

Predicting nonpayment behavior is of considerable
interest in the financial community  spanning industries such as
banking, insurance, and retail. Corporations are interested in many
issues surrounding nonpayment, ranging from various levels of delinquency
all the way to writeoffs. Typical solutions to forecasting this
behavior involve the construction of credit scoring models using
either standard linear regression and logistic regression which
attempt to classify each individual in one of two categories. If
the payment behavior was acceptable over a certain span of time,
then the dependent variable in the model might take on a value of
one. If behavior was unacceptable, the dependent variable would
be assigned a zero value. The dependent variable is then measured
against information from the credit bureau and / or the corporation
to form a statistical profile to predict future delinquencies. These
procedures often stop at the classification step.
Sometimes, as in the case of a mail order company wanting to
forecast account writeoffs for installment purchases, the classification
process may need to go further. For example, a customer may pay
100% of the actual order, 75%, 25%, or default on the order in
its entirety at some point in time. The forecast variable no longer
is simply a dichotomous variable taking on values of zero or one,
but a continuous variable. Furthermore, this variable will be
bounded between zero and one hundred, often with many observations
at one extreme. Because of the clustering problem around the limits,
ordinary least square estimates are biased and would produce numerous
negative predicted values. This bias can impact the examination
of the true relationship between the nonpayment rate and the profile
characteristics.
Tobit regression can be used as a tool to provide a more
accurate estimation of nonpayment rates and provide benchmark
comparisons using neural networks and ordinary least squares. In particular,
it has found a home in telecommunications, banking and finance, and other
industries where estimates need to be made on the amount of dollars recovered
after an account goes into the writeoff stage of the credit life cycle. In these
applications, collection models are often built using tobit regression and used
to help in the collection process.


