Forecasting of time series data is crucial for decision-making processes in many domains as it allows the prediction of future behavior. In this context, a model is fit to the observed data points of the time series by estimating the model parameters. The computed parameters are then utilized to forecast future points in time. Existing approaches integrate forecasting into traditional relational query processing, where a forecast query requests the creation of a forecast model. Models of continued interest should be deployed only once and used many times afterwards. This however leads to additional maintenance costs as models need to be kept up-to-date. Costs can be reduced by choosing a well-defined subset of models and answering queries using derivation schemes. In contrast to materialized view selection, model selection opens a whole new problem area as results are approximate. A derivation schema might increase or decrease the accuracy of a forecast query. Thus, a two-dimensional optimization problem of minimizing the model cost and model usage error is introduced in this paper. Our solution consists of a greedy enumeration approach that empirically evaluates different configurations of forecast models. In our experimental evaluation, with data sets from different domains, we show the superiority of our approach over traditional approaches from forecasting literature.