LASSO is a standard tool for high-dimensional regression, but it struggles when regressors are cointegrated—a common situation in predictive models with many economic time series. The paper's authors show that LASSO's convex penalty shrinks cointegrated coefficients too aggressively, so the estimator is not consistent even as the sample grows.

They instead analyze folded concave penalties (FCP), specifically SCAD and MCP. The paper establishes two asymptotic results: a local oracle property for the penalized regression, and a global oracle property when a penalty-level constraint is imposed. The latter guarantees that a global solution consistently estimates the true coefficients and achieves oracle-level prediction performance.

Simulation studies and an application to forecasting the euro area unemployment rate confirm the theoretical advantages. Compared with LASSO, FCP delivers better variable selection and more accurate out-of-sample predictions, suggesting a practical alternative for high-dimensional cointegrated models.