A1933
Title: What can be learned about the future
Authors: Philippe Goulet Coulombe - Université du Québec à Montréal (Canada) [presenting]
Maximilian Goebel - University of Lisbon (Portugal)
Abstract: Most of the economic forecasting literature focuses on increasing the predictability of key indicators by tirelessly developing new models and algorithms -- often met with modest success, if any. The inverse of the prototypical forecasting problem is investigated. Given an information set and a particular model, we find the transformation and combination of many variables' future realizations to maximise the composite's predictability. This is implemented through a multivariate generalization of the ACE algorithm (Alternating Conditional Expectations) that we inevitably call the MACE. The approach mechanizes many manual interventions that have populated the time series econometrics practice in recent and less recent years. Among others, notable special cases include finding predictability in tails of distributions, modeling volatilities and covariances, and core inflation. MACE also allows for the algorithmic discovery of new predictable features of the macroeconomy.
