Why
A price is not a probability unless the person paying it is risk-neutral, and almost nobody is. Once a decision-maker's utility function is concave — diminishing marginal utility of money — they will pay less than the expected value of a lottery to avoid the lottery, and that gap is a risk premium baked into every price they're willing to trade at. Skipping this distinction is the failure mode: reading a Verex mark price as "the market thinks this is 62% likely" without asking what risk premium is embedded in that 62 produces a systematically wrong estimate, worse the further the true probability sits from 50%. This lecture gives the exact formula (expected utility, not expected value) and the exact condition (concavity) that explain why.
How it works
The beauty contest: recursion converges to a Nash equilibrium
Everyone in the room picks an integer from 1 to 100; whoever is closest to two-thirds of the group's average wins (01:14, 01:29). Round 1, with no reason to expect anyone else's reasoning, guesses cluster near two-thirds of a naive average of 50, and the summary reports a winning guess of 24 (02:27, 02:35). Round 2, once players expect everyone else to already be computing two-thirds of the average, the average drops to about 21 and the winner is reported at 14 (04:15). The logic is a fixed point: if the average will be some number x, the best guess is (2/3)x, but everyone believes that too, so the average converges under iteration to (2/3)ⁿ · 50 → 0 as the reasoning steps n grow. That limit, 0, is the game's unique Nash equilibrium (0 or 1 with integer guesses, depending on the tie-break rule) (05:06, 05:21, 07:47). The lecture's aside matters as much as the math: one student wrote 100 not to win but to needle the professor (05:47, 06:11) — proof that a model is only as good as its assumption about payoffs; get the preference wrong and the equilibrium prediction is wrong however correctly you compute it (05:54, 06:01, 06:51).
Rationality is consistency, not the preference itself
Stitching together Myerson's and Aumann's definitions, game theory is "the study of interactive decision-making among intelligent, rational decision-makers" (00:26, 00:43, 09:38, 11:22). Economics takes no position on which preferences are rational: preferring chocolate to vanilla is just a fact about someone's utility, not a rational or irrational claim (12:01, 12:15, 13:01, 13:08). What can be irrational is inconsistency — choosing vanilla over chocolate with no new information, while still claiming to prefer chocolate, violates your own stated preference order (12:20, 13:13). This is the distinction the Auditor formalizes elsewhere: a preference is not a rule that can be broken, a choice can be.
Certainty: ordinal utility, order only
Under certainty the object of choice is a set of outcomes with a preference ranking — coffee C ≻ espresso E ≻ tea T, transitive by assumption (17:10, 17:51) — and "C is preferred to E" is a statement about an observed choice, not an unobservable feeling (18:47, 18:58). A utility function U representing this ranking is ordinal: only the order of the numbers carries information. U₁(C)=5, U₁(E)=4, U₁(T)=1 and U₂(C)=1000, U₂(E)=10, U₂(T)=−50 represent the identical preference, because both preserve C > E > T (20:49, 22:19, 22:35). Any strictly increasing transformation of an ordinal utility function represents the same preferences; the gaps between the numbers are meaningless (23:18, 23:55).
Uncertainty: VNM cardinal utility over lotteries
Choice under uncertainty changes the object being chosen: given a probability P of a clear day and 1−P of rain, the decision-maker isn't picking "walking" or "the subway" as outcomes, they're picking a lottery Δ(Z) — a probability distribution over outcomes (25:21, 33:38, 42:12, 44:21). Each outcome z gets a cardinal utility u(z), and a lottery's value is its expected utility: U(P) = Σᵢ Pᵢ u(Zᵢ) (45:58, 47:12). Because this is a sum weighted by probabilities, the absolute size of u now matters, not just its order (33:07, 33:47, 33:55, 42:27). Worked example: walking gives utility 7 on a clear day and 2 in the rain, so U(Walk) = 7P + 2(1−P) = 2 + 5P; the subway gives a certain utility of 5 regardless of weather, U(Subway) = 5. Setting 2 + 5P = 5 gives the indifference point P = 3/5: strictly prefer walking whenever P > 3/5 (36:21, 37:43). Change the utility numbers and the threshold probability changes with them (39:41) — the numbers are load-bearing here in a way they never were under certainty.
