← Coinbase Interview Insights

Coinbase·Data Scientist·Technical Phone Screen·Intermediate

IntermediatePrefer not to say
Jun 2026Remote

Summary

Coinbase data scientist interview that went deep into probability theory. The whole session was basically a single extended question on binomial/Poisson models with a bunch of sub-parts, which I wasn't expecting to be so math-heavy for a DS role.

Questions Asked (6)

Q1

For n independent users each linking a wallet with probability p, derive the probability that at least one user links and the probability that exactly k users link. Cover edge cases like p=0, p=1, large n, and the small-p large-n Poisson approximation.

Product Analytics & MetricsAlgorithms & Data Structures
Author's notes

The binomial formula itself was fine but I fumbled the edge cases a bit.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize that the number of users linking wallets follows a Binomial(n, p) distribution. Then derive the probability of at least one link as 1 - (1-p)^n and the probability of exactly k links using the binomial PMF. Discuss edge cases and the Poisson approximation for large n and small p.

Pro tip: Connect the math to Coinbase's business context by mentioning how these probabilities inform growth metrics, such as conversion rates and expected active users, and how the Poisson approximation can simplify modeling for large user bases with low link rates.

1. Define the random variable

Let X be the number of users who link their wallet. State that X ~ Binomial(n, p) since each user links independently with probability p.

2. Derive probability of at least one link

Use the complement rule: P(X ≥ 1) = 1 - P(X = 0) = 1 - (1-p)^n.

3. Derive probability of exactly k links

Use the binomial PMF: P(X = k) = C(n, k) * p^k * (1-p)^(n-k), for k = 0, 1, ..., n.

4. Analyze edge cases

For p=0, P(X ≥ 1)=0 and P(X=k)=1 if k=0 else 0. For p=1, P(X ≥ 1)=1 and P(X=k)=1 if k=n else 0. For large n, discuss behavior and Poisson approximation when p is small.

5. Apply Poisson approximation

When n is large and p is small such that λ = np is moderate, X approximately follows Poisson(λ). Then P(X=k) ≈ e^{-λ} λ^k / k! and P(X ≥ 1) ≈ 1 - e^{-λ}.

Key Points to Mention

  • Binomial distribution assumptions: independence and constant probability p.
  • Complement rule for 'at least one' probability.
  • Binomial coefficient and PMF formula.
  • Edge cases: p=0, p=1, and their implications.
  • Poisson approximation conditions: n large, p small, λ = np constant.
  • Business relevance: interpreting results for user growth and conversion metrics.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

Derive the expected number of users who link a wallet. Be explicit about where independence is and isn't required.

Product Analytics & MetricsAlgorithms & Data Structures
Author's notes

Linearity of expectation saves you here and I said that immediately, which felt good.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the metric precisely: the expected number of users who link a wallet, which is the sum of individual link probabilities. Then model the probability for each user using a funnel (e.g., signup → wallet creation → link) and explicitly state where independence is assumed (across users) and where it is not required (within a user's funnel steps). Finally, discuss how to estimate the parameters from data and validate the model.

Pro tip: Emphasize that independence across users is sufficient for linearity of expectation, but within-user dependencies (e.g., a user's steps) must be handled via conditional probabilities or joint distributions. This shows you understand the subtlety beyond just multiplying probabilities.

1. Define the metric and scope

Clarify what 'link a wallet' means (e.g., connecting an external wallet to a Coinbase account) and the population of users (e.g., new signups in a given period). Specify the time frame and any relevant segments.

2. Model individual link probability

For each user, model the probability of linking as a product of conditional probabilities for sequential steps (e.g., P(link) = P(signup) * P(create wallet | signup) * P(link | create wallet)). Acknowledge that these steps are not independent, so use conditional probabilities.

3. Apply linearity of expectation

The expected number of users who link is the sum of individual link probabilities. This requires independence across users only for variance calculations, but the expectation itself does not require independence across users.

4. Estimate parameters and validate

Use historical data to estimate the conditional probabilities, possibly with cohort analysis or survival models. Validate the model by comparing predicted vs. actual link counts and checking for overdispersion.

5. Discuss assumptions and extensions

Explicitly state assumptions: independence across users (for variance), no interference, and stationarity. Discuss extensions like heterogeneous user segments, time-varying probabilities, or network effects.

