Company interview guide

GSA Capital Machine Learning Interview Guide (2026)

For ML, AI and research roles: a 25-question online test, remote regression and probability rounds, a 10-minute return-prediction take-home graded on out-of-sample error, then onsite ML and portfolio rounds.

GSA Capital's machine learning interview loop has 5 reported stages (Online assessment, Round 1: regression, statistics, probability and live coding, Take-home data project and alignment call, Round 2: portfolio construction, ML and team depth, Final rounds: research fit with senior leadership); typical length: About 2 months for research roles in candidate reports; the firm says its process is thorough rather than short.

Roles covered: Research Scientist, Data Scientist, Applied Scientist and ML Engineer. The question bank holds 3 questions reported from GSA Capital interviews, 1 of them free to solve in your browser.

Try a free question

Interview process length
About 2 months for research roles in candidate reports; the firm says its process is thorough rather than short
Stages
5
Reports linked to bank questions
3
Bank report dates
Jul 2026–Jul 2026
Questions in bank
3 (1 free)
Last reviewed
October 2026

How the GSA Capital interview loop works

Research applicants usually start with a 60-minute, 25-question online test; developer applicants report a shorter coding test instead. The firm's own page then lists two remote technical interviews of up to an hour each on HackerRank or Zoom, a take-home exercise done in your own time alongside a call with the recruiting team about your goals and needs, a second round of up to about four hours of more team-specific interviews (onsite, remote if needed), and final conversations with senior or technical leadership. Compiled accounts put the take-home straight after the test; we follow the firm's order, which may vary by team. For research roles the take-home is a real modelling project: predict an instrument's 10-minute-ahead return from two price series, scored on out-of-sample error. There are no market-making or betting games in this loop; any such items are trader-track only. Detailed feedback is promised only to candidates who reach later stages.

The GSA Capital interview process, stage by stage

01Online assessment

Research track: Correlation One, 25 questions in 60 minutes, single-choice and typed-integer answers, with back and forward navigation.

Developer track: a HackerRank test with four easy coding problems and four multiple-choice questions.

The research version mixes four families. ML and statistics concepts: what fixes k-means local optima, what makes an overfitting linear classifier worse, what a one-variable residual plot cannot show, how R-squared and coefficient t-statistics relate, and which regression to run with 110 observations and 100 variables. Probability and linear algebra with typed answers, entered as the integer closest to 20p or as M + N for a reduced fraction M/N: P(3X + 4Y > 5), an eigenvalue sum, the nullity of a 5x5 matrix, real roots of a random quadratic, an 8-coin Bayes update. Brainteasers (nine dots, a molecule-collision invariant, knights and knaves). Judgment items on base rates, conjunctions, falsification, sunk cost and survivorship.

What it tests

  • Regularisation and model-choice instincts when features approach sample size
  • Fast numeric probability and linear algebra under an unusual answer format
  • Bayes updates and base-rate reasoning on short word problems
  • Spotting cognitive biases that also distort research conclusions
  • Pacing: about 2.4 minutes per question

How to prepare

  • Practise the answer conventions: p = 0.41 is entered as 8 (closest integer to 20p), and 1/3 is entered as 4 (M + N).
  • Answer the concept and judgment items first, then use the navigation to return to the heavier numeric ones.
  • On judgment pairs, pick the answer that respects base rates, judges a decision by its process and updates on disconfirming evidence.
  • Know that the trace equals the eigenvalue sum and that rank plus nullity equals the dimension; both save minutes.

Research-track format and content from compiled accounts of a 2026 sitting; one 2024 developer-role report describes the shorter coding test. The firm's process page does not list a test.

02Round 1: regression, statistics, probability and live coding

Two remote interviews of up to 60 minutes each, on HackerRank or Zoom; foundations in mathematics, statistics or computing, usually with a live coding part

The firm describes these as checks of technical foundations with limited time for your own questions. Compiled accounts of the research loop center them on regression theory and statistics argued cleanly: the ridge estimator and why X'X + lambda I is always invertible, the variance of a sum of equicorrelated variables and which correlations are impossible, the bus-stop inspection paradox, how many years until a strategy's cumulative return is positive with 97.5% probability (a CLT argument giving about 16), and how long a backtest must be to trust a Sharpe ratio. Probability classics and a coding task follow; a rolling median and a single-die versus average-of-100-dice game are reported.

What it tests

  • OLS and ridge from first principles, and when each estimate exists
  • Variance of sums, correlation bounds and length-biased sampling
  • CLT sizing for returns and Sharpe-ratio significance
  • Whether linear-regression assumptions hold for return data
  • Clean live code for a streaming statistic

How to prepare

  • Derive ridge as (X'X + lambda I)^-1 X'y and argue positive definiteness rather than quoting it.
  • For CLT sizing, write the condition as mu n >= 1.96 sigma sqrt(n) and solve for n; the Sharpe-length question has the same shape.
  • Know that the equicorrelation bound is rho >= -1/(n-1) and that expected interval length under inspection is E[X^2]/E[X].
  • Be ready to say why i.i.d. errors fail for returns and what you would do instead (robust errors, Bayesian shrinkage).

