← Upstart Interview Insights

Upstart·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Apr 2026

Summary

Upstart data scientist interview, looked like a technical phone screen focused almost entirely on probability, stats, and regression fundamentals. Nothing about the product or business, just math. Felt more like a grad school exam than a job interview.

Questions Asked (6)

Q1

A radioactive atom has a half-life of 1 day. Starting with 100 identical atoms, what is the probability a single atom survives after 10 days, what distribution describes the number of surviving atoms, and what are the expected number remaining and the probability at least one is left?

Algorithms & Data Structures
Author's notes

The single-atom survival probability is just (1/2)^10, which is clean enough.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Break the problem into four parts: compute the survival probability per atom using the half-life, identify the distribution of the number of survivors as Binomial(n=100, p), then compute the expected value and the probability of at least one survivor using the complement rule. Explain each step clearly and connect the concepts.

Pro tip: Emphasize that the atoms are independent and identical, which justifies the binomial model; also note that the probability at least one survives is not simply 100 times the individual survival probability due to dependence in the complement.

1. Compute single-atom survival probability

Use the half-life formula: after each day, the probability an atom survives is 1/2. After 10 days, the survival probability is (1/2)^10 = 1/1024 ≈ 0.0009765625.

2. Identify the distribution

Since there are 100 independent atoms, each with the same survival probability p, the number of surviving atoms follows a Binomial distribution with parameters n=100 and p=(1/2)^10.

3. Compute the expected number remaining

For a binomial distribution, the expected value is n*p = 100 * (1/1024) = 100/1024 ≈ 0.09765625.

4. Compute probability at least one survives

Use the complement rule: P(at least one) = 1 - P(none survive) = 1 - (1 - p)^100 = 1 - (1 - 1/1024)^100. Approximate or compute exactly.

Key Points to Mention

  • Half-life means the probability of survival after one half-life is 1/2.
  • Independence of atoms allows modeling the count as a binomial random variable.
  • Binomial parameters: n = number of atoms, p = survival probability per atom.
  • Expected value of binomial is n*p.
  • Probability of at least one is 1 minus probability of zero survivors.
  • Numerical approximation: (1 - 1/1024)^100 ≈ e^{-100/1024} ≈ e^{-0.09765625} ≈ 0.9069, so P(at least one) ≈ 0.0931.

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

Q2

Given the prior probability of a condition, the probability of a positive test result given the condition is present, and the probability of a positive result given the condition is absent, derive the posterior probability that the condition is present given a positive test.

Product Analytics & Metrics
Author's notes

Bayes' rule in its generic form.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the events and given probabilities clearly, then apply Bayes' theorem step by step. Derive the posterior probability formula and explain each component, emphasizing the role of the base rate and test accuracy. Finally, discuss practical implications such as the impact of low prevalence on posterior probability.

Pro tip: Always sanity-check the result with extreme values (e.g., prevalence near 0 or 1) to ensure the formula behaves as expected, and mention that this is the foundation of Bayesian inference used in many data science applications.

1. Define events and probabilities

Let A be the event that the condition is present, and B be the event of a positive test result. Define P(A) as the prior, P(B|A) as the sensitivity, and P(B|¬A) as the false positive rate.

2. State Bayes' theorem

Write Bayes' theorem: P(A|B) = P(B|A) * P(A) / P(B). Explain that P(B) is the total probability of a positive test.

3. Compute total probability of positive test

Calculate P(B) = P(B|A)P(A) + P(B|¬A)P(¬A), where P(¬A) = 1 - P(A).

4. Substitute and simplify

Plug P(B) into Bayes' theorem to get the posterior: P(A|B) = P(B|A)P(A) / [P(B|A)P(A) + P(B|¬A)(1-P(A))].

5. Interpret and discuss implications

Explain what the posterior represents and discuss how it changes with different priors or test accuracies, highlighting the base rate fallacy.

Key Points to Mention

  • Bayes' theorem and its application to conditional probability
  • Definitions of prior, likelihood, and posterior probabilities
  • Sensitivity (true positive rate) and specificity (true negative rate) or false positive rate
  • Total probability theorem for computing the denominator
  • Base rate fallacy and its relevance in real-world scenarios
  • Practical implications for decision-making in data science

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

Q3

If y = x + e where x and e are independent standard normals, what is the slope from regressing y on x, and what is the slope from regressing x on y?

Data ModelingTechnical Trade-offs
Author's notes

The first one is easy, slope is Cov(y,x)/Var(x) = 1.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify that the question asks for the population slopes from simple linear regressions, not sample estimates. Then, derive the slope for regressing y on x as Cov(x,y)/Var(x) and for regressing x on y as Cov(x,y)/Var(y), using the independence of x and e to compute the moments. Finally, plug in the values to get 1 and 1/2, and explain why they differ.

Pro tip: Emphasize that regression slopes are not symmetric and depend on which variable is treated as the response; this is a common pitfall in causal inference and predictive modeling.

1. Clarify the question

State that you are finding the population slope coefficients for the simple linear regressions y on x and x on y, assuming the standard normal distributions for x and e.