Key Points to Mention

  • Linearity of expectation: E[sum] = sum(E), which does not require independence across users.
  • Independence across users is needed for variance and confidence intervals, but not for the expected value.
  • Within-user steps are dependent, so use conditional probabilities or a joint distribution.
  • Parameter estimation from historical data (e.g., funnel conversion rates) and potential biases.
  • Heterogeneity: different user segments may have different link probabilities; consider mixture models.
  • Validation: compare predicted vs. actual, check for overdispersion, and consider time trends.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

For small p and large n, give the Poisson approximation for P(at least one user links), then derive the first-order error term including its direction.

Product Analytics & MetricsAlgorithms & Data StructuresTechnical Trade-offs
Author's notes

I blanked on the error term direction.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by setting up the problem: let p be the probability a user links, n the number of users, and assume independence. Then use the Poisson approximation with λ = np to estimate P(at least one link) ≈ 1 - e^{-λ}. For the error term, compare the exact binomial probability 1 - (1-p)^n with the Poisson approximation and expand to first order in p, showing that the Poisson approximation underestimates the true probability by approximately λp/2.

Pro tip: Mention that the error direction (underestimation) is consistent with the fact that the Poisson approximation ignores the negative binomial correlation, and that in practice for Coinbase-scale data (large n, small p), the error is negligible unless p is not truly small.

1. Define variables and assumptions

Let p be the probability that a single user links, n the number of users, and assume independent linking events. Define λ = np as the expected number of links.

2. State Poisson approximation

For small p and large n, the number of links approximately follows a Poisson distribution with mean λ. Thus, P(at least one link) ≈ 1 - e^{-λ}.

3. Write exact binomial probability

The exact probability of at least one link is 1 - (1-p)^n. Use this to compute the error of the Poisson approximation.

4. Derive first-order error term

Expand both expressions to first order in p (or λ/n) and subtract. Show that the error is approximately λp/2, meaning the Poisson approximation underestimates the true probability.

5. Interpret direction and magnitude

Conclude that the Poisson approximation is an underestimate, with error proportional to p. For small p, the error is negligible, but it grows with p.

Key Points to Mention

  • Poisson approximation: P(X ≥ 1) ≈ 1 - e^{-λ}, where λ = np
  • Exact binomial: P(X ≥ 1) = 1 - (1-p)^n
  • First-order error term: λp/2 (or np^2/2)
  • Direction: Poisson approximation underestimates the true probability
  • Assumption of independence and small p, large n
  • Practical implication: error is small for typical Coinbase-scale scenarios

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

With p=0.30 and n=5, compute P(at least one user links), P(exactly 2 users link), and the expected number of links.

Product Analytics & Metrics
Author's notes

Straightforward plug-and-chug after the theory parts.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize this as a binomial distribution problem with n=5 and p=0.30. Compute P(at least one) using the complement rule, P(exactly 2) using the binomial PMF, and the expected value as n*p. Show your work clearly and interpret the results in the context of user behavior.

Pro tip: After computing the numbers, briefly explain what they mean for the business—e.g., 'On average, 1.5 out of 5 users link, but there's a 16.8% chance nobody links, which could impact engagement metrics.' This demonstrates you can translate math into actionable insights.

1. Identify the distribution

State that the number of users who link follows a binomial distribution with parameters n=5 and p=0.30, assuming independence.

2. Compute P(at least one)

Use the complement rule: P(at least one) = 1 - P(none) = 1 - (0.70)^5. Calculate and round appropriately.

3. Compute P(exactly 2)

Apply the binomial PMF: P(X=2) = C(5,2) * (0.30)^2 * (0.70)^3. Compute the combination and powers.

4. Compute expected value

Use the formula E[X] = n * p = 5 * 0.30 = 1.5. Mention that this is the long-run average.

5. Interpret results

Briefly explain what these probabilities and expected value mean in the context of user linking behavior, e.g., likelihood of at least one link, typical number of links.

Key Points to Mention

  • Binomial distribution assumptions: fixed number of trials (n=5), independent trials, constant probability of success (p=0.30), binary outcome (link or not).
  • Complement rule for 'at least one' to simplify calculation.
  • Binomial coefficient formula: C(n, k) = n! / (k! (n-k)!).
  • Expected value formula for binomial: E[X] = n * p.
  • Interpretation of results in business context (e.g., average links per 5 users, probability of no links).
  • Rounding and precision: present probabilities to 3-4 decimal places or as percentages.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q5

Two users A and B have independent linking probabilities of 0.40 and 0.60. Compute P(A or B), P(A and B), and P(A and B given A or B).