Risk aversion: concave utility and Jensen's inequality
Two money lotteries make the point (51:54, 55:33, 52:11). Lottery A: 99% chance of \$10, 1% chance of \$0, expected value \$9.90 (52:55, 53:13, 54:36). Lottery B: 1% chance of \$1,000, 99% chance of \$0, expected value \$10 (53:23, 54:32). B has the higher expected value, yet most people pick A (54:46, 54:53, 55:07). Risk aversion is defined exactly by this pattern: a decision-maker always prefers receiving the lottery's expected value for certain over the lottery itself, u(E[P]) ≥ U(P) (55:56, 56:04, 56:12, 56:34) — Jensen's inequality applied to utility: for concave u (diminishing marginal utility of money), u(E[X]) ≥ E[u(X)] (55:25, 56:41). People maximize expected utility, not expected money, and the wedge between the two is the risk premium (57:01).
| Ordinal (certainty) | Cardinal / VNM (uncertainty) | |
|---|---|---|
| Object of choice | outcomes | lotteries — distributions over outcomes |
| What's invariant | order only | order and relative "distances" |
| Valid transform | any strictly increasing function | positive affine only: a·u + b, a > 0 |
| Formula | none needed | U(P) = Σ Pᵢ u(Zᵢ) |
Where it lands in Jayverse
- Verex: price ≠ probability. A mark price is Σ Pᵢ u(payoffᵢ) for the marginal trader, not Pᵢ itself; a risk-averse trader (concave u) compresses near-even markets toward 50% and skews long shots down — the favourite-longshot bias. Read a Verex price as a probability only with that caveat attached, and check resolution price against mark price with this gap in mind rather than treating a divergence as a bug.
- Number: a reading that backs out implied risk aversion. Given a market's price series and its resolution, fit a concave u (e.g. CRRA) that makes observed prices consistent with expected-utility maximization; the fitted curvature is a risk-aversion reading, publishable like any other Number reading, with u's functional form and fit method stated as the methodology.
- Theory neighbours: the base case for 내시균형, 경매, and 메커니즘 디자인. The beauty contest is iterated best-response converging to a Nash equilibrium, the mechanism every later strategic-interaction item assumes; risk aversion under VNM utility is the standing assumption behind 확률·기대값 items on bidding, since an auction's optimal bid depends on the bidder's risk attitude, not just their expected value.
- Auditor: "preferences can't be irrational, choices can" is the Auditor's own distinction. A rule the Auditor checks has to be a consistency rule over observed choices (does this trader's revealed behavior violate transitivity or independence), never a judgment on what payoffs someone "should" want.
- Eng: the two-lottery example is the ready-made interview answer for "explain expected utility." \$9.90-expected-value A beating \$10-expected-value B because of concavity is a 30-second answer that also shows the difference between expected value and expected utility — useful for a team-lead interview question on quant intuition.
Verified and unverified
Verified on 2026-09-19: MIT 14.12, Economic Applications of Game Theory, taught by Muhamet Yildiz, is a real MIT OpenCourseWare course; the beauty-contest / guess-two-thirds-of-the-average game is a well-known experiment (Nagel, 1995) whose unique Nash equilibrium is 0 (or 0/1 with integer constraints and specific tie-break rules); von Neumann–Morgenstern expected utility theory (1944) is the standard framework for choice under uncertainty; risk aversion is equivalent to a concave utility function by Jensen's inequality; Myerson's and Aumann's definitions of game theory, as paraphrased in the lecture, are standard citations in the field. Taken from the summary and not independently checked: the exact round-1 and round-2 winning numbers (24 and 14), every timestamp, and the specific classroom anecdote about the student who wrote 100.
Sources: YouTube — MIT 14.12 Game Theory, Lecture 1: Introduction to Individual Decision-Making · MIT OpenCourseWare, 14.12 Economic Applications of Game Theory (Muhamet Yildiz) · related items: Theory 내시균형, 경매, 메커니즘 디자인, 확률·기대값 · Invest mit-financial-markets-terms-edge, dalio-economic-machine-holy-grail · Tech #62 (agentic engineering writes the boundaries).
