Learning from Micro-Level Expert Forecasts: Real-Time Data, Regression Trees, and Bagging

Abstract:     
While macroeconomic survey forecasts are widely available at the level of individual experts, it is not clear how to optimally combine a set of forecasts to a "consensus"' prediction. This is mainly due to the characteristics of the data, such as the large-dimensional predictor space, many missing values, and potential individual and aggregate level biases of the survey forecasts. We argue that regression trees are very well adapted to these features and propose to use them as a novel forecast combination device. Our empirical analysis of data from the Philadelphia Fed's Survey of Professional Forecasters demonstrates that in combination with bagging, tree-based forecast combination outperforms equally weighted combination for the majority of time series and forecast horizons.

Keywords: Macroeconomic Survey Data, Survey of Professional Forecasters, Bagging, Regression Trees, Real-time Data

JEL Classification: C14, C23, C43, C53

working papers series