Development and Application of Super Learning Algorithms for Predicting Economic Time Series in High Dimensions

• Applicants: Dr. Philipp Adämmer; Professor Dr. Rainer Schüssler
• Subject Area: Statistics and Econometrics
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 468715873

Modern algorithms for predicting economic time series must handle vast amounts of variables such as macroeconomic indicators, survey data or text-based predictors. Key challenges are coping with varying levels of sparsity, non-linearities, instable predictive relationships and the heterogeneity of distinct predictors (e.g., macroeconomic vs. text-based predictors). While there are methods at disposal to tackle these challenges separately, no approach is yet available that can handle them at once.Inspired by methods from biostatistics, we propose super learning algorithms to handle the challenges at once in a unified framework. Constructing super learning algorithms involves compiling a library of conceptually different estimators and combining them to obtain an aggregate forecast. The key idea is to generate a robust and adaptive estimator by merging algorithms that capture different characteristics of the data generating process. In addition to offering a unified framework, another advantage of the super learner approach is that it works in a data-driven manner. Hence, no particular estimator has to be selected a priori for the given prediction problem.We propose a novel Bayesian method for combining large sets of candidate estimators. The approach counteracts estimation error in the combination weights and is applicable for combining point and density forecasts. We construct comprehensive libraries of point/density forecasting algorithms for macroeconomic and financial time series data. Our approach further allows the inclusion of yet to be developed algorithms into the ensemble. It also enables to dissect the components affecting forecast accuracy. For example, we can determine the incremental value of non-linear models and text-based predictors.

• DFG Programme: Research Grants