The secret of interlocking forces: dynamic modeling can help insurers predict losses affected by multiple factors over time.

Carol Valentic has become very frustrated with the performance reports she routinely sends her insurer and corporate clients. Running Genex Services Inc.'s workers' compensation managed-care services in the Midwest and Ontario, she wants to discuss with clients how her team is affecting their bottom-line cost of workers' comp. But her reports, and her competitors', dwell on episodes of service, with no evidence of their impact of the final cost of claims. What is needed is information on key drivers of costs and how her services address them, summed to a statement of value added. Claims outcomes appear to be a black box with no commonly used analytic key able to open it. The solution is to use an analytical process that emerged in recent decades from economics and social science: dynamic modeling.

Vendors, insurers and insureds can use this analytic process to paint a complete picture of exposures, loss cost drivers and the measures applied to control them. Dynamic modeling has already made a profound impact on catastrophe loss prediction since the early 1990s. It has the potential to trigger systemic redesign of disability insurance It can steer efforts to control health-care costs. As a bonus, dynamic modeling may be critical to controlling claims fraud.

Medical-care costs in workers' comp are an attractive target, and it is a mystery why insurers do not use it more often. Thus Valentic's concerns are a good jumping off point to explore this analytic process, which is not as foreign to risk and insurance professionals as it may appear. Dynamic modeling is a highly formalized way of representing what we often visualize conceptually when we think about complex exposures and losses.

One can't credibly explain the impact of medical control measures until one finds out what is pushing medical costs higher. The suspects behind medical-cost inflation in workers' comp fall into several camps. Medical providers may be migrating to more expensive treatments, delivering more units of treatment altogether and charging more per unit of service. Claimants themselves may be more resistant to recovery. Employers may be procrastinating at return to work, indirectly prolonging care. The plaintiff bar may be more active both in taking on a larger share of claims and promoting more medical usage by their clients.

Four Assets Needed

Each of these cost drivers is plausible, but there are advantages from having a more certain and defensible explanation. To get there one needs four assets. First is a lot of data, today more available from claims systems, and proprietary and public databases. One also needs a set of analytic tools, which has been refined by dynamic modelers for some years. Third, one has to have a very good grasp of the subject at hand--without content expertise, the modeling can go on wild goose chases. Fourth, and a hard pill to swallow, one has to have respect for the results of the analysis and not analyze by anecdote.

Maine Employers' Mutual Insurance Co. is one of a small group of workers' comp insurers with these assets. Its studied approach illustrates nuances in the meaning of "dynamic." It has been running comparisons of high-cost surgical procedures with those of private sector health plans, using dynamic simulations to model reimbursement rates closer to the health plans--comparing the cost and frequencies of surgeries.

MEMIC is also in the process of studying how surgery figures into the continuum of care for injured workers. This allows them to model how medical care dynamically cascades into further care with varying impact on loss costs and duration of disability. MEMIC, during a difficult time for the industry, demonstrates favorable financial results despite most of the book of business being written in a relatively high workers' comp cost state.

The tools for dynamic modeling come from the fields of social science and economics, where they have given rise to entirely new financial instruments that did not exist just a generation earlier--option markets and hedge funds. Researchers have been studying for decades multiple factors driving employment, spending and investment, educational achievement in developing human capital, crime, alcoholism, and other behaviors. Most Americans have, watching the evening news or observing the environs at a stoplight, done homegrown multivariate analysis on these issues. Their interpretations may include a makeshift model of how factors work among each other, dynamically over time. We enjoy our theories especially if no one is around to contradict us. We get away with it in our work when the cost of our being wrong is cheap.

What if, however, we lose big--the corporation or insurer goes bankrupt based on its failure to understand exposures and loss development? Environmental catastrophes are the chief examples of failure to use dynamic modeling. Karen Clark, founder and chief executive officer of AIR Worldwide, a risk modeling firm, lived this story.

