Predicting optical equipment failures reliably before they happen, offers the promise to significantly improve network Quality of Service (QoS) and to reduce maintenance costs. State-of-the-art prediction methods for Network Fault Prediction are not easily comparable. In this paper, a Decision-Tree based Machine Learning benchmark for optical Network Fault Prediction alarms is presented based on a real-world dataset. The Machine Learning architectures are compared with a fixed lead time between the prediction and the window for which alarms are considered. Precision, recall, F1-score metrics, a cost function representing theoretical monetary gain are used for the comparison. Additionally, a Quality of Service gain metric is proposed and used for the comparison. Perspectives for future research are proposed.