← Upstart Interview Insights

Upstart·Data Scientist·Technical Phone Screen·Intermediate

IntermediatePrefer not to say
May 2026Remote

Summary

Upstart data scientist interview was pretty much a gauntlet of probability and stats questions. Four problems back to back, ranging from radioactive decay to Monty Hall. Felt like a grad school exam more than a job interview.

Questions Asked (4)

Q1

A radioactive atom has a half-life of 1 day and decays following an exponential distribution. Starting with 100 atoms, what is the probability distribution of the number still alive after 10 days? Also compute the expected count and the probability at least one survives, then write a simulation that outputs each atom's final state.

Algorithms & Data StructuresProduct Analytics & Metrics
Author's notes

This one took me a second to set up properly.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, recognize that each atom independently survives with probability p = e^{-10} after 10 days, so the number alive follows a Binomial(100, p) distribution. Then compute the expected count as 100p, the probability at least one survives as 1 - (1-p)^100, and finally outline a simulation that tracks each atom's state using exponential random variables.

Pro tip: Emphasize the independence assumption and connect the binomial model to practical applications like customer churn or conversion rates, showing you can translate statistical theory into business insights.

1. Identify the distribution

Explain that each atom's lifetime is exponential with rate λ = ln(2) per day, so the survival probability after 10 days is p = e^{-10λ} = e^{-10 ln 2} = 2^{-10} = 1/1024. Since atoms decay independently, the number alive follows Binomial(n=100, p=1/1024).

2. Compute expected count and survival probability

Calculate the expected number alive as E[X] = np = 100/1024 ≈ 0.0977. Compute the probability at least one survives as P(X ≥ 1) = 1 - P(X=0) = 1 - (1-p)^100 ≈ 1 - (1023/1024)^100 ≈ 0.093.

3. Outline the simulation

Describe a simulation where for each atom, generate a lifetime from Exponential(λ) and check if it exceeds 10 days. Alternatively, generate a Bernoulli trial with success probability p for each atom. Output the final state (alive/dead) for each atom and optionally the total count.

4. Validate and interpret

Mention that the simulation should be repeated many times to verify the distribution, expected value, and survival probability. Discuss how this model applies to independent events in business, such as user retention or conversion.

Key Points to Mention

  • Exponential distribution and its memoryless property
  • Relationship between half-life and decay constant: λ = ln(2)/half-life
  • Binomial distribution for the number of survivors
  • Independence of atoms
  • Expected value and probability calculations
  • Simulation techniques: inverse transform sampling or direct Bernoulli trials

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

Q2

Given y = x + epsilon where x and epsilon are independent standard normals, compute the population OLS slope and intercept for regressing y on x, and then separately for regressing x on y.

Product Analytics & MetricsTechnical Trade-offs
Author's notes

The y on x regression is clean: slope is Cov(y,x)/Var(x) = 1/1 = 1, intercept 0.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify that OLS regression of y on x estimates the linear relationship by minimizing squared errors, and the population slope is Cov(x,y)/Var(x). Then compute the covariance and variances using the given model y = x + epsilon with independent standard normals, and apply the formulas for slope and intercept. Repeat the process for regressing x on y, noting the asymmetry due to error in the regressor.

Pro tip: Emphasize that the slope in regressing x on y is attenuated (biased toward zero) because of measurement error in the regressor, a key insight for causal inference. Also, mention that the intercepts are zero due to zero means, but the slopes differ, illustrating regression to the mean.

1. Define OLS population parameters

State that for regressing y on x, the population slope is β = Cov(x,y)/Var(x) and intercept is α = E[y] - β E[x]. Similarly for regressing x on y, swap roles.

2. Compute moments for y on x

Given x ~ N(0,1), ε ~ N(0,1) independent, y = x + ε. Compute Var(x)=1, Cov(x,y)=Cov(x, x+ε)=Var(x)+Cov(x,ε)=1, and E[x]=E[y]=0. Thus β=1, α=0.

