Since signal peptides play a crucial role for specifying the in-vivo fate of proteins, prediction of their existence is important for the characterization of ORFs of unknown function. To make such predictions as reliable as possible, the features of signal peptides of two important model organisms, Saccharomyces cerevisiae and Bacillus subtilis, were examined and the accuracy of current prediction methods was refined using these data. Direct optimization of the threshold values of existing methods significantly raised the predictability but the variables that were most effective for improvement were different in these two organisms. In yeast, the maximum hydrophobicity value of an 8-residue segment mainly contributed to raising the predictability to 98.5% when estimated by the cross validation procedure. In Bacillus species, the length of uncharged segment and the charges in the N-terminal region (net charge and negative charge) were combined to give a prediction accuracy of 98.2% although the data size was relatively small in this case.