CFE 2015: Start Registration
View Submission - CFE
A0293
Title: Short-term forecasting with mixed-frequency data: An appraisal of supervised factor and shrinkage methods Authors:  Boriss Siliverstovs - KOF ETHZ (Switzerland) [presenting]
Abstract: A wide variety of variable selection, dimension reduction and shrinkage methods for forecasting with mixed-frequency data is applied. In particular, we assess the effectiveness of supervised learning using the following algorithms for variable pre-selection: Least Absolute Schrinkage and Selection Operator (LASSO), partial least squares (PLS), principle covariate regression (PCovR), bagging algorithm (short for boostrap aggregation), boosting algorithm, the non-negative garotte (NNG) algorithm as well as recently suggested the three-pass regression filter. Big data become even bigger due to the use of the skip-sampling or blocking approach for frequency conversion of high-frequency variables, making the task of selecting most informative variables regarding the target variable even more challenging. Assessing which method or their combination produces most reliable results bears a direct relevance for forecasting practitioners.