How should researchers combine predictive densities to improve their forecasts? I propose consistent estimators of weights which deliver density forecast combinations approximating the true predictive density, conditional on the researcher's information set. Monte Carlo simulations confirm that...