3. Compute moments for x on y

For regressing x on y, slope β' = Cov(x,y)/Var(y). Var(y)=Var(x+ε)=1+1=2. Cov(x,y)=1. So β'=1/2, and intercept α' = E[x] - β' E[y] = 0.

4. Interpret results and discuss implications

Note that the slopes are not reciprocals (1 vs 0.5) due to error in the regressor. Explain that this demonstrates attenuation bias and the asymmetry of OLS when variables have measurement error.

Key Points to Mention

  • Population OLS formulas: slope = Cov(x,y)/Var(x), intercept = E[y] - slope*E[x]
  • Independence of x and ε implies Cov(x,ε)=0
  • Variance of y is Var(x)+Var(ε) = 2
  • Slope for y on x is 1, intercept 0
  • Slope for x on y is 0.5, intercept 0
  • Attenuation bias: regressing x on y underestimates the true relationship due to error in y acting as measurement error in the regressor

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

Q3

In the classic 3-door Monty Hall setup, what are your winning probabilities if you switch versus if you stay? Justify briefly.

Product Analytics & Metrics
Author's notes

Switching wins with probability 2/3, staying wins 1/3.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

State the probabilities clearly: staying wins 1/3 of the time, switching wins 2/3. Then justify by explaining that your initial pick has a 1/3 chance of being correct, so the other two doors collectively have a 2/3 chance; when the host reveals a goat, that 2/3 probability collapses onto the remaining unopened door.

Pro tip: Emphasize that the host's action is not random—he knows where the prize is and always reveals a goat—which is why switching doubles your odds. This shows you understand the role of information asymmetry, a key concept in data science.

1. State the probabilities

Clearly say that switching wins with probability 2/3 and staying wins with probability 1/3.

2. Explain the initial choice

Note that your first pick has a 1/3 chance of being the car, so the other two doors together have a 2/3 chance.

3. Describe the host's action

Explain that the host, who knows where the car is, always opens a door with a goat, which does not change the 1/3 probability of your initial pick.

4. Show the probability shift

Conclude that the 2/3 probability from the other two doors now rests entirely on the single unopened door, so switching wins 2/3 of the time.

Key Points to Mention

  • Initial pick probability is 1/3.
  • Other two doors collectively have 2/3 probability.
  • Host's reveal is deliberate and always shows a goat.
  • Switching captures the 2/3 probability.
  • Staying remains at 1/3.
  • This is a classic example of conditional probability and information asymmetry.

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

Q4

You roll a fair n-sided die repeatedly. What is the expected number of rolls needed to see every face at least once?

Algorithms & Data StructuresProduct Analytics & Metrics
Author's notes

Coupon collector.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize this as the coupon collector's problem and derive the expected value by summing the expected waiting times for each new face. Use linearity of expectation, where the probability of seeing a new face when k faces remain is k/n, so the expected rolls for that phase is n/k. Sum over k=1 to n to get n * H_n.

Pro tip: Mention that this is a classic problem and that the harmonic number approximation gives n ln n + γn + 1/2, which is useful for large n. Also, note that the variance is also known and can be derived similarly, showing depth.

1. Identify the problem

Recognize that this is the coupon collector's problem: collecting all n distinct faces of a fair die.

2. Define phases

Break the process into phases: phase k starts when you have n-k distinct faces and ends when you see a new face, leaving n-k-1 faces unseen.

3. Compute expected rolls per phase

In phase k, the probability of seeing a new face is k/n, so the expected number of rolls is n/k.

4. Sum expectations

By linearity of expectation, total expected rolls is sum_{k=1}^n n/k = n * H_n, where H_n is the n-th harmonic number.

5. Provide approximation and context

For large n, H_n ≈ ln n + γ + 1/(2n), so expected rolls ≈ n ln n + γn + 1/2. Mention this to show practical insight.

Key Points to Mention

  • Coupon collector's problem
  • Linearity of expectation
  • Harmonic number H_n
  • Probability of new face in each phase
  • Asymptotic approximation n ln n + γn
  • Variance or distribution (optional)

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