1: \begin{abstract}

2: This paper tackles forecast combination with many forecasts

3: or minimum variance portfolio selection with many assets. A novel convex

4: problem called $\ell_{2}$-\emph{relaxation} is proposed. In contrast to

5: standard formulations, $\ell _{2}$-relaxation minimizes the squared

6: Euclidean norm of the weight vector subject to a set of relaxed linear

7: inequality constraints. The magnitude of relaxation, controlled by a tuning

8: parameter, balances the bias and variance. When the variance-covariance (VC)

9: matrix of the individual forecast errors or financial assets exhibits latent

10: group structures --- a block equicorrelation matrix plus a VC for

11: idiosyncratic noises, the solution to $\ell_{2}$-relaxation delivers

12: roughly equal within-group weights. Optimality of the new method is

13: established under the asymptotic framework when the number of the

14: cross-sectional units $N$ potentially grows much faster than the time

15: dimension $T$. Excellent finite sample performance of our method is

16: demonstrated in Monte Carlo simulations. Its wide applicability is

17: highlighted in three real data examples concerning empirical applications of

18: microeconomics, macroeconomics, and finance.

19: 

20: {\footnotesize \medskip{} }

21: 

22: {\footnotesize \noindent \textbf{Key Words}: Forecast combination puzzle;

23: high dimension; latent group; machine learning; portfolio analysis;

24: optimization.}

25: 

26: {\footnotesize \noindent \textbf{JEL Classification}: C22, C53, C55 }

27: \end{abstract}