1: \begin{abstract}

2: This paper examines LASSO, a widely-used $L_{1}$-penalized regression

3: method, in high dimensional linear predictive regressions, particularly

4: when the number of potential predictors exceeds the sample size and

5: numerous unit root regressors are present. The consistency of LASSO

6: is contingent upon two key components: the deviation bound of the

7: cross product of the regressors and the error term, and the restricted

8: eigenvalue of the Gram matrix. We present new probabilistic bounds

9: for these components, suggesting that LASSO's rates of convergence

10: are different from those typically observed in cross-sectional cases.

11: When applied to a mixture of stationary, nonstationary, and cointegrated

12: predictors, LASSO maintains its asymptotic guarantee if predictors

13: are scale-standardized. Leveraging machine learning and macroeconomic

14: domain expertise, LASSO demonstrates strong performance in forecasting

15: the unemployment rate, as evidenced by its application to the FRED-MD

16: database.

17: \end{abstract}