By Shafi Rahman
In the era of Big Data analytics, it is fairly common to dismiss segmentation as an old methodology with minimal, if any, role to play in customer-centric decisioning. A majority of segmentation techniques are descriptive in nature and hence fail to capture the imagination of those who desire to use the latest techniques in modern crystal-ball gazing. That’s because people still think of relatively simple segmentation methods — in fact, there is a vital role for more advanced segmentation.
Segmentation of Old
The oldest technique is grouping customers based on their demographic traits. Slightly more sophisticated segmentation techniques use value dimensions instead of demographic traits. These involve identification of one or more value dimensions, followed by dividing each of the dimensions into bins, usually of equal volumes. For instance, a retailer could describe its customers by deciling them first on their yearly spend and then deciling them on total number of trips, thus giving 100 micro-segments. These techniques are relatively easy to develop and very useful for descriptive reporting and monitoring, but usually don’t provide insights beyond that point.
An improvement over these approaches is to use behavioral dimensions along with value dimensions. This allows identification of groups of customers who are behaviorally similar but differ in values, or vice versa. The approach necessitates using a data-driven technique to identify segments of customers, usually k-means clustering. The segments are much more nuanced and allow identification of the factors that generate customer value. Armed with this knowledge, sophisticated marketers use external stimuli to influence customer behavior to generate higher values. Still, these segments are not actionable, as they do not differentiate customers based on future outcomes. Another challenge faced is that k-means presupposes the number of segments and this often involves poor guesswork.
Segmentation in the Big Data World
Our approach overcomes these limitations using very powerful innovations. Using a genetic algorithm-based search technique, the optimal number of customer segments is identified. These segments have minimal overlap across segments and minimal separation within the segments. The segments are designed to provide maximal separation on certain predetermined dimensions, called the relevance-drivers. The relevance-drivers are used only during segment identification but are not needed for scoring. Hence, these are ideally suited for defining future performance metrics like future risk or future revenue.
Using our segments, an organization can not only identify behaviorally distinct groups which also differ in their historical value to the organization, but can also predict future desired outcomes like risk and revenue. Appropriate actions can then be identified for each segment based on their historic behavior and value, and expected future outcomes. These segments can be used in lieu of more expensive combinations of predictive models for tailor-made customer-oriented actions when the number of customers or accounts is relatively low, usually under a few million. These segments not only provide a powerful way of describing actionable segments for all sizes of organizations, but also provide a cheaper way to execute 1:1 customer-centric actions for small and mid-size organizations.
