Predicting Student Attrition in Competitive Programming: A Large-Scale Study Integrating Survey Insights and Global Behavioral Logs

Synthesis: A dual-layer predictive framework for understanding student attrition in competitive programming, combining large-scale Codeforces behavioral logs (n = 1,816 after filtering and balancing) with a multi-institutional psychographic survey from 10 universities in Bangladesh (n = 64 for predictive modeling).

Behavioral findings: what precedes attrition

The behavioral analysis confirmed that true attrition is preceded by an 83.71% reduction in contest participation and consistent underperformance on skill-related metrics. The authors also validate a Skill-Application Paradox: students who stop report higher theoretical confidence than their active peers yet exhibit significantly weaker practice habits (significant at p < .001 for upsolving habit and peer circle density). This disconnect between self-reported confidence and actual practice behavior is a distinctive Self-Efficacy signal for early-warning.

Predictive modeling

Machine learning benchmarks demonstrated that a Soft-Voting Ensemble achieved the strongest performance on the behavioral dataset (CV F1 = 0.737, Test Recall = 0.769), while Random Forest led on the survey dataset (CV F1 = 0.924, interpreted as a localized exploratory pilot). An engineered Intensity Ratio ranked as the third most predictive Codeforces feature, capturing independent practice effort beyond raw activity counts.

Early Warning System proof of concept

Applied as a proof-of-concept Early Warning System, the survey-trained model identified four high-risk active students whose behavioral profiles corroborated the model's predictions. This demonstrates the feasibility of flagging at-risk students from mixed survey and behavioral signals.

What this means for practice

• Instructors. Track practice habits rather than stated confidence. Students who stopped reported higher theoretical confidence than their active peers while their upsolving habit and peer-circle density were significantly weaker (p < .001), so a confidence signal alone will mislead you.
• Instructors. Act on abrupt drops in contest participation — true attrition was preceded by an 83.71% reduction — and pair experienced with newer participants while monitoring students who show irregular engagement after extended breaks.
• Designers. Build early-warning dashboards on behavioral telemetry, including engineered indicators such as the Intensity Ratio, rather than survey responses alone: the survey model's CV F1 = 0.924 was trained on only 64 students and overstates generalizability.
• Administrators. Treat the four flagged students as a proof of concept, not a deployment. Fund longitudinal validation over a 2–3 month window before such risk flags inform high-stakes decisions about real students.

Limitations

• The survey data are small (n = 73 collected, n = 64 used for predictive modeling) and limited to ten universities in Bangladesh, which the authors note may not generalize to other academic or cultural contexts.
• Attrition is a proxy: the Codeforces labels use an inactivity threshold that cannot distinguish genuine dropout from temporary academic breaks.
• Retrospective recall bias is a stated limitation: upsolving habit and thought-quit were reported by stopped students after they disengaged, which may systematically inflate the apparent separation between active and stopped cohorts on self-reported measures, and both datasets are observational, permitting no causal conclusions.
• The Early Warning System ran over only 22 active students with no follow-up data, so it is unverified whether the four flagged high-risk individuals actually attrited or declined.

Citation

Ruhan, A. I., Naeem, G. M., Rafi, R. I., Mim, S. A., Opi, N. B., Chowdhury, D. F., & Sadi, M. R. K. (2026). Predicting Student Attrition in Competitive Programming: A Large-Scale Study Integrating Survey Insights and Global Behavioral Logs. arXiv:2608.28618.