Company interview guide

PDT Partners Machine Learning Interview Guide (2026)

For ML, AI and research roles at a research-heavy systematic fund: project deep dives, clean coding, and statistics and regression from first principles.

PDT Partners's machine learning interview loop has 6 reported stages (Application & phone screen, Technical screen: project deep dive & coding, Take-home assignment, Probability & statistics, Regression, ML & time series, Onsite: open-ended mini-projects & final test); typical length: Several weeks or more; compiled accounts describe slow feedback, sometimes weeks to months between steps.

Roles covered: Applied Scientist, Research Scientist and ML Engineer. The question bank holds 12 questions reported from PDT Partners interviews, 5 of them free to solve in your browser.

Try a free question

Interview process length
Several weeks or more; compiled accounts describe slow feedback, sometimes weeks to months between steps
Stages
6
Reports linked to bank questions
12
Bank report dates
Jul 2026–Aug 2026
Questions in bank
12 (5 free)
Last reviewed
October 2026

How the PDT Partners interview loop works

A written application screened for research quality, a short phone screen, a technical screen with coding, a take-home reported at about 48-72 hours, then 45-60 minute rounds on probability and statistics and on regression, ML and time series, ending in a several-hour onsite of open-ended problems and, in some accounts, a long written test. Engineering candidates lean toward clean C++ and data-structure coding; researchers toward probability, statistics and modeling; ML candidates should prepare both. No market-making game or mental-math speed test is reported for any track. The outline rests on very few first-hand reports.

The PDT Partners interview process, stage by stage

01Application & phone screen

Written application with research questions, then a short phone call on background and fit

The application is itself a filter. The Applied ML Scientist, quant researcher and research intern postings ask for a ranked publication list, the one paper that best shows your research style, and short descriptions of a data analysis, a coding and a modeling project; the intern form also asks for code you wrote or a complex system you built. Compiled accounts then describe a brief phone call about background and interest, sometimes with a quick technical warm-up. The firm says finance knowledge is not required.

What it tests

  • Depth and quality of your research record and project write-ups
  • Why the firm and why a move into systematic investing
  • Fit with a track: applied ML research, research engineering or quant research
  • Logistics: other interviews and deadlines, relocation to New York, hybrid office days

How to prepare

  • Write each project description like a paper abstract; later rounds pick them apart.
  • Choose a representative paper you can defend line by line, not just the most cited one.
  • Have a crisp answer to why systematic investing; the forms ask it directly.

The firm's ML, research and intern postings describe the written application; the phone screen comes from compiled accounts of the research loop, not from ML-role reports.

02Technical screen: project deep dive & coding

Live video round billed as coding; the ML-engineer report had about 15 min on a past project before the algorithm problem

The one recent ML-engineer report describes a screen that opened with an implementation-level deep dive on an older production model from the resume, a small LSTM serving inference at scale, moving from infrastructure to latency to the modeling choice and ending with whether the candidate would choose it again today. The rest was an algorithm problem, listing the divisors of an integer, with a follow-up on beating square-root time. Compiled accounts of the engineering track list clean data-structure problems: merging sorted streams or iterators, a reference-counted smart pointer in C++, a stack with a bulk-increment operation done lazily, and adding numbers stored as digit lists.

What it tests

  • Implementation detail behind a production model you list: serving path, latency budget, batching
  • Justifying an architecture choice then and now, for example a recurrent model versus a transformer
  • Correct, readable code with the right data structure
  • Complexity analysis and whether a faster bound is possible
  • For engineering-leaning tracks, C++ fundamentals: ownership, RAII, reference counting

How to prepare

  • Prepare every resume project at the level of numbers: request rate, latency budget, model size, what you measured.
  • Expect a project deep dive even when the round is billed as coding.
  • After a correct solution, state the complexity and say plainly whether it can be beaten, and why.
  • If you apply as a research engineer, rehearse a smart pointer and a merge of sorted iterators in C++.

Free reported questions for this stage

One 2026 ML-engineer report (project deep dive, then a divisors problem); the C++ and data-structure items come from compiled accounts of the engineering track. Few reports.

03Take-home assignment

Multi-day take-home, reported at about 48-72 hours, mixing coding, data work and some markets and systems questions

Compiled accounts place a take-home between the screen and the technical rounds, with roughly two to three days to finish and content spanning coding, data analysis and some questions on markets and systems. The platform and exact problem pool are not public. Engineering accounts stress that it is graded like production code, not only on whether the answer is right.

