How the model works

Where the data comes from

Every college figure comes from the U.S. Department of Education College Scorecard, which is public-domain federal data reported by the institutions themselves (release of 2026-10-15). The catalogue holds 1,423 four-year colleges: every Common App member that awards bachelor's degrees — 1,079 colleges, matched by federal id to the Common App's own member list of 2026-10-18 — and every other four-year college with 500 or more undergraduates that publishes an admission rate. 1,391 publish an admission rate; the 32 that do not are, with one exception, open-admission colleges, which admit every applicant who meets their requirements — the model says so rather than estimating. 905 publish SAT ranges.

One set of figures comes from the colleges directly rather than from the federal data: their first-year applicants and admits split by residency — in-state, out-of-state and international — as each college reports them in its own Common Data Set filing. The Scorecard does not carry that split, and it is the difference between telling an out-of-state applicant to a capped public flagship the truth and telling them a comfortable average. We hold it for 307 colleges, 295 of which also give an international figure, read from the filings as collected by collegedata.fyi, which links each one to the college's original document. Every row is checked against the filing's own totals and against the Scorecard before it is kept, and a row that fails is dropped rather than repaired. Where a college has not filled those rows in, the model falls back to an estimate and says so.

Nothing is hand-curated or hand-edited. Where the data has a gap, Northstar shows a gap. A great deal of harm in this category comes from tools that quietly fill holes with plausible-looking numbers.

The core calculation

It is not a fitted model. The shape is borrowed from logistic regression — log-odds, a baseline, additive terms — but none of the numbers below were estimated from labelled outcomes. Every magnitude is a judgement, chosen to match what admissions research and published class profiles suggest, and then held still by the test suite. That is a weaker claim than a regression, and it is the true one.

The model works in log-odds space, anchored to each college's own published admission rate:

log-odds = logit(admission rate)
         + 1.15 × (composite strength, in standard deviations)
         + Σ adjustments (residency, round, major, …)

probability = sigmoid(log-odds)

Anchoring to the college's own rate means a student sitting exactly at the median admitted profile gets back roughly the published rate. Every other factor is then interpretable as “how far this student moves off that baseline”.

Everything is continuous. There are no lookup tables and no score buckets, so no small change in your input produces a large jump in your result. A 0.01 change in GPA moves the estimate by a fraction of a percentage point — as it should.

Composite strength

When a college publishes test ranges and you've entered a score:

• Test scores — 38%. Measured as a z-score against that college's own published 25th/75th percentiles.
• Unweighted GPA — 27%. Measured against a modelled admitted-GPA distribution (see the caveat below).
• Course rigor — 12%. Your actual courses, weighted by level and subject, against a full load in your own curriculum.
• Activities — 18%. Scored on scope, role, sustained commitment and recognition.
• Honours and awards — 5%. Weighted by how far the recognition reaches. Deliberately small: each is a single self-reported data point, and it is the easiest part of an application to inflate.

When a college is test-blind, or you haven't entered a score, the test weight is redistributed (GPA 43%, rigor 19%, activities 30%, honours 8%) and the confidence interval widens. You are never penalised for a college's missing data.

Rigor is judged against your own curriculum

We read your actual courses, not a count of them. A count cannot tell four studio-art courses from four calculus-and-physics courses, and universities very much can — so subjects in the core academic areas carry full weight and others carry less, with a bonus for covering breadth across maths, science, English, social science and a language.

Each curriculum is scored against its own benchmark. Three A-Levels is a complete sixth-form programme; a full IB Diploma is three Higher Level plus three Standard Level subjects. Judging either against an American expectation of five APs would mark a perfectly normal applicant as under-loaded. IB Higher Level and full A-Levels count for more than Standard Level and AS respectively, because they represent more study. If your school follows a system we don't enumerate, you can enter courses by name and they still count.

The GPA caveat — read this one

No official source publishes admitted-student GPA for US colleges. The Scorecard does not have it. So Northstar estimates a college's typical admitted GPA from its selectivity, and compares your unweighted GPA against that estimate. Anywhere this happens, the factor is labelled “estimated — see methodology”in your breakdown, and it widens your range.

What we deliberately never do is compare a weighted GPA against an unweighted average. Those are different scales, and mixing them inflates estimates for anyone at a school with generous weighting. Your weighted GPA feeds the rigor term only.

If you worked out your GPA with our calculator

The profile page has a calculator for students who have grades but have never been given a GPA. For US letter grades and percentages it uses the College Board's 4.0 chart, and the result is what a US school would call your unweighted GPA, give or take that school's own rules.

