Sicco Kooiker

Sicco Kooiker

Econometrics/ML PhD candidate at the Vrije Universieit Amsterdam. Research focuses on adaptive machine learning methods with applications in macroeconomics.

Multi-period Growth-at-Risk Forecasting with Recurrent Neural Network

Published in Working Paper, 2024

Authors: Sicco Kooiker, Lukas Hoesch, and Julia Schaumburg

Abstract:

We propose to forecast multi-horizon Growth-at-Risk using flexible dynamic sequence models. Our estimation routine combines many-to-many recurrent neural networks with an objective function that guarantees non-crossing of the resulting quantile estimates. The architecture of the model is highly flexible yet applicable in typical sample sizes of macroeconomic data sets. We establish the finite sample properties of our method in a simulation study considering a range of linear and nonlinear data generating processes. In an empirical illustration, we find that the method clearly outperforms linear quantile regression in predicting the economic vulnerability of 24 OECD countries.

Presentations:

  • ISF Conference, Dijon, France (2024)
  • National Econometrics Study Group, Rotterdam, Netherlands (2023)

Recommended citation: Kooiker, S., Hoesch, L., & Schaumburg, J. (2024). Multi-period Growth-at-Risk Forecasting with Recurrent Neural Network. Working Paper.