A0206
Title: Forecasting economic activity with a neural network in uncertain times: Application to German GDP
Authors: Boris Kozyrev - Halle Institute for Economic Research (IWH) (Germany) [presenting]
Oliver Holtemoeller - Martin Luther University Halle-Wittenberg and Halle Institute for Economic Research (Germany)
Abstract: The forecasting and nowcasting performance of a generalized regression neural network (GRNN) is analyzed. First, evidence from Monte Carlo simulations for the relative forecast performance of GRNN depending on the true but unknown data-generating process is provided. The analysis shows that GRNN outperforms autoregressive-moving average models in various practical scenarios. An additional check of fitting ARMA using simulated samples is provided. As a result, existing ARMA fitting approaches, even though in many cases yield similar to GRNN predictions, often cannot properly identify a true DGP. Later, GRNN is applied to forecast quarterly German GDP growth with a distinction between "normal" times and situations with significantly different time-series behaviour, such as during the COVID-19 recession and recovery. The specific data transformation needs to be implemented, i.e., dividing aggregated level values of each indicator by the corresponding GDP value. Then, these ratios are used to perform one-step-ahead forecasting using GRNN. After that, using actual aggregated observations within a given quarter, a set of GDP nowcasts is obtained. This algorithm has a high forecasting power, outperforming traditional nowcasting models ($AR(1)$, DFM, model averaging), especially during the COVID-19 crisis.
