Detection and interpretation of fault systems and stratigraphic features and the relationship between them are crucial for seismic interpretation and reservoir characterization. To provide better interpretation insight and to be able to extract overlooked features out of seismic data volumes, we have developed a new attribute that detects faults and other discontinuities while handling local nonstationary variations across them. First, we used predictive painting to form a structural prediction of seismic events from neighboring traces (left and right neighboring traces in 2D and neighboring traces in all directions around a reference trace in 3D) according to the local structural slopes. Then, we computed prediction residuals by subtracting each prediction from the original data, and we found the smallest prediction-error interval for each point that best represented discontinuity information at that point. The extracted fault information changed with location (spatially and temporally), and it was nonstationary. Conventional coherence measures operate on a spatial window of neighboring traces and a temporal (vertical) analysis window of samples above and below the analysis point, and they can hardly cope with nonstationarity in fault information. In contrast, in our method, neither temporal nor spatial windows were involved in coherence computation, which allowed us to honor nonstationary changes of fault information and to achieve high resolution in the vertical and lateral directions. To assess the performance of the proposed attribute, we compared it with the conventional coherence attribute over the same data set. The comparison demonstrated the effectiveness of discontinuity detection using predictive coherence and showed its value in extracting additional information from seismic data.