A1193
Title: Bayesian shape-constrained forecasting
Authors: Afshan Feroz - Macquarie University (Australia) [presenting]
Abstract: A comprehensive Bayesian method is presented to predict mortality by combining statistical modeling, machine learning, and shape-constrained estimation. The Bayesian Lee-Carter model is modified by an ensemble of ARIMA models, singular spectrum analysis (SSA) and multilayer perceptron (MLP) techniques to sex-specific French mortality. Volatility is made more stable by using discrete wavelet transforms of projected projections, estimated projections, and smoothed projections. Age-specific shape constraints guided by demographic theory are applied, such as monotonicity for adults and convexity for the elderly. The methodology shows a significant improvement in the accuracy of the prediction based on the most essential evaluation factors (RMSE, MAE, MAPE, and SMAPE). The benefits of smoothing and shape constraints are shown to be statistically significant by cross-validation, and Wilcoxon signed rank tests on residual errors. The framework is extended to project future mortality trends after 2019 to provide insight into age-structured demographic change during and after the COVID-19 epoch. It is pointed out that using domain knowledge and sophisticated forecasting methods is essential to develop more accurate and reliable mortality forecasts.
