A student-risk model travelled better after adaptation—but the smallest programme had 46 records
Across 276 student records from three programmes, covariance alignment improved several cross-program predictions. The experiment does not justify automated intervention decisions.
The short version
- What happened
- The peer-reviewed experiment used 276 student records from three programmes: 104, 126 and 46 records, with 97 on-time and 179 delayed graduations overall.
- Who it affects
- Students could receive earlier support when a model travels reliably, but false flags can stigmatise people or divert help when small datasets and structural variables are handled poorly.
- What remains unproven
- Only 276 records across three academic programmes
Keep in mind
Only 276 records across three academic programmes
Smallest target programme contained 46 records
Private dataset prevents independent replication from the published record
Target feature distributions were available during adaptation
No prospective deployment, subgroup fairness or intervention outcomes
Some model-programme combinations did not improve
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Models lost accuracy when they crossed programme boundaries
A peer-reviewed Indonesian study asked a practical question for universities with limited historical data: can a graduation-risk model trained on other programmes travel to a new one? The dataset contained 276 records across three programmes—104, 126 and 46—with 97 on-time and 179 delayed graduations. Inputs included age, gender, school of origin, admission route and grade-point averages from semesters one to four.
Four models were tested within programmes, across programmes and after CORAL adaptation, which aligns feature covariance between the source and target data. Cross-program performance usually fell. For the smallest 46-record target, random-forest F1 dropped from 0.926 within-programme to 0.737 across programmes, then recovered to 0.802 after alignment. Some programme-model combinations improved only slightly, and two combinations declined after adaptation.[1]
A better benchmark score is not a fair student-support policy
The result shows that distribution shift matters and that a lightweight statistical correction can sometimes recover performance. It does not show that any student should be labelled at risk. CORAL used the unlabelled target programme's full feature distribution, while the same fixed target records were evaluated across 30 runs. The data are private, the smallest programme contained only 46 students, and the paper did not report prospective use, calibration by subgroup or consequences of false alerts.
For advisers, the useful lesson is to test portability before adopting a model trained elsewhere—and to keep support voluntary and human-led. Gender, school background and admission route can encode structural inequality even when they improve prediction. A credible next step would freeze the model, validate it prospectively at independent institutions, publish subgroup error and calibration, compare against simple rules, and measure whether any intervention improves retention without stigmatising students. The authors reported public Indonesian funding and no competing interests.[1]
The distinction that matters
How the 46-record target programme changed the result
F1 balances precision and recall for delayed-graduation predictions; it is not a measure of student benefit.
Within programme
Train and test on similar distributions
Random forest F1 0.926
Across programmes
Train elsewhere without adaptation
F1 fell to 0.737
After CORAL
Align feature covariance to the target
F1 recovered to 0.802