Advances in macroeconometrics: (interpretable) machine learning and high-frequency data for forecasting and structural analysis

Forecasting and modelling techniques for structural analy- sis have changed through the years to cope with the com- plexity of macroeconomic systems. Recent results show evi- dence that non-parametric models such as machine learning are helping with the prediction of macroeconomic variables. On the other side, high-frequency information is widely used to provide a new source of information for structural analy- sis. This thesis contributes to all these aspects by proposing innovative approaches for forecasting macroeconomic indi- cators and providing an alternative way to make structural analysis. We first exploit the ability of an ensemble learning model combining long-short-term memory neural network (LSTM) and dynamic factor model (DFM) to detect nonlin- earities in the US GDP forecast. We also provide an inter- pretable methodological framework that uses Shapley values to generalize the data-generating process learned by neural networks and applies it to predict inflation levels. The result- ing polynomial relations between the variables provide pol- icymakers with valuable insights on the potential nonlinear relations between the evolution of future price levels and eco- nomic activity. In addition, we propose a new identification method for Structural Vector Autoregressive (SVAR) models based on nowcasted (high-frequency) macroeconomic data.