Two-thirds of insurers say their claims systems are not capable of collecting and analyzing data to improve decision-making, according to the Accenture "North American Claims Investment Survey." That's a particularly troubling statistic, as claims can consume as much as 80 percent of earned-premium dollars.
Worse, despite the vast amounts of data insurers accumulate, the industry frequently is criticized for its perceived inability to access, make sense of, and act upon that information. This can lead to lost opportunities and perpetuating inefficiencies in core business processes.
While many insurers continue to struggle with their legacy systems and accessing the data they contain, a handful of insurers-Mercury Insurance Group and Pinnacol Assurance among them-are applying predictive analytics not only to underwriting, but to claims, marketing, agency management and beyond, to reduce costs and increase profitability.
These second-generation predictive analytics applications are accessing core systems, data marts and warehouses, external data and synthetic data, and disappearing from view as they become more integrated with core applications, business intelligence and workflow or migrate to the cloud.
"The companies that are data-driven and using predictive analytics for risk selection and pricing are winning in the marketplace. It helps them in terms of profitability and market share," says Brian Stoll, senior consultant and director of Towers Watson's P&C predictive modeling practice. "One of my former bosses once said, 'Our greatest competitive advantage is that our competitors are other insurance companies.' They haven't changed and don't see any need to change. And that sort of complacency is making its way out of insurance industry."
From a conceptual level, insurance has always been about predictive analytics, which is one reason the tools were able to establish a foothold in underwriting, where even for some early adopters, they remained isolated.
"Underwriting has long been viewed as more art than science, and there was a significant change management challenge getting large numbers of underwriters to see that a quantitative tool could be of great benefit," says John Lucker, principal, Deloitte Consulting LLP, global advanced analytics and modeling leader. "There's a fear that these tools are going to replace people, which has never been the intent. What the tool can do is replace the need to touch policies during the entering process for small policies, which frankly are not worth touching."
Because more underwriting is done algorithmically, there's a need and an opportunity to incorporate data that historically has not been available electronically. A lot of underwriting information has been stored in the minds or desk drawers of underwriters, Lucker says, but now much of that data can be gathered automatically, streamlining the underwriting process and increasing its accuracy, which has created yet another opportunity: synthetic data.
"A big data source we use is synthetic data, which is actually manufacturing predictive pieces of information from internal and external data that never existed before," Lucker says.
For example, for small commercial insurers, a very good predictor of how well an agent or broker knows their customer is the physical distance that separates them. "Statistically, the further the customer is from the agent, the lower the quality of the risk," Lucker says. "The good risks tend to be marketed to by local agents, who know they are good risks. And the risks that are not being marketed to locally have to look around. And eventually they get further and further away. Plus, agents tend to be front line underwriters. To some degree they underwrite the risk first through their qualification process."
Successful predictive analytics tools for underwriting have helped pave the way for straight-through-processing for those smaller policies that don't much benefit from an underwriter's intervention, Lucker explains. "If you are an underwriter raising and lowering prices by $25, it's hard to imagine how, on average, your activity is going to be greatly beneficial," he says. That raises a secondary benefit of predictive analytics: the elimination of bias in the decision-making process.
At Mercury General, a multi-line insurer offering personal auto and homeowners' insurance through independent agents and brokers, most of the company's efforts around predictive analysis have centered on pricing. However, since the first model was rolled out on SAS Business Analytics more than five years ago, the carrier has put into production predictive analytics applications for fraud detection and marketing, and soon intends to launch models to mitigate catastrophe losses.
The goal of the initial predictive model was to better assess a premium that appropriately reflected the inherent risk in a particular customer profile, explains Robert Houlihan, VP and chief product officer for Mercury Insurance. And while the goal is the same, methodologies have improved and subsequent generations of the model have grown in sophistication and effectiveness.
"Like many other carriers, we have built pricing models using generalized-linear models, so that we are able to solve for a number of rating elements simultaneously and also are able to consider the interaction of those rating elements," Houlihan says.
