Tuesday, May 22, 2012

CLV Episode 2: Segment Migration

What bucket will you be in?

The Time Series methods discussed previously compute lifetime value based on individual performance and behavior without regard to either similarity between customers or specific drivers that impact profitability.  Sometimes we can gain more insight by pooling customer behavior and segmenting customers into cohesive groups.  This allows us to either use more information in estimating future profitability or reducing uncertainty by focusing on sameness.   These approaches also begin to help marketers think about what strategic and tactical choices they might want to make from an investment point of view.

Segmentation offers two specific advantages for estimating the future:  stability and migration.
  • Stability comes from the safety of understanding how a group behaves rather than a specific individual.  And while ‘average’ is the most dangerous word in marketing; segments provide multiple averages to use.
  • Migration reflects that people can change over time and move from one segment to another.  Life-changing events are clear boundaries between segments. 
A market segment is a group of consumers with a common need and can be reached effectively using a common set of tactics.   Similarly, a profit segment is a group of consumers whose profitability, behavior or characteristics are likewise common to one another.  Since the time series methods can be applied to segments as well as to individuals, we’ll focus on potential ways to segment and then think about migration.

Profit Only Quantiles
The concept of dividing customers into equal sized groups based on some metric is common in response analysis.   We often set campaign selection rules that ‘target the top two deciles’ – or that 20% of our database most likely to respond.

In the CLV scenario, we can create segments based simply on the distribution of individual profitability across customers.   Two commonly used methods would be deciles and quartiles. 
  • Deciles: create 10 equal sized segments of customers, each with 10% of the customers
  • Quartiles: create four equal sized segments, each with 25% of the customers.   Sometimes this will be simplified into three segments by combining the middle two.
The boundaries of quantiles are driven purely by first ranking the customers and then dividing them into equally sized buckets.   This is where the Pareto rule drives from:  20% of my customers account for 80% of my profits.  This approach works when profit is a direct result of behavior, product margin is fixed and customer characteristics don’t discriminate between profit levels.

Descriptive Nodes
Rather than using profit as the sole segmentation measure, we can use all available behavioral and descriptive data to differentiate between customer profitability.   Since we’re interested in assigning people to segments, as opposed to predicting a specific outcome, this approach leverages classification and regression trees.

The following example looks at survival rate of Titanic passengers according to demographic characteristics.  If you were female or a male in a large familial group (siblings or spouses) the odds of survival were in excess of 70%.

Descriptive nodes have a couple of distinct advantages:
  • We begin to understand what characteristics and events relate to profitability because the approach splits customers according to the attribute that does the best job of discriminating between high and low.   
  • All customers are assigned to one of a set of end segments that can be reduced to simple business "if then" rules.   While often mapped to deciles, there is no reason that the business has to fit the data to the 10% rule. 
This approach works when customer attributes or life-stage relate to behavior or when contribution margin varies greatly across the product line or channel.  

Rather than splitting on levels of a specific attribute like profit as Descriptive Nodes does, clustering is a machine-learning approach that groups consumers based on their similarity to or distance from one another.   Those who exhibit similar patterns across a host of variables are grouped together while other consumers are placed into different groups with their neighbors.  

In clustering we’re trying to find homogeneous groups and then see how they differ in terms of profitability.  There are numerous techniques to choose from, but they have all the same objective: reduce customers to groups.  This approach does carry the  risk that while distinct segments may emerge they may not have different profitability profiles.

Clustering is appropriate if unusual or complex ways of identifying 'sameness' are required.  And it is particularly useful when one should segment based on implicit or explicit consumer needs.  Because needs come and go or shift in importance, the size of the segment is critical to understand over time. There may be very profitable customers in a segment but if the underlying need evaporates, they will go the way of the carrier pigeon.

Since we are interested in the future we need to come up with ways to understand where customers will end up.  All of the segmentation approaches allow us to not only to define to which segment a customer belongs, but also the likelihood of customers migrating from one segment to another over time.

If we look at historic data we can apply a segmentation scheme at the beginning and then again at every subsequent planning period.  This allows us to understand how consumers change over time and it is this information we’re most interested in.    The following is a simple view of what we’re looking for using quantiles. 

The columns represent where someone is today and the rows are the probability she will be in a particular segment next period.  This includes the idea that someone might no longer be a customer, thereby accounting for attrition.  For example, a "Bottom Third" customer is likely to still be in the "Bottom Third" but there is a 10% chance they'll improve and a 20% chance of no longer being a customer.

Because we’re dealing with uncertainty, migration should be simulated a large number of times in order to arrive at a stable estimate for each customer.  If we didn’t simulate the answer would simply be the segment with the largest probability – which in this example is the status quo.   And if we're not sure how stable the migration rates are, we can simulate them as well.

The key is to focus on the odds of changing segments over time and use that to guide the estimate of profitability.

Previous episode: Time Series
Next episode: Predictive Models

No comments: