The successes of lesion-symptom mapping (LSM) have brought interest in predicting symptoms or deficit patterns from lesion features. This chapter covers two forms of lesion-based prediction. The first is predictive inference: a set of statistical procedures (such as k-fold cross-validation) that can be used to infer brain-behavior relationships. This improves the rigor and generalizability of claims about brain-behavior relationships compared to standard associationist inferences, which tend to be overly optimistic about the strength of lesion-symptom relationships. Predictive inference is fairly straightforward to implement, and some LSM methods—especially those based on machine learning—already use it. The second is prospective or longitudinal prediction: using lesion data to predict the degree of deficit at some future point in time or to predict the degree of response to a particular treatment. This requires collecting longitudinal data and accounting for structural and functional changes that occur over this period. Although these issues are challenging, they also present important opportunities for progress in both basic and translational neuroscience.