Expected Value and Fair-Game Reasoning for Quant Interviews

Expected value is the central tool of quant interviews, trading, and risk management. Most brainteasers, market-making rounds, and probability questions reduce to “what’s the expected value of this random variable?” or “would you take this bet at this price?” Strong candidates manipulate expected values fluently — using linearity, conditional expectations, and indicator decompositions — without grinding through full distributions. This guide covers the foundational tools, classic fair-game problems, and the tactical patterns that work in interviews.

Foundational Tools

Linearity of Expectation

The most powerful identity in probability: E[X + Y] = E[X] + E[Y], regardless of whether X and Y are independent. This decomposes complex random variables into simple pieces.

Example: “Roll 10 fair dice and sum them. Expected sum?” Each die has E[X_i] = 3.5. Total = 10 × 3.5 = 35. No need to compute the joint distribution.

Harder example: “Roll a fair die N times. Expected number of distinct values seen?” Define indicator I_k = 1 if value k appears at least once. P(I_k = 1) = 1 – (5/6)^N. Expected distinct = 6(1 – (5/6)^N). For N=6: 6(1 – (5/6)⁶) ≈ 6(1 – 0.335) ≈ 4.0.

Indicator Variables

For “expected count of X” problems, define an indicator for each potential occurrence and sum. The expectation of an indicator equals the probability the event happens.

Example: “Shuffle a deck of 52 cards. Expected number of cards in their original sorted position?” Define I_k = 1 if card k is in position k. P(I_k = 1) = 1/52 (any specific card is equally likely to land at any of 52 positions). Sum: 52 × 1/52 = 1. Independent of deck size; surprising on first encounter.

Law of Total Expectation

E[X] = E[E[X | Y]]. Condition on a useful intermediate variable Y, compute E[X | Y] for each value of Y, then average.

Example: “I roll a fair die. If I get N, I flip N coins. Expected number of heads?” Condition on the die: E[heads | die = N] = N/2. So E[heads] = E[N/2] = 3.5/2 = 1.75.

Recursion / Conditioning on First Step

For problems with self-similar structure, set up a recursion by conditioning on the first event.

Example: “Expected number of fair coin flips to get heads?” Let E be the answer. With probability 1/2 we get heads on the first flip (1 flip). With probability 1/2 we get tails and start over (1 + E). So E = (1/2)(1) + (1/2)(1 + E) → solve: E = 2.

Fair-Game Reasoning

A fair game is one where E[gain] = 0. Asking “is this a fair game” or “what would make this fair” is a constant interview frame because trading is fundamentally about identifying when games aren’t fair (you have edge) and trading them.

The basic fair-game frame

“I’ll pay you $X if a fair coin lands heads, you pay me $Y if tails. What relationship between X and Y makes this fair?” Fair: (1/2)(X) + (1/2)(-Y) = 0 → X = Y. If you’re offered $X = $11 and asked to pay $Y = $10, that’s a positive-EV bet for you of $0.50.

Stopping problems

“I roll a fair die. You can stop and take the value, or pay $1 to reroll. Optimal strategy and expected payoff?”

Set V = expected value of optimal strategy. If you reroll, expected value = -1 (the cost) + E[max(N, V)]. The optimal threshold accepts values above some K and rerolls below. Solving: K = 4, accept 4/5/6, reroll 1/2/3. Expected value = (1/2)(5) + (1/2)(V – 1), so V = 5/2 + (V-1)/2, → V = 4.

Multi-stage games

“Each round, a fair coin is flipped. If heads, you double your money. If tails, you lose half. Long-run expected wealth?”

Per round: E[multiplier] = (1/2)(2) + (1/2)(0.5) = 1.25. So expected wealth grows. But: this is the “Kelly paradox” setup. The arithmetic mean grows but the geometric mean (relevant for repeated multiplicative gambles) is sqrt(2 × 0.5) = 1, meaning typical wealth doesn’t grow. Strong candidates note both.

Classic Expected-Value Problems

Coupon collector

“Coupons of N types are randomly distributed (each type equally likely). Expected number of coupons to collect a complete set?”

Decompose: time to get the k-th new coupon, given k-1 already collected, follows geometric with p = (N-k+1)/N. Expected wait: N/(N-k+1). Sum from k=1 to N: N × (1 + 1/2 + 1/3 + … + 1/N) = N × H_N. For N=6: 6(2.45) ≈ 14.7. For N=52 (deck of cards): ≈ 236.

Drawing from urn without replacement

“Urn with 50 white and 50 black balls. Draw all 100 without replacement. Expected number of color changes in the sequence?”

Indicator: I_k = 1 if positions k and k+1 differ in color. P(I_k = 1) = 2 × (50 × 50) / (100 × 99) ≈ 0.505. Sum over 99 adjacent pairs: ≈ 50. (Equivalent to sum of indicators for “the i-th ball in the sequence triggers a change.”)

Expected longest run of heads

“Flip a fair coin N times. Expected length of the longest run of consecutive heads?” For large N: approximately log₂(N). For N = 100: about 7. (Exact computation is harder; approximation suffices for interviews.)

The 100-prisoner problem

“100 prisoners are given numbers 1 to 100 randomly placed in 100 boxes. Each prisoner can open 50 boxes; if they all find their own number, they all live. Best strategy?”

Random opening: probability all succeed = (1/2)¹⁰⁰. Vanishingly small.

Optimal strategy: prisoner k opens box k, then opens the box matching the number found, etc., following the cycle. Probability of success: probability the random permutation has no cycle longer than 50. This is approximately 1 – ln(2) ≈ 30.7%. The famous result: a clever strategy turns near-certain death into near-1/3 survival.

