The beam walk is widely used to study coordination and balance in rodents. While the task has ethological validity, the main endpoints of foot slip counts and time to cross are prone to human-rater variability and offer limited sensitivity and specificity. We asked if machine learning-based methods could reveal previously hidden, but biologically relevant, insights from the task. Marker-less pose estimation, using DeepLabCut, was deployed to label 13 anatomical points on mice traversing the beam. Next, we automated classical endpoint detection, including foot slips, with high recall (>90%) and precision (>80%). A total of 395 features were engineered and a random-forest classifier deployed that, together with skeletal visualizations, could test for group differences and identify determinant features. This workflow, named Forestwalk, uncovered pharmacological treatment effects in C57BL/6J mice, revealed phenotypes in transgenic mice used to study Angelman syndrome and SLC6A1-related neurodevelopmental disorder, and will facilitate a deeper understanding of how the brain controls balance in health and disease.