Machine learning (ML) may help to predict successful psychotherapy outcomes and to identify relevant predictors of success. So far, ML applications are scant in psychotherapy research and they are typically based on small samples or focused on specific diagnoses. In this study, we predict successful therapy outcomes with ML in a heterogeneous sample in routine outpatient care. We trained established ML models (decision trees and ensembles of them) with routinely collected clinical baseline information from n = 685 outpatients to predict a successful outcome of cognitive behavioral therapy. Treatment success was defined as clinically significant change (CSC) on the Brief-Symptom-Checklist (reached by 326 patients; 48%). The best performing model (Gradient Boosting Machines) achieved a balanced accuracy of 69% (p < .001) on unseen validation data. Out of 383 variables, we identified the 16 most important predictors, which were still able to predict CSC with 67% balanced accuracy. Our study demonstrates that ML models built on data, which is typically available at the outset of therapy, can predict whether an individual will substantially benefit from the intervention. Some of the predictors were theoretically expected (e.g., level of functioning), but others need further validation (e.g., somatization). From a theoretical and practical perspective, ML is clearly an attractive addition to more established psychotherapy research methodology. •At psychotherapy outset, it is non-trivial to predict which patients will benefit.•Machine Learning (ML) is a promising addition to psychotherapy research.•ML models can predict treatment outcome in routine outpatient care.•The best performing ML model achieved a balanced accuracy of 69%.•16 Most important predictors are sufficient; they inform future research.