A1705
Title: Advancing forecast accuracy analysis: A partial linear instrumental variable and double machine learning approach
Authors: Christoph Schult - Halle Institute for Economic Research (Germany) [presenting]
Katja Heinisch - Halle Institute for Economic Research (Germany)
Fabio Scaramello - University of Venice (Germany)
Abstract: The relationship is explored between forecast accuracy and forecast assumptions using German data and a novel empirical approach. Partial linear instrumental variable regression models are employed, combined with double machine learning methods to address issues of high-dimensional nuisance parameters and endogeneity. This innovative framework enables a more complex understanding of the relationships between forecast assumptions and forecast accuracy than traditional OLS-based analysis. The evaluation sample ranges from 1992 to 2019 and includes 1460 annual GDP forecasts and various assumptions for oil, exchange rate, and world trade. The model's inherent flexibility allows for the examination of two possible violations of assumptions of the model used in a prior study: the rationality of forecasters and the linearity of the data generation process. This research contributes to the field of forecasting by providing a more robust and flexible analysis of forecast accuracy determinants. For instance, the contribution is to the discussion regarding weak instruments and instrumental variables' validity in macroeconomic models. Further, evidence of serious differences between OLS-based estimates is found, as proposed by a prior study, and results based on DML estimates. In particular, a constant underestimation of OLS estimators of the impacts of squared assumption errors of oil price and world trade on squared forecast errors of GDP is reported.