For IB and A-Level grades it is an approximation, and you should treat it as one. Neither the IB nor the A-Level boards publish a conversion to a US GPA, and most US admissions offices read those results as they are. The calculator applies one published scale to each — the table Texas State University uses for IB applicants, and Scholaro's reference table for A-Levels — and other published scales differ from those by a few tenths of a point. Both are whole letter grades, so they are coarse: an IB 6 and an IB 7 come out the same.

A converted GPA is softer evidence than one printed on a transcript, and the model cannot tell the two apart: it treats whatever is in the GPA box the same way, and does not widen your range because the number was converted. If you have a test score, enter it too, so the estimate has something measured to lean on.

Adjustments

These are literature-derived estimates, applied in log-odds. They are not measured from Northstar's own outcomes — we don't have enough yet, and we'd rather say so than imply a precision we lack. Each one appears by name in your breakdown.

• In-state residency at public universities. Where the university publishes its own in-state and out-of-state admit counts, we use them: your pool's rate against the university's overall rate for the same year, labelled as its own figure. UT Austin, for instance, admits about one out-of-state applicant in ten against one in four overall, because Texas law reserves most places for Texans — and the estimate the model used to make for that was far too gentle. Where no figures are published, the estimate scales with selectivity: negligible at open-access publics, substantial at flagships that cap out-of-state enrolment. Applied only if you actually live in the US — a student abroad has no US state residency, however long their family lived somewhere.
• Domestic or international pool. Decided by citizenship rather than address: a US citizen at an international school is a domestic applicant. Where a college publishes its international admit count we use it, as above; this is the one adjustment that used to be a guess everywhere, and at 295 colleges it no longer is. Elsewhere: selective private universities generally admit international applicants at lower rates, partly because aid for them is far more limited; public universities recruit them for full out-of-state tuition, which offsets some of it. Almost no university publishes a separate international rate, so this adjustment is an estimate and it widens your range accordingly.
• Early Decision / Early Action. ED carries roughly a 2.1× odds ratio at selective colleges, EA about 1.3×. Part of that reflects who applies early rather than the timing itself, so treat it as an upper bound.
• The program you apply to, at that specific school. See below.
• Applying for financial aid. See below.
• Recruited athlete. Large, but only once a coach has actually committed to supporting you.
• Legacy. Applied only at private nonprofits, and flagged as a policy many colleges have now ended.
• First-generation status and class rank. Modest positive adjustments.

Ability to pay

Need-aware admission — considering whether a family can pay when deciding whom to admit — is legal and openly practised, so it belongs in the model. Three things set the size of the effect:

• Most selective US colleges are need-blind for domestic applicants, deciding admission without looking at finances. So the domestic adjustment is deliberately small.
• Nearly all are need-aware for international applicants, including ones that are need-blind domestically. Aid for international students is scarce and openly rationed, so this is where the real effect lives.
• Declining aid is never modelled as a bonus. Doing so would imply money buys admission at need-blind colleges, which it does not — being full-pay is simply the absence of the penalty.

Never skip aid you need in the hope of an edge. The money is worth vastly more than the difference here, and at a need-blind college the difference is zero. A test in our suite asserts the domestic effect stays small precisely so this page can keep saying that honestly.

For US students, your income band also drives the net-price figure on every college page. That number matters more than the odds: a private college with a $90,000 sticker price is routinely cheaper than the in-state public once aid is applied, and students rule schools out on the list price every year without ever finding that out.

The federal figure only describes students who receive US federal aid, and at a public university only in-state students. So an out-of-state student at a public sees it plus the extra out-of-state tuition. An international student, who can't receive federal aid, sees a minimum with no aid instead: the higher of the published non-resident tuition and what the least-aided US families pay (plus the out-of-state premium at a public). The exception is a short list of colleges whose own sites say they meet the full demonstrated need of the international students they admit; there an international student sees what US families at their income pay, and at the colleges that are also need-blind for international applicants, asking for aid doesn't lower the estimate.

Why race is not in the model

In Students for Fair Admissions v. Harvard (2023), the Supreme Court held that race-conscious admissions programmes violate the Equal Protection Clause. US colleges may no longer consider an applicant's race when deciding whom to admit.

So a calculator that moved your estimate up or down on ethnicity would be describing a practice that ended in 2023 — the number would simply be wrong. Worse, it would tell some students their chances are lower because of their race, which is both false and a reason not to apply. Ethnicity is optional in your profile and has exactly zero effect on your estimate, and a test asserts byte-for-byte identical results across every category at every level of selectivity.

