In recent years, car options like air-conditioning, automatic gears, car stereo, power windows, sunroof etc. were getting more and more important for car manufacturers. Especially at those car manufacturers which offer cars individually according to customer requirements, the options affect 30% - 40% of the parts by now. Therefore, very detailed planning and forecasting of options becomes more and more important. Because customer behavior concerning the options varies from model to model and from country to country, it seems necessary and reasonable to forecast each option for each model and each country separately. The resulting huge number of data sets requires an automatic forecasting tool that adapts itself to the actual data sets and that requires almost no user interaction. Because depending on the characteristics of a time series the quality of the forecasting results varies a lot, and because "N heads are better than one," the basic idea is m select in a first step the most appropriate forecasting procedures. This selection is done by a decision tree which is generated by using a symbolic machine learning algorithm. Those selected forecasting methods produce different results that are in a second step combined to get a common forecast. In this approach there are integrated univariate time series used in the first step for running the prediction as well as symbolic machine learning algorithms for generating the decision trees as well as multivariate statistical methods and neural networks used in the combination step.