For example, when pricing auto insurance, most personal lines carriers will rate whether an applicant is a homeowner or not, Houlihan says, which will interact with other rating variables, such as whether they are a standard or preferred risk. Consider a person with a high credit score, prior insurance or prior insurance with high limits, multiple policies or multiple cars, he posits.
"When you look at that kind of profile, a very high number of those people are homeowners. So the fact that you are also a homeowner doesn't add as much expository information. If you compare that information to someone who is a nonstandard risk-maybe they didn't maintain prior insurance, or had a poor credit score-the fact that they are homeowners may differentiate them from other people who are similarly situated," Houlihan says. "The idea is that you wouldn't just have a discount across the board for homeowners. You would interact that with other variables depending on whether they are a preferred risk, standard risk or nonstandard risk. It would make a difference how that one rating element interacted with all the other rating elements. And that's essentially what generalized linear models do, they solve for a myriad of ratings variables simultaneously and they enable you to adjust for those interactions."
Pinnacol Assurance, a Colorado-based workers' compensation insurer, began working on proof-of-concept models with predictive analytics vendor Valen Technologies Inc. in 2005. Since then, Pinnacol has put many of Valen's predictive analytics tools into production, including custom-built locally hosted applications, and more standardized Software-as-a-Service models.
"We were among Valen's first customers and we were getting our feet wet together," says Mark Isakson, VP of underwriting at Pinnacol Assurance. The insurer purchased Valen's UnderRight predictive pricing model in 2007 to help lower loss ratios, improve profitability and protect against bad risks. And in 2009, Pinnacol bought AuditRight, Valen's predictive model designed to help optimize claims-auditing resources.
The first generation of the predictive underwriting model was rolled out in 2008, and was designed to determine which one of Pinnacol's six tiered loss-cost multipliers (LCM) is most appropriate for an insured business, based on industry type, size and other characteristics, Isakson says.
Five rates are filed with and approved by the Colorado insurance regulator, and all of the carriers use those same rates. However, to reflect their expectations about their operating expenses and any other contingencies or profit, insurers can file LCMs to apply to those rates, Isakson explains.
Whereas a larger insurer may have multiple underwriters, and therefore a wider range of prices, Pinnacol has a single underwriting company, Isakson explains, but increases its competitiveness through tiered rates. The tiers are filed with the Colorado insurance regulator and have regulator-specified requirements to make certain the underwriting process is not discriminatory.
The underwriting model is built into the workflow, Isakson says, running behind a custom-built integrated policy-and-claims-management system.
"That's where we get an application from online or from an agent, who has filled in all the basic underwriting information related to the business and its characteristics," Isakson says. "On a new piece of business, our model is going to pick up and use that information to help underwrite and determine the loss-cost multiplier. For a renewal, similarly it uses all the policy management information in our system in addition to any loss experience we've had with the risk since underwriting it. It evaluates the loss-cost multiplier decision before any underwriting is done. The underwriter picks it up and underwrites the file from that point forward," he says.
Because insurance is such a highly regulated industry, the Pinnacol team also found itself collaborating with regulators to make certain they were comfortable with how Pinnacol was using the model, external data and variables. "That's an important part of our process and underwriting model," Isakson says. "We've met many times with our regulators on the development piece to get their feedback on what concerned them and we will be doing that again during our underwriting refresh."
For Pinnacol the biggest challenge, Isakson says, was communicating how the model works to underwriters and agents, given the model's highly proprietary nature.
"They had to adapt and we had to be able to explain the decisions our model is indicating. What it's looking at is great in the aggregate, but at the individual policy level may not make sense," Isakson says. "Where you gain efficiencies on the workflow means there's less underwriting intervention, so that's good. On the other side, the model is only as good as the information it's evaluating, and you lose some of that efficiency if you have underwriters who are touching it too much."