Learning Through Experience

In 1987, AIR Worldwide introduced to the insurance industry a catastrophe model capable of running thousands of multifactorial predictions of the economic impact of an event such as a hurricane. This modeling could not have taken place without access to computers, large databases of properties and mathematical formulas of the behavior of storms. AIR's estimates showed that the industry could experience $20 billion to $30 billion dollar hurricane losses with a reasonably significant probability, well above the expert-judgment estimates of about $7 billion at the time. When in 1992 Hurricane Andrew made landfall on the coast of south Florida, AIR produced a quick loss estimate in excess of $13 billion. It took several months for the industry to realize that they would, in fact, pay out losses of this magnitude. Some insurers failed due to excessive losses. Since then, AIR's models and those of competing firms have grown increasingly fine-grained in their predictions, thanks to better data, more proficient software, and accumulated experience.

Insurers' misjudgments about hurricane risk stemmed not from sloppy speculation, but a more disturbing cause. Insurers had relied too much on conventional actuarial techniques. For exposures and loss developments based on complex, rapidly changing social and economic trends, aggregate techniques traditionally used by actuaries can produce very misleading predictions. Property loss exposures along the South Atlantic seaboard and claim costs closely related to individual behavior come to mind.

Michael Helvacian, a Ph.D. in economics, held several key research positions at the National Council on Compensation Insurance in the 1990s. In the mid-1990s, he compared for Intratcorp claims with and without managed care, finding that managed care cases closed much faster than other cases. He showed for the first time that indemnity cost savings were much more pronounced than were medical cost savings when managed care is applied. As a consultant lot the Workers Compensation Research Institute, he studied variances in medical costs among intrastate regions in Texas. He found variances as much as 50%, due mainly to use of specific services such as vocational rehabilitation at hospitals. These findings, of practical value to insurers and regulators, could not have arisen out of traditional actuarial reviews.

He has been looking at how to predict California workers' comp loss cost inflation due to recent benefit legislation, AB 749. The impact largely depends on the future costs of permanent-partial-disability awards, a wild card of enormous importance. Permanent-partial-disability costs are affected not just by higher benefits for the existing stream of applicants, but also by the incentive for more people to apply. According to Helvacian, his model might show up to 50% higher permanent-partial-disability cost compared with traditional actuarial methods, by recognizing higher incentives of claimants to file for and receive these benefits. The total impact could be well over $1 billion annually to the state's employers and workers.

To make a prediction such as Helvacian proposes for California, an analyst would build a model, an artificial universe representing the real world. She or he would populate it with formulas describing in precise terms how a change in one driver (such as an enrichment of one benefit) would trigger a change in utilization of several benefits. The model builder will define any lagging effect. She would decide if the relationship is symmetrical; that is, if a reversal of a change in one driver restores the impacted drivers to their original position. (This may be an illusion; we never go back to where we came from.)

Having completed the model, the analyst would then apply it to a large amount of exposure information and claims data, gathered to reflect several years. The model is run, the formulas adjusted, until the formulas get close enough to predicting each new time phase in the claims data. At that point, the analyst will report how much some benefits will cost, some rising and some possibly declining. She might for instance conclude that "a 10% enrichment of benefit A will cause a 15% increase in utilization of benefit A and a 5% reduction in utilization of benefit B." She would couch her estimate with a probability of its being correct.

How Dynamic Modeling Is Different

The disciplines of statistical analysis, actuarial science and dynamic modeling overlap. The distinctions that matter are worth a thumbnail review. Statistics is a discipline of carefully defining the likelihood of your hunches and has a betting flavor to it. The end product is often not a dollar value, or quantity, but a probability that a prediction or an inference of association will prove to be correct if we knew all the facts. An example is the chances that a certain surgical procedure has a higher than average risk of surgical malpractice claims regardless of apparent abnormalities in patient demographics.

Actuarial methods look like statistics, but they are better seen as a highly evolved type of accounting, where the end product is an insurance agreement that is durable enough to survive the vast majority of unexpected developments. An insurance agreement is a contract to sell a product (financial indemnification) where the cost of goods sold is not known at the time the agreement is struck. Actuaries make extensive use of statistics to qualify their projections with probabilities of occurrence. Their key value in risk and insurance is their experience with exposures, losses and insurance provisions. This allows them to construct loss projections that tie precisely into the terms of the insurance policy or self-insurance plans.

