This study addresses the problem of rumour scarcity versus non-rumour abundance in automatic rumour detection. To tackle this issue, we portray rumour as an anomaly by showing how disproportionate is the number of rumours versus non-rumours. This imbalance is scrutinized by comparing the rate of news production versus rate of fact-check production. Then, we exploit one-class classification approach to distinguish rumour from non-rumour. One-class classification separates rumour from non-rumour via training the classifier with only non-rumour. To train the one-class classifier, we extract 33 short-term features, regarding the purpose of this research in early detection of rumours. We evaluate the performance of our model by accuracy and F-score. In terms of F-score, our model outperforms the state-of-the-art and reaches to very close proximity of highest accuracy on the same dataset.