> ## Documentation Index
> Fetch the complete documentation index at: https://raymondchang.me/llms.txt
> Use this file to discover all available pages before exploring further.

# Quant interview: probability and statistics

> Probability and statistics questions for quantitative research interviews, with worked answers: expectation, Bayes, independence, CLT and dice games.

This page collects common probability and statistics questions from quantitative research interviews. Each answer gives the intuition first, then the calculation.

> [!NOTE]
>
> **Status of this page.** Last checked on 15 November 2026. These are general question types and solution frameworks, not any firm's question bank. The numerical results have been recomputed.

## Expectation and conditional probability

> [accordion] Q1. You flip a fair coin repeatedly. What is the expected number of flips until the first head?
>
> This is a geometric distribution. With success probability p = 1/2, the expected number of flips is 1/p = **2**.
>
> As a recursion: E = 1 + (1/2) × 0 + (1/2) × E, so E = 2. Interviewers usually want to see that you can set up first-step analysis.

> [accordion] Q2. What is the expected number of flips to see HH? And to see HT? Why are they different?
>
> Use first-step analysis.
>
> **HH:** Let E0 be the expectation from the start and E1 the expectation after one head.
>
> * E0 = 1 + (1/2) E1 + (1/2) E0, so E0 = 2 + E1
> * E1 = 1 + (1/2) × 0 + (1/2) E0, so E1 = 1 + E0 / 2
> * Substituting gives E0 = **6**
>
> **HT:** After a head, another head leaves you in the same state (E1 = 1 + (1/2) E1, so E1 = 2). Then E0 = 2 + E1 = **4**.
>
> The difference: a failed HH attempt (a tail) resets you completely, while a failed HT attempt (a head) still leaves you with a useful head. This comes from how the pattern overlaps with itself.

> [accordion] Q3. A test detects a disease with 99% sensitivity and has a 5% false positive rate. Prevalence is 1%. You test positive. What is the probability you have the disease?
>
> Use Bayes' theorem:
>
> `P(D | +) = P(+ | D) P(D) / [P(+ | D) P(D) + P(+ | not D) P(not D)]`
>
> `= 0.99 × 0.01 / (0.99 × 0.01 + 0.05 × 0.99) = 0.0099 / 0.0594 ≈ 16.7%`
>
> The point is the base rate: when prevalence is low, most positives are false positives even for an accurate test. The trading analogy: a signal for a rare event (for example an abnormal spread) is mostly noise if the event itself is rare, even when the hit rate looks high.

> [accordion] Q4. Monty Hall: should you switch? Explain without just quoting the answer.
>
> Yes. Switching wins with probability **2/3**, staying wins with 1/3.
>
> Intuition: your first pick is right with probability 1/3, and that does not change when the host opens a door with a goat. The remaining 2/3 concentrates on the other unopened door. The key assumption is that the host knows where the car is and always opens a goat door. If the host opens a door at random and happens to reveal a goat, the conclusion changes (switching and staying are equally likely to win).

## Games and fair value

> [accordion] Q5. You roll a fair die and are paid the face value. You may reroll once and must keep the second result. What is the fair value of the game?
>
> Without a reroll the expectation is 3.5, so keep the first roll if it is above 3.5 and reroll otherwise.
>
> * Keep 4, 5 or 6 (probability 1/2, average 5)
> * Reroll (probability 1/2, expected value 3.5)
>
> Fair value = 0.5 × 5 + 0.5 × 3.5 = **4.25**.
>
> Extension: with two rerolls available, the threshold for the first roll becomes 4.25, so you keep only 5 or 6 (probability 1/3, average 5.5). Value = (1/3) × 5.5 + (2/3) × 4.25 ≈ **4.67**. This is backward induction, the same logic as optimal exercise of an American option.

> [accordion] Q6. What is the expected value of the maximum of two fair dice?
>
> `P(max = k) = (2k − 1) / 36` for k = 1 to 6.
>
> `E[max] = Σ k (2k − 1) / 36 = (2 × 91 − 21) / 36 = 161 / 36 ≈ 4.47`
>
> Sanity check: the maximum of two dice should exceed the single-die expectation of 3.5, and 4.47 does.

## Dependence and the CLT

> [accordion] Q7. Does zero correlation imply independence? Give a counterexample.
>
> **No.** Correlation only measures linear dependence.
>
> Counterexample: X \~ N(0, 1) and Y = X². Y is fully determined by X, so they are dependent, yet `Cov(X, Y) = E[X³] − E[X] E[X²] = 0`.
>
> Exception: if (X, Y) are jointly normal, zero correlation does imply independence. In finance, zero correlation between returns does not mean tail risk or volatility will not rise together.

> [accordion] Q8. State the central limit theorem. Where does it break down for financial returns?
>
> **Statement:** For i.i.d. random variables with finite variance, the standardized sample mean converges to a normal distribution.
>
> **Where it fails or converges slowly in finance:**
>
> * **Fat tails:** With infinite variance, or a very large fourth moment, convergence is slow and sample means and variances are unstable.
> * **Dependence:** Returns are not independent. Volatility clustering makes the effective sample size smaller than the nominal one.
> * **Non-stationarity:** The distribution changes over time, so history may not represent the future.
>
> Practical consequence: normal-based confidence intervals and VaR understate tail risk.

> [accordion] Q9. A strategy has an annualized Sharpe ratio of 1.0 over 4 years of daily data. Is the mean return statistically significant?
>
> The t-statistic is roughly the annualized Sharpe times the square root of the number of years: `t ≈ 1.0 × √4 = 2`. That is right at the edge of 5% significance (two-tailed critical value about 1.96).
>
> Caveats:
>
> * This assumes independent daily returns with constant variance. With autocorrelation or volatility clustering, standard errors are understated and you should use HAC (Newey-West) standard errors.
> * If the strategy was picked from many candidates, t = 2 is not enough and you need a multiple-testing correction. See [Regression, time series and backtesting](https://raymondchang.me/quantitative-research/interview/statistics-backtesting.md).

> [accordion] Q10. What does a p-value of 0.03 mean, and what does it not mean?
>
> **Meaning:** If the null hypothesis is true, the probability of observing a result at least this extreme is 3%.
>
> **It is not:**
>
> * The probability that the null hypothesis is true
> * A 97% probability that the effect is real
> * A measure of how large or economically meaningful the effect is
>
> Note that p-values depend on sample size, so with enough data even a tiny effect is significant. Report effect size and confidence intervals alongside the p-value.

## Related pages

* [Derivatives pricing and Greeks](https://raymondchang.me/quantitative-research/interview/derivatives-greeks.md)
* [Regression, time series and backtesting](https://raymondchang.me/quantitative-research/interview/statistics-backtesting.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.