Predicting Student Attrition in Competitive Programming: A Large-Scale Study Integrating Survey Insights and Global Behavioral Logs
We compared the recorded behavior of competitive programmers who became inactive on Codeforces with that of active users of similar skill. From 6,927 accounts collected via the public Codeforces API in June 2026, we excluded 888 accounts with incomplete submission histories and restricted both cohorts to a peak rating below 1500. This yielded 3,421 active accounts and 908 accounts with no submissions for at least three months, labeled as attrition. Behavioral differences between the two groups were small. Attrition accounts accepted smaller shares of their attempts overall and on mathematics and data-structure problems (rank-biserial r from -0.16 to -0.12), and solved fewer problems per rated contest (median 5.95 vs. 7.80; r = -0.13). Struggle time, upsolving, and recovery after rating drops did not differ significantly after correcting for multiple testing. Two pre-specified models trained on eight observation-window-independent features achieved cross-validated AUROCs of 0.596 (logistic regression) and 0.611 (random forest), yielding predicted probabilities barely above the base rate. Across random samples, stricter filtering rules, alternative definitions of inactivity, and tenure-matched cohorts, AUROC ranged from 0.58 to 0.66. Restricting features to activity from the 6- or 12-month windows ending three months prior to each user's final submission produced AUROCs of 0.59 to 0.65, indicating that minor behavioral differences precede inactivity but remain too weak to reliably identify individual at-risk users. All data and code are publicly archived, with a single command reproducing every reported result.