Dynamic modeling, like actuarial methods, uses statistical methods to point out the degree of probability to a finding. However, its most valuable service in risk and insurance is predicting outcomes affected by multiple factors through time. The best use of dynamic modeling lies in its tools to predict "how much" as the world turns, such as how much permanent-partial-disability costs in California will grow when all major factors are analyzed, or what New Orleans' potential hurricane losses might be as various preventive measures are introduced into the city's ever more valuable real estate. Because of their training in graduate level economics, many modelers call themselves econometricians.

The distinctions between these disciplines and their practitioners are vanishing, with the one exception that chartered actuaries will continue to dominate the work of predicting losses for regulatory premium rate setting or for individual insurance contracts. The dynamic modeling enterprise will continue to grow in potency. In the area of short- and long-term disability, it may be about to turn the entire disability risk and insurance field upside down.

Disability Findings

Human resource information systems have improved to the extent that there is now much better quality of detailed nonoccupational disability data from employers. Until now, short- and long-term disability programs have been designed, funded and sold on the basis that exposures can be successfully handled without considering other absences. This now appears to be false assumption. Not only are short- and long-term disability claims in an important way driven by other types of employee absences, but these other absences cost employers so much that disability should be now seen as a relatively minor subset of a larger exposure: unscheduled employee absences overall.

This revelation, the outgrowth of extensive modeling of large, detailed data sets by Nucleus Solutions, of Arlington, Va., puts short- and long-term disability insurers in a position analogous to American railroads when truck and airplane cargo traffic was emerging. Originally, railroads felt they practically owned long-haul cargo transportation. Today they are a subset of a much larger, more complex web of transportation. In like manner, short- and long-term disability insurance may become a subset of broad scale absence risk management.

Among health-benefit plans, there is a high potential value of identifying covered persons whose health is about to deteriorate. In one instance, the health-care costs of 2,000 persons rose in one year from less than 1% of a plan's paid benefits to 24%. To a large extent, nonfatal chronic diseases are driving health-care costs, and behavior patterns of individuals are the nursing beds of these diseases. The goal is to predict deterioration, locate a very small set of people at risk, and stop it. A Boston area firm, Medical Scientists, has written into predictive models the epidemiology underlying these rapid changes in health and medical spending. It is applying these models to locate persons at risk, for example, for the Florida Medicaid program. This modeling goes far beyond the conventions of actuarial forecasting in health care in the understanding of disease progression and intervention.

Claimant fraud detection and investigation await vigorous application of modeling. In this instance, modeling helps insurers in two ways. First, it sorts through the myriad number of variables associated with claims to find indicators of fraudulent behavior. An example of an auto fraud indicator is the presence of several unrelated passengers in a car experiencing a collision with another car. The consulting firm of Correlation Research found after extensive analysis of many auto claims that this is a reliable enough indicator to warrant almost by itself special investigation. Modeling by its impersonal nature also helps the insurer fight claimant fraud by deriving these indicators in an objective manner, lessening risk of human bias and assuaging concerns of many about racial or ethnic profiling.

Dynamic modeling of exposures and losses will likely grow in use if only because the quality of exposure and claims data is improving. Its potential to transform risk and insurance is proven by the case of catastrophic exposures. It can greatly affect how insurers and corporations monitor risks. That is the secret of its future.

Before You Begin

Using dynamic modeling to understand the effects of changing forces on losses requires four assets:

* Data. A great deal of data is needed. Fortunately, it is available from claims systems and public and private databases.

* Analytic tools. These come from the fields of social science and economics and have been refined by dynamic modelers for some years.

* Content expertise. Without a good grasp of the subject at hand, modelers can go on wild goose chases.

* Respect for the results. This may be the most difficult asset to maintain. One has to have respect for the results of the analysis and not analyze by anecdote.

Peter Rousmaniere is a Boston-based consultant and frequent writer on risk and insurance.