<p>When data are assimilated in a reservoir history matching problem, not all resulting models are equally reliable for forecasting the future behavior. Not only are many of the model assumptions incorrect, but the assimilation methods themselves are imperfect. In fact, even in the perfect model scenario in which all assumptions are correct, the data assimilation algorithm itself may be imperfect, and the posterior ensemble from a given data assimilation procedure may be much different from the correct posterior. In these cases, it is not the correctness of the scenario that is of ultimate importance, but rather the usefulness of the individual model realizations for determining forecast quality. In this paper, the forecast realizations are stacked after history matching to improve predictability. The effectiveness of the approach is demonstrated on applications with redundant observations for which the fit to observed data provides a good surrogate for the fit to true data, and for problems with noisy fitness landscapes that allow minimizations to models with poorly matched data. Although the goal is not to identify the correct scenario, incorrect scenarios were either eliminated or heavily downweighted in some cases. Computation of the stacking weights requires only the solution of a constrained linear optimization problem and does not require alteration of the data assimilation methods or high computational cost.</p>