\magnification=1200
\baselineskip=20pt
\nopagenumbers
\font\big=cmr12 scaled \magstep2

\centerline{\bf STANFORD UNIVERSITY}
\centerline{\bf DEPARTMENT OF STATISTICS}
\centerline{\big SPECIAL SEMINAR}
\bigskip

\baselineskip=12pt
\centerline{Host: Trevor Hastie}
\bigskip

\centerline{2:15 p.m., Friday, November 5, 1999}
\centerline{Sequoia Hall Rm. 200}
\centerline{}

\bigskip
\baselineskip=15pt
\centerline{\sl Edwin P. D. Pednault}
\centerline{\sl Mathematical Sciences Department}
\centerline{\sl IBM T. J. Watson Research Center}
\centerline{\sl Yorktown Heights, New York}
\bigskip
\centerline{\bf A Machine Learning Approach to Insurance Risk Modeling}
\centerline{\bf }
\bigskip


The IBM DecisionEdge for Insurance product announced in November 1998
includes software for constructing tree-based models of insurance risk for
personal lines property and casualty insurance.  The algorithms that are
employed differ from other tree-based modeling techniques, such as CART,
CHAID, C4.5, SPRINT, etc., in that insurance risks are modeled using joint
Poisson/log-normal statistical models in the leaves of the trees that are
constructed, and actuarial credibility constraints (i.e., confidence bounds
on parameter estimates) are used to guide the tree building process.  The
joint Poisson/log-normal
models take into account insurance-specific concepts such as earned
exposure (i.e., the amount of time that a policy is actually in force
during a given time interval), and the status of each claim (i.e., whether
a claim is closed and paid versus open with final action pending).  Neither
of these concepts are accounted for in standard tree-based algorithms.  Our
method is thus tailored to the specific statistical properties of insurance
data, and it has so far yielded risk models that are more predictive of
actual risks than those
obtained using conventional tree-based algorithms.

Edwin Pednault joined IBM in 1996, where he is a Research Staff Member in
the Data Abstraction Research Group.  From 1986 to 1995, he was a Member of
Technical Staff at AT\&T Bell Laboratories.  Edwin received a Ph.D. in
electrical engineering from Stanford University in 1987, a M.S. in computer
science from Stanford in 1981, and a B.Eng. in electrical engineering from
McGill University in 1979.  His current research interests center on
statistical learning theory and its application to automated predictive
modeling.

\bye