Product Analytics & MetricsAlgorithms & Data Structures
Author's notes

The conditional at the end is where people slip up and I nearly did.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify that A and B are independent events with given probabilities. Then apply the addition rule for P(A or B) and multiplication rule for P(A and B). Finally, use the definition of conditional probability to compute P(A and B | A or B).

Pro tip: In a data science interview, always state your assumptions (e.g., independence) and explain how you would validate them with data. This shows you think beyond the math and consider real-world implications.

1. Clarify independence and notation

Confirm that A and B are independent events, meaning P(A and B) = P(A) * P(B). Define the events clearly: A = user A links, B = user B links.

2. Compute P(A and B)

Since A and B are independent, multiply their probabilities: P(A and B) = 0.40 * 0.60 = 0.24.

3. Compute P(A or B)

Use the addition rule: P(A or B) = P(A) + P(B) - P(A and B) = 0.40 + 0.60 - 0.24 = 0.76.

4. Compute P(A and B | A or B)

Apply the conditional probability formula: P(A and B | A or B) = P((A and B) and (A or B)) / P(A or B). Since (A and B) implies (A or B), the numerator is P(A and B). Thus, 0.24 / 0.76 ≈ 0.3158.

5. Interpret the results in context

Explain what these probabilities mean for Coinbase: e.g., the chance that at least one user links is 76%, both link is 24%, and given at least one links, the chance both link is about 31.6%.

Key Points to Mention

  • Independence assumption and its implications
  • Addition rule for probability of union
  • Multiplication rule for independent events
  • Definition of conditional probability
  • Interpretation of results in a business context (e.g., user linking behavior)
  • Potential need to validate independence with real data

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q6

If independence between users A and B is not assumed, but you know P(A or B)=0.70 with marginals 0.40 and 0.60, what is the feasible range for P(A and B)? Then give the corresponding range for P(A and B given A or B), and explain what dependence structure produces each boundary.

Product Analytics & MetricsAlgorithms & Data StructuresTechnical Trade-offs
Author's notes

This was the hardest part and I definitely didn't nail it.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, use the inclusion-exclusion principle to express P(A and B) in terms of the given probabilities, then apply the constraints that probabilities must be between 0 and 1 and that P(A and B) cannot exceed the marginals. Next, compute the conditional probability range by dividing the intersection range by P(A or B) = 0.70. Finally, interpret the boundaries in terms of dependence structures: maximum overlap (A subset of B) and minimum overlap (mutually exclusive with the union fixed).

Pro tip: Emphasize that the conditional probability is just a linear transformation of the intersection probability, so the range scales by 1/0.70. Also, note that the dependence structures correspond to extreme positive and negative dependence, which is crucial for understanding risk in financial data.

1. Apply inclusion-exclusion

Use P(A or B) = P(A) + P(B) - P(A and B) to solve for P(A and B) = 0.40 + 0.60 - 0.70 = 0.30. This is the only possible value given the union and marginals, so the range is a single point.

2. Check feasibility constraints

Verify that 0.30 is valid: it must be ≤ min(0.40, 0.60) = 0.40 and ≥ max(0, 0.40+0.60-1) = 0. Since 0.30 satisfies these, it is feasible. Thus the range for P(A and B) is [0.30, 0.30].

3. Compute conditional probability

Calculate P(A and B | A or B) = P(A and B) / P(A or B) = 0.30 / 0.70 ≈ 0.4286. Since the intersection is fixed, the conditional probability is also fixed at approximately 0.4286.

4. Explain dependence structures

The single value corresponds to a specific dependence structure: since P(A and B) = 0.30, which is less than P(A)P(B)=0.24? Actually 0.30 > 0.24, so events are positively dependent. The structure is such that the overlap is exactly 0.30, which is the only possibility given the union and marginals.

Key Points to Mention

  • Inclusion-exclusion principle: P(A or B) = P(A) + P(B) - P(A and B)
  • Feasibility bounds for intersection: max(0, P(A)+P(B)-1) ≤ P(A and B) ≤ min(P(A), P(B))
  • Conditional probability definition: P(A and B | A or B) = P(A and B) / P(A or B)
  • Dependence structure interpretation: positive dependence when P(A and B) > P(A)P(B), negative when less
  • The given union and marginals uniquely determine the intersection, so no range exists
  • In product analytics, such constraints help validate metric definitions and user overlap assumptions

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.