Actuarial Science and Financial Mathematics seminar series
Hong Beng (Ben) Lim
Chinese University of Hong Kong
Room: M3 3127
The Promise of Neural Tangent Kernel Regression in Actuarial Science: Application to Variable Annuity Valuation and Mortality Forecasting
There has traditionally been a steep learning curve associated with using neural networks. Even a fitting basic network requires one to tune hyperparameters with respect to the optimization routine and the network architecture, dissuading many actuarial practitioners from considering it as part of their toolkit. This is despite such networks having a strong track record in predictive and classifcation performance, both inside and outside of the actuarial field. One promising avenue towards alleviating this issue is the development of the Neural Tangent Kernel (NTK) in broader machine learning. It has been shown that fitting kernel regression using this kernel is equivalent to training an infinite-width neural network, hence combining the ease of use and optimization of kernel regression with the predictive power of neural networks. This talk aims to illustrate the potential ways that NTK regression can be used in actuarial science. We illustrate the advantages of such a methodology, as well as discuss hyperparameter tuning via an application to variable annuity valuation, and discuss how the methodology may be scaled to larger datasets through an application to mortality forecasting.