here are several relevant layers of uncertainty that characterise econometric models routinely used for policy making. First and foremost,  there is intrinsic uncertainty about the future conditional on a model and parameters. Then there is extrinsic uncertainty about model parameters conditional on a model. Then there is uncertainty about models conditional on a set of models. In addition there is unconditional uncertainty, when all models considered are false. In this paper we incorporate all four levels of uncertainty and we assess the improvements in the quality of prediction that we get by doing so. We provide a practical example based on the joint combination of a DSGE model, a Bayesian VAR and a dynamic factor model for a set of
US macroeconomic time series. In our paper we find that :
1.Taking into consideration parameter uncertainty is most relevant in periods of unusual data.
2. A pool with equal weights provides predictions of superior quality with respect to prediction obtained with individual models
3. We introduce a measure of value of each of the models being combined that can be decomposed across sub-periods and this measure provides important indication regarding the usefulness of the individual models.