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

Competitive programming develops algorithmic reasoning in computer science students, yet sustained participation is fragile and many withdraw after stalled progress or thin mentorship. Dropout has been studied extensively in degree programs and online courses, but competitive programming attrition has not been examined as a behavioral phenomenon in its own right. This paper reports a two-layer study combining platform telemetry with a localized survey. We retrieved activity traces for 6,927 Codeforces accounts and reduced them to a balanced set of 1,816 users, half persistently active and half showing frustrated withdrawal, matched on peak rating so that disengagement is not confounded with skill. We additionally audit the features for temporal leakage, removing three columns whose predictive weight derives from when a user stopped rather than what they did. In parallel we surveyed students at ten universities in Bangladesh. Attrition follows an 83.7 percent fall in contest participation and a 42.7 percent drop in problems solved monthly. Students who stopped practised significantly less between contests while rating their own ability no lower than active peers, a divergence we term the Skill-Application Paradox. Random Forest reached a cross-validation F1 of 0.682 on the audited behavioral data and 0.832 on the survey layer.