Insurance claims professionals are pioneers in the use of predictive data analytics. Well before the term “Big Data” was coined, claims examiners were digging into the data within filed claims to unearth kernels of wisdom. These insights illuminated ways to reduce claims duration and costs, returning injured or ill employees to health and to work on a faster timetable.
Now a more robust form of predictive analytics is at hand for enhancing the way claims are handled. Unlike previous analytical processes that were focused entirely on an organization’s structured internal claims data, systems have now been advanced to mine the wealth of unstructured and external data that heretofore have been challenging to analyze.
By using predictive analytics to sift through these millions of data sets, claims specialists can uncover complex patterns, trends and correlations that otherwise would pass undetected. By using algorithms to make sense of these associations and putting this knowledge in the right hands quickly, specific actions can be taken to improve the outcome of a claim for individuals and businesses.
All types of insurance claims will benefit from the application of predictive analytics. The technology provides insights about the likelihood of future claim events, ferrets out instances of fraudulent claims activity, and ensures appropriate treatments that expeditiously return employees to work.
Like all things that sound almost too good to be true, the devil is in the details. Superior predictive analytics depends on the breadth and quality of data captured. The best models accumulate a vast array of applicable internal, external, structured and unstructured data, turning this mountain of words and figures into information to support decisions with practical value.
Few organizations have greater access to this data or a wider scope of analytical capabilities than insurance companies, given the breadth of internal claims information at their disposal and their unparalleled knowledge of risk. But, not all insurers are comparable in their ability to offer comprehensive claims insights.
Much of the claims information analyzed to date has been internal structured data—the information residing in pre-defined fields within the claim file, or in line-item spreadsheet format. By analyzing the different line items and comparing these data sets to each other, interesting correlations may arise. For example, the distance that an employee drives to work may be associated with an increase in workers’ compensation claim severity.
Such eye-opening linkages are not always intuitive. In other words, they often pass by unnoticed, as the ability of the human brain to imagine unique correlations—without “doing the math”—is slim. Obviously, the more data sets for comparison purposes, the greater the likelihood of uncovering additional patterns and associations.
This is where both external and unstructured data come into the picture. External data is all of the information on a subject that exists on the Internet – pieces of information culled from studies, empirical observations, news items, regulatory and legal data, and millions of other statistical sources. Much of this information can be correlated with claims data to produce interesting insights. Knowledge of changing weather patterns in relation to slip and fall accidents is a case in point.
Unstructured information, as it relates directly to a claim that has yet to close, is also critical. However, since this information is not structured, a predictive model may fail to accumulate and assess it. In a workers’ compensation or bodily injury claim, for instance, this data may include the claims examiner’s notes; the different medical services, medications and physical therapy that were prescribed to the claimant; and medical bills paid to date.
Assuming this unstructured data is captured and analyzed in concert with a claim’s structured information, interesting correlations may develop. Perhaps someone describing a particular type of back strain in the claim notes is likely to develop more serious medical complications than people describing another type of back strain. Prescribing the same treatment regimen to both sets of claimants and expecting the same outcome is overly optimistic.
Although not as vast as external data, unstructured records comprises a broad array of useful information, such as emails, texts, transcribed telephone recordings and anything in a Word document, such as complaint letters. By accessing and analyzing this information, companies can make more informed decisions of specific actions to take that will help improve the outcome of a claim.
This information has always been available, of course. But without automated tools, a claims examiner would need to manually sift through every detail. And without algorithms, they would be frustrated in attempting to uncover unusual patterns of claims activity.
Actions Through Algorithms
Algorithms are the crystal ball of analytics–an algorithm turns claims data into useful information. But first the data needs to be found and extracted before it can be refined and used.
Data mining can ferret out how many times a particular word, phrase or combination of words pops up in different sets of data. Using a carefully constructed algorithm, this information can then be compared to other data sets on, for example, claim duration. The analytics may indicate that the phrase “out of breath” is related to claims that took longer than 180 days to close.
Predictive analytics also can be applied to discover the possibility of fraud. For instance, the unstructured data in a police report’s notes of an automobile accident may indicate that the accident was a rear-end collision, which may not neatly align with the size of a bodily injury claim filed in the accident’s aftermath. Mining additional unstructured data from the car repair shop may confirm that the accident was a minor fender-bender. Other unstructured data may indicate that the average speed of vehicles on the street where the accident occurred is no more than 15 miles per hour at that particular time of day, due to traffic conditions. A clear picture of potential fraud emerges.
These are obvious examples, of course. It is the less intuitive findings of predictive analytics that offer the more unique opportunities to improve a claim’s outcome. For example, the analytics may indicate that a particular type of claim involving certain individuals and specific circumstances has a higher than average chance of resulting in a lawsuit. Knowing the potential for expensive litigation, the claims examiner can develop a more informed strategy.
The more data captured in the predictive model, the more verifiable the mathematical conclusions. Internal data is a significant starting point for developing analytical solutions.
Adding unstructured data like the analysis of millions of claims notes to these data sets enhances the opportunity to assess the likelihood of future events. Consequently, executive decision-makers will have greater confidence to invest in actions that will make a decided difference.
Decisions Backed By Data
Predictive models alone are not enough to attain actionable knowledge. The analytics must integrate with a company’s existing processes to have a positive impact on claim outcomes.
Each organization is unique, with different loss histories, a varied range of both open and closed claim files and disparate ways of conducting business. Failing to appreciate these differences, an off-the-shelf predictive modeling tool won’t make much of a difference. The best solutions are integrated with a company’s business model and workflows.
Since the production of data is ongoing, predictive models must reflect this dynamism. New claims data and evolving external and unstructured information must be input into the model. And the actions taken to reduce claim duration and costs based on this analysis must be monitored to ensure they are achieving the intent.
Predictive models aren’t a replacement for seasoned claims professionals or corporate risk managers. People will always be needed to evaluate the mathematical conclusions produced by the analytics, applying their intellect and expertise to this intelligence to make more informed decisions.
The bottom line is that you can’t manage what you don’t measure. By measuring only certain characteristics of a claim, only half the picture will emerge to moderate its impact. With sophisticated predictive analytics, a fuller picture appears, telling a much richer story.
Scott Henck is a senior vice president for global claims analytics, business intelligence and actuarial insights at Chubb and responsible for Chubb 4D, the company’s proprietary analytics solution driving better claims outcomes for insureds. Scott can be contacted at firstname.lastname@example.org.