Publicação

Forecasting mortality rates with Recurrent Neural Networks

Ver documento

Detalhes bibliográficos
Resumo:Forecasts of age-specific mortality rates are a critical input in multiple research and policy areas such as assessing the overall health, well-being, and human development of a population and the pricing and risk management of life insurance contracts and longevity-linked securities. Model selection and model combination are currently the two competing approaches when modelling and forecasting mortality, often using statistical learning methods. This paper empirically investigates the predictive performance of Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM) architecture in jointly modelling and multivariate time series forecasting of age-specific mortality rates across the entire lifespan. We empirically investigate different hyperparameter choices in three hidden layers LSTM models and compare the model’s forecasting accuracy with that produced by classical age-period and age-period-cohort stochastic mortality models. The empirical results obtained using data for Portugal suggest that the RNN with LSTM architecture can outperform traditional benchmarking methods. The LSTM architecture generates smooth and consistent forecasts of mortality rates at all ages and across years. The predictive accuracy of the LSTM network is higher for both sexes, significantly outperforming the benchmarks in the male population, an interesting result given the added difficulties posed by the mortality hump and higher variability in male survival functions. Further investigation considering other RNN architectures, calibration procedures, and sample datasets is necessary to confirm the robustness of deep learning methods in modelling human survival.
Autores principais:Bravo, Jorge Miguel
Assunto:Mortality forecasting RNN LSTM Deep learning Pensions Insurance Information Systems and Management Management Information Systems Management of Technology and Innovation Information Systems Computer Science Applications
Ano:2021
País:Portugal
Tipo de documento:documento de conferência
Tipo de acesso:acesso aberto
Instituição associada:Universidade Nova de Lisboa
Idioma:inglês
Origem:Repositório Institucional da UNL

Registos relacionados

Registos relacionados