| David Annis, Ph.D. & Associates Quantitative Solutions to Vexing Marketing Problems |
|
Announcing a New Product .
. .
Understanding "Market Structure" Our new software quantitatively describes "market structure," i.e. it determines which product attributes are most important to customers. It works like this: Customer records for each date and purchase category are scrutinized to obtain a rank ordering of attributes. For example if a customer makes two purchases of soft drink, we might see that both were 12 packs, one was Diet CokeTM and the other was FrescaTM so 12 pack would be ranked first and {Diet Coke/Fresca} tied for second. The complete customer history is an aggregate of these orderings (with ties ? ties are important, but difficult to deal with mathematically) that are used by our copyrighted, proprietary algorithms to obtain attribute rankings for each customer using extensions of the ranking and selection methods developed in my doctoral dissertation at Purdue(1) as well as my work on voting systems conducted while I was on the faculty at the Naval Postgraduate School(2). The compelling feature of modeling based on rankings, rather than on raw data and some regression model, is that rankings do not require assuming a parametric data model which likely will be wrong due to the complexity of the structure(3). That is the first stage of our proprietary two-stage system. The second stage organizes all of those household rankings of each product attribute to produce a detailed view of the broader market. This stage uses a proprietary, hierarchical classification algorithm to cluster customers in the multidimensional attribute-space according to their behavior. Technically this involves a recursive algorithm that identifies the product attribute that reduces weighted, degrees-of-freedom-corrected, average intracluster distance between households. An attribute can be really important to customers, but if it's ubiquitous, it's not a product differentiator. (For example, consider power steering on a car ? everyone expects it, so no one says they're looking for it since they assume they'll get it by default). Our methodology finds splits in the data that identify customer-similarity. Customers are similar when they like this feature but don't like that one; customers that like one or the other feature are similar. After each data-split, our algorithm looks at the remaining attributes for that branch in the tree and repeats the process until all attributes have been accounted for. The result is a robust, well-defined, customer-preference mapping that cannot be achieved using conventional analytical methods.
Product Licensing For information on product licensing, contact
Who is David Annis, Ph.D.? (1) "A New Statistical Model Combining Strength and Binary Choice, with Applications to Paired Comparison Problems." (David H. Annis, Ph.D. dissertation, Purdue University, December 2003) (2) "Wu, S. S. and Annis, D. H. (2007), "Asymptotic Efficiency of the Majority Rule Relative to Rank-Sum Method for Selecting the Best Population," Journal of Statistical Planning and Inference, 137 (6), 1838?1850. (3) Regression models are vulnerable to misspecification, under-fitting, and over-fitting. Misspecification can result, for example, by using an additive term when a multiplicative one would be more appropriate, or using a value when its logarithm is more useful. Under-fitting is caused by omitting important (but unknown) terms in the model. Over-fitting is a result of model terms that appear significant but in actuality arise only from data noise. Over-fit models appear to describe the existing data well, but exhibit poor predictive performance. Since our copyrighted algorithms do not use regression models, these problems are moot.
Copyright ? 2007 David Annis, Ph.D. |