What it tests

  • Turning a loosely specified data or modeling task into a structured analysis
  • Production-quality code: readable, tested and reproducible
  • Clear write-up of assumptions, results and limitations
  • Basic reasoning about markets and systems, without finance expertise

How to prepare

  • Treat it like code a colleague will maintain, with a short README covering how to run it.
  • Spend the extra time on validation and on what could break the result, not on a marginally better score.
  • Read up on basic market mechanics so a markets question does not stall you.

Only compiled accounts of the firm's research and engineering loops describe the multi-day take-home; no first-hand ML-role report mentions one.

04Probability & statistics

45-60 min over video with a researcher

A round on probability, expectation, combinatorics and estimator reasoning. Compiled question sets show Bayesian updating (is a coin that showed five heads in a row fair), coin pattern races, coupon-collector and gambler's-ruin style stopping problems, integrals such as the Gaussian, and statistics questions on bias and variance, combining two estimates, order statistics, inverse-transform sampling and bounded-memory distinct counting.

What it tests

  • Bayesian updating with a stated prior and a clear posterior
  • Expectation via conditioning or recursion, including stopping and first-passage problems
  • Clean combinatorial counting
  • Estimator properties: bias, variance and how to merge two estimates
  • Sampling and streaming estimation: inverse transform, order statistics, distinct-count sketches

How to prepare

  • Set up the conditioning or recursion out loud before reaching for a known result.
  • State the prior explicitly in any fair-coin question and show how the answer moves with it.
  • For each estimator, give its bias and variance and when you would prefer another.

Free reported questions for this stage

Format and content from compiled accounts of the firm's quant research loop and its reported question set; no first-hand ML-role report covers this round.

05Regression, ML & time series

45-60 min research discussion, whiteboard style

The best-documented block of the research loop. Expect to derive least squares, give ridge and lasso closed forms, explain what collinearity does to coefficient variance, read the geometry of fitted values, and diagnose residuals for heteroskedasticity and autocorrelation. The time-series strand covers EWMA variance and half-life, stationarity tests and block bootstrap; the ML strand covers interpreting logistic-regression coefficients, inverting or shrinking a high-dimensional covariance matrix, PCA with missing data and comparing forecasting models for daily returns.

What it tests

  • Deriving OLS, ridge and lasso from the objective and explaining each term
  • Collinearity, variance inflation and how shrinkage trades bias for variance
  • Residual diagnostics: heteroskedasticity, autocorrelation, stationarity
  • Time-ordered validation: walk-forward splits, purging and embargo, leakage
  • Linear algebra behind regression: projections, eigenvalues, covariance inversion

How to prepare

  • Practise the regression derivations on a whiteboard while explaining each step aloud.
  • Connect every formula to a picture: projection onto the column space, eigenvalues shrunk by ridge.
  • Say why random k-fold leaks on time series and describe walk-forward with a purge and embargo.
  • Know when a high-dimensional covariance needs shrinkage before you invert it.

Free reported questions for this stage

Format and content from compiled accounts of the firm's quant research loop and its reported regression and ML questions; no first-hand ML-role report covers this round yet.

06Onsite: open-ended mini-projects & final test

Several hours with researchers; open-ended problems and, in some accounts, a long standard test at the end

Compiled accounts describe final rounds built from very open-ended problems, each becoming a small research project under time pressure, often with researchers whose background is close to yours who also probe your own work. Typical shapes are finding a real signal in deliberately noisy data, calibrating it and judging whether an improvement is statistically real. Some accounts end with a long, more standard test; the hardest variants of every earlier topic can appear.

What it tests

  • Structuring an ambiguous modeling problem: target, features, validation, failure modes
  • Finding signal in noisy data and judging whether an improvement is real
  • Defending your own research choices to a specialist
  • Stamina and composure across several hours and hard follow-ups

How to prepare

  • Practise talking through an open-ended modeling problem for 30 minutes without a fixed answer.
  • Before claiming a gain, say how many variants you tried and how you would correct for that.
  • Know your representative paper well enough to say what you would do differently now.

Format from compiled accounts of the firm's quant research onsite; the postings list whiteboard collaboration as a skill but describe no rounds. No ML-role report covers it.

