A0196
Title: Local predictability in high dimensions
Authors: Philipp Adaemmer - University of Greifswald (Germany) [presenting]
Sven Lehmann - University of Rostock (Germany)
Rainer Alexander Schuessler - University of Rostock (Germany)
Abstract: A novel time series forecasting method is proposed, which is designed to handle vast sets of predictive signals, many of which are irrelevant or short-lived. The method transforms heterogeneous scalar-valued signals into candidate density forecasts via time-varying coefficient models and, subsequently, combines them into a final density forecast via time-varying subset combination. The approach is computationally fast, because it uses online prediction and updating. We validate our method through simulation analyses and apply it to forecast daily aggregate stock returns as well as quarterly inflation, using over 12,000 and over 400 signals, respectively. We find superior forecasting performance and lower computation time for our approach compared to competitive benchmark methods.
