RepostedAkuna Capital logo

Akuna Capital · Phone screen → Onsite

In process

Akuna Capital QR-ML Interviews: 11 Experiences (Part 4 of 6)

Quantitative researcher · UnknownReported Mar 2, 2020
Shared byjanebeingjane

Akuna Capital quant researcher and QR-ML interviews from 2019–2020: 11 reports on probability, ML theory and live coding in final rounds.

At a glance

  • Process (2019–2020): phone screens, sometimes on CodePair, then onsite or final rounds with math, ML and coding (H43–H53).
  • Asked in these reports: probability in 10 of 11, ML in 7, statistics in 6, coding in 5; examples include SVM regression and SGD (H46), Newton's method (H47) and a live statistics monitor (H49).
  • Difficulty: one QR-ML interview centered on proofs rather than code (H50).

Coverage: 11 Akuna Capital quantitative research reports from March 2019 to March 2020 (QR intern, QR, QR-ML). All are more than three years old, so they sit in History below.

In this series

  • Part 1: 2022–2025 screens, coding and the newest reports.
  • Part 2: 2021–2022: coding, simulation and probability.
  • Part 3: 2020: root finding, regression and shuffles.
  • Part 4 (current): 2019–2020: ML and final rounds.
  • Part 5: 2018–2019: probability and SVMs.
  • Part 6: 2017–2018: JQR, ML and projects.

Newer reports: Akuna quant technical interviews, 2025–2026.

What came up

Probability was nearly universal, and the ML-track roles added model theory and optimization. Coding showed up as live class design and matrix tasks. Every question stays in its entry.

Common questions

What ML questions did Akuna ask QR-ML candidates?

SVM regression and SGD (H46), a quadratic fit and Newton's method (H47), k-means assumptions and time-aware feature selection (H53).

Did Akuna quant final rounds include coding?

Several did: live class coding (H44), matrix rotation (H45) and a live statistics monitor (H49).

History (before 2024)

H43: Regression assumptions and a constrained walk

Report date: 2020-03-02 · Role: Quantitative Research Intern

Technical phone

  1. What assumptions does linear regression make?
  2. A probability question involved a constrained random walk. The transitions, boundaries and requested quantity were not retained.

H44: Concept questions followed by live class coding

Report date: 2020-02-22 · Role: Quantitative Researcher, ML

The first phone covered the resume, programming, ML and optimization. A second phone involved live OOP coding.

First phone

  1. The first interview included programming knowledge.
  2. ML knowledge was another part of that conversation.
  3. Optimization was discussed as well.

Second phone

  1. The second phone was a live OOP exercise, emphasizing program structure and coding fluency. The candidate did not identify the class or domain.

H45: Sphere sampling, matrix rotation and four short finals

Report date: 2020-01-05 · Role: Quantitative Research Intern

A referral led to a phone, CodePad and then a two-hour video onsite: four interviewers, 30 minutes each, with almost no gaps. The candidate struggled with the last options discussion and was rejected.

First phone

  1. Using one-dimensional random generators, how would you generate uniform points on the surface of the three-dimensional unit sphere? This was an oral method question. The candidate believed their first approach was wrong but still advanced.

CodePad

  1. Rotate a matrix by 90 degrees, first using extra storage and then in place. In Python, each row of the list-of-lists needed its own copy.

Final: coding

  1. A final-round coding task was described as system design. Its identity was forgotten, but the candidate remembered close attention to efficiency.

Final: mathematics

  1. What is the greatest number of distinct n-dimensional vectors that can have equal pairwise angles? Prove the result. The candidate used a contradiction involving covariance-matrix eigenvalues; vector-length and direction assumptions were not specified in their account.

Final: statistics

  1. The ML discussion included generalized least squares.
  2. Latent space was another topic.

Final: options

  1. There were also options questions, without the contracts or calculations.

H46: SVM regression, SGD and an unusually worded digit question

Report date: 2019-12-04 · Role: Quantitative Researcher, ML; seniority not stated

The live phone lasted half an hour. The invitation had arrived roughly two weeks after the earlier assessment.

Technical phone

  1. How does an SVM perform regression, and how does that compare with Ridge regression?
  2. Explain SGD: how it works, its problems, acceleration and what happens with nonconvex objectives.
  3. Another ML-theory question was not understood or retained.
  4. Break a stick randomly into three pieces. Find the expected length of the longest piece.
  5. What is the probability that the tens digit of a power of 2 is 1? The account says tens digit, not leading digit, and does not specify how the exponent is sampled.

H47: A quadratic fit, fair bits and Newton’s method

Report date: 2019-11-25 · Role: Quantitative Researcher, ML

The first full-time phone lasted 30 minutes and focused on mathematics and statistics. The candidate reported failing it.

Technical phone

  1. Can a quadratic curve fit three given points in a plane? The candidate said the issue was existence rather than a unique fit.
  2. Without a coin or other objects, how would you produce equally likely 0 and 1 outcomes? The candidate suggested taking the parity of a number from 1–10, but did not supply a random selection mechanism.
  3. Explain Newton's method, then why it can converge to zero. The exact function or root setup was not given.

