Early Prediction of Student Dropout in a Large-Scale Cram School Using Longitudinal Test Performance Data

Student dropout is a major challenge in large-scale supplementary education, where early identification of at-risk learners is essential. Although prior educational data mining studies have explored dropout prediction using learning logs, long-term test-based data common in offline settings remain underexplored. This industry paper reports a real-world application of dropout prediction using longitudinal standardized test data from over 3,000 junior high school students in a large Japanese cram school. We engineered interpretable features capturing achievement dy-namics, including recovery efficiency after score declines and trend consistency. Random Forest models trained on a previous cohort were validated on an independent cohort. The best-performing model achieved approximately 89\% accuracy and a dropout F1-score of 0.80, enabling prediction before the final academic year. Feature importance analysis showed that recovery from performance declines was a stronger indicator of dropout risk than absolute achievement. These results demonstrate that longitudinal test data can support accurate and interpretable dropout prediction at scale, offering practical value for early intervention.