Case Study

Timothy Paris and Michael Frings November 2020

Suppose you are responsible for modeling policyholder behavior at your company, which is a large issuer of variable annuities. The guarantee features common to these products make them a complex combination of risks, and your company uses a capital markets hedging strategy based on the policyholder behavior models that you provide. When future policyholder behavior corresponds closer to your models, hedges are more efficient with a lower risk of hedge “breakage” for your company. Hedge breakage directly affects earnings, which is a key performance indicator (KPI). As a result, you and your manager, the chief actuary, have a strong vested interest in doing everything reasonably possible to minimize hedge breakage.

Based on your own company’s data, you use predictive analytics techniques to develop models for policyholder surrender behavior. These models provide a reasonable fit to your company’s historical data, and based on out-of-sample testing, you expect that future actual-to-expected ratios will be 97 percent to 103 percent of the model average surrender rate of 5 percent.

Suppose also that after researching a range of alternatives, you determine that you could access a much larger and high-quality data set for similar products across the industry and, using the same predictive analytics techniques on this new data, you can consistently improve actual-to-expected ratios by just 1 percent (i.e., to 98 percent to 102 percent). This would correspond to 0.05 percent (1%*5%) annual improvement in surrender rates, which would be worth about 0.60 percent in present value over the life of the product.

Using a common reference basis of 15 percent annualized equity market volatility and two standard deviations, the improvement in hedge breakage would be 0.18 percent (0.60%*15%*2) of the hedge notionals. The hedge notionals for your company are about $3 billion, so the potential present value of benefits of this new data is $5.4 million.

The cost to access the new data is about $50,000 annually. You would do the bulk of the work yourself using your existing predictive analytics framework, so it would not entail any material increase in other costs. So, the present value of costs for this new data over the life of the product is about $600,000.

You make this cost-benefit case to your manager, the chief actuary. While she sees its strong merits and is authorized to approve a $50,000 cost within her annual budget, she decides to present it to the enterprise risk management committee for approval due to the long-term recurring nature and magnitude of this project. Since the $5.4 million present value of benefits is nine times greater than its costs, the approval is granted.

Return to Demonstrating the Value of Data