Build notes · Projections
How we project the next IPEDS release
What a Superiora projection is, how method 3.0 chooses one method for each school’s path and one method for each market year, how it says how accurate each choice was, when a range is shown, what we tested and rejected, and the scorecard we will publish after each release.

IPEDS reports each year one to two years after the year it describes. A college that plans next year’s budget reads figures about a year that is already over. Academia Superiora projects a set of figures for every institution it covers, from the next IPEDS release through the three years after it, and the outlook page projects the markets themselves: totals and medians for the nation, each sector and each state’s sectors, for each of those four years. This post says how a projection is made, how we say how accurate it was, when it carries a range, what we tested and rejected, and how we will score it.
What a projection is
A projection is our number for a figure IPEDS has not yet released: an institution’s Fall 2025 enrollment, for example, before the 2025-26 release reports it. Wherever it appears it is labelled Projected, never with an IPEDS year. It names its edition, the method version and the date the set was issued together (v3.0 · 12 Nov 2026), and beside it the data it ran on (data through 2024-25). It never replaces a reported figure, and it never enters a comparison’s verdict.
Every projection carries the same sentence: Academia Superiora’s projection from various publicly available data, and we are focused on continuously refining our methods. Not published or endorsed by NCES or IPEDS.
Each figure targets a declared period. A Fall enrollment figure in the 2025-26 release describes Fall 2025. A finance figure describes the fiscal year that ends in 2025. A graduation rate describes the cohort that entered about six years before. When an institution’s history stops short of the latest release, we make no next-release projection for it and say no current filing.
How a method earns its place
A backtest chooses every method. For each family of figures we went back to every year from 2017 to 2024, gave each candidate only the years before it, and compared its projection with what IPEDS later reported. The score is the error scaled by the series’ own typical change from one year to the next, read three ways: over all years, inside the COVID years (2020 and 2021), and outside them. The candidates are last year’s value carried forward; the school’s own damped trend; a move toward similar schools; the school’s market trend; a model learned across many schools’ histories; and blends of these.
Last year’s value is the yardstick, and in method 3.0 it is never the answer for everyone. A tested method replaces it unless it is measurably worse, and where every tested method is worse, the least-bad one is shown with a label that says how much worse. Last year’s value comes back only for one school whose own record shows the chosen method clearly worse there; that school’s figure then says no change expected.
The rules, in plain words
- One method for each school’s path. A school’s FY2025 to FY2028 path is made by one method, chosen for the figure on hundreds of schools, so a path does not jump from one method to another between years.
- A cap on each year’s move. Each year’s change is held to what similar schools, and the school itself, have done in 95 of 100 past years. Measured on schools the cap was not built from, real one-year moves went beyond it about 5 to 8 times in 100. A year the cap held says so.
- A school falls back only on its own record. Last year’s value is used for a school only when its own past results, weighed against all schools’ results, show the chosen method clearly less accurate there. It is never the method for everyone.
- Every label quotes the worst reading. How much a method beats last year’s value is read over all years, inside the COVID years and outside them, and the label quotes the worst of the three: beats last year’s value, as accurate as last year’s value, or less accurate than last year’s value, with the size.
- When the windows disagree, the label says both. If a method beats last year’s value over all years but is less accurate in the COVID years, the label says so in one line, for example beats last year’s value by 14 % overall; about 2 % less accurate in 2020–21. It never claims more than the weaker window allows, and it never hides the weaker window.
- Markets pick the best method each year. A state, sector or national figure is projected as a market, directly, never by adding up schools. Each of its four years takes the method that tested best for that year, and a market never shows last year’s value carried forward: where nothing beat it, the least-bad method is shown and labelled less accurate.
- Each finance figure has its own method. Revenue, expenses, assets and ratios are each chosen on their own, on 299 schools with every figure of each.
- Operating revenue leads. The revenue path we show is operating revenue; investment return is shown as reported, because it swings too much from year to year to project.
- A figure that has been zero every year is not projected. Where an institution has reported zero in every year, its page says reported zero every year — not projected, rather than projecting another zero.
- A share is never projected on its own. A share or ratio of counts we project, such as the share of students in one group, is worked out from the projections of its parts, so it always agrees with them. It carries no range.
- One figure, one projection. Where IPEDS publishes the same figure in two places, one is projected and the other reads it.
- Public institutions’ state appropriations take the figure’s own method for all four years at the school. The markets read state budgets for FY2025 and FY2026.
- A market-share scenario sits beside each school’s projection, never in place of it: if you keep your current share of your state and sector market, here is where the market’s projected path takes you. It is labelled as a scenario and carries its own measured accuracy.
