پیش بینی سهم بازار با استفاده از مقادیر پیش بینی رفتار رقابتی : نتایج تجربی بیشتر

Forecasting is an important marketing activity for evaluating the expected performance of alternative marketing plans, especially in order to predict earnings, sales or market shares. The purpose of this paper is fourfold. Firstly, we develop and evaluate alternative econometric approaches to predict competitors’ future actions. Secondly, the forecasting performance of attraction models is compared to those of linear and multiplicative market share models not only if competitors’ actions are known a priori but also if competitors’ actions are forecasts. Thirdly, the effects of alternative structural specifications of attraction models on the forecasting accuracy are investigated. Finally, we reinvestigate the impact of OLS estimation versus GLS estimation on the forecasting performance. The adopted empirical methods account for the interdependence of marketing instruments. We also allow for competitive reactions up to 10 periods ago and introduce a new approach concentrating on so-called marketing events characterizing directly the contemporaneous choice of several promotional activities within a brand. Analyzing weekly scanner data from three markets we find that attraction models outperform the share predictions of the linear and multiplicative models even if competitors’ actions are forecast. This result is valid on the market and brand level. In addition, response models outperform the naive model on the market level irrespective of whether competitors’ actions are known a priori or if they are forecasts. On the brand level the superiority of response models over naive models diminishes though it still exists. With respect to the best method of predicting competitors’ actions it turns out that parsimonious specifications like autoregressive price predictions or binary logit models perform conveniently.

Forecasting is an important marketing activity for evaluating the expected performance of alternative marketing plans, especially in order to predict earnings, sales or market shares. In the last two decades an intense debate has focused on the advantages and benefits of using market share response models to forecast market shares. Brodie, Danaher, Kumar and Leeflang (1998) have developed principles to guide market analysts. They summarized the key criteria for judging under which circumstances market share response models are useful for forecasting. Strong support exists in favour of market share response models over the naive model when: (1) the sample size is rather large with approximately 100 periods for calibration and validation; (2) when strong current effects of the marketing instruments are present; (3) when market share models with brand-specific response parameters are estimated; and (4) when store-level scanner data (disaggregated data) rather than aggregated market-level data are used. In addition, a recent study by Kumar (1994) indicates that an attraction model with brand-specific parameters for each marketing instrument outperforms linear and multiplicative models and that GLS estimation should be preferred to OLS estimation if the data are not homoskedastic. When competitors’ actions are forecast, however, the results are not that clear. For example Kumar (1994) shows that the performance of market share response models is superior to that of naive models even if competitors’ actions are forecast. However when autocorrelated errors and heteroskedasticity are introduced to the values for competitors’ variables, the performance of naive models appears to be better than that of econometric models. When introducing smaller errors, the performance of naive models is comparable with attraction models estimated by GLS and better than linear and multiplicative models. With larger errors in competitors’ variables, the performance of naive models is better than all types of market share response models. In contrast to these results Brodie and Bonfrer (1994) show that when competitors’ actions are forecast, a market share response model does not consistently outperform the naive model on the brand or market level (in 10 out of 20 cases the naive model shows higher forecasting accuracy). Brodie and Bonfrer use store-level scanner data and linear and multiplicative market share response models with brand-specific parameters. In contrast to this, the study of Alsem, Leeflang and Reuyl (1989), which is based on bimonthly data for six brands from three markets, reveals that using predicted values of competitive marketing behavior may provide better market share predictions than when using observed values of competitive behavior. Their results are surprising because predicted values of competitors’ actions by definition contain a prediction error. In addition to that, the fully specified market share response model fails to outperform the naive model even when competitors’ actions are known a priori. With reference to this Danaher (1994) shows that a naive model is likely to be preferred in most market share forecasting situations when competitors’ actions are forecast. He has developed a criterion to evaluate the conditions when market share response models are useful for forecasting on the assumption that competitors’ actions are also forecast. The performance of market share response models in contrast to naive models depends on the number of observations and parameters, the number of brands in the market and the fit of the market share response model. With