Incorporating Industry Stylized Facts into Mortality Tables: Transfer Learning with Monotonicity Constraints

Abstract

Developing mortality tables can be challenging when a firm lacks credible experience data or the expertise to adjust such rates, especially for smaller life firms or underwriters serving the senior life settlements industry. The recently proposed entity embedding neural network (EENN) methodology aims to resolve this issue by extracting features from industry tables reflecting such information for use in fitting mortality rates for such firms. Notably, such fitted rates do not incorporate desired monotonicity constraints e.g. with respect to attained age. We extend the EENN methodology to do so, which has the added benefit of regularizing rates and, as we show, improving predictive performance.

Keywords: Mortality modeling, transfer learning, entity embeddings, neural network, monotonic regression, monotonic network