Profile: What are the limitations of Profiling?

Only categorical variables can be used to profile i.e. variables that are presented with a coded pick list. This includes Date, Numeric and Currency variables that have been banded.

When choosing variables you need to be aware of relationships within the data.  As an example,  the fields 'Town', 'Postcode' and 'Economic Region' are correlated. They are all representations of the location of a business. In general, when choosing which variables to use you should only include one of each of these sets of fields.

If you use multiple correlated fields, you multiply the strength of that information in the PWE calculation and bias the result.  

The most likely problem with profiling is that you do not distinguish between true characteristics of the set of records you profile and correlations that are present because of the way the data has been collected or processed.

Profiling always needs a set of records to start from – it is not a self starting process.  In marketing terms, you must start the process by making an offer to some potential customers before you can profile the responses and build a model to use in the next round of marketing.

Profiling is a regressive technique when used with modelling and selection – it will always look for a subset of characteristics that apply to the data you analyse.  Beware that you do not end up spiralling into a smaller and smaller set of prospects and lose sight of the bigger picture.  Some marketers pull other groups (for example a random selection of low scoring records) into their promotions in order to introduce “new blood” into the profiling and marketing process.

Profiling is limited by the base universe data and by the available variables.  For example, the most useful variable might be “how likely are you to change bank”, but this variable is unlikely to be available.

 

Return to Profile: Overview