Mortality rate predictions for Italy in 2020 using Lee-Carter model and Recurrent Neural Networks
Predict total italian mortality rate in 2020 with the following models:
- Lee-Carter
- Bayesian Lee-Carter
- Shallow LSTM
- Shallow time-distributed LSTM
- Deep LSTM
- Deep GRU
The dataset contains observations of mortality rate from 1974 to 2020 for 24 different age classes, ranging from age 0 to age 199 (source: ISTAT). Each age class x
is predicted using the two adjacent age classes x-1
and x+1
as features in the RNN models.
The shallow LSTM and time-distributed LSTM are used to produce a one-step-ahead forecast for 2020, while the remaining models are used to produce a 10-step-ahead forecast for the period 2010-2020.
The Lee-Cater model produces the best results in terms of training and test forecast error.
Launch a Jupyter notebook with R kernel using the base Docker image jupyter/r-notebook:r-4.1.2
:
- Build the Dockerfile
docker build -t notebook .
- Run on http://localhost:8888/
docker run --rm -p 8888:8888 -v ${PWD}/R:/home/jovyan/work/R notebook
References: Richman, R., & Wuthrich, M. V. (2019). Lee and Carter go machine learning: Recurrent neural networks.