respect to the results of Brodie and Bonfrer, 1994 and Danaher, 1994 and Kumar, 1994 and Brodie et al., 1998 conclude that there is a priority for research to investigate alternative procedures of forecasting competitors’ actions. On the basis of these results the purpose of this paper is fourfold: 1. To develop and evaluate alternative econometric approaches to predict competitors’ future actions. Time series models and econometric models that account for the contemporaneous choice of other marketing instruments are applied to predict competitors’ actions. We also allow for competitive reactions within the last 10 weeks. Adding additional insight into the lag structure of competitive reactions and their impact on the forecasting accuracy is important because Leeflang and Wittink (1996) and Brodie et al. (1996) have shown that manufacturers and retailers tend to under- or overreact with their price and promotion activities to competitors’ actions. In addition to that, we introduce a new approach that explicitly allows for the modeling of the contemporaneous choice of several promotional marketing instruments within a brand by transforming the single marketing instruments to marketing events. 2. To compare the forecasting performance of attraction models to that of linear and multiplicative market share models both when competitors’ actions are known a priori and also when competitors’ actions are forecast. The benchmark model is a naive model. There are two reasons to do this. Firstly, if competitors’ actions are known a priori, the forecasting results can be directly compared with previously published research. Secondly, if the competitors’ actions are forecasts we must test how alternative response model specifications affect the forecasting accuracy. Following Kumar (1994, p. 296) it is important to understand the consequences of errors in the predictor variables. The theoretical superiority of attraction models to linear models can diminish in this case because errors in predictor variables get multiplied in attraction models, compared to being added in linear models. The analysis is performed on the market and on the brand level. The analysis on the market level provides information about the general forecasting accuracy of the market share models whereas the analysis on the brand level will guide the selection of the market share models with respect to brands’ market shares and brands’ promotional activities. Models that perform well on the market level may provide poor market share predictions for a certain brand (and vice versa). In addition, the analysis on the brand level will report the cases in which the market share predictions of a naive model are outperformed by a market share response model. 3. To investigate the effects of the structural specification of market share attraction models on forecasting performance. Kumar (1994) shows that market share models with brand-specific parameters for each marketing instrument generally outperform simple effects models. But Kumar does not consider structural specifications that allow for cross-competitive effects. Especially in consumer goods markets asymmetric cross-competitive effects may influence the competition between the brands. These effects are documented in several studies, for example by Allenby and Rossi, 1991, Bemmaor and Mouchoux, 1991, Blattberg and Wisniewski, 1989, Carpenter et al., 1988, Cooper, 1988, Grover and Srinivasan, 1992, Krishnamurthi and Raj, 1988 and Krishnamurthi and Raj, 1991. In connection with this, Chen, Kanetkar and Weiss (1994) compare the forecasting performances of fully crossed attraction models with those of differential effects attraction models. Their results, which are only valid if competitors’ actions are known a priori, reveal that the forecasting performance of more complex (fully crossed) attraction models is inferior to that of differential effects models. While this result is not surprising with respect to the sensitivity of fully cross-effects models to collinearity this study provides an alternative cross-effects model that allows for only a limited set of a priori determined cross-effects. 4. To reinvestigate the impact of OLS estimation versus GLS estimation on the forecasting performance. Again, if competitors’ actions are known a priori our results can be directly compared to previously published research. However, if competitors’ actions are forecast the impact of OLS or GLS estimation on the forecasting accuracy is still unclear. The paper is organized as follows. After an assessment of the previous research on the forecasting performance of market share models in Section 2, our research approach is outlined in Section 3. We discuss alternative approaches to predict competitors’ actions and competing market share response models that are used for predicting market shares. We briefly consider criteria for the comparison of forecasting methods and provide a short description of our data. The empirical results are then presented in Section 4. Results on forecasting competitors’ actions and market shares are discussed separately. With respect to the latter we distinguish competitors’ actions to be known or to be forecast. The forecasting accuracy is evaluated both on the market and on the brand level. We take the naive model as a benchmark specification and we also investigate the impact of OLS and GLS estimation on the forecasting performance. The paper concludes with the key results and an overview of future research in Section 5.

