Enrollment Melt Prediction

Enrollment melt is the stretch between an admitted student who has signaled intent, often with a deposit, and a student who is actually registered when the term starts. A model does not close that stretch. It ranks live files so a person can intervene before the enrollment deadline.

Ship a quality flag: a student identifier, a risk band or score, and cites to the deposit, aid, and engagement rows the model used. If those cites are missing or contradict each other, leave the cell empty. Empty means the file is too thin to score, not that the student is safe to ignore or safe to drop.

A high flag means this file needs a human look. It does not mean the seat is already gone. Do not pull a wait-list name, cut a section, or stop working the student because a percentile looks grim. Predictive yield scoring still owns admit-to-deposit risk. Melt work starts after intent is on file and runs until your published enrollment deadline or census.

Campus CRM and SIS suites such as EAB, Slate, Technolutions, Salesforce, and Ellucian already store the events you need. Treat them as systems of record, not as a ranked shortlist. The contract is the same in any of them: first-party cites in, counselor decision out.

Build the score from rows a counselor can open next to the flag.

Deposit cites: paid, waived, or pending deposit; deposit date versus the deadline; refund or cancellation request; housing or orientation holds that usually travel with the deposit. A deposit that never posted is not melt. That is an incomplete yield file. Keep it out of this queue, or label it, so staff do not mix never committed with committed and cooling.

Aid cites: packaging status, verification flags, outstanding documents, SAI or EFC as stored, offered versus accepted awards, and any documented unmet need. Do not infer need from ZIP code or high school when aid has no FAFSA or institutional application on file. If the aid cell is thin, leave the aid cite blank. Do not fill it with a proxy.

Engagement cites: first-party events only. Portal logins, required checklist items, campus visit or admitted-student day attendance you recorded, counselor notes, email or SMS replies you logged, housing application, immunization or placement holds. Purchased intent scores and social listening are not cites. If the timeline is blank, the engagement cite stays empty. Do not impute disengaged from silence in a channel you never used.

The artifact enrollment receives is the rank plus those three cite groups, each present or empty. A counselor should see a plain sentence: deposit posted on this date, verification incomplete, last portal login on that date. If you cannot produce that sentence from source systems, do not display a score. When a new score lands, update the flag. Do not erase counselor notes.

The score orders a work queue. Enrollment owns what happens next.

A scheduled job writes melt-risk flags for deposited, not-yet-enrolled students into the CRM task list. Each task carries the cites. Staff sort by deadline proximity, then by risk, then by program capacity where seats are scarce. A person opens the file, reads the cites, and chooses the intervention: a call, a checklist nudge, an aid-office referral, a wait-list conversation if the student is truly releasing the seat, or no action because the cites are stale or the student already registered overnight.

Do not auto-send a campaign from the percentile. Keep multi-channel enrollment nudge templates next to the queue so the counselor can pick a channel that matches the cite: SMS for a missing housing contract, email for a document list, phone when the file is messy. The nudge is a tool the person selects.

Illustrative path, not a measured case. An admitted student in a competitive health program has a deposit on file, a campus-visit event, and a housing application started but not submitted. Aid shows verification incomplete and a packaged award the student has not accepted. The model flags high melt risk and cites those facts. The enrollment manager does not cancel the seat. They assign a counselor who calls, confirms intent, walks the housing step, and routes the file to aid for verification. If the student says the package will not close the gap, the next step is gap analysis, not a silent extra scholarship from the model. If the student is simply slow on paperwork, the counselor logs that and leaves the award untouched.

The same score can mean nudge the checklist, open an aid conversation, or do nothing today. Only the person with the file can tell which.

Melt models overweight engagement drop-off because logins are easy to count. Unmet need is slower to read and easier to mishandle.

If the aid cite is a documented gap between cost of attendance and the current package, do not treat it as a communications problem. Run financial aid gap analysis, or hand the file to staff who will, before you spend another reminder sequence. A student who cannot make the number will not stay because you emailed them twice.

Gap analysis still does not award. It tells you whether professional judgment, a payment plan, a work-study reminder, or a constrained institutional award is even in policy. Scholarship optimization ML can help aid leadership see how scarce dollars sit across the class. It must not consume the melt score as an automatic award trigger. Melt risk is not a merit signal, and it is not a need calculation.

Watch this failure: operations sees high melt plus unmet need and pushes a one-click scholarship from the enrollment desk. That bypasses packaging rules, treats similar files unequally, and teaches the shop that money is the only lever. Keep award authority in aid. Keep the melt flag in enrollment. Connect them with a referral and an SLA, not a write from the score into disbursement.

Two other failure modes show up as soon as the flag looks official.

Treating the flag as a lost seat. Overbook logic, wait-list pulls, and section cuts should not key off melt risk alone. Students flagged today still enroll. Students unflagged still disappear. Use the queue to work files. Use registration and census reports to count seats. If you need a capacity buffer, set it from enrollment operations your registrar already trusts, not from a live risk percentile.

Auto-awarding from the score. Binding a dollar amount to a melt band, even with a bulk approve at the bottom of a list, is still the model awarding. Require an aid officer to open the package, see the gap analysis, and enter the award in the SIS. If that step feels slow, fix the referral SLA, not the autonomy of the score.

A quieter failure: ignoring empty cells. A high score with blank aid cites is a guess. Hide or quarantine those rows until admissions or aid fills the record. Quality means the flag is only as strong as the cites behind it.

A melt-risk task is usable only when the counselor can read deposit, aid, and engagement cites in one place, choose an intervention or a no-action reason, and leave a note the next refresh will not erase. A dashboard melt percentage is not the deliverable, and this page does not give you a target rate. Census already measures the class. If a cite says unmet need, start with gap analysis. If the cell is thin, leave it empty.