University of Waterloo Faculty to Mathematics researchers have developed a new method that enables large insurers to reduce the time spent estimating the financial liabilities of their portfolios from days to hours while achieving high accuracy.
A study details the new method which significantly reduces computational time, but still estimates the financial liability of variable annuity portfolios accurately for business purposes.
A variable annuity is a variety of popular insurance products that offer policyholders benefits linked to the financial markets performances. Variable annuities provide great flexibility to satisfy different policyholders’ needs. But, at the same time, such flexibility leads to great complexity when it comes to evaluating the insurer’s liability.
“For large insurers, simulating the liability for every variable annuity in a large portfolio can take days,” Ben Feng, a professor in Waterloo’s Department of Statistics and Actuarial Science, explained. “Our new method employs machine learning methods to reduce the evaluation time to within a few hours, but still achieve high enough accuracy for business purposes.”
The new method works by dividing the portfolio evaluation process into three main stages.
Firstly, unsupervised methods, such as clustering algorithms, are used to select a few representative policies in the portfolios. Simulations are then run for only these representative policies. Finally, supervised learning methods are used to estimate the liability of all policies in the portfolio and that of the whole portfolio. Most importantly, the new method only requires minimal changes to what insurers are currently doing, such as adding a random sampling step prior to clustering and identifying complicated representative policies prior to running any simulation.
“Each of the three steps in isolation only uses simple and existing methods,” said Feng. “But the main innovation lies in the innovative ways to combine these simple methods. And the combined method works surprisingly well in both efficiency and accuracy.”
The study, Efficient simulation designs for valuation of large variable annuity portfolios, authored by Waterloo’s Faculty of Mathematics researchers Feng, Zhenni Tan and Jiayi Zheng, was published recently in the North American Actuarial Journal.