Forecasting enrollment and program viability: where prediction meets cabinet-level decisions

Student-level prediction (who will return next term, who will graduate within six years) gets most of the attention in the predictive analytics conversation. It is also the part of the conversation furthest from where Provosts, CFOs, and trustees actually spend their time. The questions on a cabinet agenda are different in scope and different in shape: how many students will enroll next fall, which programs are quietly losing money, which of our peers are likely to consolidate or close.

This post is the guide to that second cluster of questions. The model families are different (mostly time-series, not classification). The data is different (institution-level rather than student-level). The honest limits are different too. For the student-level use cases that anchor most predictive work, the retention deep-dive and the graduation deep-dive are the companion pieces.

Almost every leadership-level predictive question at a higher-education institution sorts into one of three buckets. Each one calls for a different model family and a different evaluation standard.

Enrollment forecasting is the most mature predictive use case in higher education and the one where IR teams have the most institutional muscle memory. The question is also deceptively simple: given a long history of headcount, plus what we know about admitted-student pipelines and demographic trends, how many students will be on the roster next fall.

The model families that fit the question are time-series methods, not classification. Three are worth knowing.

ARIMA (AutoRegressive Integrated Moving Average) is the workhorse. It models how a number evolves over time based on its own history and is the right starting point when you have many years of relatively stable data. ARIMA is well-understood, well-tooled, and forgiving of small datasets. It is the model to beat, not necessarily the model to settle on.

Prophet is Facebook's open-source forecasting library, designed specifically for the kind of time series that has trend changes, holidays, and seasonality. It handles missing data gracefully and produces forecasts with intuitive confidence intervals. It is a strong default when ARIMA feels too rigid.

LSTM (Long Short-Term Memory) is a deep-learning model that fits complex sequence data with multiple interacting variables. It is the right tool when enrollment is driven by many simultaneous factors (county HS graduation rates, competitor pricing, financial-aid changes, demographic shifts) and the simpler models are leaving signal on the table. LSTM is also overkill for most use cases, and the engineering cost is real.

The right answer is usually to start with ARIMA, benchmark Prophet on the same data, and only move to LSTM if the gap between the simpler models and the actual outcomes justifies it.

Yield prediction is the bridge between admissions and enrollment forecasting. The question is not "how many students will enroll" but "of the students we admitted, which ones will actually deposit and arrive."

The model family for this is classification, not time-series. Each admitted student is a row, the outcome is enrolled or not, and the features include the obvious admissions data (HS GPA, application timing, aid offer, distance from campus) plus competitor enrollment data from the National Student Clearinghouse where it can be pulled.

The same gradient-boosted approach that works for retention works well here, and for the same reasons: structured tabular data, hundreds of features, class balance that varies but is generally tractable. The XGBoost vs LLM post covers why this model family fits.

Yield prediction is also where data sharing across institutions creates real value, and also where partner-only model designs hit their hardest limit. A model trained only on one institution's admitted students may have less signal than a model trained on a regional consortium. The trade-off is straightforward and worth explicit conversation in any procurement.

Program-level margin and viability is the analysis Provosts and CFOs ask for, do not fully trust, and end up running anyway. The mechanics are not hard. Per-program revenue (tuition plus aid attributable to majors in the program) minus per-program cost (faculty load, dedicated facilities, program-specific support). The output is a margin per program, ideally rolled up over multiple years to smooth one-time effects.

The predictive question on top of the descriptive one is: which programs are likely to continue, improve, or deteriorate in margin, given current enrollment trends and labor-market signals. That is a regression or time-series problem on a small number of rows per program, and the modeling has to be honest about the noise.

Two failure modes to watch for. First, treating program margin as a pure financial calculation ignores cross-subsidy: a humanities program with a thin margin might be the prerequisite path for a much larger pre-professional pipeline. The model output is the start of the conversation, not the decision. Second, forecasting program-level revenue with only a handful of years of data is risky. Small programs are noisy by construction, and tight confidence intervals on small-program forecasts should be treated with suspicion.

Beyond institution-internal data, the same toolkit applies to public IPEDS data for sector-level questions. Two come up often.

Enrollment migration patterns between institution types (college-to-university, university-to-community-college, public-to-private) are largely time-series problems on the IPEDS migration tables. ARIMA and Prophet both handle this well, and the data is fully public.

Closure and accreditation risk screening is a classification problem on a combination of IPEDS financial indicators, enrollment trends, and demographic projections. A model trained on the history of closed and accredited-loss institutions can flag the institutions whose profile most resembles them. The output is not a verdict (closure is driven by board decisions and political factors a model cannot see), but it is a useful screening signal for boards, consultants, and rating agencies.

For IR leaders, the IPEDS-based work is also valuable as a content and credibility asset. The audiences your institution wants to be in front of (other IR teams, conference programs, sector publications) work in IPEDS daily, and original analysis on public data tends to land well.

Different questions justify different levels of model investment. A rough guide.

For total fall enrollment forecasting, a well-tuned ARIMA or Prophet model that comes within roughly 2 to 3% of actuals on a one-year horizon is operationally useful. Beyond that horizon, confidence bands widen quickly and the right move is to report ranges, not point estimates.

For yield prediction, a model that ranks admitted students by enrollment probability with an AUC in the 0.80 to 0.85 range is genuinely useful for tailoring outreach, aid offers, and yield events. Going higher requires either more features or cross-institutional data sharing, and the marginal value drops fast.

For program-margin forecasting, "good enough" is mostly about uncertainty management. A model that produces honest confidence bands (often quite wide for small programs) is more useful than one that produces tight intervals the audience will overweight.

For IPEDS-based closure-risk screening, the goal is rank-ordering, not point prediction. A model that puts the genuinely at-risk institutions in the top decile is doing its job, even if it cannot tell you which specific institutions will close in which specific year.

Four limits are worth surfacing before any of this work reaches a board or cabinet meeting.

Clema's work on institution-level prediction (enrollment, yield, program viability, IPEDS-based screening) follows the same operating principles as the student-level models: open-source tooling, honest confidence bands, partner-only training, and an explicit owner for the output. The model selection follows the question (time-series for forecasting, classification for yield, regression for program margin), not the other way around.

For the broader strategic case, the reactive-to-proactive playbook is the companion piece. For the cost and vendor picture (which matters as much for institution-level analytics as it does for student-level), the cost and vendor reality post covers the math.

NCES IPEDS, use the data. Facebook Prophet, forecasting at scale. statsmodels, ARIMA documentation. National Student Clearinghouse Research Center. WICHE, Knocking at the College Door demographic projections. Higher Education Demand Index, Carleton College.