Generalizability and Transferability of Course Workload Models Across Institutions

Accurately predicting course workload is an emerging approach to identifying institutional workload trends and supporting students in mitigating risks such as excessive stress and course dropout. However, replicating these predictive models across different institutional contexts remains underexplored. In this study, we investigate whether a workload-prediction model developed at a large public research university on the United States West Coast can be generalized to a small private college on the United States East Coast, focusing on the multidimensional aspects of workload: time load, mental effort, and psychological stress. Our findings confirm that credit hours alone cannot capture students' workload experiences at both universities. Instead, rich features derived from learning management system (LMS) and enrollment data explain substantially more variance in perceived workload at both institutions. However, applying a model trained in one context "out-of-the-box" to another led to systematically biased predictions, often underperforming naive baselines. Despite these challenges, we demonstrate that fine-tuning with even a small amount of local data (20-30 courses) can substantially reduce prediction error, aligning model performance with locally trained alternatives. Combined models trained on data from both institutions achieve competitive accuracy compared to single-institution models, suggesting the viability of an institution-general approach to workload prediction. These results highlight the promise and complexity of transferring course workload analytics across contexts. Leveraging richer data features-beyond credit hours-consistently improves workload predictions, yet local institutional calibration remains crucial. Our study thus offers practical insights into developing scalable, data-driven models for workload management and underscores the potential for multi-institutional collaborations to reduce the burden of new data collection. All data analysis code used for this study is publicly available at https://github.com/CAHLR/cla-replication-jedm.