What it does do is show you how each campus's student body actually breaks down, from federal enrolment data. That is the whole of it. It is not given to the essay coach — the coach is handed your prompt, your draft and its word limit, and nothing else about you.

The ruling did explicitly preserve one thing: you may write about how your background has shaped you, tied to your character or what you would bring. That essay counts in a way a checkbox never did. But it is yours to write, and Northstar does not read this field to help you write it.

Majors: what we know, and what nobody publishes

Your major is stored per college, not once on your profile, because that is how applications actually work — you might apply to computer science at one university and economics at another, and a program's competitiveness is a property of that pairing.

The list of programs at each university is real: it comes from the federal field-of-study data, along with how many students graduate from each per year and what those graduates earn four years out. Admission rates by major are a different matter. We checked every field the federal endpoint returns, across all programs at a large university: there is no admission-rate-by-major field. It does not exist. Any product showing per-major admission odds for hundreds of colleges is inventing them.

So every by-major figure you currently see is an adjustment by field: computing, engineering and nursing are capacity-constrained at most universities and admitted more competitively than the school-wide rate implies, which is well documented at that level of granularity. That adjustment is labelled modelled, and the page says plainly that the school does not publish by-major rates.

There is a second path that is built but carries no data yet, and it is worth stating plainly rather than implying otherwise. A handful of universities do publish their own by-major rates — the University of California system, UT Austin, Illinois, Georgia Tech, Purdue. Where one of those figures is entered, it overrides the field-level estimate entirely and is labelled as the school's own measurement. None have been entered. The table is empty, so today that override never fires for anybody. Each entry is data to be read off a university's own publication by hand and dated, and a wrong one would be worse than no entry — so it stays empty until someone does that work.

Why you get a range

The range is not decoration. It starts from an irreducible floor — essays, recommendations, and whatever a college happens to need in a given year are genuinely invisible to us — and widens with every piece of missing information, when a college is test-blind, and when the college is highly selective (outcomes there are inherently noisier). Each estimate says which of your inputs would narrow it.

How wide it is drawn is a judgement, and we changed it in October 2026. The band used to reach one standard error either side of the estimate; it now reaches two-thirds of one, so it is about a third narrower than it was. Students testing the product found the old ranges too wide to act on — “15 to 66%” is not an answer to “should I apply?” — and the owner chose to narrow them. The cost is real and we would rather name it: a narrower band is wrong more often, in that the actual outcome will fall outside it more often than it did. Neither width is calibrated against results, because nothing here is yet, so neither is the correct one; when enough students report back we will publish how often outcomes fell inside the band and set its width from that rather than from judgement. The tiers read the low end of the band, so the change also makes “likely” and “safety” a little easier to reach.

If your range is still wide, that is the honest answer, not a failure of the tool. Filling in more of your profile narrows it, because the uncertainty actually shrinks — with one exception we record in the code rather than smooth over: telling us you live outside the US widens the range at a college that publishes no international admission rate, because our adjustment there is doing more guessing than the rest of the model. At the colleges whose own figures we hold, it does not.

What this model gets wrong

Stated plainly, because you deserve to weigh it yourself:

• It cannot read your essays or recommendations. At selective colleges these carry enormous weight. This is the single biggest source of error.
• Admitted-GPA figures are modelled, not measured. See the caveat above.
• Adjustment factors are estimates from published research, not measured from Northstar users. They are directionally right; the exact magnitudes are not precise.
• It has no institutional priorities. Colleges shape a class — by geography, intended major, financial need, and much else. None of that is visible in public data.
• Federal data lags. The Scorecard reflects a recent but not current admissions cycle, and admission rates have moved quickly in recent years.
• It is not calibrated against outcomes yet. Once enough students report back, we intend to publish a calibration curve — “when we said 40%, X% were admitted” — and to correct the model where it is off. Until we can publish that, treat these as informed estimates rather than validated probabilities.

What we don't use

• Race and ethnicity. Optional on your profile, never modelled. A test asserts the estimate is byte-for-byte identical across every category. It is used to show a campus breakdown and for nothing else — in particular, the essay coach is never given it.
• Demonstrated interest. Real at some colleges, but not measurable from public data, so we leave it out rather than guess.

Reading the tiers

• Safety — the low end of your range is at or above 80%.
• Likely — the low end is at or above 55%.
• Target — best estimate of 30% or better. These should be most of your list.
• Reach — best estimate of 10–30%.
• Far reach — under 10%, which describes most applicants to the most selective colleges, including very strong ones.

These thresholds are deliberately stricter than is common. A 40% chance is not a “match”; it's closer to a coin flip, and calling it a match is how students end up in April with no acceptances.

Browse colleges →