Time series forecasting is challenging as sophisticated forecast models are computationally expensive to build. Recent research has addressed the integration of forecasting inside a DBMS. One main benefit is that models can be created once and then repeatedly used to answer forecast queries. Often forecast queries are submitted on higher aggregation levels, e.g., forecasts of sales over all locations. To answer such a forecast query, we have two possibilities. First, we can aggregate all base time series (sales in Austria, sales in Belgium ...) and create only one model for the aggregate time series. Second, we can create models for all base time series and aggregate the base forecast values. The second possibility might lead to a higher accuracy but it is usually too expensive due to a high number of base time series. However, we actually do not need all base models to achieve a high accuracy, a sample of base models is enough. With this approach, we still achieve a better accuracy than an aggregate model, very similar to using all models, but we need less models to create and maintain in the database. We further improve this approach if new actual values of the base time series arrive at different points in time. With each new actual value we can refine the aggregate forecast and eventually converge towards the real actual value. Our experimental evaluation using several real-world data sets, shows a high accuracy of our approaches and a fast convergence towards the optimal value with increasing sample sizes and increasing number of actual values respectively.