تعیین مقدار مناسب از داده ها برای طبقه بندی مصرف کنندگان برای اهداف بازاریابی مستقیم

This article examines the impact of using incremental amounts of purchasing data on the ability to classify consumers in consumer packaged goods categories for direct marketing purposes. Building on the work of Rossi, McCulloch, and Allenby (1996), who focused on the impact of three information sets—(a) demographics only, (b) demographics and one purchase made by a consumer, and (c) demographics plus an entire purchasing history of a consumer—we examine the impact of each additional purchase, starting with no purchasing information (i.e., demographics only) through 20 purchases. Using two different classification models, a Multinomial Logit model and an Artificial Neural Network model, we examine the sensitivity of classification accuracy to each additional purchase. We use these results in a profitability analysis of a hypothetical direct marketing campaign to determine the optimal number of purchases to use for classification in the category studied. The findings suggest an optimal number of purchasing observations exists for classification and targeting purposes and this optimal number falls between one purchase and a “history” of purchases as studied by Rossi et al. Our findings illustrate the importance of conducting a sensitivity analysis to identify the optimal amount of purchasing data to use when classifying consumers for the purpose of a direct marketing campaign.

Data collection in consumer packaged goods (CPG) categories has grown dramatically over the past decade due to (a) technological advances that have made it possible for marketers to collect, store, and analyze such data and (b) consumers’ willingness to let marketers collect data on their purchasing behavior. Despite this growing amount of available data, direct marketing using such data is not yet widespread.This is partially due to the expenses of using such data for micromarketing purposes, which may include administrative costs, as well as the general costs of managing and analyzing the information (Kahn & McCalister, 1997). Given these expenses, it would be useful for marketers to comprehend exactly how much purchasing information truly is necessary to understand consumers’ purchasing behavior and preferences.For example, while it may be relatively inexpensive to only use demographic information to profile a set of potential customers, andmany marketers do this, it is well-documented in the direct marketing literature and the choice literature that purchasing information,while expensive, is a far better predictor of behavior than are demographic variables alone (Fader & Lattin, 1993; Guadagni & Little, 1983;Gupta & Chintagunta, 1994; Rossi et al., 1996;Schmittlein & Peterson, 1994). As such, a tradeoff between the costs of using additional purchasing information to profile prospective customers and the benefits of improved targeting accuracy must be made in order to determine the optimal amount of information to use when classifying consumers for direct marketing purposes.However, no study has examined this problem based on incremental purchasing information.Therefore, we are interested in examining the impact of using incremental amounts of purchasing data on the ability to classify and effectively target potential customers in a CPG category.

The purpose of this article is to demonstrate the importance of determining the optimal amount of purchasing information to use when classifying consumers in CPG categories for direct marketing purposes. Our procedure advocates the following steps: (a) use the purchasing histories of consumers for which you have a substantial amount of data on which to classify their preferences,(b) determine the optimal amount of purchasing data necessary to maximize profits for a direct marketing campaign, and (c) only purchase–use this optimal amount of data when classifying a new set of customers to determine the appropriate targeted message to send to each.In demonstrating this procedure we test two classification models, an MNL model and an ANN model, to see how classification accuracy improves with incremental purchasing information on consumers and how this improvement in accuracy affects the profitability of a direct marketing campaign. We find that the ANN model consistently outperforms the MNL model in classification accuracy and therefore is the model on which most of the discussion of the results is focused.Our empirical investigation reveals that a direct marketer for the market share leader in the CPG category studied (baby diapers) can optimize returns from a direct marketing campaign by using more than one purchasing observation but less than the entire purchase history for consumer classification (three purchases to be exact). This finding is an improvement over that of Rossi et al. (1996), who found that using a “purchasing history” to guide direct marketing decisions is better than using just one purchasing observation, but who did not conduct a sensitivity analysis to each additional purchase in between these two alternative data sets as we do here. Furthermore, our results provide some interesting insights into the relationship between market share and classification accuracy.Specifically, we find that larger share brands are more capable of identifying their loyal customers with less purchasing data than are brands with lower market share. This is most likely due to the fact that brands with lower shares will have fewer observed purchases per consumers given visibility to a specified amount of purchasing information. As such, it may take more purchasing observations to identify a consumer who is loyal to a small share brand than it is to identify one who is loyal to a large share brand.Our approach is useful for a variety of different direct marketing contexts. It is relevant for managers of products in mature CPG categories who wish to buy purchasing information (from a third-party provider) on consumers in their category to target their competitors’ customers but who are unsure how much information to purchase to accomplish this task. Our approach is also useful for managers of products in new CPG product categories or categories where customers are easily identifiable as entering the category for the first time (e.g., baby products,pet products, pharmaceutical products, etc.). In these cases, a marketer wishes to determine the optimal amount of purchasing information to observe on a new prospect to know which to target, and how to target those deemed as attractive,to maximize profits from a direct marketing campaign. Finally, our work is relevant for retailers who wish to target customers with incentives to increase store traffic and profits.For example, a grocery store manager’s objective is to increase sales of high margin brands in his or her store. However, without profiling his or her customers, a manager will be unsure of whether to target an individual with a loyalty reward incentive for the brand or with a brand switching incentive. Our approach can be applied to this problem to help grocery store managers determine the amount of purchasing information to use when classifying store patrons in order to determine how to target those consumers to maximize store profits.Given the framework of our study, there are numerous extensions for future research. To begin, our data has no visibility to promotional activity. It would be nice to replicate this study on a data set that contains promotional variables to see how the results would change for a situation where marketing-prone consumers are identified. It would be interesting to focus on and investigate consumers who are incorrectly classified. If generalizations can be made about the relation between incorrect classification and true classification, then perhaps misclassified consumers can be successfully targeted despite the fact that they are misclassified. It might also be worthwhile to gain a better understanding of how the optimal number of observations changes given different direct marketing costs,response rates, or POST(SCR) values. Finally, it would be interesting to further investigate differences across brands in their ability to identify their loyal consumers given limited purchasing information and whether this is driven more by market share or brand loyalty. We hope that this research will encourage others to continue researching some of these topics.