Business intelligence for student retention in higher education

Institutional context and problem scope

A university with 18,000 enrolled students across 6 campuses and 47 programs faced a first-year dropout rate of 24%. Dropout was not uniform: it was concentrated in the first semester of engineering and health sciences programs (where it reached 31%), in first-generation university students (38% first-year dropout), and in peripheral campuses with less support infrastructure.

The detection process was entirely reactive: the earliest risk signal the institutional system captured was cumulative absenteeism above 30% — an indicator that in most cases manifests when the student has already decided to leave. The tutoring team's interventions arrived on average 67 days after the student had begun academically disengaging.

The economic cost of dropout was substantial: each percentage point reduction in dropout rate represented approximately 180 students retained, with a tuition revenue impact of $2.7M annually per point. The social cost was even greater, given the institution's critical role in the social mobility of first-generation university students in its region.

Data sources and 360-degree profile construction

The project began with an audit of available information systems. The university had five independent record systems: an LMS (Moodle), an academic ERP (Banner), a library system, a tuition payment portal, and a student welfare system with records of psychological and social support requests. No system was integrated with the others, and information about the same student was fragmented across five databases without a standardized common identifier.

The first month of the project was dedicated to data integration and cleaning. A unique student identifier was built that unified the five sources, and incremental ingestion pipelines with quality validation were designed. The process revealed that 34% of LMS records had no direct correspondence with the academic ERP due to user identifier inconsistencies — a problem that required a reconciliation process based on name, date of birth, and program.

• LMS (Moodle): weekly accesses, active time per course, completed vs. pending submissions, forum participation, and downloaded resources
• Academic ERP (Banner): partial and final grades, attendance by subject, records of dropped or frozen courses
• Library system: physical and digital accesses, material loans by subject area, study room usage
• Tuition portal: payment status, overdue days, installment plan requests, and scholarship records
• Student welfare: support requests, active social work referrals, attendance at soft skills workshops
• Satisfaction surveys (institutional NPS): applied in week 4 and week 10 of the semester, with sentiment analysis of open comments
• External context data: distance from student's residence to campus, public transport dependency, family employment situation

Modeling methodology and feature selection

Three analysis cohorts were built using 3 years of historical data (9 semester cohorts, 54,000 student-semesters). For each student, whether they had dropped out within the analyzed semester was labeled. The target variable was binary: formal dropout within the semester (enrollment withdrawal or non-payment leading to administrative removal).

The feature selection process was iterative and guided by specialists from the tutoring and student affairs team. 127 candidate features were evaluated in three categories: academic (grades, attendance, course load), engagement (LMS activity, library, institutional events), and contextual (financial status, socioeconomic profile, distance). The 35 final model features were selected combining statistical importance (SHAP values from the baseline model) and operational relevance validated with the tutoring team.

A relevant finding from the feature analysis was that LMS digital engagement signals had higher predictive power at 6-8 weeks into the semester than partial grades, which were only available from week 10-12. This meant the model could generate useful alerts 4-6 weeks before the traditional academic system generated any signal.

• Academic features (12): cumulative GPA, attendance rate by subject, course load vs. historical performance, at-risk subjects by program
• LMS engagement features (8): weekly accesses, active time, submission rate, activity trend in last 2 weeks
• Financial features (6): overdue days, debt amount, installment plan history, current scholarship type
• Welfare features (5): support requests, active referrals, attendance at support instances
• Contextual features (4): first-generation university student, distance to campus, transport mode, employment during studies
• Predictive NPS feature: institutional satisfaction score estimated before the formal survey, updated weekly based on engagement and context

Churn models and alert system architecture

Four model families were evaluated on the development dataset: L2-regularized logistic regression (interpretable baseline), Random Forest, XGBoost, and a 3-layer feed-forward neural network. The model selected for production was XGBoost for three reasons: best AUC-ROC in stratified cross-validation (0.847 vs. 0.831 for second best), greater performance stability across different historical cohorts (lower AUC variance between cohorts), and native SHAP value support for per-student explainability.

The alert system was configured with three risk levels: high (dropout probability > 65%), medium (35-65%), and low (< 35%). The high-level threshold was calibrated to maximize coverage of actual dropouts (recall) with an acceptable false positive rate for the tutoring team: on average each tutor receives between 8 and 12 high-risk alerts per week, a manageable volume that allows personalized intervention.

The sentiment analysis module processes open comments from NPS surveys using a BERT model fine-tuned on a corpus of 12,000 manually annotated comments from Latin American university students. The model classifies comments into 8 categories (academic difficulty, economic problems, health issues, lack of sense of belonging, transport problems, teaching dissatisfaction, family problems, and positive/neutral). These categories guide the type of intervention the tutor should prioritize.

