Ground reaction forces are an important measure in gait analysis but traditionally require force plates and laboratory equipment to measure. For this project, I worked with a teammate to develop machine-learning models that predict a runner's vertical ground reaction force directly from 3D motion-capture data. We combined and standardized two public running-biomechanics datasets, developed a data-processing pipeline to identify and extract individual foot strikes, and trained and evaluated deep neural network and spatial-temporal graph neural network models across different running conditions. Our best models achieved R² values of approximately 0.95, demonstrating the potential to eventually estimate running forces from video-based pose data without force plates.