Predicting the upcoming alarm state helps industrial plant operators take timely actions to avoid the negative effects of abnormal conditions. This paper proposes a new data-driven method for trend forecasting in a process variable. The proposed method has two important features: (i) alarm states are predicted in real-time whenever a new trend occurs in the current time sequence, which helps industrial operators take timely actions at the early stage of abnormal conditions; (ii) alarm states are predicted with quantitative probabilities and credibility intervals. By contrast, existing early warning methods are either suitable for specific scenarios where a short prediction step length is needed or cannot provide a quantitative measurement of prediction results. The proposed method firstly captures qualitative trends in the current and historical time sequences and clusters these trends into several categories. Secondly, current and historical trend sequences are obtained according to the cluster results. Finally, recent trends in the current trend sequence are chosen, and alarm states are predicted based on the historical trend sequence having the same sub-sequence. A Bayesian estimator is exploited to calculate the posterior probabilities and their credibility intervals for predicted alarm states. The optimal predicted alarm state is the one with the largest lower bound of posterior probability as a quantitative measurement of prediction accuracy. The effectiveness of the proposed method is illustrated via numerical examples based on a well-accepted benchmark of the Tennessee Eastman process.