Multiple sclerosis (MS) is marked by heterogeneous disease activity, progression, and therapeutic response. Here, we developed a prognostic score based on machine learning using randomized clinical trials (RCTs) and observational datasets. The score's ability to predict short-term prognosis, expressed as absolute risk, was assessed on the French population in the context of all common MS therapeutic scenarios. Nine industrial RCTs and one prospective cohort from the French MS registry were used to develop several types of multilabel binary classifiers designed to predict the two-year risk of relapse, advent of new brain T2 lesions, and sustained disability worsening, as well as the respective yearly risks. Model evaluation prioritized calibration of probabilistic predictions over discriminatory capacity. Virtual cohorts simulated from the model predictions were analyzed to determine how the predictive score captured clinically meaningful information, such as the efficacy of different therapeutic classes. The model with the best calibration was evaluated externally on the population-based cohort of the French MS registry. Random forest modeling optimally captured the time-course of MS risks. In the evaluation dataset, calibration shifted with underconfident predictions of relapse and overconfident predictions of new brain T2 lesions. Nevertheless, unadjusted average therapeutic class efficacy on MRI activity generalized well. At external validation, discriminatory capacities were modest: AUC = 0.67, 0.75, and 0.58 for relapse, new brain T2 lesions, and sustained disability worsening, respectively. Based on a panel of variables currently available during routine care for MS patients, we propose a score predictive of short-term therapeutic response to commonly prescribed therapeutic classes. Predictions of MRI activity generalized well across the common therapeutic scenarios. The model's probabilistic approach, emphasizing prediction certainty rather than the prediction itself, captured clinically useful information for the selection of disease-modifying treatments.