We examine the problem of forecasting the spatial extent of a just-occurred traffic incident’s impact and the travel delay induced by it at certain future time points. We present and evaluate a machine learning-based solution for the above problem. The proposed solution is based on a standard classification model that takes in a variety of input features that include the incident attributes and features derived from traffic sensor data. We evaluate several versions of the solution by varying the classification model, the number of impact classes, the type of training data, and the time at which the prediction is made. This is done by conducting a series of experiments using a real-world traffic incident dataset along with the corresponding traffic sensor data. In particular, we investigate the issue of class imbalance in the incident dataset, the disparity in the class-wise prediction accuracies, the benefit of taking the incident’s early impact into account, and the relative importance of the input features. The findings of this study are potentially insightful to practitioners and researchers in the field of intelligent traffic management.