Algorithm 2: Super Learner
* For each algorithm k:
* Perform V-fold cross validation, obtaining cross-validated predicted values Zk;
* Fit on full data O, obtaining $\widehat{\Psi }{(P)}_k$;
* Index a proposed family of convex combinations of the k algorithms by α;
* Select $\widehat{\alpha }$ to minimize E0L(O, Ψ(P)), which can be shown is solved by estimating:
logit(P̂(Y = 1|Z)) =α1Z1 + ... + αkZk;
* Save $\widehat{\Psi }{(P)}_{\mathit{SL}}$, the final estimator of Ψ(P0) = P0(Y = 1|C), constructed as:
$\widehat{\Psi }{(P)}_{\mathit{SL}}={\widehat{\alpha }}_1\widehat{\Psi }{(P)}_1+\dots +{\widehat{\alpha }}_K\widehat{\Psi }{(P)}_K$.

Note: The entire super learner algorithm above is itself externally cross-validated to obtained cross-validated performance metrics.