A Non-Parametric Model for Forecasting Life Expectancy at Birth Using Gaussian Process
Gaussian Process Regression (GPR), a Bayesian nonparametric machine learning modelling technique, is gaining interest in recent times in a wide range of fields as a practical and powerful approach. To plan for economic services for any nation, projections of future Life Expectancy (LE) are required. In our research, we have proposed a model to forecast LE using GPR up to 2040. Initially, we sub-categorized countries into four sections based on income level. Then we treated LE at birth for different countries as a time series to create our model. Among the data of 165 countries we have, we used 27 countries' 60 years of LE data (1960-2019) to optimize and visualize the performance of our model. In our model, we tried to maximize log-marginal-likelihood (lml) for each prediction, while optimizing the hyper-parameters of our models. We further verified our model by using cross-validation; fitting the model into 40 years of data and validating on the other 20 available. The results we got from our prediction model demonstrated the subtle increase of LE over the years, which varied depending on the income groups. We have made the data processing and model development code publicly available via GitHub to carry forward this research.