Biomechanical models driven by motion capture (MoCap) offer unprecedented insight into musculoskeletal (MSK) function and aid clinical decision-making. However, traditional MSK models are computationally expensive, laborious to implement, and require meticulously curated inputs. Such models are being complemented by machine learning (ML) methods for user-friendly, real-time predictions, which, however, have often lacked rigorous implementation and assessment. Choosing a fit-for-purpose ML technique is fraught with trade-offs. Here, we show the comparative implementation of nine ML techniques on relatively understudied human upper-extremity MSK modelling from optical MoCap input data (non-invasive gold standard). We identified and investigated model selection and accuracy, generalisability, robustness (to instrumentation errors, soft-tissue artefacts, and anatomical landmark misplacement---inherent to optical MoCap systems), model complexity, transferability (from intact-limbed participants to ‘mimicked’ transradial prosthesis usage), and interpretability. We also undertook the first assessment of data sufficiency using learning curves and the carbon footprint of training/inference. We found convLSTM to be the optimal ML technique, which efficiently learns the spatial and temporal aspects of MoCap data, while random forest offers a computationally-efficient alternative with minimal accuracy trade-off. This novel holistic characterisation helps lay the methodological foundation towards better deployment (via increased interpretability and robustness) of ML pipelines in biomechanical studies. Finally, we provide best practices and a reporting guideline (LearnABLE) for systematic implementation and transparent reporting of ML techniques, aiding their development to better complement and improve traditional MoCap-driven biomechanical modelling.