Predicting the future past. How useful is machine learning in economic short-term forecasting?

Many macroeconomic variables are first available when they have already become part of the past. At the same time, there is a large set of figures that is indeed available without the same delay. This thesis use machine learning models to transform real-time data into forecast of future releases about the current and near future state of the economy, asking if such methods produce more accurate forecasts than conventional econometric models. The results indicate that there are no single superior forecasting device. The closest to a "free lunch" is an unweighted average of the simple benchmark models