Stacking: linearly combine leave-one-out prediction of predictors $\hat{f}_{m}$ with weights $w$ ($w$ optimized over training set)