Operating revenue
Operating revenue is total revenue less investment return and, at public institutions, less capital money that does not pay for a year’s operations. For public institutions reporting under GASB it is F1B25 less F1B17 (investment income), F1B20 (capital appropriations), F1B21 (capital grants and gifts) and F1B22 (additions to permanent endowments); for private institutions reporting under FASB it is F2D16 less F2D10. An institution with no investment-return line has operating revenue equal to its total revenue less those capital lines, and a year where the result is zero or below is left out. It takes total revenue’s fiscal year. One year’s investment return can be a large gain and the next a large loss, so a total-revenue path swings with markets that no backtest can see; operating revenue does not. Method 3.0 beats last year’s value by 3 percent on operating revenue at schools, and by 21 to 44 percent in the markets.
How accurate each market year is
| Figure | First year | Second year | Third year | Fourth year |
|---|---|---|---|---|
| Operating revenue | beats last year’s value by 21 % | beats last year’s value by 28 % | beats last year’s value by 39 % | beats last year’s value by 44 % |
| State appropriations | beats last year’s value by 15 % | beats last year’s value by 5 % | beats last year’s value by 13 % | beats last year’s value by 15 % |
| Applicants | beats last year’s value by 22 % overall; about 3 % less accurate in 2020–21 | beats last year’s value by 22 % overall; about 8 % less accurate in 2020–21 | beats last year’s value by 4 % | beats last year’s value by 35 % |
| Admission rate | about 7 % less accurate than last year’s value | about 18 % less accurate than last year’s value | about 25 % less accurate than last year’s value | beats last year’s value by 9 % |
| Fall enrollment, all students | beats last year’s value by 2 % | about 6 % less accurate than last year’s value | about 5 % less accurate than last year’s value | beats last year’s value by 3 % overall; only as accurate as last year’s value in 2020–21 |
| Fall enrollment, undergraduates | beats last year’s value by 3 % overall; only as accurate as last year’s value outside 2020–21 | about 9 % less accurate than last year’s value | about 9 % less accurate than last year’s value | about 3 % less accurate than last year’s value |
| Graduation rate within 150% of normal time | beats last year’s value by 2 % overall; only as accurate as last year’s value in 2020–21 | about 2 % less accurate than last year’s value | about 9 % less accurate than last year’s value | about 4 % less accurate than last year’s value |
| Net tuition revenue per FTE student | as accurate as last year’s value overall; about 4 % less accurate in 2020–21 | about 12 % less accurate than last year’s value | about 16 % less accurate than last year’s value | about 22 % less accurate than last year’s value |
| Figure | Label |
|---|---|
| Admissions (applications, admits, enrolled) | as accurate as last year’s value overall; about 2 % less accurate in 2020–21 |
| Finance, each figure on its own (rolled up) | beats last year’s value by 4 % |
| Operating revenue | beats last year’s value by 3 % |
| State appropriations | beats last year’s value by 8 % |
| Graduation rates | beats last year’s value by 3 % overall; about 1 % less accurate in 2020–21 |
| Admissions, the silver table | as accurate as last year’s value overall; about 0.7 % less accurate in 2020–21 |
| Fall enrollment | about 3 % less accurate than last year’s value |
| Students by race and sex (demographics) | about 0.9 % less accurate than last year’s value |
| First-time students | about 6 % less accurate than last year’s value |
| Staff headcount (EAP) | about 1 % less accurate than last year’s value |
| Staff FTEs | about 2 % less accurate than last year’s value |
| Faculty counts | about 0.5 % less accurate than last year’s value |
| Faculty by tenure status | about 2 % less accurate than last year’s value |
| Faculty salaries | beats last year’s value by 14 % overall; about 2 % less accurate in 2020–21 |
| Student aid (IPEDS SFA) | beats last year’s value by 8 % |
| Average aid and aid shares | beats last year’s value by 7 % |
| Net price by family income | beats last year’s value by 4 % |
| Completions by award level | about 0.7 % less accurate than last year’s value |
| Completions per program | beats last year’s value by 6 % |
A school’s own label can differ from its figure’s: it adds the school’s own record, weighed toward all schools’. The labels that say less accurate are printed as they were measured. They are why each label is shown at all.
What 3.0 projects beyond the headline figures
Besides enrollment, admissions, graduation and finance, version 3.0 projects first-year retention and the cohort it is measured on, students by race and sex, staff and faculty counts, faculty salaries, student aid and net price, completions by award level, and each program’s completions where the program has at least six years of history and at least five completions a year. Each figure first takes its family’s method; a figure gets its own method only where its own record, tested on held-back years and weighed toward its family’s, is clearly better, moves off last year’s value in most years, and earns a range as often as the family’s method does. Most of these figures are labelled less accurate than last year’s value, and the labels say so.
