Functional regression with randomized signatures: An application to age-specific mortality rates

Abstract

We propose a novel extension of the Hyndman-Ullah (HU) model to forecast mortality rates by integrating randomized signatures, referred to as the HU model with randomized signatures (HUrs). Unlike truncated signatures, which grow exponentially with order, randomized signatures, based on the Johnson-Lindenstrauss lemma, are able to approximate higher-order interactions in a computationally feasible way. Using mortality data from four countries, we evaluate the performance of the novel HUrs model compared to two alternative HU model versions. Our empirical results show that the proposed HUrs model performs well, particularly for Bulgarian and Japanese data.