Welcome to the ARIA 2021 Annual Meeting Agenda App. You are seeing this in simple view. Click on any session to see Presenters, Discussants and a Session Description. Click on Presenters to find a single presenter and their sessions. If you find an error, please email gphillips@aria.org. Registrants have been added and will continue to be updated until 7/29/21 when registration closes. Zoom links will be shared directly with all participants two days prior to the conference. Please make notifications@sched.com a "safe sender" in your email system. We will be sending messages throughout the conference through this medium. 
Back To Schedule
Monday, August 2 • 10:45am - 12:15pm
1E1 Metamodeling for Variable Annuity Valuation What works and what does not

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

Feedback form is now closed.
Xiaochen Jing, University of Wisconsin - Madison

Variable Annuities have become popular retirement products with various options of guarantees, but their complex design also make liability management a difficult task for insurers. There have been several dozen papers published in the past years on exploring the use of statistical learning and metamodeling approaches for Variable Annuity valuation and risk management in the quantitative finance and insurance literatures. However, they all focus on specific techniques in the context of synthetic data. In this paper, I investigate the effectiveness of metamodels with different experimental designs (sample selection methods) and metamodel forms (machine learning methods) in the context of a large set of empirical Variable Annuity contracts. In particular, I use textual analysis to extract value-related information from these contracts and develop a flexible and comprehensive simulation-based scheme for Variable Annuity valuation. I find that (1) real variable annuity contracts are very complex and their liabilities are difficult to evaluate. And (2) the overall performance of a metamodel depends on the employed machine learning methods as well as the sample size—though not substantially on the sampling method. Both improve performance at the cost of longer runtime.


Thorsten Moenig

Temple University


Xiaochen Jing

University of Wisconsin - Madison

Monday August 2, 2021 10:45am - 12:15pm EDT

Attendees (4)