ABSTRACT
Competition in the insurance market today is intense and orthodox pricing mechanisms are already failing. Any risk model overestimates the level of risk of some groups of policies and underestimates the risks of others. Overcharging low-risk policyholders drives them away to seek lower rates from competitors, thereby reducing revenue and market share. Undercharging high-risk policyholders attracts similar high-risk customers, thereby driving up claims and losses. To remain competitive, insurers must constantly improve the assessments of their true levels of risk and price policyholders accordingly.

In the U.S. auto insurance industry, many insurers have been slow to improve their modeling and pricing approaches with the new analytic tools they have at their disposal. Traditionally, the industry relied on loss ratio models, used limited external data, had a compartmentalized model structure and used variables that were difficult to obtain accurately.

Led by a few insurers, the industry is headed towards pure premium modeling at individual coverage level using extensive external data and variables that can be verified. These companies are also improving their data collection process so that they have access to richer information to make better pricing decisions. They are using advanced software tools such as SAS, CART and MARS, and new modeling techniques like Poisson and Gamma regression for accurate analysis. They are also focusing on more accurately allocating non-loss related expenses such as acquisition expenses into the pricing structure as opposed to traditional fixed load methods.

While the industry continues to face challenges such as regulatory restrictions on use of certain variables, difference across states and data integration across the organization, companies that embrace these new techniques while managing the market challenges will emerge winners in the next few years at the expense of those who continue to rely on outdated methods that can lead to further erosion of market share and adverse risk selection.

With increased data collection ability, improved database management techniques, and advanced statistical tools, companies today can mine large quantities of data and perform complex statistical analysis to drive optimal pricing decisions.
There are five key levers companies need to master to effectively leverage the power of data and analysis in order to drive profitable growth through improved pricing:

(A) DATA PREPARATION
For insurance pricing, large datasets need to be analyzed.
A large auto insurer could have gigabytes of data on millions of policies spanning over 8-10 years. Therefore, optimal data preparation and processing techniques should be followed to avoid long processing times.
There are five steps to effective data preparation:

(B) MODELING APPROACH
The traditional method for constructing risk models involved segmenting the overall population of policy holders into a collection of risk groups based on a set of factors, such as age, gender, driving distance to place of employment, etc. This assumed that the resulting segments are homogeneous in terms of risk. The introduction of credit scores was a significant first step in moving beyond class plan variables in developing different pricing tiers but mostly as overlays of existing class plans.

Over time with the development of new tools and data mining techniques, the focus shifted to loss ratio modeling. This method appears simple but is often misleading. A ratio greater than 1.00 indicates that more losses have been paid than premiums that would have been earned at current rate level, suggesting that an increase in the premium rate is needed. The converse also applies.

This method is most appropriate when the units exposed
to loss are homogenous and do not change substantially over time. But given the nature of the industry and competitive offers available to customers, the assumption of having uniform unit exposure is often violated.

The most advanced companies are now moving beyond this method towards pure premium (claim frequency X claim severity) modeling which by definition is free from this fundamental assumption fallacy¹.

At the same time, complex statistical techniques like Poisson and Gamma regression are now being used. In general, number of claims in a given time period should satisfy the following: