We introduce a new data oversampling technique, the Negative Calibrated Generative Adversarial Network (NCGAN). In fraud detection, high-quality data oversampling is the key to successful classifications. When dealing with ultra imbalanced data in binary classification scenarios, the similarity of synthetic positive data to negative data is an undesirable property. Our method focus on alleviate such undesirable property by calibrating synthetic positive samples with negative samples. We formulate our NCGAN in mathematical notations and conduct experiments to validate that NCGAN is universally effective for linear models, tree models and deep learning models. Compared with traditonal method, our NCGAN will increase the precision and F1 score for machine learning models.