Loading…
The science
QuantAdmit runs on decades-old, peer-reviewed decision science — the same tools used in physics, finance, marketing, and medicine. Here is where each one came from, how it developed, and exactly how we use it.
Estimating outcomes by simulating chance many times over.
Origin. Invented in the 1940s by Stanislaw Ulam and John von Neumann at Los Alamos, with Nicholas Metropolis coining the name (after the Monaco casino). Faced with intractable physics, Ulam realized it was easier to simulate random trials thousands of times than to solve the equations exactly.
Development. It became a foundational tool across physics, finance, and operations research — anywhere outcomes are uncertain and interact. Modern computing turned thousands of trials into millions, making it the standard way to quantify risk distributions rather than single point estimates.
In QuantAdmit. QuantAdmit simulates thousands of full admissions seasons across your college list. Instead of a single 'chance', you see the whole distribution — the probability of at least one admit, the expected number, and the shutout risk — with outcomes linked by an applicant-strength correlation, so a strong applicant's results move together.
Key figures: Ulam · von Neumann · Metropolis (Los Alamos, 1940s)
Inferring what you truly value from the choices you make.
Origin. Grew from the 1964 'conjoint measurement' work of mathematical psychologist R. Duncan Luce and statistician John Tukey, which showed how to recover the weight of each attribute from how people trade them off.
Development. Paul Green and V. Srinivasan brought it to marketing in the 1970s, where it became the dominant method for pricing and product design — because stated preferences ('I want it all') diverge from revealed preferences (what you'll actually pick when forced to trade off).
In QuantAdmit. QuantAdmit's choice exercise asks you to pick between hypothetical colleges. From those trade-offs it estimates how much you really weight cost vs. selectivity vs. outcomes vs. fit — your revealed preferences — which then feed your personalized value scoring across schools.
Key figures: Luce & Tukey (1964) · Green & Srinivasan (1970s)
Your odds across the whole list — not one bet at a time.
Origin. Harry Markowitz's 1952 'Portfolio Selection' (Nobel Prize, 1990) reframed investing around the trade-off between expected return and risk across a whole portfolio, not one asset at a time.
Development. That mean-variance thinking spread far beyond finance — to any decision about spreading effort across uncertain options, where what matters is the joint outcome, not each bet alone.
In QuantAdmit. Your application list is a portfolio of uncertain bets. We simulate thousands of admissions seasons — each school a weighted coin-flip at its admit odds — to show the full distribution of outcomes and, most importantly, your shutout risk: the chance of zero admits. It's a lens on the list you have; to build the optimal list, see the application-portfolio problem below.
Key figures: Markowitz (1952, Nobel 1990)
The optimal set of schools to apply to — solved from first principles.
Origin. In 2006, economists Hector Chade and Lones Smith formalized the problem of 'simultaneous search': a student facing application costs and uncertain admissions must choose a whole portfolio of schools to apply to at once. They proved you should value a list by the expected quality of the best school that admits you — and build it by adding whichever school most improves that expectation.
Development. Their result explained, from first principles, why rational applicants diversify across reach, match, and safety schools, and it seeded a literature on how students actually build lists. But the theory deliberately takes two things as given: your preferences over schools, and your admission probabilities.
In QuantAdmit. QuantAdmit's 'Choose your 20' optimizer is a direct implementation of Chade–Smith's value function — the expected utility of your best admit — extended with exactly the two things the theory leaves open: your preferences (measured by conjoint) and your correlated admit odds (from your profile, via Monte Carlo). We then add the real-world constraints the theory abstracts away: the Common App's 20-school cap, one binding Early Decision, and a penalty for piling on correlated reaches.
Key figures: Chade & Smith, “Simultaneous Search” (Econometrica, 2006)
Scoring options that differ on many things at once.
Origin. Formalized in Ralph Keeney and Howard Raiffa's 1976 'Decisions with Multiple Objectives', building on the expected-utility foundations of von Neumann and Morgenstern (1944).
Development. MAUT became the backbone of formal decision analysis in medicine, public policy, and engineering — wherever a choice must weigh several incommensurable criteria transparently.
In QuantAdmit. QuantAdmit scores each school as a weighted sum of normalized attributes (selectivity, cost, and more) using your conjoint-derived weights, so 'fit' becomes an explicit, comparable number rather than a gut feeling.
Key figures: von Neumann & Morgenstern (1944) · Keeney & Raiffa (1976)
Pulling bold estimates back toward what's known — honestly.
Origin. Herbert Robbins introduced empirical Bayes in 1956; Charles Stein's 1961 result (the James–Stein estimator) proved the counterintuitive fact that shrinking individual estimates toward a common mean improves accuracy.
Development. Shrinkage is now standard wherever per-unit estimates are noisy — sports analytics, genomics, small-area statistics — because it tempers overconfidence from thin evidence.
In QuantAdmit. QuantAdmit treats each school's published acceptance rate as a prior and shrinks your modeled odds toward it — more aggressively at the most selective schools, where individual stats are least determinative. That's why a strong profile reads as a realistic mid-teens chance at a 4%-admit school, not a coin flip.
Key figures: Robbins (1956) · Stein / James–Stein (1961)
Deep dive
Our optimizer isn’t a heuristic. It solves the exact problem economists Hector Chade and Lones Smith posed in 2006 — and that problem has an elegant answer.
You can apply to a limited number of schools. Each admits you only with some probability, and you’ll ultimately enroll at the best one that says yes. Which set should you apply to? Apply only to dream schools and you risk getting in nowhere; apply only to safe ones and you waste your shot at something better. The right list balances the two — but “balance” has to be made precise.
Chade & Smith showed the value of a list is the expected utility of the best school that admits you. Rank your schools from most to least preferred; the chance you end up at school i is the chance it admits you and every school you prefer rejects you:
ui is how much you’d value attending school i, pi its admit probability, and the product runs over every school you prefer to it. A shutout contributes nothing — so the formula rewards a shot at a school you love and the safety net of landing somewhere, in a single number.
You build the list greedily: repeatedly add the school that most raises V. Early on a safety adds a lot — it guarantees the fallback. Once you’re likely to land somewhere, the marginal value shifts to reaches, which can only help by beating what you already have. The reach / target / safety mix isn’t a rule of thumb; it emerges from maximizing V.
The theory takes preferences and probabilities as given. We measure them: conjoint recovers your ui from real trade-offs, and your profile plus a correlated Monte Carlo model gives the pi. Then we extend the model to the real world: the Common App’s 20-school cap, one binding Early Decision (an odds boost we spend where it raises V most), a diversification penalty so near-identical reaches don’t all count, and optional floors on your reach/target/safety mix and the number of schools you’d love.
It’s a model. The probabilities are estimates, not certainties; the utilities are yours, measured imperfectly; and “you enroll at your best admit” is a simplification — aid, visits, and gut can change it. Treat the recommended 20 as a rigorous starting point to react to, not a verdict.
The methods are established; their application to an inherently uncertain process is not a guarantee. Every output is a model estimate. See how we combine them, and their limits, in the methodology.