Over the
past several months, many of my clients and prospective clients have
turned their attention to "mining" their customer databases.
Data mining is the umbrella term for processes designed to identify
and interpret data for the purpose of discerning actionable trends and
formulating strategies based on those trends.
As firms
scrutinize their spending on marketing activities, they begin to focus
on their data mining capability. How can they learn more about customers,
use that information to make appropriate offers to customers, and understand
which offers succeed?
My last
article focused on distributing customer information across an enterprise
for use in analysis and marketing. Information about customers is gathered
from a variety of sources across the enterprise, assembled in a consistent,
reliable and usable format, and provided at an appropriate level throughout
the firm. Once a firm begins to use customer information to make decisions,
they may begin to develop more sophisticated means of using customer
data.
Data mining,
data exploration and knowledge discovery are all terms that create an
image of the demanding and sometimes tedious search to uncover insights
that are neither obvious to competitors nor easy for competitors to
duplicate. Customer relationship management depends on data analysis
activities to uncover directions and opportunities and highlight warning
indicators for CRM initiatives.
CRM uses
data mining to understand how to reach out to and communicate with customers.
Data analyses can range from simple, intuitive determination of who
to contact, when and where to applying complex algorithms in real-time
to deliver customized responses to customer-initiated interaction. Let's
review two broad categories of data analysis and see how they might
be used to prioritize CRM initiatives.
Descriptive
Analysis
Not all
data mining relies upon complex statistical analytics. Segmentation
and clustering techniques are commonly used to group customers by shared
characteristics to highlight patterns that can be used in developing
marketing plans.
Basic segmentation
is often used to group customers by easily identified, mutually exclusive
characteristics such as demographics or product ownership or usage.
Segments can be as simple as females versus males or females over 55
years old versus those under 55 years old. As long as the grouping leads
to insights, which can be used to drive marketing initiatives, it can
be a segment.
"Clusters"
is often used to describe mutually exclusive sub-segments according
to a list of pre-selected characteristics, usually those thought to
be key indicators of consumer behavior. Large firms often use geo-demographic
clusters to target brand marketing. Some firms use "value clusters"
to drive marketing activities based on the current or potential value
of a customer group.
Non-exclusionary
segments require more sophisticated analytic techniques and allow customer
behavior to drive the creation of segments. In a non-exclusionary segmentation
schema, a customer may spend as if affluent on one product type, say,
travel, and not spend at all on associated products such as room service.
These spending patterns might place the customer in two segments.
Other types
of descriptive analyses include market basket analysis, which links
products together based on customer purchase behavior, and clickstream
analysis, which uses behaviors such as web browsing, site path, shopping
and shopping cart abandonment to describe customer activities on a given
web site.
Predictive
Modeling
Predictive
modeling is a powerful data mining tool using statistical methods to
compare and contrasct customers on a wide variety of factors. Predictive
modeling determines which factors are highly correlated and measures
the degree of correlation and statistical reliability. The result of
a predictive model is a mathematical formula or score that may be applied
to customers to predict likely behavior.
There are
several common types of predictive models. Univariate models test a
single factor against a series of other factors to see which has the
highest correlation. For instance, product purchase may be tested against
age, income, computer usage, pet ownership or any other factor to discover
which attribute has the highest association.
CHAID or
CART analyses create decision trees of the most predictive attribute
combinations by testing multiple factors against each other. These tree
analyses are popular for their easy-to-describe, visual output relating
predictive attributes. Each attribute adds branches to the tree. For
instance, branches predicting product purchase may include age groups
of under 25, 25 to 55, 55 and older. Each age branch will have a percentage
associated with it, such as "the under 25 node (or cluster) has
a 60% likelihood of purchasing the product."
Multivariate
regression analysis tests multiple factors against one another to generate
a score that indicates the probability of displaying the targeted behavior.
In a multivariate regression, several attributes will be combined to
predict the outcome. For instance, product purchase may be highly correlated
with age, somewhat correlated with income and negatively correlated
with computer use. Each of these attributes will be required for every
customer you wish to score.
A neural
network or neural net is a type of sophisticated analysis that imitates
the workings of the human brain by learning from each observation. Like
a multivariate regression, a neural net generates a score that indicates
the probability of displaying the targeted behavior. Neural nets are
often run in conjunction with other predictive modeling techniques,
as the analysis performed is pretty "black box" and the results
difficult to explain.
Using
Predictive Models
Models
can be used to predict response to a targeted offer. Individual customers
or businesses may be scored on their likelihood to respond to an offer.
The model scores may be used to run economic and what-if scenarios.
Risk models
may be used to determine the likelihood of default or non-payment and
they typically relay on credit bureau data. These models require a fairly
long time frame to validate. Attrition models also require a longer
time horizon to validate. These models identify customers at risk of
defecting.
A simple
segmentation, requiring significant effort understanding customer value,
may be one of the most effective ways to use descriptive analyses and
predictive modeling. Firms can create a two-by-two matrix and assign
customers to a quadrant based upon their current and potential value.
CRM initiatives may be organized around the customers in each quadrant.
Quaero LLC calls this customer value segmentation strategy the MUST®
segmentation.
Quadrant
I: High current value/high potential value - Maintain. Depending
upon the industry, the most profitable 10% of customers may represent
between 50% and 80% of a firm's profits, so losing a customer from this
group may be very costly. On-going retention or loyalty efforts should
be aimed at the customers in Quadrant I.
Quadrant
II: Low current value/high potential value - Upgrade. These customers
may increase in value through cross-selling and account management efforts.
Perhaps these customers have not received appropriate offers in the
past or they may be delaying purchases. Efforts should be aimed at increasing
the depth and breadth of the relationship of each Quadrant II customer
with the firm.
Quadrant
III: High current value/low potential value - Study. In some segmentation
matrices, the recommended strategy for Quadrant III customers is to
milk them for current revenue. We recommend studying these customers
to determine which ones can be converted to profitable clusters in the
future and how they can be converted.
Quadrant
IV: Low current value/low potential value - Table. We assume that
you cannot focus on every segment at once, so we suggest tabling Quadrant
VI customers while your firm works to improve relationships with customers
in other quadrants. Some experts recommend proactively ending relationships
with Quadrant VI customers, others focus on information gathering to
determine ways to convert unprofitable customers to profitable customers.
The combination
of good customer information, data mining, and technology enables companies
to better understand their customer base and communicate with them more
effectively. Once a firm is actively using customer information to make
decisions about how, when and what to market to customers, they often
increase the volume of targeted customer contacts. This increase leads
many firms to look for new ways to automate mining and marketing processes
to make the most of their newfound learnings about customers.