One of Mercury's later predictive analytics projects focused on fraud referrals, which evolved from the rules-based system they used previously. The predictive model considers data from Mercury's proprietary Next Gen Claims System, Houlihan says, which it supplements with policy data pulled from Mercury's enterprise data warehouse.
"Historically, we considered claims-related information to establish a system of red flags," Houlihan says, such as a lag between when the accident happened and when it was reported, or the accident occurring right after the policy was purchased, or the number of claimants. "The model was able to evaluate these indicators to determine which were most predictive of fraud. More importantly, the model considers additional variables from our data warehouse that were not available within the claims system. How long was the individual a customer of Mercury? Did the individual have other policies with Mercury? These additional variables were as predictive, or more predictive, than information just related to the claim."
Mercury runs the model weekly, Houlihan says, generating a report that is sent to Mercury's special investigations unit. Subsequently, the investigations are more frequently successful and the recoveries are potentially larger than when they used the heuristic, "gut-feel" approach, Houlihan says.
"Fraud detection is an area that we never intend to fully automate," Houlihan says. "There always needs to be some level of human involvement to review suspicious cases. Fraud can morph over time, so it is important to consider those human observations and periodically update the models."
Fraud detection is an important domain for predictive analytics, but such claim-scoring models also lend themselves to subrogation and salvage.
"It's a great way to look at leakage more broadly, and the reality is that the models are not that different," says Christina Colby, VP business information management for insurance at Capgemini. A successful implementation of those models also generates money that can be used to fund more initiatives, she adds.
"There are two pockets of opportunity. One is to look at closed claims that are still within the statute of limitations as well as open claims you can pursue," Colby says. During the claims handling process, claims can be scored on the likelihood of recovery, and by accessing past patterns of recovery, prioritized and automatically referred to the recovery team, increasing the return on those efforts. "If you can discover that cash within your organization, you earn the right to do things where the business case may be more complicated," Colby says.
Deloitte's Lucker says that for workers' compensation insurers, predicting claims severity is a particularly promising application for analytics.
Historically, rules-based systems were used to determine how claims were processed, but there was not a statistically confident predictor of whether a claim was average or bad. "Based on rough numbers, 80 percent of claims costs are the result of 20 percent of the claims," Lucker says. "But how do you segment claims into piles, so you can find those worst 20 percent and handle them? In the past, it was done with tribal wisdom or gut instinct within the claims organization."
To accomplish that differentiation, a typical workers' compensation claim predictive model may contain as many as 80 variables to account for comorbidities, such as high blood pressure or diabetes, for example, and those comorbidities have a statistical relationship to the length of time it takes for an injured person to return to work, or the severity of their medical problems.
"You're predicting this claim is going to be worse than that claim and why, using different medical and behavioral attributes of a claimant," Lucker says. "If we are both at work and wrench our backs, there's a statistical relationship between those comorbidities and 75 or 80 more that are going to allow you to go back to work faster than me."
Lucker says predictive analytics are providing improvements of 8 percent to 12 percent in such instances. "If you can improve a claims result by 10 percent and you have a billion dollars in losses, that's a $100 million solution," Lucker says.
Pinnacol's AuditRight model determines what sort of audit, be it by phone, mail, Web or face-to-face, will be performed on a specific claim.
"With 55,000-plus policies and a limited number of people, that's an important distinction," Isakson explains. "We can do a lot more mail-in audits than out-in-the-field audits. So it's really about how we allocate and optimize the use of those resources."
That determination used to be a manual, rules-based process. The auditor would pull up-coming audits, review them and, based on claims data and personal experience, determine which would be conducted in the field. The rest were done by phone or mail.
"A rules-based engine might just say 'if the policy is under $5,000 and in these classes of business, then we are going to do a mail-in audit.' A predictive model is more dynamic," Isakson says.
By accessing audit results from the past five to seven years, plus some external data, such as economic indicators and what sectors of the economy are growing, the predictive model helps Pinnacol understand where things may be changing in front of them, rather than behind them, Isakson says.
