A1548
Title: A sparse dynamic factor approach to mortality modelling
Authors: Jianjie Shi - Monash University (Australia) [presenting]
Abstract: Dynamic factor models (DFM) are an appealing and effective tool for handling a large number of time series. Despite its popularity among empirical macroeconomists, there are still some challenges that have made the DFM less favourable in practice. One challenge is the concern of over-parameterization, especially if there are many latent factors. This problem becomes even more acute when faced with high-dimensional time series. Another challenge is model interpretability which is often considered unattainable due to the lack of identifiability. Motivated by the finite mixture model, we develop a new sparse dynamic factor model (SDFM) that can achieve sparsity and enhance interpretability by classifying a set of time series into several different groups. The idea of SDFM is to simultaneously build several single-factor models while keeping the dependence structure among those factors via a dynamic mechanism. By applying the SDFM to French mortality data, we fit and forecast age-specific mortality rates parsimoniously. We compare the forecasting performance of this model against the ordinary DFM. Our results show that the sparse dynamic factor model generally provides superior forecasts when applied to French mortality data.