In this paper, we propose a microfounded framework to investigate a panel of forecasts (e.g. model-driven or survey-based) and the possibility to improve their out-of-sample forecast performance by employing a bias-correction device. Following Patton and Timmermann (2007), we theoretically justify the modeling of forecasts as function of the conditional expectation, based on the optimization problem of individual forecasters. This approach allows us to relax the standard assumption of mean squared error (MSE) loss function and, thus, to obtain optimal forecasts under more general functions. However, different from these authors, we apply our results to a panel of forecasts, in order to construct an optimal (combined) forecast. In this sense, a feasible GMM estimator is proposed to aggregate the information content of each individual forecast and optimally recover the conditional expectation. Our setup can be viewed as a generalization of the three-way forecast error decomposition of Davies and Lahiri (1995); and as an extension of the bias-corrected average forecast of Issler and Lima (2009). A real-time forecasting exercise using the Brazilian Focus survey illustrates the proposed methodology