An in-depth knowledge of customers and prospects is essential to stay competitive
in today's marketplace. Therefore, data is now often collected at a detailed
level for a large number of customers. Segmentation is the act of breaking down
this large customer population into segments, in which the customers within
the segment are similar to each other, and those being in different segments
are dissimilar. A "true" segmentation will be collectively exhaustive
and mutually exclusive, i.e. every customer in the database must fall into exactly
Segmentations are valuable, since they allow the end user to look at the entire
database from a bird's eye perspective. A common use of segmentation analysis
is to segment customers by profitability and market potential, also known as
"Share of Wallet" analysis.
There are several reasons to use segmentations in business. One can use it for
distribution channel assignment in sales, i.e. which segment will be addressed
by which channel. Or segments may be identified and used as a framework to communicate
for strategic purposes, that there are differences between customers, that could
be captured at a very high level. Ideally, one would like to know each customer
individually. Segmenting the customer base up to the level of individual customers,
lays the basis for one-to-one marketing. One-to-one marketing is the ideal marketing
strategy, in which every marketing campaign or product is optimally targeted
for each individual customer. For other concepts, like profitability, it does
not make sense, to define in the worst case 100,000 segments or levels of profitability.
Instead one would restrict oneself to e.g. four or five and assign the customers
to these clusters.
There are two possible ways how segmentations can be performed. Either humans
do the segmentation "by hand", a process also known as "profile
analysis", or the segmentation is created data-driven by analysis of the
customer data. Data-driven segmentations are performed most often by using a
variety of DM and statistical techniques, which fall into two categories: predictive
segmentations and clustering. Predictive segmentations create segments with
some goal in mind, e.g. segments of high and low value customers, according
to their buying habits of a product or service.
In contrast, clustering has no particular goal in mind but is merely a way to
give more structure to the data. Clustering techniques pull interesting commonalities
from the data, which help to organize the data. These segments do not present
customers who e.g. bought the product or turned in the rebate, but generally
have similar characteristics. For address lists it is typical, that there are
some predefined clusters e.g. households with high income and no kids. The people
that fall into one particular cluster do not necessarily buy products at a higher
or lower rate than people in different clusters. Nevertheless, the clusters'
definitions are helpful because they provide high-level data organization in
a consistent way.
In most situations in the process of customer acquisition, the goal is to turn
a group of prospects into actual customers of some product or service. In general,
there are different kinds of customers, and it may take a long time, if ever,
before a customer becomes a valuable customer. Therefore, when modeling customer
acquisition processes, one should distinguish between response and activation
models. Response models predict if a person, who got an offer will react somehow.
In contrast, activation models predict if a person will accept and fully use
the offered product or service.
Response models are among the first types of models which a company starting
with analytical CRM tries to develop. Due to the rather low time requirements
in contrast to other DM goals, response modeling is a good candidate for quick
win projects in CRM. If no targeting has been done in the past, a response model
can boost the efficiency of marketing campaigns by increasing responses and/or
reducing mail expenses. Reported figures in the literature indicate, that response
rates can be increased by up to 100 %.
The goal of a response model is to predict the "response behaviors"
of the prospects who got an offer for a product or service. The model can be
based on either the past behavior of a similar population, or on the behavior
of a subsample of the target population, which was used for a test mail, or
on some logical substitute. The response behavior defines a distinct kind of
customer action and categorizes the different possibilities so that they can
be further analyzed.
A response can be received in several ways, depending on the offer channel.
An e-mail offer for example can direct the responder to reply by e-mail, phone,
or mail. When compiling the results, it is not only important to monitor the
response channel and action, but also to manage duplicates. There are situations
in which a company may receive more than one response from the same person.
This is especially common if a prospect receives multiple or follow-up offers
for the same product or service that are spread over several weeks or sent by
Activation models are models that predict if a prospect will become a full-fledged
customer. These models are most applicable in the financial services industry.
For example, for an insurance policy prospect to become an active customer,
the prospect must respond, be approved, and pay the initial premium. If the
customer never pays the premium, he or she actually ends up costing the insurance
company more than a non-responder.
There are two ways to build an activation model. One method is to build a model
that predicts response and a second model that predicts activation given response.
The final probability of activation from the initial offer is the product of
these two models. A second method is to use one-step modeling. This method predicts
the probability of activation without separating the different phases.
Which of two methods is preferable depends on available data. If one prepares
a mailing action for acquisition of new customers, it is common practice that
one buys address lists. This lists will be combined with data about possible
prospects which one has already. If the whole data set contains a rich set of
attributes and its data quality is high, both methods are applicable. Otherwise,
one is obliged to use the two step approach.
Customer Up and Cross Selling
Cross selling is the process of offering new products and services to existing
customers. One form of cross selling, sometimes called up selling, takes place,
when the new offer is related to existing purchases by the customers. Using
DM for cross selling helps to predict the probability or value of a current
customer buying these additional products. Doing a DM analysis for cross selling
is more than doing the analysis required for single product customer acquisition
several times, i.e. for each additional product. Here, the key is to optimize
in addition the offerings across all customers and products. This achieves the
goal to create a win-win situation, in which both the customer and the company
Analyzing previous offer sequences with DM methods can help to determine what
and when to make the next offer. This allows to carefully manage offers to avoid
over-soliciting and possibly alienating customers.
Attrition or churn is a growing problem in many industries. It is characterized
by the act of customers switching companies, usually to take advantage of a
better offer or just to get some incentives. Attrition is defined as a decrease
in the use of a product or service. For bank accounts, attrition is the decrease
in balances on which interest is being earned. Churn is defined as the closing
of one account in conjunction with the opening of another account for the same
product or service, usually at a reduced cost to the consumer. This must not
necessarily be a competitor, e.g. a customer of the "traditional"
business division may switch to the e-business division of the same company.
For DM activities there are several opportunities. One type of analysis predicts
the act of reducing or ending the use of a product or service after an account
has been activated. At Swiss Life for example, the serious business problem
of life insurance surrender falls into this category. The benefit of DM would
be to build an understanding of what the factors are that indicate high risk
customers. Using the DM results, substantial money can be saved by targeting
specifically at-risk customers
Without knowing the value of one's customers, it is hard to determine what
the optimal marketing efforts would be. Data mining can be used to predict customer
profitability under a variety of different marketing campaigns. Profitability
in turn can be analyzed under different angles. First, there is the profitability
which is associated with a particular product or service, and the risks, which
may arise during the lifetime of this particular product or service. And second,
one may consider the profitability and risks of the customer life cycle.
Approval or risk models are unique to certain industries that assume the potential
for loss when offering a product or service. The most well-known types of risk
occur in the banking and insurance industries. Banks assume a financial risk
when they grant loans. In general, these risk models attempt to predict the
probability that a prospect will default or fail to pay back the borrowed amount.
For the insurance industry, the risk is that of a customer filing a claim.
The risk of fraud is another area of concern. Fraud detection models are assisting
banks for example in reducing losses by learning the typical spending behavior
of their customers. If a customer's spending habits change drastically, the
approval process is halted or monitored until the situation can be evaluated.
Net present value models and lifetime value models attempt to predict the overall
profitability of a product or customer. i.e. a person or a business, for a predetermined
length of time. The values are often calculated over a certain number of years
and discounted to today's value. Since market shares vary over time, companies
are looking for opportunities to profit more from their existing customer base.
As a result, many companies are expanding their product and service offerings
in order to cross-sell or up-sell their existing customers. This approach is
exactly creating the need for models which go beyond the net present value to
the lifetime value of a customer.