Bag of Tricks

Pauli doesn't want to portray predictive analytics as a victim of the economic crisis. On the contrary, she believes analytics has gained more acceptance among insurers because of the crisis. "All IT budgets are under investigation at this point as people try to figure out what's what," she says. "But among the vendors we know of with analytics solutions, discussions with claims folks are moving forward."

Joel Appelbaum, chief analytics officer for programs and direct markets with Zurich, maintains insurers in the commercial lines space are at an early stage in achieving value from modeling. "I see opportunities for green products based on predictive models," says Appelbaum.

He understands the problems insurers face, though. "You have to be somewhat concerned with investments," he says. "Legacy systems and interfacing with them are a challenge that also is holding back initiatives."

Carriers have to be willing to invest in time and effort. Also needed is a commitment from the top. "So often we rely on growing the top line by traditional marketing and increasing commissions, but the harder, more disciplined approach sometimes takes an investment in time to prove itself," says Appelbaum. "This is kind of the new kid on the block."

NEW DIRECTIONS

Insurers are discovering a variety of ways to advance operations with their analytics solutions. Appelbaum points out it can be used to enable growth, target customers and business partners, explore underwriting opportunities, and reduce expenses.

He asserts customer relations analytics offers real opportunity through the utilization of resources to know when to be there for customers. "If a model can predict when customers are most likely to have claims, the models can notify us when we should deliver risk-engineering services," says Appelbaum. "You could be delivering services at the right moment for risk engineering or premium audit. Different events can be used to model when a customer would benefit from having a physical audit on site or when a customer could benefit from a phone audit. Both would probably meet the criteria of conducting an audit, but one might cost you more to deliver and be more appreciated by the customer. Some customers would rather have a phone call. Models also can help you identify that and help you deploy your scarce resources in the most efficient and customer-friendly manner."

MARKETING MODELS

State Farm uses analytics throughout the enterprise, including in the marketing of products to policyholders, thanks to a product from SAS, according to Eric Webster, vice president of marketing for the carrier. This begins with a customer's propensity to buy another product. "We've got that down to a granular level in terms of who is likely to buy which product," says Webster.

State Farm also has branched out to explore a customer's propensity to defect to another insurer, which Webster describes as a standard analytic treatment.

The next step for the carrier is to use all the marketing models in concert with each other, targeting customers with the best possible offer to send their way, based on each policyholder's needs and desires.

"We're trying to go in the direction of looking across the suite of models and determining what is the right [marketing piece] to send them right now–based on their propensity for defection, purchases, lifetime value, and other functions," says Webster.

He indicates the exciting part of this step from the agency standpoint is there is little work for the agents. "The agents appreciate not having to worry about what is going out," says Webster. "If someone is calling in, [the agent] can look up what was sent in the past, but the agent doesn't have to try to figure out what the right piece to send to each customer may be."

NEW APPROACH FOR AGENTS

All this may sound like it would be a
shoo-in with agents, but Webster admits many agents are accustomed to what he calls the "Product A, Product B, and Product C approach."

However, from a customer standpoint, continues Webster, it allows the State Farm marketing staff to optimize what a customer is getting. "It is much more powerful and consumer relevant," he says. "We are seeing great results."

Webster also wants his team to determine how to pivot to the next product in the middle of the application stream. "As I'm typing in information–family composition, customer address, miles to work, or whatever the case may be–on the fly we need to recalculate all the propensity-to-buy models so we know when, where, and what to offer the customer, possibly during that transaction or pending a follow-up for the agent. Our customers appreciate getting material that's relevant to them and not a bunch of what they think of as junk mail. That's where we are trying to head. "

Webster reports State Farm is receiving positive feedback from the agents as to what Webster's staff is doing. At the same time, the carrier is looking to become more precise with its actions.

"We are sending out more direct mail than we used to, but we are being smart about how we do it, too," he says. "The more agents can focus on what they need to do–talking to customers and focusing on their problems–the better. Analytics have been our key for enabling a lot more automation and process streamlining for the agent to make that easy button possible."

FRAUD PREVENTION

MetLife Auto & Home has been using technology in fraud prevention for six years. The carrier scores auto and home claims upon first notice of loss and continues to score them throughout the claim history, according to John Sargent, director of the special investigative unit, MetLife Auto & Home.

