A statistical method for finding the optimal preictal period to be used in epileptic seizure prediction algorithms is presented. As supervised machine learning methods need labeled training samples, the adequate selection of preictal period plays a key role in the training of an efficient classifier employed in seizure prediction. The proposed method uses amplitude distribution histograms of a candidate feature extracted from electroencephalogram (EEG) signals. The method is evaluated on 135 hours of intracranial EEG (iEEG) recordings related to 27 epileptic seizures.