Every spring, millions of high-school seniors type their grades into an admissions calculator and wait for a probability. We catalogued more than fifty of these tools — from the billion-user platforms to the single-developer GitHub repos — and compared them against the academic literature. Most are GPA-and-SAT lookup tables. None of them model the market.
Every commercial chancing engine answers the same question: given my profile, what are my chances at School X? It is an individual-prediction problem. Each student-college pair is treated in isolation — as if no other applicants existed.
A small academic literature asks something different: given the entire applicant pool, the seat constraints, and the round structure, what happens? That is a market-simulation problem. The answer depends on who else applies, who else gets in, and who else enrolls.
Almost every tool in the marketplace is on the left side of this divide. Two are on the right.
First, look at what each tool actually inputs.
The marketing copy makes everything sound rich and AI-powered. The feature lists tell a different story. Across the major consumer tools, almost everything reduces to GPA, SAT, and a handful of demographic flags.
Hooks — legacy, recruited athlete, donor — are the most consequential factors at selective colleges, and almost no consumer tool models them. Round structure, demonstrated interest, yield protection, waitlist dynamics: the same blank squares, again and again.
The columns shift only on the right side of the matrix, where academic models sit. Arcidiacono's 370-variable Harvard regression and our agent-based simulation are the two columns with consistent fill.
Now zoom in on the major consumer engines.
CollegeVine is the market leader: 75 input factors, 1,500 colleges, free. Its calibration check — 50 percent predicted versus 48.1 percent actual — is the only published accuracy number in the consumer space. The trade-off is that it cannot distinguish a legacy athlete from a first-generation applicant with the same stats.
Scoir is the only commercial platform that produces separate ED, EA, and RD predictions — architecturally the closest to a real admissions process. But it is institution-licensed, not student-facing.
Niche, PrepScholar, and Parchment compress everything to GPA and a test score. Naviance shows scattergrams from your own high school but doesn't predict at all — and Mulhern (2021) found those scattergrams discouraged high-achievers from applying to selective schools by roughly half.
Then a wave of AI startups arrived.
Between 2023 and 2026, more than twenty new AI-powered chancing tools launched. The headline numbers are extraordinary: 98.2 percent, 90 percent, 85–90 percent. Every one is self-reported. Not one publishes a calibration table.
Most are thin wrappers around an LLM or a small ML model trained on whatever data the founder could scrape. None of them model hooks. None of them model rounds. None of them model market dynamics.
The most interesting entrant is Ultra, Y Combinator-backed and founded by former admissions officers. It explicitly simulates the AO evaluation process — closer in spirit to our approach than any other startup.
The real coefficients live in the academic literature.
The Students for Fair Admissions v. Harvard trial pried open the most detailed empirical admissions model we have. Peter Arcidiacono's expert report estimated a binary logistic regression with 370+ controls on Harvard's 150,000 applicants from the classes of 2014–2019. Pseudo R2 was 0.56 — excellent for a logit.
The odds ratios are staggering. Recruited athletes were roughly 5,075 times more likely to be admitted, all else equal. Legacies were 8.5 times more likely. The Dean's Interest List — the donor track — exceeded 7×. Asian-American applicants were 0.63× — the disadvantage that the Supreme Court ultimately struck down.
Our simulation imports these magnitudes as logit-space hook multipliers, scaled by tier. The athlete coefficient is smaller because Arcidiacono's model includes Harvard's separate athletic rating; we fold that into a single recruited-athlete multiplier.
Espenshade's earlier work translated the same effects into SAT points.
Thomas Espenshade's 2004 study of 124,374 applications to ten selective colleges produced a famously memorable framing: express each admissions advantage as the SAT-point increase that would produce the same admit probability.
On the 1600 scale, recruited athletes were worth +200 points. Legacies +160. African-American applicants, +230. Hispanic applicants, +185. Those numbers became the canonical reference for hook magnitudes for the next two decades.
Two warnings about reading them today. First, they are pre-SFFA — race advantages have been judicially eliminated and in our model are reversed. Second, no commercial chancing tool has fully incorporated the post-2023 reality. The factor weights you see in CollegeVine and its peers still reflect a regime that no longer exists.
A different academic tradition went straight to agent-based simulation.
Sean Reardon's 2016 paper in JASSS is the published model that most resembles ours. Same architecture: a population of student-agents, a set of college-agents, a multi-step matching process. Same goal: understanding stratification as a system-level outcome rather than an individual probability.
The differences are calibration depth and scope. Reardon used 8,000 stylized agents, two attributes per student (resources and caliber), 40 generic colleges with 150 seats each, and a single application round. Our simulation runs ~4,000 agents with thirty attributes against 55 named colleges with real Common Data Set numbers, across six rounds.
Reardon's headline finding — that a 0.3 correlation between resources and caliber drives most observed stratification — anticipated the ALDC findings that would dominate the 2020s admissions literature.
All of which leaves one quadrant of the map sparsely populated.
Plot every tool on two axes. On one, the number of factors the model uses. On the other, whether it simulates the market — seat constraints, round sequencing, yield competition, waitlist resolution, summer melt.
The upper-left is crowded: dozens of high-factor tools that treat each applicant in isolation. The lower-right has Reardon's ABM — high system-fidelity but low factor count. The upper-right is almost empty.
That is the quadrant our simulation occupies. About thirty factors per applicant, all six admissions rounds modeled, seat constraints, phantom applicants, demonstrated-interest interaction, waitlist dynamics, summer melt, Monte Carlo confidence intervals. Not a chancing engine. A market simulator.
The full report — with sources and odds-ratio tables — is on the research site.