The aim of this paper has been to further investigate the conditions when market share response models are superior to naive models in terms of forecasting accuracy. We focus our attention especially on the performance of market share response models when competitors’ actions are also forecast. With the exception of the work by Alsem et al., 1989, Brodie and Bonfrer, 1994 and Danaher, 1994 and Kumar (1994) this realistic assumption has not been considered beforehand. In our paper alternative approaches for predicting competitors’ prices and promotional marketing instruments have been developed and introduced to the problem of forecasting market shares. Time series models as well as econometric models that account for the contemporaneous choice of marketing instruments have been applied to predict competitors’ actions. Additionally predictions of competitors’ actions have been considered for the case of competitive reactions. We allow for competitive reactions within the last 10 weeks by taking the results of Leeflang and Wittink (1992, 1996) as well as of Brodie et al. (1996), who show that manufacturers and retailers tend to overreact or underreact with their marketing instruments to competitors’ actions. This approach to predict competitors’ actions is new in this research field. Furthermore, we have introduced a new approach for transforming the contemporaneous choice of several promotional marketing instruments to so-called marketing events. With respect to the performance of the different approaches to predict competitors’ actions within each marketing instrument we found the following results: 1. Simple approaches like autoregressive price predictions or binary logit models provide accurate forecasts of competitors’ actions. 2. If the relative competition across markets increases and if in this way the promotional intensity increases, as is the case in our Markets A and B versus Market C, the event specification appears to be an appropriate competitor. With respect to the forecasting performance on the market level we found the following results: 1. Market share predictions of the response models outperform those of the naive model, irrespective of whether competitors’ actions are known a priori or if they are forecast. 2. In case competitors’ actions are known a priori the linear model is consistently outpredicted by the multiplicative model and the attraction models. 3. The comparison of the forecasting accuracy of the multiplicative model with that of the attraction models reveals that models’ performance is determined by market characteristics, i.e. by the promotional intensity and share variability. The multiplicative model and the differential effects model perform equally well and dominate an attraction model with cross-effects in markets in which the relative promotional intensity is lower than in the other markets. The CCHM model provides the most accurate forecasts on the market level in those markets in which the promotional intensity is higher than in the other markets, indicating that this modeling approach can better represent the competition within the market. 4. If competitors’ actions are also forecast simple approaches to predict the marketing activities of the competitors appeared to be sufficient in order to provide useful forecasts. As such the autoregressive price predictions and the usage of binary logit models outperform complex approaches that account for the contemporaneous choice of marketing instruments. 5. Modeling approaches that allow for competitive reactions within the last 10 weeks do not improve the forecasting accuracy on the market level. With respect to the forecasting accuracy on the brand level we found the following results: 1. Market share response models clearly outperform the brand-specific share predictions of the naive model in those markets that exhibit greater promotional intensity. However, the advantage of the response models over the naive model — though still existing — diminishes in our third market, which has the lowest promotional intensity across our three markets. If competitors’ actions are forecast the market share response models still outperform the predictions of the naive model, but the naive model shows itself to be slightly better than if competitors’ actions are known a priori. Again the level of promotional intensity is a moderator for the performance of alternative strategies to predict competitors’ actions. 2. Simple approaches like autoregressive price predictions and binary logit models to predict promotions provide sufficient input for market response models to estimate future market shares in markets with relatively lower promotional intensity. But the multiplicative model using econometric approaches that account for the contemporaneous choice of marketing instruments and causal reactions within the last 10 weeks performs better than using simple approaches to predict competitors’ actions in those markets that exhibit more promotions. However, this result cannot consistently be observed with all market response models. It is even inverted with the cross-effects attraction model. The comparison of the forecasting accuracy across models reveals that in general terms the differential effects attraction model provides good brand-specific market share predictions even if competitors’ actions are forecast. Our findings are in agreement with Kumar (1994) but in contrast to Brodie and Bonfrer (1994), who do not demonstrate an overall advantage of market share response models over the naive model if competitors’ actions are forecast. However, Brodie and Bonfrer (1994) consider only linear and multiplicative response models, whereas this study also investigates the forecasting performance of attraction models, a differential effects model, and an attraction model which additionally allows for potential cross-effects. On the whole the attraction models outpredict the share predictions of the linear and multiplicative model even if competitors’ actions are forecast. This result is valid for the market and the brand level. Hence, these results do not confirm Kumar’s explanation that the theoretical superiority of the attraction models over linear models can diminish because errors in predictor variables get multiplied in the attraction model compared to being added in the linear model (Kumar, 1994, p. 296). A general advantage of the more fully specified attraction model, the CCHM model, is not consistently established. The promotional intensity appears to be a moderator of the model performance. The differential effects model outperforms the CCHM model in those markets in which the promotional intensity is lower. In addition to that, we can ascertain that GLS estimation improves the forecasting performance on the market and brand level, though we cannot justify this result in Market A. On the whole, this study adds additional insight into the forecasting accuracy of market share response models. We focussed our investigation on the forecasting accuracy when competitors’ actions are forecast. In general, simple approaches like autoregressive price predictions or binary logit models are suitable to predict competitors’ actions. However, the comparison of the forecasting accuracy on the market level and on the brand level reveals some contradictory results across markets. The discussion of the empirical results emphasizes that the promotional intensity within a market may be a moderator for the performance of alternative market share response models. Future work on the forecasting accuracy of market share response models should investigate the forecasting accuracy of the market share models in markets that exhibit differential promotional intensity and that may be characterized by a differential degree of competitiveness. Future work may also establish the forecasting performance of different models where competitors’ actions are also forecast and when longer forecast periods such as six or eight weeks are chosen. The results of the present study can provide valuable guidelines for selecting models to predict competitors’ actions.