Data Analytics & Research
Predicting academic performance with supervised machine learning
Tertiary institution research project · client withheld under confidentiality · 9 weeks
A supervised learning study predicting student academic outcomes from behavioural, demographic and assessment data, delivered with a full interpretability and limitations report.
Reported with CI
Cross-validated accuracy
Point estimate published with confidence interval, never as a bare figure
Full handover
Reproducibility
Cleaned dataset, analysis scripts and codebook delivered
Accepted
Review outcome
Methodology cleared faculty review without a rework cycle
The challenge
- Assessment, attendance and demographic records sat in three unlinked systems with inconsistent student identifiers.
- Previous internal analysis produced a single accuracy figure with no documentation, so no reviewer could reproduce or defend it.
- The output had to survive faculty review, which meant assumptions and limitations mattered as much as the score.
How we approached it
Record linkage and a written codebook
Sources were reconciled into one analysis table with a documented matching rule for ambiguous identifiers, and every derived variable was defined in a codebook released with the deliverable.
Model selection against a baseline
Logistic regression served as the interpretable baseline; gradient-boosted trees and a random forest were evaluated against it under identical stratified cross-validation folds.
Interpretability before deployment
Feature importance and partial dependence were reported for the selected model so the review panel could see which behavioural signals drove predictions, not just how often they were right.
Limitations stated in writing
Class imbalance, cohort-specific effects and the absence of socio-economic covariates were documented as constraints on generalisation.
What changed
- The department received a re-runnable pipeline rather than a static report.
- Early-risk flagging moved from anecdote to a documented, auditable rule.
- The codebook became the institution's internal template for later studies.