Multiscale decomposition of Big Data time series for analysis and prediction of macroeconomic data: a recent approach

Abstract

The problem of the extraction of the relevant information for pre- diction purposes in a Big Data time series context is tackled. This issue is especially crucial when the forecasting activity involves macroeconomic time series, i.e. when one is mostly interested in finding leading variables and, at the same time, avoiding overfitted model structures. Unfortunately, the use of big data can cause dangerous overparametrization phenomena in the enter- tained models. In addition, two other drawbacks should be considered: firstly, humandriven handling of big data on a case-by-case basis is an impractical (and generally not viable) option and secondly, focusing solely on the raw time series might lead to suboptimal results. The presented approach deals with these problems using a twofold strategy: i) it expands the data in time scale domain, in the attempt to increase the likelihood of giving emphasis to possibly weak, relevant, signals and ii) carries out a multi-step dimension reduction procedure. The latter task is done by means of crosscorrelation functions (whose employment will be theoretically justified) and a suitable objective function.