In this work an algorithm called PREFIX based on performance data retrieved by OSC (Optical Supervision Channel) boards with the aim to predict fiber faults in OTN networks is presented. Data pre-processing techniques including data cleaning, data ordering, feature engineering, dataset rebalancing, data standardization and Principal Component Analysis (PCA) are proposed to leverage raw performance over time series. Moreover, a Random Forest machine learning model has been exploited for prediction purpose. The whole tests were performed using 6-month long real traces obtained over a live optical network composed by 300 optically amplified unidirectional links. Results are presented over ROC curve and trade-offs between precision and recall.