Abstract. Support vector machines (SVM) are a widely used state-of-the-art classifier in molecular diagnostics. However, there is little work done on its overfitting analysis to avoid deceptive diagnostic results. In this work, we investigate the important problem and prove that a SVM classifier would inevitably encounter overfitting for gene expression array data under a standard Gaussian kernel due to the built-in large data variations from DNA amplification mechanism in the transcriptional profiling. We have found that SVM demonstrates its own special overfitting characteristics on array data, in addition to showing that feature selection algorithms may not contribute to overcoming overfitting, and discussing overfitting in biomarker discovery algorithm.