The firm's process page describes two remote interviews of up to an hour. A 2026 research-intern report had a probability round then coding; two older research reports cite ridge and regression-assumption questions.

03Take-home data project and alignment call

Done in your own time with your own tools; research roles get a return-prediction project filled into two fixed functions and scored on out-of-sample mean squared error.

A call with the recruiting team runs in parallel.

From 10-second-bar prices for two instruments, X and Y, predict Y's price change 10 minutes ahead using only data up to time t; the target column is already outlier-truncated. You fill a template in which one function fits parameters from a training file and another returns one forecast per row of a test file; extra files and libraries are allowed, and scoring uses held-out sets of similar but not identical shape. A past submission used about 33 engineered features (lags, rolling mean and volatility, spread and ratio, momentum and oscillator indicators, volatility ratios, z-score spikes), median-based winsorising, train-only standardising and a cross-validated lasso with time-series splits, and was still judged short of the bar. The firm's advice is to keep it simple and think about design, tests and documentation.

What it tests

  • Causal feature construction with no look-ahead
  • Time-ordered validation and train-only preprocessing reused at test time
  • Regularisation and feature selection at very low signal-to-noise
  • Clean, reproducible code inside a fixed interface
  • Judgment about where to spend limited effort

How to prepare

  • Check every feature for look-ahead; one leak inflates validation and ruins the held-out score.
  • Use expanding or rolling time splits, and fit winsorising and scaling on training data only, then save those parameters.
  • Drop the last 10 minutes of each day from training, since their forward targets are cut off.
  • Prefer a stable, well-regularised linear model you can explain over a complex one; document why each step is there.

The firm's process page describes a take-home in parallel with an alignment call. Task, data and scoring come from compiled accounts of a past research take-home; no first-hand ML-role report.

04Round 2: portfolio construction, ML and team depth

Up to about 4 hours of team-specific interviews, onsite (remote if needed), with researchers from the team you may join

Deeper, more focused interviews that also show what the role would look like. Compiled accounts describe a portfolio round: mean-variance weights given an alpha signal and a covariance matrix (objective, constraints, closed form), and how to estimate covariance when assets outnumber observations (sample, exponentially weighted, PCA factor models, graphical lasso). Recent loops add ML: why the top principal components can have near-zero R-squared against returns, what tree feature importance and SHAP values measure, and why a network is sensitive to its initialisation and how batch norm and dropout behave at test time. Expect follow-ups on your take-home.

What it tests

  • Mean-variance optimisation with and without a budget constraint
  • Covariance and precision-matrix estimation when N is large relative to T
  • Interpreting PCA, tree importance and SHAP on noisy financial features
  • Deep-learning basics: initialisation, normalisation and train versus test behaviour
  • Extending or criticising your own take-home model

How to prepare

  • Write the objective as w'mu - (gamma/2) w'Sigma w; unconstrained weights are proportional to Sigma^-1 mu, then add the Lagrange multiplier for weights summing to one.
  • Have a ranked menu of covariance estimators and say when the sample covariance breaks down.
  • Low R-squared on top principal components often means the signal is nonlinear or lives in small-variance directions; know that a one-feature SHAP value reduces to the regression beta times the centred feature.
  • Know what batch norm uses at inference and why dropout is switched off at test time.

Free reported questions for this stage

The firm's process page describes up to four hours of team-specific interviews. Content from compiled accounts of the research loop; two ML problems are tagged to the firm in 2026, but no ML-role report.

05Final rounds: research fit with senior leadership

Conversations with senior or technical leaders; mostly discussion of your work, vision and fit rather than set puzzles

The firm says these talks cover your expertise, longer-term growth and what you would bring, and that the decision is made here. Compiled accounts describe a deep dive on your take-home method and past projects: why each modelling choice, how you validated it and what you would do next. One research candidate reported no technical questions at all, only detailed questions about their own strategies; another was told personality fit is weighed alongside skill.

What it tests

  • Defending modelling and validation choices under questioning
  • Candour about limitations and what you would try with more time
  • Motivation for systematic research and long-term ambitions
  • How you work with colleagues and handle failure

How to prepare

  • Treat it as a research review of your take-home: features, horizon, validation scheme and why simpler beat complex.
  • Bring questions about culture, team structure, growth and failure cases; the firm suggests exactly these.
  • Be honest about weaknesses and show how you learn; the firm lists this as a tip.

The firm's process page describes final talks with senior leadership. One research-role report had only detailed questions on the candidate's own strategies; the rest is compiled accounts.