Federal Student Aid’s Title IV reports come out a year before IPEDS’s student aid survey. For undergraduate Pell grants (recipients and dollars) and undergraduate federal loans (count and dollars), next year’s figure is last year’s moved by the school’s change in those reports. We could test this on one year only, 2023–24, where it was 58 to 82 percent more accurate than last year’s value, so it is labelled tested on 2023–24 only and carries no range. The years after it continue from that moved figure, moving year to year as the family’s method does, so the path never falls back to a level drawn from last year’s value. Those years carry no range either, and their labels say that their accuracy was measured from a reported year, not from a moved one. Average federal loan size keeps its family’s method.
Federal Student Aid data
State budget data: SHEEO Grapevine
States enact their budgets for a fiscal year months before IPEDS reports what each public institution received. The state-appropriations markets read the change in each state’s support for higher education for FY2025 and FY2026, as the budgets were published by the day the edition was issued; later years follow the market’s own trend.
How a range is earned
An institution’s projection shows a likely range only where past tests earned it. For each figure’s method we built the range from errors on years the method had not seen, separately for small, medium and large institutions, and counted how often the real result fell inside. Each year’s range is built from errors at that distance only, so it widens step by step. In v3.0 that share is between 0.766 and 0.798 by family, depending on the distance. The label prints the share as measured, about 79 in 100 past results fell inside, never a round 80 percent. Where fewer than 78 in 100 fell inside, no range is shown. A history that has been the same every year has no range, because nothing was measured for it.
The market figures on the outlook page carry no range. No range for a market was tested on held-back years, so none is shown.
What we tested and rejected
| Candidate | What we measured | Outcome |
|---|---|---|
| TimesFM, a pretrained forecasting model | Worse than last year’s value in every family one year ahead: Fall enrollment 0.952 against 0.858, admissions 1.025 against 0.833 | Rejected in v1.0 |
| Shrinking toward the peer group’s trend, alone | No family cleared the bar. Admissions 0.898 against 0.914 sat inside the tie band; graduation rates got worse, 0.869 against 0.833 | Rejected in v1.0; kept only inside the blends of 3.0 |
| Prices and wages (CPI-U and the ECI for colleges), for finance flows | 3.0 percent better overall, but no better in the COVID years | Rejected in v2.0 |
| State and county unemployment (BLS LAUS), for Fall enrollment | 1.6 percent worse with the state rate, 2.5 percent worse with the county rate | Rejected in v2.0 |
| Federal student aid recipients (FSA Title IV), for Fall enrollment | Not testable: we hold four award years, too few to test against earlier years | Not tested yet |
| State budgets leading public schools’ state appropriations for FY2025 and FY2026 | On 147 public schools the figure’s own method was 20.5 percent better than last year’s value; the state-budget lead, 15.6 percent | Withdrawn at schools in v3.0; kept for the markets |
| A separate method for each year of a school’s path | A school’s path changed method from year to year and could jump | Replaced by one method per path in v3.0 |
The changelog
| Version | Date | Status | What changed |
|---|---|---|---|
| 3.0 | 2026-11-10 | draft: the first public edition, v3.0 · 12 Nov 2026, ships with the next release | One method for each school’s path, chosen per figure, with a yearly cap; last year’s value only as one school’s fallback; labels that quote the worst reading; markets projected directly with the best method each year; operating revenue leads; the market-share scenario beside each school’s projection; and, in the same version before release, demographics, staff, salaries, student aid, completions and each program’s completions, with a method chosen for each figure where its own record earns one, shares worked out from their parts, labels that name both windows when they disagree, federal aid data moving next year’s Pell and federal loan figures, and first-year retention. |
| 2.1 | 2026-11-10 | superseded before release; never shipped | Added state budgets two years ahead for public institutions’ state appropriations, as a labelled exception; showed four years ahead. Everything else was last year’s value. |
| 2.0 | 2026-11-10 | superseded before release; never shipped | Added state budgets for public institutions’ state appropriations, one year ahead. Everything else was as in 1.0. |
| 1.0 | 2026-11-09 | superseded before release; never shipped | The measured baseline: the target periods, the backtest, the ranges. Last year’s value won every family. |
The scorecard
Each projection is written to a ledger before the release it predicts, and nothing in the ledger is ever rewritten. When IPEDS publishes 2025-26, we will score every projection against the figure as first released and publish the result here after each release, one scorecard per edition. We never average one score across method versions, and an older version’s scores stay published under its own name. There is no scorecard yet, because the release v3.0 predicts has not come out. The scorecard scores the next-release year; the later years are scored when their own releases come.
What we refuse to do
- Name an institution on the outlook page. It shows totals and medians only, and leaves out any group with fewer than five reporting institutions or one that a single institution dominates.
- Present a projection as a figure NCES or IPEDS published or endorsed.
- Show a range we did not test, or round a measured share up to 80 percent.
- Hide a method that was less accurate than last year’s value: its label says so.
- Let a projection stand in for a reported figure, or a scenario stand in for a projection.
- Use a source whose licence does not allow our use.
