Quickly becoming an insurance industry
best practice, predictive modeling
is already doing for workers’ comp what
credit scoring, another form of predictive
modeling, did to revolutionize pricing for
private-passenger auto insurance.
The key is the consistency it offers, said Joseph Wells, head
of underwriting for workers’ compensation and accident and
health at Zurich. “The underwriters are getting the same infor-
mation and analyzing it in a more consistent way. Consistency
really helps when managing an underwriting book.”
It also saves underwriters research time, allowing them to give
more time to customers, distributors, and less experienced under-
writers. “Having that consistent way of bringing in information,”
added Wells, “allows for a more seamless transfer of knowledge.”
Quickly becoming an insurance industry best practice, pre-
dictive modeling is already doing for workers’ comp what credit
scoring, another form of predictive modeling, did to revolution-
ize pricing for private-passenger auto insurance. But getting
into predictive modeling isn’t a slamdunk (see box on right).
To save even more, workers’ comp carriers are turning to
predictive modeling to detect fraud and improve the claims
process. “Once you get started, so many other areas present
themselves,” said Curtis Gary Dean, an actuarial professor at
Ball State University.
Looking Back With Premium Auditing
The workers’ comp industry has long held that an estimated
15 percent to 20 percent of employers misreport information to
lower their insurance costs. Premium auditing locates incorrect
information and ensure policyholders are paying their fair share.
“Typically, the way that carriers will choose which policies
to audit on-site is to first look at the premium of the policy,”
said Janine Johnson, director of analytics for Insurance Services
Office Inc. (ISO). The organization launched a premium audit
predictive modeling tool in 2010 that reveals which policyholders are most likely to need the attention of an on-site audit and
which are more likely to require a large premium adjustment.
Carriers traditionally chose a dollar-amount threshold—say
$5,000 to $10,000 or more—in annual premium and audited
those risks on-site. But “premium by itself is not a sufficient
segmentation strategy,” she said.
Thanks to predictive modeling, carriers can detect potential misreporting using other factors. “We are trying to segment
two risks with the same premium because they don’t necessarily
have the same characteristics, so you should treat them differently from the audit perspective,” Johnson said.
BARRIERS of ENTRY
There’s no predictive modeling without data gold—sufficient amounts
of clean and appropriate information necessary for revealing predictive
relationships.
“If you are trying to differentiate risk in workers’ compensation
and workers’ compensation risks are fairly diverse, you really have to
have a lot of data points around the risk,” said Janine Johnson, director
of analytics at Insurance Services Office Inc. “If you don’t, you will get
erroneous predictions.”
Workers’ comp data are just hard to come by. “Every time you
ask for something, there is a cost to that,” said Curtis Gary Dean, an
actuarial professor at Ball State University. Whatever data are collected
aren’t necessarily entered electronically, either, said Peter Wu, a
director at Deloitte Consulting.
Data extraction is a challenge. The more centralized the insurer,
the more centralized the data, added Wu. When carriers have different
pieces of information in different parts of the organization, data can
be hard to obtain, requiring numerous requests for information
technology (IT) support.
Using new technology with legacy systems also can make the
process more difficult because they aren’t as advanced or as efficient as
modern systems, he observed.
Barriers go beyond data. Investment in technology infrastructure
and talent is a requirement as well, Wu said, and that’s one reason
many companies are behind. Getting qualified IT professionals
is a barrier. “IT people know how to collect data, but they lack
experience in supporting what is needed to do predictive modeling.”
Underwriters can feel threatened that predictive modeling
will diminish their roles, said Dean. At Zurich Financial Services,
underwriters resisted predictive modeling because “changing
underwriting habits can be challenging,” said Joseph Wells, head of
underwriting for workers’ compensation and accident and health.
“We had to show the underwriters what is in it for them.”
Despite the barriers, predictive modeling will continue to
move forward, simply because the potential return on investment
is just too enticing.
Predictive modeling uses carrier and nontraditional data
from outside sources to differentiate between risks. Doing so
saves carriers money by putting resources into risks that actually
need auditing. “You’ll end up with a more accurate representation of the true exposures of the risk,” she added.
It took five years to recruit development partners and de-
velop the model, said Johnson, which ISO did in response to
carriers’ interest. “What we are trying to figure out is the most
cost-effective method for conducting the audit.”
Non-insurance data, such as wages and employment in-
formation in government databases, are part of the model to
reveal discrepancies in payroll information. “We are trying to
build a profile of a risk and look at how the risk has changed
over time,” said Johnson. The variables include location,
places of operation, payroll and wages, the claims associated
with the policy, and claims history.
Some policyholders go to “great expense” to avoid an adverse experience modification, she said. They might change
their company in name only, not in operations, to get a clean
slate for the experience modifier. Others switch between carriers to hide migration of classification codes to exchange a more
