1: \begin{abstract}

2: A typical predictive regression employs a multitude of potential regressors

3:  with various degrees of persistence while their signal strength in

4: explaining the dependent variable is often low. Variable selection

5: in such context is of great importance.  In this paper, we explore

6: the pitfalls and possibilities of LASSO methods in this predictive

7: regression  framework with mixed degrees of persistence.  In the presence

8: of stationary, unit root and cointegrated predictors, we show that

9: the adaptive LASSO asymptotically breaks cointegrated groups although

10: it cannot wipe out all inactive cointegrating variables. This new

11: finding motivates a simple but novel post-selection adaptive LASSO,

12: which we call the \emph{twin adaptive LASSO} (TAlasso), to fix variable

13: selection inconsistency. TAlasso's penalty scheme accommodates the

14: system of heterogeneous regressors, and it recovers the well-known

15: oracle property that implies variable selection consistency and optimal

16: rate of convergence for all three types of regressors. In contrast,

17: conventional LASSO fails to attain coefficient estimation consistency

18: and variable screening in all components simultaneously, since its

19: penalty is imposed according to the marginal behavior of each individual

20: regressor only. We demonstrate the theoretical properties via extensive

21: Monte Carlo simulations. These LASSO-type methods are applied to evaluate

22: short- and long-horizon predictability of S\&P 500 excess return.\\

23: \\

24: \\

25: \\

26: Key words: cointegration, machine learning, nonstationary time series, 

27: shrinkage estimation, variable selection\\

28: JEL code: C22, C53, C61

29: \end{abstract}