Predictive Yield Scoring

A predictive yield score is a cited prior from your own closed inquiry-to-enroll history. For this prospect's attributes, how often did similar inquiries apply, and how often did similar admits enroll. The enrollment manager uses that prior to rank recruiter outreach. The score does not admit anyone, deny anyone, or set a scholarship.

Empty is a valid quality outcome. If the matching cell is thin, the probability fields stay empty. Do not backfill with a national benchmark, a vendor default, or a rounded guess. A missing prior is cleaner than a precise-looking number built on a handful of records.

Every scored row needs a cite the counselor can open: cycle years, geography or high-school cluster, intended program, and closed-record count. If staff cannot see who the model treated as "like this person," the number is not ready for the floor.

Train on people whose funnel is finished. An inquiry that never applied is a negative for application probability. An admit who declined or melted is a negative for enrollment probability. Open inquiries still in this cycle are not labels.

Keep the two events in separate columns. Application probability and enrollment probability answer different questions. A single "interest" score hides whether the model is reacting to inquiry volume, counselor coverage, or deposit behavior.

Use fields you already store and can explain: inquiry source, intended major, geography, first-generation flag, visit or event attendance, counselor assignment, and timing relative to your deadlines. Do not ingest credit, immigration, or health data. Do not use proxies for race or income unless counsel has approved that use for outreach ranking only.

Hold out a recent closed cycle and score it as if the outcomes were unknown. Check calibration in the cells you actually staff: territory, intended college, first-generation status, inquiry source. If a cell is sparse, mark it unscorable rather than averaging it into a neighboring cell.

Refresh when a cycle closes, not continuously on live inquiries. Mid-cycle learning from people you have not finished recruiting rewards whoever already got the phone time.

Look-alike prospect discovery sits upstream. It finds people who resemble enrolled students. Yield scoring ranks people already in your inquiry pool. A look-alike hit with a thin local cell still gets a human review, not an automatic skip.

For each prospect, write three things into the CRM when you can support them: an application prior, an enrollment prior (or a clearly labeled conditional prior), and a cite that names the cohort and the count.

If the count is too small to trust, leave the probability fields blank. Staff will invent a story around any number you show them, including 0.00 and 1.00.

Do not paste an unpublished vendor propensity into the same fields. EAB, Slate, Technolutions, Salesforce, and Ellucian appear in this stack as inquiry capture, CRM, scoring add-ons, or SIS handoff. Treat that class as systems of record and workflow, not as oracles. If a platform returns a score without a first-party cite you can open, park it in a vendor column. Your yield prior stays empty until closed history can support it.

Conditional yield for inquiries who are not yet admits is a planning aid. Label it as yield if admitted, given this cell. Never let that number travel into an admit committee packet or an aid grid.

Scholarship optimization ML uses overlapping features and a different decision. Keep the models separate. A high yield prior is not a reason to cut aid. A low yield prior is not a reason to buy the student.

You own the queue. The model proposes; you assign.

Pull inquiries with a cited application prior in a range your team can actually cover, plus every inquiry whose cell is empty. Empty cells are work, not noise. First-generation inquiries, new high schools, and late sources show up empty more often. Those people still get a human look.

Attach capacity. If a territory is already overloaded, the score is a tie-breaker inside that territory, not a license to abandon the rest of the map.

Write the reason on the assignment. A cite from closed inquiries in this program and region, last two cycles, is an instruction. "Hot" is not.

Keep a protected slice of counselor time for low-score and unscored inquiries you still want. If only high-score inquiries get visits, next year's model learns coverage, not demand.

One path, not a measured case: a first-generation inquiry arrives from a high school that has sent you almost no students. The application-prior cell is empty. A counselor who sorts only on score never sees the name. A manager who requires empty cells on the weekly list assigns a call, a visit offer, and a note that this record must not inherit a neighboring suburban rate. If the student later applies, the enrollment prior stays empty until you have a real admit-to-enroll cell for that cluster. Nothing in that sequence is an admit decision.

After deposit, stop using recruit-stage yield as a live priority. Enrollment melt prediction is the downstream model for people who have already said yes. Mixing melt risk back into the inquiry score double-counts summer behavior and punishes students whose summer jobs look like low engagement in a CRM.

Treating a thin cell as a precise probability. A handful of closed records that enrolled, or a handful that did not, are not a rate you can brief leadership on. They are anecdotes with a denominator. If you display a precise prior from a thin cell, counselors will treat it as a forecast. Hide the number. Show the count. Leave the prior blank until the cell is thick enough that a single family cannot swing it.

Skipping a first-generation inquiry because the score is low. First-generation students are often thin in your history because your history is the coverage you already gave. A low or missing prior is a data statement, not a judgment that the student will not come. Put those names on the unscored and low-coverage list. If you systematically skip them, the next model will learn that first-generation inquiries do not yield, when what it learned is that nobody called.

Using yield as an admit veto. Yield scoring is a recruit-stage quality prior for outreach. It is not an academic forecast and not a substitute for the admit review. Academic success prediction lives in a later stage, with different labels (persistence, progression), different features, and a different owner. Do not import a recruit yield score into a deny recommendation, a waitlist rank, or an aid-packaging rule. If a committee asks for likelihood they will come, answer with capacity and mission, not with a CRM field.

Watch three more shortcuts: scoring incomplete forms as a prior; retraining on unlogged recruiter overrides as if they were outcomes; using another institution's benchmark as your cite. First-party inquiry and enroll history, or empty.

Required CRM fields: scored versus unscored, cite (cohort and count), owner, last human touch, and a flag that this number is for outreach only. Optional: a short manager note when you override the queue for mission, athletics, first-generation coverage, or geographic mix.

Review the unscored pile every cycle with admissions leadership, not only with the CRM admin. If entire high schools, intended majors, or first-generation records remain empty, that is a coverage and data-quality problem, not proof those students will not enroll.

When you change a rule (minimum cell size, allowed features, whether conditional yield is shown before admit), record the change date.

Dashboards from this vendor class will keep arriving. Your standard does not change with the logo: closed inquiry-to-enroll labels, a cite on every number, empty when thin, and a human who still owns who gets the next call.