Student Dropout Prediction Model — Private University Group, Pune, India
The Situation
A private university group running 6 campuses across Maharashtra and Karnataka was losing between 11% and 14% of enrolled students to dropout each academic year. The numbers varied by campus and by programme, but the pattern was consistent: by the time a student's struggle was visible to an academic advisor, the student had usually already decided to leave.
The data to catch this earlier existed. Attendance records, LMS login frequency, assignment submission patterns, internal assessment scores, and fee payment histories were all being collected across all 6 campuses. None of it was connected. Each system sat with a different administrative team and was reviewed, when it was reviewed at all, in isolation.
The university's accreditation body had flagged dropout rates in two consecutive annual reviews. That made this a compliance issue as much as an academic one.
What Amorisoft Did
Amorisoft started with a 3-week data audit across all 6 campus systems. Seven years of historical student records were extracted, cleaned, and linked at the individual student level across attendance, LMS, assessment, and fee payment data. This produced a dataset of 41,000 student records with full academic lifecycle coverage.
The predictive model was trained to identify dropout risk at the 8-week mark before a student's last recorded academic activity. That window was chosen after analysing historical dropout timelines: 8 weeks was the point at which risk signals became statistically reliable while still giving advisors enough time to act.
Feature engineering drew on 24 variables across 4 data categories. The two strongest predictors turned out to be LMS login frequency in weeks 3 to 6 of a semester and the gap between assignment due dates and submission timestamps. Attendance, which the university had assumed would be the primary signal, ranked fourth.
The model outputs a weekly risk score per student, visible to academic advisors through a dashboard integrated into the university's existing student information system. Advisors receive a flagged list each Monday morning. No student data is visible to anyone outside the student's own campus.
Campus-level calibration was applied after initial training because dropout patterns varied between engineering programmes and management programmes significantly enough to affect model accuracy. A single uncalibrated model produced 71% accuracy. Campus-programme calibration brought it to 81%.
Results
Dropout rates across all 6 campuses dropped by 34% in the first full academic year after deployment. The model's at-risk predictions were confirmed accurate in 81% of cases reviewed by academic advisors. Of students flagged and contacted by advisors, 62% were retained through to the end of their enrolled programme. The 8-week warning window held in production, matching the timeline established during training.
The university submitted the dropout reduction figures to its accreditation body in the annual review following deployment. The flag raised in the two prior reviews was not repeated.