Your School ERP Already Knows Which Students Are Slipping

Every school runs on a system that already knows more about its students than any single teacher can hold in their head. Attendance is marked twice a day. Every assessment result is entered. Every disciplinary note, every fee transaction, every timetable change lands in the same database. Then, once a month, someone exports a slice of it into a PDF, emails it to the head, and it dies there.

The claim here is narrow and hard to argue with: schools do not need a new data source to identify a student sliding off track. They need to stop letting the one they already have rot in monthly reports. That holds for a district in Boston, a multi-academy trust in London, and a school group in Kathmandu alike, because the underlying data model is the same everywhere — a register, a gradebook, a behaviour log.

The signals that predict, and the ones that just look like data#

The research here is unusually settled. Decades of longitudinal work across US districts converged on three families of indicator — attendance, behaviour, and course performance, the “ABCs” — as the strongest school-held predictors of whether a student finishes. The NCES Forum Guide to Early Warning Systems documents the thresholds districts settled on: missing 10 percent or more of instructional time, any course failure in the transition year into secondary school, a GPA under 2.0 in the middle grades, two or more behaviour infractions.

The most instructive finding comes from the UChicago Consortium’s work on the Freshman On-Track indicator, which found that on-track status at the end of the first year of high school is more predictive of graduation than race, socioeconomic status and prior achievement combined. Almost all students with a B average at the end of that year graduate; roughly a quarter of those with a D average do. Course attendance proved far more predictive of course failure than test scores.

Read that again, because it kills most of what gets sold as student analytics. Test scores are weaker than the register. Demographic risk profiles are weaker than what the student did last week. And the things vendors love to surface — engagement scores derived from learning-platform clickstreams, sentiment ratings, dashboards ranking students by a composite index nobody can decompose — are noise dressed as insight until somebody demonstrates they add signal over attendance and grades on that school’s own cohort.

A second filter cuts the feature list fast: is the signal actionable? Nobody can intervene on a postcode or a household income band. They can intervene on three missed Mondays.

Leading and lagging is the whole game#

A failed term is a lagging indicator. By the time it exists, the student has already lost the term. A three-week attendance slide is a leading indicator — it is happening now, and it is reversible now.

Most school reporting collapses this distinction because everything is aggregated to the cadence of the slowest signal. Attendance is captured daily and then thrown away into a termly percentage — a lagging summary of a leading signal, which is close to the worst of both worlds. The fix is a modelling decision, not a technology one:

• Match the window to the signal. Attendance gets rolling windows measured in days and weeks. Course performance gets assessment-level tracking, not term averages. Behaviour gets counts over a trailing period.
• Track trajectory, not level. A student at 92 percent attendance who was at 99 percent a month ago is a more urgent case than one sitting steadily at 88 percent. Level tells you status; slope tells you what is changing.
• Require persistence. One absence is noise. A threshold crossed and held for a defined number of consecutive periods is a signal. Without a persistence rule, the flag list churns every morning and staff learn to ignore it.

Hospitals worked this out earlier. A Hospital Management System that surfaces a deterioration score computes it from observations taken hours apart, not from the discharge summary.

The model has to be readable, and that is a requirement#

There is a strong temptation to reach for gradient boosting over four hundred features. Resist it. Interpretability here is not a nice-to-have you trade against a couple of points of AUC — it is load-bearing.

Cynthia Rudin’s argument in Nature Machine Intelligence applies almost verbatim: for high-stakes decisions, build a model that is inherently interpretable rather than a black box you then attempt to explain after the fact. Post-hoc explanations are approximations of the model, not accounts of it, and they can be confidently wrong in exactly the cases you care about.

Three reasons this bites harder in schools than elsewhere. The output is handed to a form tutor who has to act on it, and a tutor who cannot see why a student surfaced will not act, or will act on the wrong thing. The flag is contestable: a parent is entitled to ask why their child was singled out, and “the model said so” is not an answer any school should be willing to give. And staff need to be able to override, which requires visible reasoning.

The interpretable option is also genuinely competitive here. The ABC threshold rules are a model — a rule-based one, derived from exactly the longitudinal analysis a modern approach would run. A logistic regression over a dozen well-chosen features, with coefficients a head of year can read, is a defensible ceiling for most schools. Spend the effort you save on feature quality and on the loop that follows the prediction.

