International Symposium on Nonlinear Theory and its Applications
Nonlinear and Nonparametric Models for Forecasting the US Gross National Product
Siddharth Arora, Max A. Little, Patrick E. McSharry,
PDF download (101.9KB)
Nonlinear models, like the self-exciting threshold autoregressive (SETAR) model and Markovswitching autoregressive (MS-AR) model have been proposed for modeling the Gross National Product (GNP) time series. Both SETAR and MS-AR, however, require estimation of a large number of parameters relative to the small amount of GNP observations. While modeling the training data reasonably well, these models tend to overfit and perform poorly in terms of out-of-sample forecasting. The aim here is to investigate the efficacy of a novel parsimonious nonparametric and nonlinear model, which can outperform SETAR and MS-AR in terms of out-of-sample GNP forecasting accuracy. As it is important to quantify forecast uncertainty, leading to well informed policy-making, we generate both point and density forecasts. We evaluate point forecasts using the root mean square error (RMSE) and mean absolute error (MAE), while density forecasts are evaluated using the continuous ranked probability score (CRPS).