Ground reaction forces (GRFs) are crucial for understanding movement biomechanics and for assessing the load on the musculoskeletal system. While inertial measurement units (IMUs) are increasingly used for gait analysis in natural environments, they cannot directly capture GRFs. Machine learning can be applied to predict 3D-GRFs based on IMU data. However, previous studies mainly focused on vertical GRF (vGRF) and isolated movement tasks. This study aimed to systematically evaluate the prediction accuracy of convolutional neural networks (CNNs) for 3D-GRFs using IMUs from single and multiple sensor configurations across various movement tasks. 20 healthy participants performed six movement tasks including walking, stair ascent, stair descent, running, a running step turn and a running spin turn at self-selected speeds. CNNs were trained to predict 3D-GRFs on IMU time series data for different configurations (lower body [7 IMUs], single leg [4 IMUs], femur-tibia [2 IMUs], tibia [1 IMU] and pelvis [1 IMU]). Prediction accuracies were assessed based on leave-one-subject-out cross validations using Pearson correlation (r) and relative root mean squared error (relRMSE). Across all tasks, CNNs predicted vGRF most accurately (r = 0.98, relRMSE ≤ 7.44 %), followed by anterior-posterior GRF (r ≥ 0.92, relRMSE ≤ 14.24 %), with medial–lateral GRF being the least accurate (r ≥ 0.74, relRMSE ≤ 29.46 %). CNNs predicted vGRF consistently across tasks, with similar accuracy for multi IMU (average r = 0.98, average relRMSE:​ 6.06 %) and single IMU configurations (average r = 0.98, average relRMSE:​ 6.88 %), supporting single IMU configurations for vGRF in practical applications. During cutting maneuvers, the lower body configuration reduces the relRMSE for mlGRF (5.23–12.46 %) and apGRF (1.53–3.16 %) compared to single IMU configurations. However, for mlGRF and apGRF during cutting tasks, lower body configuration improve accuracy, highlighting a trade-off between simplicity and performance.