A model trained on history learns history’s inequity#

This is the part that gets skipped, and it is the part most likely to cause harm.

A model fitted on historical outcomes learns which students the institution previously failed. If a group was under-supported for a decade, the model faithfully learns that membership of that group predicts poor outcomes, and encodes a self-fulfilling loop. The behaviour indicator is the sharpest example: a disciplinary record is a log of adult decisions about referral, not a neutral measurement of student conduct. Any model weighting it heavily is partly modelling staff behaviour.

Fairness is also not a property you establish once. A nationwide study of evolving bias in dropout prediction presented at ACM Learning @ Scale examined how model bias shifts over time, and other work has documented uneven predictive performance across institutions — which is how a system meant to close gaps ends up widening them.

What mitigation looks like in practice, stripped of the ethics-statement register:

• Evaluate by subgroup, and lead with false negatives. Overall accuracy hides everything. The equity-critical metric is the false negative rate — the students who needed help and were not flagged — measured separately for every group you can lawfully measure.
• Exclude protected attributes and their tight proxies from features; retain them for evaluation. You need the attribute to audit, not to predict.
• Calibrate thresholds per site. A trust-wide threshold tuned on the aggregate will systematically over- or under-flag individual schools.
• Discount the behaviour indicator. Weight attendance and course performance ahead of it, and audit referral patterns as a separate exercise.
• Re-audit on a schedule. Cohorts change, policies change, and a model validated in 2024 is an assumption in 2026.

A prediction with no owner changes nothing#

The strongest available evidence on early-warning systems is not about prediction quality at all. A randomised controlled trial run by REL Midwest assigned 73 high schools — some 37,000 students — to implement the Early Warning Intervention and Monitoring System or continue usual practice. After one year, chronic absence was lower in EWIMS schools than in control schools, 10 percent against 14 percent, and course failure was 21 percent against 26 percent. There was no statistically significant difference in the share of students earning enough credits to be on track, and suspension rates were identical.

The lesson is not that prediction works. It is that a process works: a named team, a defined review cadence, students assigned to specific interventions, progress monitored afterwards. The prediction is the cheapest component in that list.

So build the loop into the software, not into a policy document nobody reads. Every flag gets an owner, a due date, a recorded action, and a follow-up that closes the case or escalates it. The system’s primary job is case management; scoring is a feature of it. And size the flag rate to the staff you actually have — if the model surfaces a third of the cohort, nothing happens to any of them.

Minors’ data raises the bar, not lowers it#

Two rules matter more than the compliance checklist. Purpose limitation: data collected to keep a child in school must not leak into admissions decisions, fee-collection pressure, or marketing. Access control: a risk flag visible to the entire staff room stops being a prompt and becomes a label that follows the student around, and it should never be shown to the student as a score.

Jurisdictions differ in detail and agree in substance. In the UK and EU, a school’s lawful basis here is generally public task rather than consent, which makes transparency and data minimisation the operative duties rather than a consent banner. In the US, FERPA governs education records but covers less of a minor’s privacy in K-12 than people assume. Wherever you are, write down the retention period and enforce it — risk history from four years ago is not a live signal, and keeping it is pure downside.

The ERP is a data platform, or it is a filing cabinet#

None of this is achievable on a student information system that treats data as something to be stored and printed. That is the real constraint in most schools, and it is a vendor decision, not a technical one: no event-level history, no queryable interface, a reporting module whose only output is the monthly PDF, and a data model that makes “attendance by student by week for the last six weeks” a support ticket rather than a query.

The alternative worth holding out for is an ERP built as a data platform with a thoughtful interface on top — events retained at event grain, an analytical layer that answers trend questions in milliseconds, an API the school owns, and export rights that are real. Once that exists, the early-warning system is a modest piece of work: a few dozen lines of feature logic, an interpretable model, a case-management screen. Operational Automation handles the routine end of the loop — nudging an unclosed case, batching the weekly review pack — while judgement stays with staff. That sequencing is the honest answer to most AI implementation questions in education: fix the data platform first, and the intelligent layer becomes almost boring to build.

The data is already there. It is being collected today, by people who will never see it again. That is the waste worth fixing.

---

The register you already keep is the best predictor you have — the problem is that it stops at the PDF. Tell us what your student data is trapped in at /#contact.