Economic Forecasting in a Data-Rich Environment for Actuarial Applications Using Neural Networks

With the advent of Big Data and machine learning/AI technologies, actuaries can now develop advanced models in a data-rich environment to achieve better forecasting performance and provide added value in many applications. Traditionally, economic forecasting for actuarial applications is developed using econometric models based on small datasets including only the target variables (usually around 4-6) and their lagged variables. This paper explores the value of economic forecasting using deep learning with a big dataset (FRED Database) consisting of 121 economic variables and their lagged variables covering periods before, during, after GFC, and during COVID (2019-2021). Four target variables considered in this paper include inflation rate, interest rate, wage rate, and unemployment rate, which are common variables for social security funds forecasting. The proposed model combines dimension reduction via principal component analysis (PCA) and Neural Networks (including CNNs, LSTMs, and fully-connected layers), which is suitable for economic forecasting in a data-rich environment. The results show that the proposed model consistently outperforms the benchmark vector auto-regression (VAR) model, although the level of benefits varies across different economic variables and forecast periods. Using residual bootstrapping, this paper provides prediction intervals to quantify the prediction uncertainty. To provide explanations for the black-box Neural Networks, this paper uses SHAP values to understand how different economic variables influence the prediction outcome. The model performance is demonstrated using a social security fund forecasting application.