is the process of mathematically combining data about
past performance to make predictions about future events.
A simple example of such a model is a baseball player's
Data about past performance (times at bat and number
of hits) is combined into a mathematical formula (hits
divided by times at bat) to estimate the probability
that the next time at bat will be a hit.
This batting average tells you what you can expect
from a player "on average" for their next
time at bat. It also allows you to determine which of
two players has a better chance of hitting the ball
the next time they are up.
An insurance score works the same way. Data about
past performance such as number of previous MVR violations,
length of time since most recent claim and deductible
amount is combined in a formula to determine the expected
loss ratio of a policy. Such a loss ratio score allows
you to determine which of two risks is the better.
Underwriters have been building models since the start
of insurance. They have taken information about past
performance and made assessments as to what is likely
to happen during the period of the policy. This process
becomes Analytics when statistical theory, mathematical
formulas and computers are used to process massive amounts
of data to come up with predictions or scores.
Models can be built to predict
· Loss Ratio
· Claim Frequency
· Professional Liability
· Propensity to Renew
· Propensity to Churn
· Probability of Responding to a Marketing Campaign
· Probability of Fraud
· Probability of MVR violation
Modeling Case Study
Consider the costs and benefits of ordering motor
vehicle reports (MVRs) on drivers as a part of the underwriting
process. Ordering an MVR on every driver will ensure
that every sur-chargeable offense will be found, but
the ordering cost will be incurred on every driver.
If an MVR score is used to predict who is likely to
have a sur-chargeable offense, all policies or applications
could be scored and those least likely to have such
offenses would not have MVRs ordered. A typical MVR
model could result in not ordering on about 7.5% percent
of all policies. The savings from not ordering on this
group is greater than that missed on sur-chargeable
offenses. This savings drops immediately to the bottom
This same score could be used by a carrier that only
orders on some of their policies. In this situation,
every policy would be scored in the MVR model and only
those most likely to have violations would be ordered.
In this way the "hit rate" for those with
violations would be sharply increased while ordering
costs could be kept constant.