Within this product are three separate functions, according to Sargent: a modeling piece that looks at the claim and prior claims the insurer knows were fraudulent; an identity search piece that looks at the data from the current claim and compares that against various external and internal data sources to see whether anything matches; and a business rules component to look for the red flags that may exist within the actual claim.

David McMichael, assistant vice president of actuarial, MetLife Auto & Home, reports the SIU was looking for yet other potential solutions for a predictive modeling component.

"Our group has developed some expertise at doing predictive modeling, starting out from a pricing and underwriting perspective–things such as proprietary credit modeling, underwriting referral models, and so forth," says McMichael. "We've been looking for more opportunities within the company where that technology can help us make better business decisions."

Tweaking the models to fit specific needs is a significant consideration for insurers because most products are custom systems, points out Pauli. "You would need someone with custom capabilities either to build it for you or there are actuaries who think they can build some good models," she says. "They have been at it long enough with reinsurance and cat models. We urge companies to go with vendors that have expertise, but some of the bigger companies with a lot of actuaries do their own thing, too."

PRICING PREDICTIONS

Predictive modeling is essential for HomeWise Insurance Group because the company's sole focus is insuring homeowners in hurricane-exposed states. "When you don't price your product correctly, bad things happen," says Dale Hammond, president and CEO of HomeWise. "Either you don't write business you'd like to write, or you write business you shouldn't be writing because it is underpriced and you end up losing money."

The modeling product HomeWise is using was developed by FICO and Millennium Information Services and currently has multiple uses for HomeWise in different states, but the long-term intention, as the carrier continues to validate the tool, is to use it in pricing.

"We use it in pricing in Louisiana and South Carolina now as part of our classification system that leads into our rating algorithm," says Hammond. "We need to continue to make sure it does what we need it to do. Our early indications are it does, and it does it well. We would like to integrate it into our Florida rating classification system, which then would feed into the algorithms."

The expectation is the information provided will be nearly as valuable as credit in classifying risk, Hammond hopes, but without some of the social issues attached, which credit scoring has.

"It's really related to the structure you are insuring, what the characteristics of the structure are, and how history has demonstrated structures with those characteristics have performed from a profitability standpoint," he says.

Understanding the catastrophe exposure under consideration is essential, points out Hammond. Otherwise an insurer is going to be over-exposed and underpriced very quickly. "We came to this by building a model that was driven by analytics–cat models, reinsurance models, all factor in to how we price our risks and how we select our risks," he says.

Pricing also is the greatest area of opportunity to leverage the benefits of predictive modeling for Pinnacol Assurance, according to Mark Isakson, assistant vice president of Pinnacol, a workers' compensation carrier. "It was somewhat of a new area for us to venture into," he says. "It also was an opportunity to learn on the front end about the benefits and implications for an insurance carrier, particularly with workers' compensation."

By examining Pinnacol's book of business and understanding the power of the tools and techniques the carrier utilizes, Valen Technologies provided the modeling expertise and Pinnacol provided the insurance expertise. "We saw it as a chance to bring some consistency to our pricing model on the front end and leverage a predictive tool as opposed to a binary or linear approach that didn't contemplate anything looking forward or any of the variables we used in pricing," says Isakson.

Pinnacol was able to evaluate its book of business and database of detail-level transactions and metrics on the policy and claims side in various combinations. "Before, we could look at something in its own environment, but we weren't utilizing the power of our own database to tell us what to expect going forward and how an individual risk might behave from an exposure standpoint," says Isakson. "What Valen was able to help us with is taking the information and doing the analysis in a dynamic environment and providing something new from several pieces of information. It gave us a better lens to view our book of business and place our exposures in the appropriate pricing mechanism."

COMPILING DATA

"I always say whoever has the most data wins," says Webster. State Farm has all the pertinent transaction data from existing policies. With that, if State Farm doesn't know specifics about an individual, the carrier will add census data substitutions.

"Between all that we end up with transaction history, policy and purchase history, demographics and psychographics–all together they form a pretty powerful combination," says Webster.

Pauli has cautioned carriers against thinking they have to have their data in perfect shape before launching an analytics solution, but organizing data and putting it into data dictionaries is an important place for carriers to start. "The more data you bring to it, the better it is," she says.

Data about an individual customer or product sits in many different places–underwriting systems, claims systems, financial systems–so data preparation is essential, maintains Pauli. She also believes cultural preparation is important, as well.

"If you are going to bring in predictive modeling and the analytics that goes with it, people are suspicious," she says. "It's not necessarily intuitive, and people think it's only there to replace human capabilities."