H48: Random walks, CodePair and a later onsite

Report dates: 2019-10-11, 2019-11-25 · Role: Quantitative Researcher

The first phone covered the resume, mathematics/statistics and computer science. Another phone followed before the onsite.

The Boston onsite took place the week before the November 25 account. It included presentation, mathematics, coding and ML discussions and ended unsuccessfully.

First phone

  1. Discuss random walks in one, two and three dimensions: recurrence, transience and related calculations.
  2. Find the intersection of all the circles. The required output representation and treatment of boundaries were not recorded.
  3. Some optimization methods were also discussed.

Second phone / CodePair

  1. Remove duplicates.
  2. Print every power of 2 between a and b.

Onsite: presentation

  1. Present a project and answer questions about it.

Onsite: mathematics

  1. A t-test question involved a biased coin. The hypothesis and observations were not retained.
  2. Find the correlation between the number of heads in eight flips and the number in the first four of those flips.
  3. Use U(0,1) to sample a point on a circle.
  4. There were further probability, statistics and physics questions that the candidate could not reconstruct.

Onsite: coding existing classes

  1. Complete functionality in existing classes. The candidate clarified that this was coding; the domain and functions were not shared.

Onsite: optimization

  1. The ML conversation named Adam, Newton's method, SGD and other optimization methods.
  2. What does primal–dual mean?

Onsite: further probability

  1. A random walk moves one position left or right around a circle of N points. Starting at the origin, what is the distribution of the last point visited? The transition probabilities were not specified.

H49: Serial-number inference and a live statistics monitor

Report date: 2019-11-14 · Role: Quantitative Researcher, ML

There were two phones, with the second on CodePair. The candidate received the second invitation the day after the first phone and a rejection the day after coding.

First phone

  1. You see serial number N on one of X papers. What is your estimate of X? Then observe N/2 and update it. The prior and sampling rules were not given.

Second phone

  1. Remove duplicates, then discuss time and space complexity. The candidate later said there were no special requirements.
  2. Use a list comprehension to generate powers of 2 between 2 and 16.
  3. Generalize the list comprehension to powers of 2 between integer bounds x and y.
  4. Build a monitor with add-number, maximum, mode and mean operations.

H50: The QR-ML interview centered on mathematics and proofs

Report date: 2019-10-21 · Role: Quantitative Researcher, ML

The QR–ML conversation centered on statistics and mathematical reasoning.

QR technical phone

  1. The QR-ML phone included two U(0,1) variables and their sum X+Y; the requested quantity was only briefly described.
  2. Correlation and independence were also discussed.
  3. An unexpected proof question took the candidate by surprise. They later remembered using Jensen's inequality twice and another inequality, but not the proposition.

H51: A probability-and-mathematics phone

Report date: 2019-08-14 · Role: Quantitative Researcher, ML

Resume discussion led into a probability-and-mathematics phone. The candidate identified the process as new-grad hiring but had forgotten the individual questions.

Technical phone

  1. The candidate remembered a phone interview made up of probability and mathematics, but had forgotten the individual questions.

H52: Two phones and three research-focused final rounds

Report date: 2019-03-12 · Interview dates: 2019-03-01, 2019-03-06 and 2019-03-08 · Role: Quantitative Research Intern

The first phone was March 1, the second March 6, and three 30-minute final interviews were on March 8, 2019. The result was communicated March 11.

First phone

  1. Cut the unit circle at three random points. Find the expected length of the arc containing (0,1), and explain why it is greater than 2π/3.
  2. How would you sort an almost-sorted list? This account did not define a displacement bound k.

Second phone

  1. Remove duplicates while preserving their order.
  2. Implement a class with insert, get_mean, get_max and get_mode.

Final interview 1

  1. Project a fixed n-dimensional vector onto a random k-dimensional subspace. Find the expected projected vector, with fixed k and component-by-component expectation. The target was the vector. For clarification, the candidate used n=2, k=1, e=(0,1) and a line in direction (1,1), writing b=(√2/2,√2/2) and asking for (E[b1],E[b2]). That written example has a projection arithmetic error.

Final interview 2

  1. A final-round interviewer went through the candidate's research and its difficult parts.

Final interview 3

  1. Another interviewer also asked about research and resume projects.

H53: K-means assumptions and time-aware feature selection

Report date: 2019-03-08 · Role: Quantitative Researcher

The phone lasted about 30 minutes. The interviewer was a few minutes late after another interview, then covered model choices, feature selection and statistics.

Technical phone

  1. The resume led to K-means: its principle, assumptions, cautions, an application example and what needs to be done.
  2. Given features, time and labels, how would you select features? The candidate proposed a decision tree and information gain and was asked for other approaches, including time-related considerations.
  3. The candidate later recalled Bernoulli-distribution basics and derivations, without the individual numerical questions.
  4. There were other statistics and distribution questions that they did not remember.

Keep practicing

In this series

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.
0 comments

Comments

0

No comments yet — be the first to share what you know.