A1027
Title: Broken adaptive ridge in time series regression: The TS-BAR
Authors: Orest Prifti - University of Rome Tor Vergata (Italy) [presenting]
Gianluca Cubadda - University of Rome Tor Vergata (Italy)
Luca Margaritella - Lund University (Sweden)
Abstract: The broken adaptive ridge (BAR) estimator is extended to high-dimensional, multivariate autoregressive time series models where the number of variables can exceed the sample size, addressing the curse of dimensionality. The BAR is a sparsity-inducing technique that iteratively reweights the L2-penalty to mimic the selection properties of the non-convex L0 penalization. Additionally, BAR enjoys a grouping effect -where highly correlated covariates are automatically jointly selected- as well as the oracle property, thus performing asymptotically as if the true model were given. In sparse, high-dimensional settings, Monte Carlo simulations show how the BAR can perform better than the LASSO and its variants. Its superior feature selection and forecasting precision are further confirmed by two empirical applications in macroeconomics and finance.
