Ensemble forecast of analyses (EFA) couples a classical data assimilation method with sequential aggregation of ensemble forecasts. The assimilation method produces analyses whenever new observations become available. At the same time, an ensemble of given simulations is generated in order to forecast future time steps. The objective of EFA is to forecast the upcoming analyses (to be generated with the future observations) with a linear combination of the ensemble of forecasts. The weights of the linear combination depend on the ensemble simulation, on time and on the state component. Hence, for any forecast variable and at any location in space, EFA linearly combines the ensemble of forecasts to produce one single forecast of the upcoming analysis for the given variable at the given location. In previous work, aggregation was carried out by machine learning algorithms that are well adapted to operational forecasting as they enjoy robustness properties. The approach is indeed applied on the French national air quality forecasting platform, Prév'air. Nevertheless, the method does not provide uncertainty estimations. Instead of machine learning, we propose to apply a minimax filter on the aggregation weights, which allows us to preserve the forecast performance while introducing uncertainty estimation along with the forecasts. The approach forecasts the upcoming analyses, which are the best a posteriori representation of the real state in some sense, significantly better than any model. A Kalman filter can also be applied on the aggregation weights; we will discuss how it compares to the minimax filter in this context. The method will be illustrated for the forecast of peak ozone concentration fields over Europe, based on an ensemble of chemistry-transport models.