Comparison of prediction methods for macroeconomic variables through AI

The prediction of macroeconomic variables is used daily both by economic policy makers and by industries that seek to consider the effects of the national economy in their sales projections, credit rating models, customer segmentation, etc. These predictions are calculated using macroeconometric, machine learning, and more recently, deep learning models. This paper reviews the history and state of the art of two main methods of prediction of economic aggregates: structural and non-structural models. In this context, several models are estimated through auto-ARIMA and recurrent neural networks methodologies that seek to predict gross domestic product, quarterly accumulated inflation and the urban unemployment rate for the Ecuadorian case until the fourth quarter of 2019. The results suggest that, in the case of gross domestic product, the ARIMA methodology is a better fit; whereas, for inflation and unemployment, recurrent neural networks obtain better error metrics. At the end, possible improvements are presented that continue this line of research, through the optimization of hyperparameters and the assembly of models.

• Main Author: Porras Encalada, Hugo Boanerges (author)
• Format: article
• Language: spa
• Published: 2021
• Subjects: ARIMA; LSTM; Predicción macroeconómica; Redes neuronales recurrentes; Series de tiempo
• Online Access: https://estudioseconomicos.bce.fin.ec/index.php/RevistaCE/article/view/322