Pauli scoffs at such suggestions, though, pointing out the value analytics provide to support decision-making, to get rudimentary tasks off people's desks, and to move transactions along.

"I think the cultural preparation–thinking about business in different terms–is something people have to think about, as well," she says. "It's hard to change–probably the hardest thing to change–but I think it's thinking about these technologies in different ways."

Hammond cautions insurers need to use good judgment and not rely totally on software. "Data really can tell you anything you want it to tell you. So, you have to apply common sense to what you are doing and make sure, if something gives you an answer that doesn't seem intuitively correct, you go back and test it to understand if you gathered the right data or interpreted things incorrectly," he says.

One of the strengths of HomeWise since the insurer was formed in 2005 is everything it has done has been focused on getting the data correct, according to Hammond. "We don't have to go back and look at 30 years of data. We've been in a very fortunate position with a lot of help from Millennium to gather accurate data on everything we write," he says.

"We validate that in multiple ways–information from the agent, from the insured, from Millennium, or other data sources–and if we have inconsistencies, we can validate them through the local tax assessor's
department. We sometimes use our claims department to look at local exposures. We're a very data-driven organization."

DATA TRACKERS

Tracking data is as much a challenge for Zurich as it is for most carriers, notes Appelbaum. "In some instances, we have good data, but we need to organize it and store it in a way for the predictive models to utilize it," he says. For example, a carrier might collect risk-engineering reports that are written in text, a format that is difficult to mine, explains Appelbaum.

"There are numerous challenges and fantastic opportunities to get better with the data," he says. "Many times we are cobbling data from different sources–a claims system or a premium system. In one system, we may have coded effective dates in one methodology and in another system a different methodology. When we want to send those fields to predictive modeling, the data format can be challenging. We might have collected the data [internally], but we might have to reformat it."

Pinnacol was able to go back nearly a decade for historical transaction details on the policy level, claims level, and all the interrelationships between those variables, in order to get an extensive overview of what its accounts and book of business has done historically, according to Isakson.

"Valen had the horsepower and the technological savvy to process [the data] and sit down with us," says Isakson. "It was a very collaborative process once we got the data to Valen and allowed [the vendor] some time to comb through the data and tell us what it saw."

MetLife Auto & Home uses a horizontal data-mining work bench called PASW Modeler from SPSS, explains McMichael. "The functionality is there, but the real challenge is shifting from a problem where you are both the modeler and have the domain expertise to a situation where you are just the modeler and someone else understands the business better than you do," he says. "That's a strong shift in how you think about things."

Every problem the MetLife team tackles involves a new data set, points out McMichael. "All the work needed to familiarize with that data can be an iterative process with the modelers working with the IT systems people and the domain experts," he adds. "You need all three aspects to determine how to use the data in a model."

Most data used by MetLife Auto & Home is internal data. "There's not much third-party data within the model," says Sargent. "The scoring engines rely on some of the third-party data, but the model primarily is made up of our own claims history and some text-mining capability we utilize from our own data."

McMichael believes additional third-party data would be helpful in the modeling, and the carrier is working to incorporate some of it. "It is being used in the other prongs of the approach–business rules and identity search," he says. "Would it add some lift to the predictive models? Yes. But it's quite a bit of work to incorporate that data within the other internal data sources we have."

The carrier has a fair amount of claims to look through from a historical perspective, McMichael remarks, but in fraud investigations the carrier tries to use as much recent data as possible because patterns shift over time. "You have to have a fair amount of volume, as well, but it's always a fine trade-off between how current the information is and what the volume needed is for some of the data-driven approaches."

TO THE FUTURE

With his title of chief analytics officer, Appelbaum isn't sure whether such a title constitutes a trend, but he nonetheless believes it is imperative for insurers to have someone focusing on the value and the possibilities for analytics throughout the enterprise.

"Organizations that centralize analytics and use it as a competitive advantage are the ones that will succeed as opposed to the ones that have a little analysis done here and a little done there," he says. "The [insurers] that are most successful centralize it and make it a competency. That's what we're trying to do."

As far as expanding the use of analytics within MetLife Auto & Home, McMichael explains the approach within the enterprise has been ad hoc to this point. "We hear about [problems] people are looking at, and we suggest they take a look at the analytics technology," he says. "More [business users] are coming to us with problems they have. We started out drumming up business, but now we have a lot of people coming to us."