Nowcasting with Machine Learning Methods

This master's thesis aims to nowcast the Czech GDP growth and compare several machine learning regression models' performances. While GDP is the most important macroeconomic indicator, there is a publication lag of around 70 days in the Czech Republic, and it is published quarterly. This publication lag hinders rapid decision-making and effective policy implementation. For this reason, while the early estimation of GDP is crucial, many government agencies and Central Banks are working and using models to nowcast GDP. There are many monthly independent variables in GDP nowcasting models. The MIDAS method suggested in the literature was used for the variables explained at different frequencies to work in a regression model in the most effective way. Machine learning regression models were run after the MIDAS application, estimates were created for train and test sets, and errors were obtained. Performance improvements to machine learning models have been realized with hyperparameter tuning. While this means finding the optimal values of the specific parameters for all models, the amount of error is noticeably reduced in the results obtained from the train and the test set. This thesis compared the performances of all machine learning models used on tables and figures.