Questions and themes reported most often

  • Deep dive on a past production model: serving, latency and why that architecture (1 reports, technical screen: project deep dive & coding)
  • List all divisors of an integer, then ask whether it can beat square-root time (1 reports, technical screen: project deep dive & coding)
  • Add two large numbers stored as digit arrays (1 reports, technical screen: project deep dive & coding)
  • Merge sorted streams cleanly through a generic iterator interface (technical screen: project deep dive & coding)
  • Is a coin that showed a run of identical flips fair? Bayesian update (probability & statistics)
  • Coin pattern race: which sequence appears first (probability & statistics)
  • Derive the least-squares estimator, then ridge and lasso (regression, ml & time series)
  • What severe collinearity does to coefficient variance (regression, ml & time series)

What PDT Partners asks about

  • Statistics & regression28%
  • ML modeling & time series22%
  • Coding & algorithms18%
  • Probability & expectation16%
  • Research & project deep dive16%

PDT Partners interview prep checklist

  • Written the application project descriptions and picked a representative paper you can defend
  • Rehearsed a numbers-level deep dive on one production or research project
  • Derived OLS, ridge and lasso and explained collinearity on a whiteboard
  • Worked Bayesian, pattern-race and stopping-time probability problems aloud
  • Explained walk-forward validation, purging and multiple testing on a noisy signal
  • Solved clean algorithm problems with a complexity follow-up, in C++ if on the engineering side
  • Talked through an open-ended modeling problem on noisy data for 30 minutes
  • Asked the recruiter about the take-home and the length of each round

Tips

  • Treat the written application as the first interview: the project descriptions and representative paper shape later questions.
  • Rehearse a numbers-level deep dive on each resume project and be ready to say what you would change today.
  • Derive OLS, ridge and lasso on paper until you can do it while talking.
  • Review collinearity, shrinkage, residual diagnostics and the geometry of least squares.
  • Drill time-ordered validation: walk-forward splits, purging, embargo and leak-free preprocessing.
  • Keep core algorithms sharp and always finish with complexity and a word on whether it can be improved.
  • For research-engineer roles, practise clean C++: ownership, RAII and a reference-counted pointer.
  • Budget a full working weekend for the take-home and grade yourself on code quality, not just results.
  • Skip market-making games and mental-math drills; the firm says finance knowledge is not required.

See the rest of the PDT Partners loop

The later stages, the questions reported for them and how to prepare are in Pro.

Unlock with Pro

Practice the PDT Partners loop

Work the stages in order: every reported question for this company, grouped by round, with solutions in the app.

PDT Partners interview FAQ

How many interview rounds does PDT Partners have for ML roles?

Candidates report 6 stages: Application & phone screen, Technical screen: project deep dive & coding, Take-home assignment, Probability & statistics, Regression, ML & time series and Onsite: open-ended mini-projects & final test. A written application screened for research quality, a short phone screen, a technical screen with coding, a take-home reported at about 48-72 hours, then 45-60 minute rounds on probability and statistics and on regression, ML and time series, ending in a several-hour onsite of open-ended problems and, in some accounts, a long written test. Engineering candidates lean toward clean C++ and data-structure coding; researchers toward probability, statistics and modeling; ML candidates should prepare both. No market-making game or mental-math speed test is reported for any track. The outline rests on very few first-hand reports.

How long does the PDT Partners ML interview process take?

Reported timelines run Several weeks or more; compiled accounts describe slow feedback, sometimes weeks to months between steps.

What does PDT Partners ask in ML interviews?

By reported question mix: Statistics & regression 28%; ML modeling & time series 22%; Coding & algorithms 18%; Probability & expectation 16%; Research & project deep dive 16%. The most repeated themes: Deep dive on a past production model: serving, latency and why that architecture; List all divisors of an integer, then ask whether it can beat square-root time; Add two large numbers stored as digit arrays.

How should I prepare for the PDT Partners ML interview?

Treat the written application as the first interview: the project descriptions and representative paper shape later questions. Rehearse a numbers-level deep dive on each resume project and be ready to say what you would change today. Derive OLS, ridge and lasso on paper until you can do it while talking.

Are there free PDT Partners interview questions?

Yes. LeetML has 12 questions reported from PDT Partners interviews, and 5 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources