Arrowstreet Capital Interview Questions: OA, CodePair & the Written Exam

Arrowstreet Capital is a Boston-based systematic global equity manager founded in 1999 by Bruce Clarke, Peter Rathjens, and Harvard economist John Y. Campbell, who co-directed research at the firm through 2023. It manages roughly $290B+ for institutional clients as of early 2026 and runs long-only and long/short equity strategies built on quantitative models. That academic-economist DNA shows up directly in the interview: this is one of the most econometrics-forward hiring processes in quant finance — closer to a PhD qualifying exam in applied statistics than to an HFT speed gauntlet. What follows is the funnel as candidates who went through the 2025–26 cycle describe it, anonymised and cross-checked across several accounts; the stage-by-stage version with practice sets lives on the Arrowstreet interview process page.

The reported interview process

Pooling candidate reports from the 2025–26 cycle with earlier Glassdoor accounts gives a consistent four-stage funnel for quantitative researcher roles. Exact ordering varies by year and seniority — some accounts add a recruiter call and video technical rounds between the coding screen and the superday — but the core sequence is:

Glassdoor's quantitative researcher reviews rate the difficulty around 3.2 out of 5 — demanding but not extreme. Reports remain thinner than for Citadel or Jane Street, so treat any single account as one data point rather than a fixed script. Compensation chatter from the same candidates — a high base salary and a very small headcount — is hearsay and unverified. If you want a baseline for how these screens work generally, start with our guide to quant online assessments.

Stage 1: the online assessment

The OA is finance math plus regression judgment. The 2027-cycle sitting a candidate captured runs on a HackerRank-style platform: numeric “complete the blanks” questions on sums of normals, a likelihood-ratio test between two decks, factor-model and tangency-portfolio Sharpe ratios, beta to an equal-weighted portfolio, a constrained cone volume and the KKT system of an execution problem, alongside “Pick ONE” / “Pick ONE or MORE” MCQs on dummy-variable specification, adding quadratic terms, calendar-month train/test splits, seasonal time-series models and what a t-stat of 4 implies. Rounding instructions are explicit (“rounded down to the nearest integer”, “in %”) and the countdown allows about 4 minutes a question. Optimization shows up more here than at most funds — one earlier report specifically mentions linear programming. Our optimization question bank and linear algebra problems map onto this directly, and the timed replica runs all 13 captured questions with verified answers.

Stage 2: the HackerRank CodePair screen

Three candidates describe the same session: a live CodePair link, one researcher on the call, and conceptual questions interleaved with coding. The concept questions are not trivia — they are “what does this print?” items that separate people who use pandas from people who understand it:

• Python indexing and slicing. Negative indices, slice bounds past the end of the list, step semantics like a[-3:1:-1], and which operations copy versus alias versus view (a list slice copies; b = a aliases; a NumPy basic slice is a view). Work it through in Python Slicing Semantics: Bounds, Steps, Aliases and Views.
• @staticmethod vs @classmethod. What each receives, calling through an instance versus the class, and — the part that trips people — what cls is bound to when an inherited classmethod is called through a subclass. Our Base/Child “what does this print” problem covers every reported angle.
• pandas. merge with duplicated keys (row counts multiply per key, and the notional double-counts — worked example), groupby agg versus transform (shapes, alignment, and the all-NaN column trap), rolling with min_periods and center=True (including which variant leaks the future), and resample of irregular timestamps with closed and label (boundary ticks and empty bins). Every output in those problems was produced by actually running the snippet.
• Two LeetCode easy/medium problems. The one candidates remember is counting the words in a list with exactly k distinct letters — a one-pass hash-set problem with a “can you do it without a set?” follow-up. It is on our judge as Count Words With Exactly k Distinct Letters, with a medium companion in the same spirit, Longest Anagram-Free Run of Words (canonicalise, then slide a window).

The tone is conversational. State the complexity and the edge cases (empty list, k = 0, repeated letters) before being asked, write the brute force first, and narrate while you type.

Stage 3: what the written exam covers

The superday opens with a paper-based exam — the stage nobody prepares for properly. It is mainly statistics and asset pricing, and it asks for closed-form derivations rather than computation. Pooling four candidates' recollections, the topics were:

• Mean-variance portfolio theory. The tangency portfolio and why its weights are proportional to $\Sigma^{-1}(\mu - r_f\mathbf{1})$; with numbers in Tangency Portfolio for Two Assets (two equal-Sharpe assets that still get 2:1 weights).
• Risk parity / equal risk contribution. Marginal risk contributions, the ERC condition $w_i(\Sigma w)_i = w_j(\Sigma w)_j$, the two-asset closed form $w_1\sigma_1 = w_2\sigma_2$, and the proof that inverse-volatility weights are ERC only when the correlation matrix has equal row sums — Risk Parity: Two-Asset Closed Form and the ERC Condition.
• Regress X on Y versus Y on X. The two slopes are $r\,s_y/s_x$ and $r\,s_x/s_y$; their product is $R^2$; the fitted lines are not inverses of each other and coincide only when $|r| = 1$. Numbers in Regressing Y on X Versus X on Y.
• Autocorrelation and effective degrees of freedom. For AR(1) residuals with autocorrelation $\rho$, the effective sample size is $N_{\text{eff}} \approx N\,\frac{1-\rho}{1+\rho}$; a t-statistic of 1.73 on 300 days becomes 1.0 at $\rho = 0.5$. Derivation and the Newey–West connection in Effective Sample Size Under AR(1) Autocorrelation.
• Log utility. Expected utility of a gamble, the certainty equivalent (the geometric mean of outcomes), the maximum price you would pay (which is not the certainty-equivalent gain), and the growth-optimal sizing version — Log Utility: Certainty Equivalent, Maximum Price and Growth-Optimal Sizing.
• One asset-pricing problem-set item. Pricing a payoff under a given utility, SDF or CAPM style — the kind of question a first-year PhD asset-pricing course sets. Our finance question bank covers the standard forms.

Writing a derivation out longhand under time pressure is a different skill from talking through it, and it is the skill this exam isolates. The interviews that follow tend to pick up whatever you struggled with on paper.

Stage 4: the interview loop

After the exam, a chain of back-to-back conversations with several researchers, lunch in the middle. The through-line in nearly every report is linear regression and econometrics: OLS assumptions, interpreting coefficients, regression pitfalls, and time-series concepts, alongside general probability and statistics and a deep dive on your own research. Job postings for the PhD researcher track list the expected toolkit explicitly: probability, statistics, linear regression, time-series analysis, linear algebra, calculus, optimization, and portfolio theory, with fluency in Python, R, MATLAB, or similar. Coding in the loop is reported as light — the CodePair screen has already done that filtering.

A representative regression question

We won't attribute specific questions to Arrowstreet, but omitted-variable bias is exactly the kind of problem the reported loop targets. Here is a representative practice example.

You regress returns $y$ on a single signal $x_1$, but the true model is $y = \beta_1 x_1 + \beta_2 x_2 + \varepsilon$. What does your estimated coefficient converge to?

The short regression estimator picks up the effect of the omitted variable through its correlation with the included one:

$$\hat{\beta}_1 \xrightarrow{p} \beta_1 + \beta_2 \, \frac{\mathrm{Cov}(x_1, x_2)}{\mathrm{Var}(x_1)}$$

The follow-ups write themselves: When is the bias zero? What sign is it if both signals are positively correlated and both predict returns positively? Why does this matter when you add a new factor to an existing alpha model? If you can walk that chain fluently — formula, intuition, portfolio-research consequence — you're at the level this interview expects. Drill more of these in our regression question bank.

How to prepare, stage by stage

1. OA: portfolio algebra and rounding rules. Mean and variance of linear combinations, tangency weights, likelihood ratios, Lagrange conditions; read the rounding instruction twice. Run the timed replica once at pace and once untimed.
2. CodePair: make pandas outputs automatic. Build a six-row toy frame and predict — then run — merge with duplicated keys, agg versus transform, rolling(3, min_periods=1), rolling(3, center=True), and resample('1min', closed='right', label='right'). Do the same for slicing expressions and a classmethod/staticmethod hierarchy. Then solve two easy/medium string or list problems out loud, complexity and edge cases first.
3. Written exam: derive on paper, timed. Five derivations, each from a definition: tangency weights from maximising the Sharpe ratio; ERC weights from $w_i(\Sigma w)_i$; both regression slopes from $r$ and the standard deviations; $N_{\text{eff}}$ from the double sum of $\rho^{|s-t|}$; certainty equivalent and price from $E[\ln W]$. No notebook, no autocomplete — longhand until the algebra runs without pausing.
4. Loop: rebuild OLS from first principles. Derive the estimator, know the Gauss–Markov assumptions and what breaks when each fails (heteroskedasticity, autocorrelation, multicollinearity), argue the sign of omitted-variable bias from the correlations, and be able to explain $R^2$ traps out loud. Our statistics bank, probability questions and time series bank cover the reported range.
5. One research story in depth. A project you can take from data to signal to portfolio, including what you would do differently. This is an academic-leaning firm; depth beats breadth.

Culturally, Arrowstreet interviews are consistently described as courteous and low-hostility — "nice to talk to, no harsh or weird questions," as one Glassdoor reviewer put it. The bar is depth of understanding, not performance under abuse. Candidates comparing systematic managers should also look at how AQR's process runs — the two firms draw from a similar academic-quant candidate pool.

Ready to drill? Work through the thirteen Arrowstreet CodePair and written-exam problems linked above (they are collected on the Arrowstreet interview process page), then the regression question bank and statistics problems that map to the interview loop, and browse the full QuantVault problem bank — 2,800+ questions with worked solutions, around 400 of them free.