Traffic light (TL) classification is an important feature for automated driving, and it requires correct color separation of the TL signals captured using cameras. A key camera component for the color separation performance is the color filter array (CFA). For common automotive-specific CFAs, we have observed unsatisfactory performance for TL color separation, which indicates the need for an optimization. Based on typical scenarios for TL classification and a set of recorded TL signals, we evaluate the performance of common automotive CFAs. For a quantitative evaluation, we propose a suitable color distance metric. We also propose a method for optimization of the CFA and show that using this method, reference color separation performance can be achieved, trading in only a small amount of sensitivity.