Big data and advanced analytics have expanded the actuarial toolkit, at least in theory.

But the theory has been a challenge to put into practice. It can be hard to figure out which data set to tap and which analytical tools to apply and which problems to solve. CAS members heard ideas on how to approach predictive analytics at the recent CAS Ratemaking and Product Management seminar at a session titled “Applying ‘Big Data’ Analytics in the Insurance Sector.”

The insurance industry seems interested in analytics, according to A.M. Best’s most recent insurer survey. Nearly 32 percent named advanced analytics as the greatest opportunity before the industry, far more than No. 2 — increasing the use of mobile apps. Applying big data to insurance problems was No. 3.

Clearly, insurers see the value of sophisticated analysis, and as quantitative experts, actuaries would seem likely to deliver the promise of the data into management’s hands.

But wielding the new tools is a tricky business. During the session, seminar speaker Jason Rodriguez, data analytics manager at Willis Towers Watson, offered ideas on how to approach a big data project as well as tips on how to manage one.

By “big data,” Rodriguez referred to enormous data sets that can present challenges in extracting. An example: text messages, where researchers must sift through and combine information that has not been formatted or standardized, such as great = grate = gr8. In other words, the data is dirty.

The data is so complex, he said, “It really requires us to use more advanced techniques … to extract information.”

Often machine learning is used to find useful signals in noisy data. It is an analytic system that can learn as it analyzes and is used to clean up the dirty data. That’s a step up from traditional programming, in which the computer consistently responds to the data based on a set of rules. Rodriquez explained that with machine learning, you can train an algorithm to distinguish between cats and dogs based on pictures of each, then let the computer by itself classify a new set of pictures. Machine learning tools are critical in creating order from the relative chaos that big data presents.

Once it is tidied up, big data is ready for predictive analytics, the branch of advanced analytics used to make predictions about unknown events.

A company that wants to undertake a project has a series of decisions to make early on. One of these is to decide which business problem to address. Analytics can help with most insurance functions — underwriting policies, detecting fraud or developing sales leads, for example.

Naturally, actuaries think of the potential to price insurance, but analytics could help set claims reserves or guide claims processes, too. Topic modeling is a text-mining tool that extracts meaning from a block of text. Rodriguez said that a topic modeling tool could scour claims notes looking for words like “hospital” or “emergency” and create new predictors that could be used to estimate claim settlement values or the probability that a claim becomes much more complicated over time, which would let a company quickly assign the trickiest claims to the most seasoned adjusters.

“Advanced analytics would help you improve business decisions based on the data available,” Rodriguez said.

Clearly, insurers see the value of sophisticated analysis, and as quantitative experts, actuaries would seem likely to deliver the promise of the data into management’s hands. But wielding the new tools is a tricky business.

The company also has to know what it hopes to accomplish. Rodriguez outlined three typical goals for pursuing advanced analytics opportunities:

• Discovery — exploring your data to learn more about your customers, the goal of which is to extract knowledge, not optimize or minimize some measure. A typical application would be to help product developers learn ways to segment customers.
• Optimization — finding a way to improve a defined business process, like improving a rating plan or increasing claims handling efficiency. Here, Rodriguez said a company has to be sure that implementing the results of the analysis would be feasible — for example, a pricing model would have to be able to fit in a point-of-sale rating engine, or a claims triage model would have to be able to be deployed within the claims workflow and systems.
• Automation — teaching a machine to follow a human process. You “take your business decision and turn it into code,” he said, creating decisions “that don’t need to be monitored except with minimal human input.”

Usually a company has several ideas for a project. To pick the best one, Rodriguez recommended considering the following:

Project viability: Is there management buy-in? How complex is the project, and how well will you be able to assess its success? What resources (data, subject experts) are available?

Supporting data: What internal data can you use? If you need to add external data, how well can it link up with your data?

Potential impact: How does the project fit into the overall business strategy? How long will it take before the project yields tangible results?

Implementation issues: Will stakeholders accept the implementation solution? How difficult will it be to implement the solution in existing technology?

The most common problems, Rodriguez said, relate to data and expertise. Sometimes analysts can’t get the data they need, particularly if it is administered by a third party. Sometimes internal data isn’t accurate enough, or there isn’t enough to carry out the analysis. Other times, analysts don’t have the right sort of training.

In managing the project, Rodriguez recommended what is known as an agile approach. A traditional project takes place over many months, with a lengthy planning phase followed by a lengthy building phase and a lengthy review phase. An agile approach breaks the project into several small pieces, each of which has its own (shorter) planning, building and review phases.

The advantage of the agile approach is that each component project is delivered sooner, bringing value to the company quickly. This ensures that effort is efficiently translated to valuable results even if the company’s needs change during the course of the project.

“You deliver value quickly,” Rodriguez said, “then add features [in later phases] to continue to improve.”

Rodriguez emphasized the need to choose an analytical approach and technology spend that fit the problem at hand, keeping in mind that mature technologies exist to deal with some increasingly common problems.