Reformulate the stacking optimization problem

[thomaswiemann]

The formulation of the stacking optimization problem for the MSPE minimizing ensemble weights in ensemble_weights is unlikely to be computationally efficient. Consider an alternative approach, for example, via non-negative least squares.

[thomaswiemann]

The package causalDML also implements an ensemble procedure (see here). Weights are solved for using the nnls package. This seems like a good approach.

[thomaswiemann]

To further simplify, one could compute the ensemble weights without any constraints (i.e., "weights" only in name then, as potentially negative and mustn't sum to one) via least squares.

As an alternative to the inital implementation for the fully constraint set of weights (i.e., all weights in [0, 1] and sum(weights) = 1), use a quadratic programming solver such as gurobi.