transformer

TRACE predicted courses and grades jointly, cutting grade error without establishing intervention benefit

Paul Savala

arXiv preprint

Savala introduces TRACE, a transformer that represents courses by semester and jointly predicts a student's next course set and corresponding grades. On ten years of institutional data, joint training reduced mean absolute grade error by nearly 50 percent compared with the same architecture predicting grades alone and outperformed LSTM and graph baselines. External validity, fairness, calibration, and intervention effects remain open.