The Tweedie Index Parameter and Its Estimator: An Introduction with Applications to Actuarial Ratemaking
Loading...
Date
2018-06
Authors
Temple, Seth David
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oregon
Abstract
Tweedie random variables are exponential dispersion models that have power unit
variance functions, are infnitely divisible, and are closed under translations and scale
transformations. Notably, a Tweedie random variable has an indexing/power param-
eter that is key in describing its distribution. Actuaries typically set this parameter to
a default value, whereas R's tweedie package provides tools to estimate the Tweedie
power via maximum likelihood estimation. This estimation is tested on simulations
and applied to an auto severity dataset and a home loss cost dataset. Models built
with an estimated Tweedie power observe lower Akaike Information Criterion rela-
tive to models built with default Tweedie powers. However, this parameter tuning
only marginally changes regression coefficients and model predictions. Given time
constraints, we recommend actuaries use default Tweedie powers and consider alter-
native feature engineering.
Description
87 pages