• Primary churn model: XGBoost with 35 features, AUC-ROC 0.847 in cohort-stratified cross-validation
• Probability calibration: Platt method to convert scores into well-calibrated probabilities by risk level
• Sentiment analysis: BERT fine-tuned on Latin American university corpus, 8 detectable problem categories
• Predictive NPS: regression model estimating student satisfaction score before formal survey, updated weekly
• Tripartite alert system: three risk levels with thresholds calibrated for manageable intervention volume per tutor
• Score update: weekly recalculation of dropout probabilities with previous week's data, with trend history per student

Validation and anticipation horizon calibration

Model validation was performed with backtesting across 9 historical semester cohorts (3 years, 54,000 student-semesters). For each cohort, system behavior was simulated: the model was trained on prior cohorts and its ability to identify that cohort's dropouts using only data available in week 6 of the semester was evaluated.

The backtesting results showed the model could identify 71% of actual dropouts in each cohort in week 6 of the semester, with a precision of 58% (meaning 58% of high-risk alerts corresponded to students who actually dropped out). The 29% of dropouts not identified by the model in week 6 were predominantly students dropping out for low-predictability reasons with available data: acute personal crises or unanticipated external events.

The average alert anticipation relative to the formal dropout event was 45 days in the validation dataset. In 23% of cases, anticipation exceeded 60 days — the window in which scholarship and financial support interventions have the highest probability of resulting in retention.

Implementation and tutor workflow

The retention dashboard is deployed as a web application accessible from any device, with institutional authentication via SSO. Each tutor sees exclusively their assigned students, with the list ordered by current risk level. For each student at high or medium risk, the dashboard shows: the risk score and its trend over the last 4 weeks, the 3 features contributing most to the current score (expressed in plain language, not technical values), the most likely problem category according to prior survey sentiment analysis, and a suggested action panel with the history of prior interventions.

The tutor workflow was designed with three principles: minimum friction for intervention recording (the tutor can record the outcome of a student conversation in less than 90 seconds), sufficient context for the first conversation to be productive (the tutor arrives at the meeting with relevant academic and welfare context pre-processed), and closing the feedback loop (each recorded intervention updates the intervention effectiveness model by problem type and student profile).

• Risk dashboard: student list ordered by score, with 4-week trend and top-3 risk factors in plain language
• Per-student context panel: academic activity, financial situation, welfare history, and prior survey sentiment in a single view
• Intervention recording workflow: maximum three fields, recording time < 90 seconds, no redundant required fields
• Automatic alerts to program directors: when the percentage of high-risk students exceeds the configured threshold by campus or program
• Optional student notifications: proactive tutoring invitation messages configurable by the student affairs team
• Integration with scholarship system: automatic alerts to the scholarship office when a high-risk student has an active scholarship that could be lost due to performance

Results and effectiveness analysis

Results were measured by comparing the first-year dropout rate in the post-deployment semester against the equivalent semester of the prior year, controlling for context variables (total enrollment, program mix, economic environment). The first-year dropout rate was reduced from 24% to 16.6%, an absolute reduction of 7.4 percentage points (-31% relative).

The effectiveness analysis by intervention type showed that financial support interventions (accelerated access to scholarships or payment plans) had the highest impact on students identified with more than 45 days of anticipation (67% post-intervention retention rate), while academic tutoring interventions were more effective for students identified between 20 and 45 days before the dropout event (54% retention rate). This finding allowed optimizing the intervention protocol by alert type and anticipation time.

• First-year dropout rate reduction: from 24% to 16.6% (-31% relative, -7.4 absolute points)
• Average alert anticipation: 45 days before the formal dropout event
• Model coverage: 71% of dropouts identified in week 6 of the semester with available data
• Post-intervention retention rate: 67% for financial interventions with >45 days anticipation, 54% for academic tutoring
• Reduction in average time of tutor's first intervention: from 67 days to 12 days since the start of disengagement
• Estimated economic impact in first operating semester: retention of 1,260 additional students equivalent to $3.4M in tuition revenue

Lessons learned and ethical considerations

The most important technical lesson was that the quality of digital engagement features outperforms academic features as early predictors. This conclusion goes against institutional intuition (which tends to value grades as the primary risk indicator) and required a careful communication process with the executive team for them to accept allocating early intervention resources based on LMS activity before grades were available.

The most relevant ethical consideration was the design of the consent and privacy process. Student data is processed under the enrollment contract and institutional data use policies, but the team decided to go beyond the legal minimum: an explicit opt-out process was implemented (any student can request their data not be used for the retention alert system), and psychological welfare data is used only as aggregated features (problem category, not the content of sessions) to protect the confidentiality of the therapeutic relationship.

• ✓LMS digital engagement features are earlier predictors than grades: prioritize their ingestion quality from the start
• ✓The 45-day anticipation operational target defines week 6 as the critical scoring threshold — any data latency above 5 days invalidates this target
• ✓Alert volume per tutor is a critical design parameter: more than 15 high-level alerts per week exceeds the capacity for personalized intervention
• ✓Psychological welfare data requires a higher level of protection than the rest of the profile — it must be used only as aggregated features
• ✓Intervention effectiveness feedback is a necessary condition for improving the model over time
• ✓System communication to students must be framed as support, not surveillance: language and channel matter as much as content