Questions and themes reported most often

  • Ridge closed form and why X'X + lambda I is always invertible (1 reports, round 1: regression, statistics, probability and live coding)
  • Is linear regression right for returns? i.i.d. errors, autocorrelation and a Bayesian alternative (1 reports, round 1: regression, statistics, probability and live coding)
  • Detailed questioning on your own strategies or research (1 reports, final rounds: research fit with senior leadership)
  • CLT sizing: years until a positive return and how long a backtest must be (round 1: regression, statistics, probability and live coding)
  • Variance of a sum of equicorrelated variables and impossible correlations (round 1: regression, statistics, probability and live coding)
  • Inspection paradox: expected wait at a bus stop (round 1: regression, statistics, probability and live coding)
  • Rolling median of a stream; estimator bias such as the German tank and n-1 (round 1: regression, statistics, probability and live coding)
  • Lasso or ridge when observations barely outnumber variables (online assessment)

What GSA Capital asks about

  • Regression, shrinkage & statistical inference28%
  • Applied ML on time-ordered data (take-home)24%
  • Probability & linear algebra18%
  • Portfolio construction & covariance estimation10%
  • Coding10%
  • Research discussion & judgment10%

GSA Capital interview prep checklist

  • Sat a timed 25-question set mixing ML concepts, typed-integer probability, brainteasers and bias items in 60 minutes
  • Can derive OLS and ridge and explain why X'X + lambda I is invertible
  • Can size a backtest or a years-to-positive-return question with the CLT, and explain the inspection paradox
  • Built a leak-free 10-minute-ahead forecaster with time-ordered validation and train-only winsorising and scaling
  • Can write mean-variance weights and compare sample, exponentially weighted, factor and graphical-lasso covariance estimates
  • Can explain low-R-squared principal components, SHAP versus betas, and batch norm and dropout at test time
  • Have one project story with numbers, validation, a failure and what you would do next

Tips

  • Make regression your first priority: OLS and ridge derivations, when estimates exist, and why the usual assumptions fail on returns.
  • Drill the online test's answer format (closest integer to 20p, M + N for fractions) on timed probability sets.
  • Rehearse CLT and Sharpe-ratio sizing arguments until you can write them in two lines.
  • Build one end-to-end short-horizon forecasting pipeline with causal features, time-ordered splits and train-only preprocessing before you see the take-home.
  • Keep the take-home simple and documented; the firm says it values sound judgment over over-engineering.
  • Learn the portfolio basics: mean-variance weights, and covariance estimators for many assets and short histories.
  • Connect ML answers back to linear models: SHAP to betas, PCA to factor models, regularisation to shrinkage.
  • Prepare to defend every take-home and project choice in front of senior researchers, including what failed.
  • Skip trader-track prep such as market-making games and speed arithmetic; none is described for this loop.
  • Use the alignment call for the deeper questions; the firm says time in Round 1 is limited.

See the rest of the GSA Capital loop

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

Unlock with Pro

Practice the GSA Capital loop

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

GSA Capital interview FAQ

How many interview rounds does GSA Capital have for ML roles?

Candidates report 5 stages: Online assessment, Round 1: regression, statistics, probability and live coding, Take-home data project and alignment call, Round 2: portfolio construction, ML and team depth and Final rounds: research fit with senior leadership. Research applicants usually start with a 60-minute, 25-question online test; developer applicants report a shorter coding test instead. The firm's own page then lists two remote technical interviews of up to an hour each on HackerRank or Zoom, a take-home exercise done in your own time alongside a call with the recruiting team about your goals and needs, a second round of up to about four hours of more team-specific interviews (onsite, remote if needed), and final conversations with senior or technical leadership. Compiled accounts put the take-home straight after the test; we follow the firm's order, which may vary by team. For research roles the take-home is a real modelling project: predict an instrument's 10-minute-ahead return from two price series, scored on out-of-sample error. There are no market-making or betting games in this loop; any such items are trader-track only. Detailed feedback is promised only to candidates who reach later stages.

How long does the GSA Capital ML interview process take?

Reported timelines run About 2 months for research roles in candidate reports; the firm says its process is thorough rather than short.

What does GSA Capital ask in ML interviews?

By reported question mix: Regression, shrinkage & statistical inference 28%; Applied ML on time-ordered data (take-home) 24%; Probability & linear algebra 18%; Portfolio construction & covariance estimation 10%; Coding 10%; Research discussion & judgment 10%. The most repeated themes: Ridge closed form and why X'X + lambda I is always invertible; Is linear regression right for returns? i.i.d. errors, autocorrelation and a Bayesian alternative; Detailed questioning on your own strategies or research.

How should I prepare for the GSA Capital ML interview?

Make regression your first priority: OLS and ridge derivations, when estimates exist, and why the usual assumptions fail on returns. Drill the online test's answer format (closest integer to 20p, M + N for fractions) on timed probability sets. Rehearse CLT and Sharpe-ratio sizing arguments until you can write them in two lines.

Are there free GSA Capital interview questions?

Yes. LeetML has 3 questions reported from GSA Capital interviews, and 1 of them are free to solve in the browser; they are listed by stage above.

Related interview guides

Official sources