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| Keynesian beauty contest | 케인즈 미인대회(남들의 평균 선택을 예측해서 맞히는 게임) · 재귀적 추론을 보여주는 고전 예시. "runs the Keynesian beauty-contest guessing game live" |
| Nash equilibrium | 내시 균형(누구도 혼자 바꿔서 더 나아질 수 없는 선택 조합) · 전략적 상호작용의 기본 해 개념. "the game's unique Nash equilibrium" |
| iterated reasoning / recursion | 반복 추론 / 재귀("상대가 무엇을 생각할지를 생각하는" 과정) · 평균이 0으로 수렴하는 논리를 설명할 때. "The logic is a fixed point" |
| fixed point | 고정점(반복 적용해도 값이 변하지 않는 점) · 수렴 논증의 수학 용어. "The logic is a fixed point" |
| ordinal utility | 서수적 효용(순서만 의미 있는 효용) · 확실성 하 선택 이론의 표준 개념. "A utility function U representing this ranking is ordinal" |
| cardinal utility | 기수적 효용(크기 자체가 의미 있는 효용) · 불확실성 하 선택 이론의 표준 개념. "cardinal utility for choice under uncertainty" |
| VNM (von Neumann–Morgenstern) utility | 폰 노이만-모르겐슈테른 효용(기대효용 이론의 공식 이름) · 1944년 이래 표준 프레임워크. "von Neumann–Morgenstern (VNM) cardinal utility" |
| lottery | 복권(결과에 대한 확률분포) · 불확실성 하에서 실제로 선택되는 대상. "a lottery Δ(Z) — a probability distribution over outcomes" |
| expected utility | 기대 효용(확률로 가중한 효용의 합) · 기대값과 구별되는 핵심 개념. "U(P) = Σᵢ Pᵢ u(Zᵢ)" |
| expected value (EV) | 기댓값(확률로 가중한 결과값 자체의 합) · 기대 효용과 대비되는 개념. "expected value \$9.90" |
| risk aversion | 위험 회피(복권보다 확실한 기댓값을 선호하는 성향) · 이 글 전체의 핵심 정의. "Risk aversion is defined exactly by this pattern" |
| concave (concavity) | 오목(함수가 아래로 볼록한 형태, 한계효용 체감을 뜻함) · 위험 회피의 수학적 조건. "This holds precisely when u is concave" |
| Jensen's inequality | 옌센 부등식(오목함수에서 u(E[X]) ≥ E[u(X)]가 성립한다는 부등식) · 위험 회피를 증명하는 도구. "which is Jensen's inequality applied to utility" |
| diminishing marginal utility | 한계효용 체감(가진 게 많을수록 추가분의 가치가 줄어듦) · 오목성의 직관적 의미. "diminishing marginal utility of money" |
| affine transformation | 아핀 변환(a·u + b 형태의 선형 변환) · 기수적 효용이 허용하는 유일한 변환. "positive affine only: a·u + b, a > 0" |
| risk premium | 위험 프리미엄(기댓값과 지불 의사 사이의 차이) · 위험 회피가 가격에 남기는 흔적. "the wedge between the two is the risk premium" |
| favourite-longshot bias | 우승 후보-롱샷 편향(강자는 과소평가, 약자는 과대평가되는 가격 왜곡) · 예측시장 가격이 확률과 어긋나는 대표적 패턴. "the classic favourite-longshot bias" |
| mark price / resolution price | 마크 가격(현재 거래가) / 정산가(최종 확정가) · Verex 같은 예측시장에서 괴리를 설명할 때. "resolution price should be checked against mark price" |
| transitivity | 이행성(A≻B, B≻C이면 A≻C) · 선호가 일관되기 위한 최소 조건. "transitive by assumption" |
| CRRA (Constant Relative Risk Aversion) | 불변 상대위험회피(위험 회피 정도가 부에 비례해 일정하다고 가정하는 효용함수 형태) · Number 읽기에서 위험 회피를 역산할 때 쓰는 표준 함수형. "fit a concave u (e.g. CRRA)" |