The variable annuity with a guaranteed lifetime income benefit (VA/GLWB) was and remains a masterpiece of flexibility. It lets contract owners keep their investment and income options open indefinitely. Since the early 2000s, that formula has helped close over a trillion dollars in sales.
But by giving clients lots of options, insurers made it harder for their own actuaries to predict how policyholders would use their contracts. That in turn, made it difficult to set prices and anticipate reserve requirements. Overestimates in surrender rates, for example, have led to sudden, expensive and unwelcome calls for extra reserves.
Moody’s red-flagged this danger three years ago. Insurers “misestimated and underpriced the lapse rates on this product, as policyholders held on to their policies at a greater rate than the insurance companies anticipated,” a Moody’s analyst wrote in mid-2013.
“This miscalculation forced insurers to take significant, unexpected earnings charges and write-downs over the past year and a half,” he noted. In November 2013, to name one instance, lower-than-expected lapse rates forced Prudential to take a $1.7 billion charge.
To better understand the factors that have driven their contract surrender rates in the past and to help them predict those rates more accurately in the future, a group of 18 large VA issuers, ranging from AIG to Voya, has been sharing anonymous client data with Ruark Consulting, a Simsbury, Conn., actuarial firm that specializes in this specific topic.
For several years, Ruark has been producing a series of “experiential” studies of actual VA contract owner behavior. This year, the firm and its “advisory council” of annuity issuers turned to the relatively new discipline of predictive analytics to build models that can predict policyholder behavior and help each company’s actuaries make more accurate assumptions. That project, now being implemented at several of Ruark’s client companies, could help annuity issuers hit their targets and meet the expectations of senior management, regulators, analysts and investors.
Predictive analytics
“Policyholder behavior assumptions used to be a relatively quiet corner of these products. They weren’t reviewed or challenged all that often,” Ruark president Tim Paris told actuaries from VA issuers during a recent webinar. “But income utilization behavior, and partial withdrawal behavior on GLWBs, are big question marks hanging over this line of business. How we answer them will help determine the profitability of the business over the next couple of years.”
Predictive analytics has been used, variously, to forecast the effects of global warming, to infer the creditworthiness of new loan applicants from a few bits of personal data, or to complete the URLs that you begin typing into your browser’s address bar. IBM is using its Watson technology to help broker-dealers predict which advisor will be most compatible with a new client, or to identify clients who are dissatisfied and likely to jump to another firm.
When it comes to applying predictive analytics to annuity policyholder behavior, insurers can’t do it—or can’t do it very effectively—on their own. Though each VA issuer has its own actuaries, and some insurers even have their own predictive analytics teams, their data is limited to their own experience, and some of their products are too new (and contract owners too young) to have established behavior patterns around withdrawals and income rider utilization. That’s why issuers have reached outside their walls to collaborate with Ruark and their peers.
“Even large companies have holes in their data and can’t make statistically credible assumptions,” Paris told RIJ. “We can bring our tools to bear on a very large data set.” Ruark says it now has 50 million contract years of monthly data going back to 2007, provided by 20 VA/GLWB issuers. For a separate study, it has compiled 10 million contract years of data from a dozen providers of indexed annuities with GLWBs.
Four factors drive surrenders
During a webinar in April, Ruark presented its plans for creating a predictive model for annuity surrender rates. Creating a model is both an art and a science. It involves choosing from among several existing techniques of statistical analysis, which, in laymen’s terms, means deciding which equations to use. Once those choices are made, the analysts must decide which types of data to plug into the equations.
From ten years’ worth of industry experience studies, Ruark has learned that a contract owner’s likelihood to surrender a contract is determined primarily by four factors: The number of years left in the contract’s surrender period; the type of living benefit rider on the contract; the “in-the-moneyness” of the contract (i.e., the extent to which the guaranteed benefit base exceeds the current market value of the contract assets), and the size of the policy.
Interpreting the exact roles of these factors isn’t easy, however. “We understand intuitively that surrender rates will spike as people come out of the surrender period and level off,” Paris (below left) said. “But we now know that the surrender rate also depends on the value of the guarantee, which goes up and down with the markets,” he added. “When we saw surrender rates go down after the financial crisis, we assumed it was an in-the-moneyness effect. Then the equity market recovered and the guarantees weren’t so deep in-the-money, but surrender rates stayed low.”
Now that Ruark has established its predictive model and settled on the four most important factors to run through it, the actuarial firm hopes to apply the analytic tool to creating benchmarks. Going forward, the benchmarks will reveal the accuracy of each company’s forecasts and show each issuer whether its experience falls within industry norms or not.
“The results will vary by company because each company has a different mix of living benefits types, different ages of policyholders, and a different history of actual-to-expected results. The benchmark will show every company’s position relative to its peers,” Paris told RIJ, noting that the results are blind—each firm’s data remains confidential.
“Each company will then be able to explain to their analysts and stakeholders why their experience is different and to ask themselves if they want it to be different,” he added. The benchmark results can also give companies early warning signs, which will enable them to change their product distribution, for example, sooner rather than later.
“If you have more data and better analysis, then you won’t have to be as conservative in your use of capital,” Paris said. “You could also weed out products that are being used in undesirable ways. It brings clarity to the business. We’re talking about products that could be in force for as long as 40 or 50 years but have been around for only about ten years. So we try to squeeze as much insight out of the existing data as we can.”
© 2016 RIJ Publishing LLC. All rights reserved.
