Objective: To design a VCG lead system dedicated to the analysis of atrial fibrillation. Methods: Body surface potentials during atrial fibrillation were simulated by using a biophysical model of the human atria and thorax. The XYZ components of the equivalent dipole were derived from the Gabor-Nelson equations. These served as the gold standard while searching for methods for the derivation of the vectorcardiogram from a limited number of electrode positions and their transfer coefficients. Six electrode configurations and dedicated matrices were tested using 10 different episodes of simulated AF and 25 different thorax models. Results: RMS-based relative estimation error of the VCG using the Frank electrodes was 0.39. An adaptation of the 9 electrodes of the standard ECG, including one electrode on the back, reduced the error to 0.24. Conclusion: The Frank lead system is suboptimal for estimating the VCG during AF. Alternative electrode configurations should include at least one electrode on the back.