Accueil > Papers in english > 11 tips for good segmentation

11 tips for good segmentation

écrit par René Lefebure

19 août

Constructing a segmentation is a complex matter to put in place for a beginner in statistics or in marketing.

In effect, unlike a score where one creates a model for a behavior (purchase, fraud, departure, etc…) from descriptive or behavioral data, there is not one segmentation to discover but rather the “best segmentation” to choose. We don’t know what to look for but we must try to find it with techniques such as the analysis of primary components, of correspondences, of canonical forms, etc.

There are « marketing » principles for good segmentation (accessible, substantial, etc.) but they give no indication as to “how to create the segmentation”.

To help young marketers, here are several rules to follow taken from my lengthy experience.

1- Confirm the need

The first step in a segmentation is to understand the objectives and expectations of the company with regard to this segmentation : improving communication, refurbishing a product, improving sales performance, etc. Often, the goals are not formulated clearly and sometimes they differ within a same organization : the CEO does not have the same expectations as does the Director of Marketing… it is thus important to prioritize and validate the objectives (even if you have to force a decision!).

2- Select «good» data

The selection of data is a critical phase in the process which can distort the entire analysis. One has to navigate between 2 temptations :

a. laziness : take only available data and/or traditional data and thus have a segmentation that prescribes what is already known (deception upon delivery),

b. ambitiousness : take into account all the data in believing naively that the statistical techniques will sort out the « good variables” (and a host of possibilities).

The will « to do » quickly may lead to taking into account those data available immediately and thus to miss the objective of decrypting a market. The selection of “good data” also goes through a balanced approach to data collection by “concept” (client / purchase / point of purchase / activity / channels / etc…) with a weighing of the different concepts… if you have 50 variables on purchases and 1 variable on client age you will obtain a segmentation … on purchase… useless for a master file! (At most, the RFM is more pertinent).

3- Select «significant clients»

To construct a segmentation, one must attach importance to understanding « client behavior ». It is thus important to avoid over-representation in the data base of clients that have disappeared, or of client levels that are “non- capturable” with present offers. A file containing 25% inactive clients (with perfect lines of 0 or of unknowns) will lead any statistical technique to consider as representative a primary behavior of inactivity and will consider other clients as “satellites” of this behavior (the buyers). In this context, the buyers appear almost as aberrations… who become difficult to organize. To want to explain “life” as an aberration of “death” is not a good starting point. If you analyze a file with 20% inactives and 15% customers acquired through game mechanisms… you understand the risk of making a segmentation that is useless and unstable.

4- Clean the variables

There are always aberrant data which can have important impact on the direction of the analysis axes (the famous force of the axes). The “incomplete” or “aberrant” points can be the result either of input errors, the mix of heterogeneous populations : to mix “international companies” and “individuals” in the analysis of banking flows will result in a de facto link between executives with 250,000€/year and revenues of 15,000€… a fictive grouping while for example, far from the billions in Total’s flow. The

Daily life allows us to appreciate that there are importance differences in the 2 salary profiles above!

5- Select « active variables »

A step in segmentation consists in evaluating the correlations and the dependency between variables to attempt to distinguish “dimensions”. A dimension is that much stronger than the numerous variables which contribute to it. One must know how to progressively exclude certain variables, highly correlated to others, to allow other dimensions to emerge. The selection of active variables must be based on qualitative notions (levels of information), the attainment cost (age is easier to obtain than revenue), communication (the logarithm of sales revenue is more difficult to interpret than the number of children) so as to “reduce” the number of principal axes… and to bring out the “emerging axes”. This weight will shock the “maniacs of the power of representation” but “grands crus” segmentation like fine wine has a bouquet.

6-Know how to «delve into the dimensions»

Reading of the first dimensions often reveals little new information. One will obviously find … the data input and the intuitions one had before beginning. … we crunched through tons of data to extract the obvious. To be sure it is reassuring because the marketing and sales teams are not destabilized (sometimes they don’t like to be upset) but we cannot limit ourselves to the reading of the first 4 or 6 axes but must know how to look a little further to discover emerging behavior in axes 8 to 12. It is necessary to continue the investigation to identify the variables and the individuals who make up this emerging and specific behavior. Thus, to understand the “usage of the Mobile Web” in a population of cell phone subscribers … is never in the first 5 axes but to know how to read it early on allows one to stay a step ahead of the market.
7- Define the size and the number of segments

The selection of the number of groups is often guided by several constraints:

a. the ability for remembering the publics… beyond7 to 8 segments, it becomes difficult for non-experts to retain the number of groups,

b. the ability for differentiation … to create 12 segments … to put into place 2 separate advertising strategies is over-segmenting,

c. fixed costs : to have segments that are too narrow makes them inoperable from an economic perspective but, careful, if a segment represents 1% in number but 20% in revenues… it justifies itself de facto as to lose it could mean the end of the business.

So size is not only a question of number, one must also view the size of the segments on dimensions of sales, margin, cost, visits, etc.

8- Construct an understandable allocation tree

It is necessary to understand why an individual belongs to a segment so as to put in place an allocation program in the data bases (and to list the segment code in the call centers, for example) because this permits following and interpreting migrations. To be in segment A in 2007 and then to move to segment B in 2008 must be understood as a modification of one or n criteria. My philosophy for allocation has progressively improved over time with the evolution of data mining techniques, the simplest currently being the use of decision trees to build allocation programs but other possibilities exist today (comparative scores, semantic tags under SPAD).

9- Test and mesure stability

It is necessary to evaluate the stability of the segmentation over time before communicating it. It is important to test the segment’s structural distribution over 2 or 3 periods (to go from 15% in 2007 to 4% in 2008 in segment A … is suspicious) as well to understand the migrations between segments. Movements that are too important may reveal problems with the conception (and thus the operability of the segment).

10- Know how to communicate

It is important to strike the minds of potential users. To paraphrase a French periodical, the shock of “names” and the weight of “mappings”. The statistical techniques must disappear to give body and blood to the segments and the dimensions. One must know how to name them, to position them, to interpret them and to make them come alive in the eyes of the future users. A good segmentation sells itself, and its success is measured by the fact that it

Decided liked feels trihexyphenidyl happy purse forever quality.

escapes the hand of the statisticians. When the segment of the “opportunists or the eclectics” moves into the language of the sales teams… the game is won!

11- Have a guiding line

Between each step, re-read the elements expected in phase 1 : are the objectives still present? In my experience, the taking charge by n actors (with differing objectives) in a segmentation scheme may in the end lead to a “compromise” segmentation, that is to say, one that has lost all its flavor. Read and re-read the objectives!

In statistical studies, to create a segmentation remains, despite everything, « a work of art » in the noblest sense of the term because one’s spirit and soul must go into the reading and interpretation of the data.

Today, I remain sceptical before the « automatic segmentations » that form groups but do not bring answers to strategic problems.

One day perhaps the techniques will prove me wrong but for the present, I admit to having as much pleasure in creating segmentations even if “I could have gone” faster. The time it takes for a segmentation to be accepted is always much longer than its construction, but that’s another story.

Pas encore de commentaire

Faire un commentaire