The construction industry, for example, is very cyclical and can grow or shrink quickly depending on macroeconomic indicators, Isakson says, and the number of claims should rise and fall accordingly. "As you see recent trends in certain areas, the predictive model might look at that dynamically and identify that as emerging or slowing, and identify the appropriate resources," he says. Pinnacol then can deploy an appropriate number auditors to perform the right number and type of audits, limiting expenses and increasing their effectiveness.
While the predictive model now performs the analysis and makes a recommendation, an auditor can override that model, based on expertise and experience, to make any fine-tuning decisions the model may not have contemplated, Isakson says.
On the marketing side, Mercury also uses predictive analytics to determine how it compensates and incents agents, offering different commissions for different types of customers or products, for example, to determine an optimal mix that could help Mercury more efficiently grow the business.
The marketing model considers different policy types and how the volume of that business would respond to increases or decreases in agent compensation. "If you increase your compensation, that increases your expenses, but you also tend to write more business because you've provided that incentive," Houlihan says. "We look at the incremental business that we wrote: Does it generate enough expected profit? Is it more than the incentive we've just provided? If it is, that would suggest we should increase the incentive even further."
Other insurers that sell through independent agents are using predictive analytics for agency management, which holds great potential for predictive analytics, says Mike Reilly, managing director insurance at Accenture. He explains that while insurers may have traditionally doled out agency dollars to the biggest agencies, more-sophisticated carriers are now using analytics to predict agent profitability and investing in those agents accordingly.
Consider a mature agent in a steady-state run mode, as opposed to young and hungry agent, Reilly says. "Do I want to be spending my extra marketing dollars on the guy who is probably not going to grow? Or maybe on someone I have less penetration with, or who has a relatively small book now, but has that potential to grow? Those analytics are one-and-done analytics: I understand the problem better and I treat it anew."
A good data set is critical to doing good analysis, as is back testing, Houlihan says. "We do a lot of tests to see if the output is better than what we had before. One of the common steps is to take the database of customers, hold out a random sample, build the models on a majority of the data, and then test the model against that hold-out set," he says. "You would also test your current rating scheme on that holdout set. There are a number of validations one can do to make sure you got a better answer, that you took a step forward and not a step backward."
The availability of predictive analytics also has changed how Mercury manages risk. "We have moved [closer to] a model where we manage risk more by pricing rather than through selection," Houlihan says. "If a profile presents a higher risk, rather than deeming that unacceptable, we try to address that by charging a slightly higher rate." In the past, the company would have tried to underwrite in such a way as to not accept those risks.
Next, Houlihan plans to look at models to control exposure to catastrophic risks. The model could, for example consider how the concentration of policies affects potential losses in the event of various catastrophe types, such as wildfire, hurricane and severe storms. Mercury could then adjust prices to mitigate that risk. "Insurance carriers need to think beyond the traditional applications of analytics to topics such as pricing and leverage the opportunity to drive more data-based decision-making through all aspects of the business," he says.
Predictive analytics is now a core competency for Pinnacol, Isakson says, and more models are coming. "We'd like to target risk management services on our book of business, find where that is going to provide the greatest return and where the greatest need is," he says. "We have invested in better tools to convert from a data warehouse structure to more of a business intelligence environment, where we are not only able to leverage predictive analytics, but have better software and tools to help us use that data. For insurance, when you are trying to set rates and identify opportunities in your book of business to better drive outcomes, that's a core competency."
Pinnacol's commitment to further developing their predictive analytics capabilities was recently underscored with the hiring of a staff predictive modeler, "a kind of quasi-actuarial statistician," Isakson says, who will guide them through the creation of the next generation of the underwriting model and then develop models for claims and safety services.
Register or login for access to this item and much more
All Digital Insurance content is archived after seven days.
Community members receive:
- All recent and archived articles
- Conference offers and updates
- A full menu of enewsletter options
- Web seminars, white papers, ebooks
Already have an account? Log In
Don't have an account? Register for Free Unlimited Access