Macroeconomic Forecasting: Econometrics vs. Machine Learning
Machine Learning for Social Scientists, CERGE-EI2026
Abstract
Compares ten models for GDP growth and inflation nowcasting in the Czech Republic and the United States: AR, ARMA, VAR and BVAR against Ridge, Lasso, XGBoost, Random Forest, a feedforward network and an LSTM. Over an expanding-window evaluation for 2014–2026, regularised regressions and tree ensembles cut GDP forecast errors by 45–55% relative to the AR benchmark from 2020 onward but overfit in stable times. Classical time-series models dominate inflation nowcasting, and the LSTM consistently underperforms.
The Length of the Pre-Doctoral Period and Academic Performance
Statistics, CERGE-EI2026
Abstract
Fixed-effects estimation on a panel of five million observations, with a critical evaluation of the identification strategy and possible extensions. Presented in the Applied Microeconomics Research Seminar Series.