If you wait until the last minute to
buy car insurance
, are you a bigger risk than a customer who purchases it 10 or more
days in advance?
Are 16-year-old drivers in Miami more or less likely to crash
their cars than 16-year-old drivers in Minot, N.D.?
Does a history of filing a lot of auto insurance claims mean
you're also likely to file home insurance claims?
These are among countless questions insurers are considering as
they search for new clues to predict risk. The process is part of
an evolving science called predictive modeling, which has
revolutionized how the industry prices auto and home insurance
"Its intent is to predict the future by making mathematical
sense of the past," says Steven Armstrong, a fellow of the Casualty
Actuarial Society and global chief actuarial officer for consumer
insurance at Chartis Insurance.
"Predictive modeling" uses statistical and analytical techniques
powered by technology to sift through millions of pieces of data.
It looks for patterns and clues that indicate how likely customers
are to file claims.
The mathematical theories that make predictive modeling possible
have been around for decades, says Eric Huls, vice president of
quantitative research and analytics at Allstate. "But there wasn't
a machine powerful enough to apply them to large data sets. What's
really advanced is computer power."
Add to that the stiff competition among insurers and the growing
amount of detailed data available, and you have the perfect
conditions for predictive modeling to take off in the last decade,
says Claudine Modlin, a senior consultant at Towers Watson, a
global research and consulting firm.
"Today data is literally everywhere," says Armstrong.
If insurers can't collect certain information directly from you
they can buy from vendors. Armstrong says he's heard of predictive
modelers scouring data about what people buy at retail stores to
see if there are any connections between shopping habits and risk.
He's unaware of any insurance companies actually using that data to
Predictive modeling gives insurance companies the ability to
consider a massive number of variables in more combinations than
they could a generation ago. Back then, Armstrong says an insurer
would probably base an auto insurance premium on 15 different
characteristics. "Today 40 or 50 variables might be considered."
(Here's more on
how insurance companies decide if you're a good
Insurers are also using data with greater detail, such as the
specific birth dates of customers versus broad age ranges, such as
55-plus, Modlin says. In home insurance, some companies are looking
at how characteristics of the neighborhoods and houses -- square
footage, number of bathrooms, whether the plumbing has been updated
-- can predict losses.
Traditionally insurers have used that type of information to
determine how much home insurance coverage a customer needs, Modlin
says. Now they're looking at how such data might correlate with
risk. Does the median age of residents in the neighborhood predict
theft risk, for instance? What about the neighborhood's
The science also allows insurance companies to isolate the
correlation of each variable with risk, and then consider how that
correlation might be different in certain circumstances, Huls
For example, an insurer can look at how the combination of your
age and where you live correlates with risk, Huls says. With the
ability to isolate variables as well as examine many combinations
of them, insurance companies can set prices that more closely
reflect the likelihood that you'll make an insurance claim.
All of this attention to detail may seem intrusive, but
improving pricing accuracy is good for you, as well as the
industry, Modlin says.
"Nobody wants to pay more than they have to, and at the very
least they want it to be commensurate with their risk," she says.
"You don't want to be a good driver who's paying more to subsidize
a bad driver."
Before predictive modeling, rating structures were similar among
insurers, Modlin says. Now the structures are much more complex and
sophisticated, and it's not as easy for one company to understand
another insurer's rating methods.
Looking for the next big predictor
Predictive modelers can approach their work in a couple of ways.
They can create hypotheses for what predicts risk and test them or
feed data into computers and let the machines identify
Insurance companies are on the hunt for anything new that can
help them price insurance more accurately. With that information,
Modlin says, insurers can attract low-risk customers from the
competition, which boosts their profits.
"It's like trying to find the needle in the haystack," Armstrong
says. "The fun of predictive modeling is finding that needle."
Among predictive modelers' biggest finds was the connection
between credit history and the risk for filing auto and home
insurance claims. Actuaries don't have to show how one factor
causes another, only a correlation.
use of credit history for pricing insurance
is controversial. Insurers say customers with poor credit will file
more claims. But some consumer advocates say the practice is unfair
to people who have suffered financial setbacks and that it
disproportionately affects low-income people and minorities. Some
states have put limits on using credit history to price
The industry learned an important lesson from the backlash,
Modlin says. Even though the connection between bad credit and risk
is clear, insurers must do better communicating internally and with
the public about how they use such information.
A new source of data for predictive modeling is real-time
driving behavior collected through usage-based insurance programs.
Customers who enroll in usage-based programs, such as Progressive's
Snapshot, agree to plug in a telematics device in their cars, which
records key driving habits, such as mileage, frequency of hard
braking and time of day when driving. Customers with less-risky
habits earn discounts on their car insurance rates.
"Predictive modelers are going gaga over telematics because it
provides so much data, and the data is so rich," Armstrong
Insurance companies are expanding the role of predictive
modeling beyond pricing and into other areas, such as marketing,
claims handling and fraud prevention. Which claims should be
investigated for fraud? How will getting a damaged car to a body
shop one day sooner affect the size of the claim? Which customers
are most likely to shop for a new insurer when their premium goes
The questions are endless.