Conditional Expectation Tactics

The “secretary problem” framing

“You see N candidates one at a time, must hire one immediately on seeing them, can’t go back. Strategy to maximize probability of hiring the best?”

Optimal strategy: reject the first N/e candidates outright; hire the next one better than all rejected. Probability of success: 1/e ≈ 36.8%. Connects to fair-game reasoning because hiring early bets on average quality; hiring late bets on having seen enough info.

Optimal-stopping framing

“Stock starts at $100, each minute moves up $1 or down $1 with equal probability. You can sell at any time but must sell within an hour. Optimal stopping strategy and expected sale price?”

Without discounting: E[sale price] = $100 by martingale logic; the random walk is a martingale and any stopping rule gives the same expected value. Surprising result tested often.

Pricing-Adjacent Expected Value

Bid-ask reasoning

“A coin is flipped 10 times. The number of heads is H. I’ll pay you H². What would you pay to play?”

Expected payoff = E[H²] = Var(H) + E[H]² = 10(0.25) + 25 = 27.5. So a fair price is $27.50; you’d pay less to make money.

Fair odds

“I roll two fair dice. What odds should I offer on the sum being 7?”

P(sum = 7) = 6/36 = 1/6. Fair odds are 5:1 against. If I’m offered odds of 6:1 against (i.e., I pay $1 if sum ≠ 7, you pay $6 if it equals 7), expected gain = (1/6)(6) + (5/6)(-1) = 1/6, which is positive. Bet.

Common Pitfalls

Confusing E[1/X] with 1/E[X]. Generally not equal (Jensen’s inequality). Don’t naively swap.
Computing full distributions when expected value suffices. Linearity of expectation is your friend; use it before grinding through joint distributions.
Forgetting the variance component. “What’s the expected value of squared returns?” requires variance, not just mean.
Mistaking arithmetic for geometric mean. For multiplicative compounding, geometric mean (or log-mean) matters; arithmetic mean overstates.

Interview Tactical Patterns

Pattern 1: Decompose into indicators

If the question is “expected count of X,” try to define an indicator for each potential X and sum. This works for any problem with countable potential occurrences.

Pattern 2: Condition on a useful variable

If a problem has a natural intermediate quantity (the result of a roll, the time of an event, the size of a set), condition on it.

Pattern 3: Set up a recursion

If the problem has self-similar structure (“expected time until X” with restarts, “optimal strategy” with reset behavior), let E be the answer and write an equation in E.

Pattern 4: Identify a martingale

For random-walk-style problems, ask: is the quantity a martingale? If so, expected value at any stopping time equals the starting value (assuming standard conditions hold).

Pattern 5: Sanity-check with extremes

Compute the expected value at extreme parameter values to check formula correctness. If you derive E[X] = N(N+1)/2 for some quantity, check N=1 (should give 1) and N=2 (should give 3).

Practice Strategy

Weeks 1–2: work through 30–50 expected-value problems from Crack’s Heard on the Street or Zhou’s A Practical Guide to Quantitative Finance Interviews. Focus on fluency with the foundational tools.

Weeks 3–4: harder problems — coupon collector variants, conditional expectation, multi-stage games. Time yourself: target 5 minutes per problem with clear reasoning.

Weeks 5+: mock interviews. Have a peer present novel problems and grade your reasoning, not just final answers. Talk through your thought process out loud.

Frequently Asked Questions

How is expected value different from mean?

Same concept, different terminology. The “mean” of a random variable’s distribution is its expected value. “Mean of N data points” is the sample mean (an estimator of the true expected value). Interviewers use the terms interchangeably; you should be comfortable with both. The distinction that matters is between arithmetic mean (additive averaging) and geometric mean (multiplicative averaging), which affects long-run multiplicative-payoff problems.

When should I use conditional expectation vs direct computation?

Use conditional expectation when there’s a natural intermediate variable to condition on (the outcome of a coin flip, the size of a set, the time of an event). Use direct computation (full distribution) when the variable is simple enough that conditioning adds complexity rather than removing it. As a rule of thumb: if the problem has multiple stages or recursive structure, condition. If it’s a single-step computation, don’t.

How do I avoid silly errors in expected-value problems under interview pressure?

Three habits: (1) write down what you’re computing in words before computing it (“Let E = expected number of flips until two heads in a row”). (2) Sanity-check with simple cases (“If only one heads needed, E should be 2; my formula better give 2 in that case”). (3) Talk through linearity decomposition explicitly when using it (“Total = sum of these indicators; expected total = sum of indicator probabilities”). Silent computation under pressure is where errors happen.

Do quant traders actually compute expected values in their day-to-day work?

Constantly, but usually with technology assistance. Manual mental computation of EV happens in market-making (quick fair-value updates as new info arrives), in evaluating proposed trades or strategies, and in interviews. Production trading systems compute EV at scale via models. The interview frame tests whether you have the underlying intuition to evaluate models, propose strategies, and reason about risk; the day-to-day arithmetic is automated for most quant work.

What’s the connection between expected value and trading?

Direct. A trade is a bet; the bet has an expected value (positive if you have edge, negative if you don’t). Market makers quote prices that, on average, capture a small positive EV per trade through bid-ask spread. Quant strategies are sequences of trades whose expected total return must exceed transaction costs to be profitable. Interviewers test EV reasoning because everyday trading reasoning sits on top of it. Strong candidates can articulate any trade in EV terms; weak candidates struggle to translate vague intuitions into computable EV.