Forecasting, 3, 463478. Litterman, R. (1980). A Bayesian procedure for forecasting with vector autoregressions. Working Paper, MIT. Litterman, R. (1986a). A statistical approach to economic forecasting. Journal of Business and Economic Statistics, 4, 14. Litterman, R. (1986b). Forecasting with Bayesian vector autoregressions ve years of experience. Journal of Business and Economic Statistics, 4, 2538. Lui, L. -M. (1987). Sales forecasting using multi-equation transfer function models. Journal of Forecasting, 6, 223238. Makridakis, S. et al. (1982). The accuracy of extrapolation (time series) methods: results of a forecasting competition. Journal of Forecasting, 1, 111153. Montgomery, D., & Weatherby, G. (1980). Modeling and forecasting time series using transfer function and intervention methods. AIIE Transactions, 12, 289307. Moriarty, M. (1985). Transfer function analysis of the relationship between advertising and sales: a synthesis of prior research. Journal of Business Research, 13, 247257. Moriarty, M., & Salamon, G. (1980). Estimation and forecast performance of a multivariate time series model of sales. Journal of Marketing Research, 17 (November), 558564. Palda, K. (1964). The measurement of cumulative advertising effects. Englewood Cliffs, NJ: Prentice-Hall. Prothero, D., & Wallis, K. (1976). Modelling macroeconomic time series (Series A: General) (Part 4). Journal of the Royal Statistical Society, 139, 468486. Runkle, D. (1987). Vector autoregressions and reality. Journal of Business and Economic Statistics, 5, 437442. Sims, C. (1980). Macroeconomics and reality. Econometrica, 48, 148. Sims, C. (1988). Bayesian skepticism on unit root econometrics. Journal of Economic Dynamics and Control, 12, 463474. Sims, C., Stock, J., & Watson, M. (1990). Inference in linear time series models with some unit roots. Econometrica, 58, 113 144. Spencer, D. (1993). Developing a Bayesian vector autoregression forecasting model. International Journal of Forecasting, 9, 407421. Sturgess, B., & Wheale, P. (1985). Advertising interrelations in an oligopolistic market: an analysis of causality. International Journal of Advertising, 4, 305318. Takada, H., & Bass, F. (1988). Analysis of competitive marketing behavior using multiple time series analysis and econometric methods. Working Paper, University of California, Riverside. Theil, H. (1971). Principles of econometrics. New York: Wiley. Tiao, G., & Box, G. (1981). Modeling multiple time series with applications. Journal of American Statistical Association, 76, 802816. Tiao, G., & Tsay, R. (1983). Multiple time series modeling and extended sample cross-correlations. Journal of Business and Economic Statistics, 1, 4356.

110

F.F. Ribeiro Ramos / International Journal of Forecasting 19 (2003) 95110 Zellner, A., & Palm, F. (1974). Time series analysis and simultaneous equation econometric models. Journal of Econometrics, 2, 1754. Biography: Francisco Fernando RIBEIRO RAMOS is a lecturer in the Faculty of Economics at the University of Porto (Portugal) with primary research interests in modelling and forecasting of time series, with a specic focus on Marketing Econometrics, where he has published widely.

Tiao, G., & Tsay, R. (1989). Model specication in multivariate time series. Journal of The Royal Statistical Society, Series B, 51 (2), 157213. Todd, R. (1984). Improving economic forecasting with Bayesian vector autoregression. Quarterly Review, Federal Reserve Bank of Minneapolis, Fall, 1829. Umashankar, S., & Ledolter, J. (1983). Forecasting with diagonal multiple time series models: an extension of univariate models. Journal of Marketing Research, 20 (February), 5863. Zellner, A. (1979). Statistical analysis of econometric models. Journal of the American Satistical Association, 74, 628643.