Cant impact on fi rm pro fi tability gupta et al 2004

cant impact on fi rm pro fi tability: Gupta et al. ( 2004 )
fi nd that 1 % improvement in retention leads to an increase of fi rm value by 5 %.
Therefore, the focus of CRM research has been to exactly identify the factors of
dormancy at a particular point in time as it helps the fi rm to design strategies to
serve the customer better in order to prevent inactivity. The primary objective of
consumer dormancy scoring has therefore been to develop methods and techniques
to precisely identify the potential attrition in a cost-effective manner.
This chapter contains contributions from S. Raja Sethu Durai, Madras School of Economics,
Chennai, India.
© Springer Science+Business Media Singapore 2016
S.N. Bhaduri and D. Fogarty, Advanced Business Analytics ,
DOI 10.1007/978-981-10-0727-9_2
19

Historically, fi rst-generation models have dominated the CRM analytics scenario,
in which the consumer-scoring methods are aimed at classifying customers into
“ good ” and “ bad ” defection classes according to their probability to go dormant by a
given time in the future. Several different modeling techniques have been attempted
to develop such classi fi er discriminant analysis, logistic regression, partitioning tree,
mathematical programming, neural networks, expert systems and genetic algorithms,
Markov chain, and to mention a few. Survey chapters by Hand and Henley ( 1997 ),
Rosenberg and Gleit ( 1994 ), Thomas ( 1998 ) provide a useful summary of such
techniques. However, despite the plethora of such techniques, logistic regression has
become the workhorse in scorecard developments. Albeit its simplicity and rea-
sonable prediction power for discrete events, logistic regression (and other similar
classi fi er) has signi fi cant disadvantages. Therefore, the fi rst-generation models have
served a signi fi cant value addition to businesses but often pose a signi fi cant con-
ceptual and implementation challenges. Firstly, these models primarily focus on the
“ outcome ” of an economic process, e.g., attrition, dormancy, payoff, or retention and
hence completely ignore the data-generating process behind the consumer behavior.
Second is the absence of endogenous “ cost – bene fi t analysis ” for the proposed model.
Most often or not, these models are implemented with an ad hoc average estimate of
cost or bene fi ts. Therefore, without an estimate of “ customer level loss or bene fi t, ” it
tends to overestimate or underestimate the potential bene fi t of the proposed model.
A major limitation of logistic regression (and other similar classi fi er) is that it
ignores the simple observation that customers go dormant with different balance
characteristics in their life cycle and hence undermines the dynamic element in
customer inactive behavior. Therefore, as we explore the dynamic element in the
inactive behavior, one questions not if a customer would go dormant, but if they,
what would be the cost to the business from loosing the customer? Though this is a
far more dif fi cult question than to provide a binary answer of yes/no, it has several