Template-type: ReDIF-Article 1.0
Author-Name: Alexandra Bozhechkova
Author-Email: bojechkova@ranepa.ru
Author-Workplace-Name: RANEPA; Gaidar Institute for Economic Policy
Author-Name: Urmat Dzhunkeev
Author-Email: dzhunkeev@gmail.com
Author-Workplace-Name: Universite Catholique de Louvain; Universite Paris 1 Pantheon-Sorbonne
Title: CLARA and CARLSON: Combination of Ensemble and Neural Network Machine Learning Methods for GDP Forecasting
Abstract: This paper is devoted to the use of ensemble and neural network machine learning methods, in particular, CatBoost and LightGBM methods and convolutional neural networks to forecast the GDP. The study uses a vintage database to identify the impact of statistical revisions on the accuracy of models. Our results show that a combination of neural network methods retains a predictive advantage over the benchmark models - a first-order autoregression, a dynamic factor model, and a Bayesian vector autoregression - across a panel of countries, including the pandemic crisis periods, based on preliminary and revised data. According to the econometric test for confidence in the set of models, convolutional and recurrent neural networks are among the most accurate methods for GDP forecasting. Revisions of the statistical data lead to an increase in the root mean square errors of the benchmark models and the ensemble and neural network machine learning methods.
Classification-JEL: C52, C53, E37
Keywords: gradient boosting, neural networks, GDP, BVAR, dynamic factor models, vintage data
Journal: Russian Journal of Money and Finance
Pages: 45-69
Volume: 83
Issue: 3
Year: 2024
Month: September
DOI:
File-URL: https://rjmf.econs.online/upload/iblock/c8b/ocamyrdyrdfpgmlimx00ohmt26wql976/CLARA-and-CARLSON-Combination-Methods-for-GDP-Forecasting.pdf
Handle: RePEc:bkr:journl:v:83:y:2024:i:3:p:45-69