A0293
Title: Boosting XGBoost: Using the panel dimension to improve machine-learning-based forecasts in macroeconomics
Authors: Johannes Frank - Friedrich-Alexander University Erlangen/Nuremberg (Germany) [presenting]
Jonas Dovern - Friedrich-Alexander University Erlangen-Nuremberg (Germany)
Abstract: The short-time dimension of commonly used macroeconomic data sets presents challenges for the estimation of machine learning models designed for real-time business cycle monitoring. We consider panel data to increase the data set available for training and nowcasting US unemployment using extreme gradient boosting and neural networks. The underlying idea is that dynamics between variables and across time at the state level are similar to each other and to the dynamics at the national level. We use data pooling in combination with weight sharing that accommodates some cross-sectional heterogeneity. This approach facilitates parameter regularization and safeguards against overfitting. We find that this soft pooling approach improves forecast accuracy at the national level and reduces both the variance and the mean of the RMSE distribution across states. Thus, leveraging regional information in a panel data framework with suitable regularization techniques addresses data scarcity in macroeconomic nowcasting effectively.