2. Recall the slope formula

For a simple linear regression of Y on X, the population slope is β = Cov(X,Y)/Var(X). Similarly, for X on Y, the slope is Cov(X,Y)/Var(Y).

3. Compute moments

Given y = x + e, with x and e independent standard normals, compute Cov(x,y) = Var(x) = 1, Var(y) = Var(x) + Var(e) = 2.

4. Calculate slopes

Plug in: slope of y on x = 1/1 = 1; slope of x on y = 1/2 = 0.5.

5. Interpret and discuss

Explain that the slopes differ because regression is asymmetric; the slope depends on which variable is the response. Mention implications for modeling and causality.

Key Points to Mention

  • Population slope formula: β = Cov(X,Y)/Var(X) for Y on X.
  • Independence of x and e implies Cov(x,e)=0, so Cov(x,y)=Var(x).
  • Var(y) = Var(x) + Var(e) because x and e are independent.
  • Standard normal: Var(x)=1, Var(e)=1, so Var(y)=2.
  • Regression slopes are not symmetric: slope(y~x) = 1, slope(x~y) = 0.5.
  • This asymmetry is important in causal inference and when interpreting regression coefficients.

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

Q4

In the Monty Hall problem, after the host opens a losing door, should you switch or stay, and what is the winning probability for each strategy?

Algorithms & Data Structures
Author's notes

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

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly stating that you should switch, with a 2/3 winning probability, versus 1/3 if you stay. Then explain the reasoning using conditional probability or a simulation-based argument, emphasizing that the host's action provides information. Finally, connect it to how you would validate such probabilistic reasoning in a data science context, e.g., via Monte Carlo simulation.

Pro tip: Mention that the host's behavior is crucial: he always opens a losing door and always offers the switch. If the host's strategy were different, the probabilities could change. This shows you understand the assumptions behind the problem.

1. State the answer clearly

Begin by directly answering that you should switch, and give the probabilities: 2/3 for switching, 1/3 for staying.

2. Explain the intuition

Describe the intuitive explanation: initially, your chosen door has a 1/3 chance, so the other two doors collectively have 2/3. When the host reveals a losing door, that 2/3 probability collapses onto the remaining unopened door.

3. Formalize with conditional probability

Use Bayes' theorem or a probability tree to show that P(win|switch) = 2/3 and P(win|stay) = 1/3, given the host always opens a losing door and offers the switch.

4. Address common misconceptions

Clarify that the probabilities are not 50-50 because the host's action is not random; he knows where the prize is and deliberately avoids it.

5. Connect to data science practice

Mention that this problem illustrates the importance of understanding data-generating processes and that you can verify the result via Monte Carlo simulation, a common technique in data science.

Key Points to Mention

  • Switching wins with probability 2/3, staying wins with probability 1/3.
  • The host's action is not random; he always opens a losing door and offers the switch.
  • The initial choice has 1/3 chance; the other two doors have 2/3 combined.
  • Conditional probability or Bayes' theorem can formalize the reasoning.
  • Monte Carlo simulation can empirically confirm the probabilities.
  • The problem highlights the importance of assumptions and information in probabilistic reasoning.

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

Q5

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 Structures
Author's notes

Classic 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.

Q6

What is a likelihood function in parametric modeling, and how does it differ from a probability statement?

Data Modeling
Author's notes

A probability statement treats parameters as fixed and asks how likely the data is.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the likelihood function as a function of the parameters given fixed observed data, emphasizing that it is not a probability distribution over parameters. Then contrast it with a probability statement, which treats parameters as fixed and data as random, and highlight the inversion of roles. Use a simple example like coin flips to illustrate the difference concretely.

Pro tip: Mention that the likelihood is not a probability density over parameters, so integrating it over parameters does not yield 1—this subtle point often trips up candidates and shows deep understanding. Also, connect it to why we maximize likelihood (MLE) rather than compute probabilities of parameters.

1. Define likelihood function

State that the likelihood function L(θ | x) represents the plausibility of parameter values θ given observed data x, and is proportional to the probability of the data as a function of θ.

2. Define probability statement

Explain that a probability statement P(x | θ) treats θ as fixed and x as random, describing the distribution of data under known parameters.

3. Contrast the two

Highlight that in likelihood, data is fixed and parameters vary, while in probability, parameters are fixed and data varies. Emphasize that likelihood is not a probability distribution over θ.

4. Provide a concrete example

Use a simple example, such as flipping a coin with unknown bias θ, to show how the likelihood of θ given observed heads/tails differs from the probability of observing heads given θ.

5. Connect to modeling and inference

Explain how likelihood is used in parametric modeling for estimation (e.g., MLE) and inference, and note that it is not normalized like a probability density over parameters.

Key Points to Mention

  • Likelihood is a function of parameters given fixed data; probability is a function of data given fixed parameters.
  • Likelihood is not a probability distribution over parameters—it does not integrate to 1 over the parameter space.
  • The likelihood function is proportional to the probability of the observed data under the model.
  • Maximum likelihood estimation (MLE) finds parameter values that maximize the likelihood.
  • In Bayesian inference, the likelihood is combined with a prior to form a posterior, which is a probability distribution over parameters.
  • The distinction is crucial for avoiding misinterpretation of likelihood values as probabilities of hypotheses.

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