Machine Learning Quant Interview Questions: Overfitting, Cross-Validation & Feature Leakage

What quant researcher interviews actually test about ML — and why textbook answers fail on financial data.

Updated July 3, 2026 · QuantVault

Machine learning questions in quant interviews are not Kaggle questions. Nobody at Two Sigma cares whether you can recite the XGBoost objective from memory. What they care about is whether you understand why techniques that work beautifully on ImageNet blow up on financial data — where the signal-to-noise ratio is brutally low, the data is non-stationary, and a single leaked feature can make a worthless model look like a money printer. This guide covers the questions that actually come up, the adaptations interviewers expect you to know, and one fully worked example of the single most common trap.

Why quant interviews ask ML questions

At systematic funds — Two Sigma, D.E. Shaw, Citadel, G-Research — the quant researcher's day job is building predictive models on noisy data. The interview is a proxy for that job. The failure mode they are screening against is specific: a researcher who trains a model, sees a great backtest, and ships something that loses money live. So ML rounds probe three things: do you understand generalization at a mechanical level (bias-variance, regularization), do you know how validation must change for time-ordered data, and can you smell leakage in a described pipeline. This is a researcher-track topic more than a trader-track one — if you're unsure which loop you're in, see our trader vs researcher breakdown.

The core toolkit you must have cold

Bias-variance decomposition. For squared loss, $\mathbb{E}[(y - \hat{f}(x))^2] = \sigma^2 + \text{Bias}^2[\hat{f}] + \text{Var}[\hat{f}]$. You should be able to derive this in two minutes and explain why financial data pushes you toward high-bias, low-variance models (ridge, shallow trees) — the noise term $\sigma^2$ dominates.
Regularization as a prior. Ridge is a Gaussian prior on coefficients, lasso a Laplace prior. Expect the follow-up: why does lasso zero out coefficients and ridge doesn't? (Corners of the $\ell_1$ ball.) This connects directly to regression questions, which often appear in the same round.
Cross-validation mechanics. Know k-fold, know its i.i.d. assumption, and know why that assumption is violated by overlapping return labels — which motivates purged and embargoed CV (López de Prado's formulation is the standard reference).
Trees vs linear models. When does a gradient-boosted tree beat ridge on returns data? Usually only with strong feature engineering and careful regularization, because trees happily memorize noise at low signal-to-noise.
Evaluation metrics. Accuracy is nearly useless; know information coefficient, out-of-sample $R^2$ (and why an OOS $R^2$ of 0.5% can be a great model), and Sharpe as a model metric.

Where textbook ML breaks on financial data

Textbook practice	What it breaks in finance	The fix interviewers want
Random train/test split	Future data trains the model that predicts the past	Strict temporal split; walk-forward validation
Standard k-fold CV	Overlapping labels leak across fold boundaries	Purged k-fold with an embargo period
Fit scaler/imputer on full dataset	Test-set statistics leak into training	Fit all preprocessing inside the training fold only
Tune until test metric peaks	Multiple testing turns the test set into a training set	One held-out final set; deflate for number of trials
Assume stationary distribution	Regimes shift; 2019 features mean something else in 2022	Rolling retraining; monitor live vs backtest decay

Worked example: the backtest overfitting question

“A researcher tests 100 independent signals, each with zero true predictive power, on one year of daily data. What in-sample Sharpe ratio do you expect the best one to have?” This exact structure appears in QR interviews because it tests probability, statistics, and research hygiene at once.

With $T$ years of data, the annualized Sharpe estimate of a zero-skill strategy is approximately Gaussian with standard error $\sqrt{1/T}$. For $T=1$ year, each of the 100 measured Sharpes is roughly $\hat{S}_i \sim \mathcal{N}(0, 1)$. The question is then the expected maximum of $N=100$ i.i.d. standard normals. The standard asymptotic:

$$\mathbb{E}[\max_i Z_i] \approx \sqrt{2\ln N} - \frac{\ln\ln N + \ln 4\pi}{2\sqrt{2\ln N}}$$

Plugging in $N = 100$: $\sqrt{2\ln 100} = \sqrt{9.21} \approx 3.03$, and the correction term is $\frac{1.53 + 2.53}{2(3.03)} \approx 0.67$, giving

$$\mathbb{E}[\max_i \hat{S}_i] \approx 3.03 - 0.67 \approx 2.4$$

So the best of 100 worthless signals shows an in-sample Sharpe around 2.4 — better than most real production strategies. The interviewer wants the number and the moral: selection bias grows like $\sqrt{2\ln N}$, so every strategy you evaluate raises the bar a discovery must clear. Strong candidates mention the deflated Sharpe ratio or a Bonferroni-style correction as the practical response. The underlying math is pure expectation and order statistics, so it doubles as a probability question.

The traps that fail candidates

Answering leakage questions generically. “Don't use future data” is table stakes. Name the subtle versions: survivorship-biased universes, restated fundamentals, features built with full-sample normalization, labels computed over horizons that overlap the test fold.
Reaching for deep learning. Saying you'd throw a neural net at daily returns signals inexperience. With $\sim$2,500 daily observations and near-zero signal, regularized linear models are the honest default; say why.
Ignoring stationarity. Any answer about validation on financial data should acknowledge regime change — this is where ML rounds bleed into time series questions on stationarity and autocorrelation.
Not knowing the stats underneath. Overfitting questions are hypothesis-testing questions in disguise. If your p-value and multiple-testing fundamentals are shaky, patch them with the statistics bank before touching ML prep.

Practice the real questions

Reading about leakage is not the same as catching it under pressure. Work through our machine learning interview question bank with fully worked solutions, shore up the foundations in the statistics bank, and browse the full problem library — around 400 of the 2,800+ problems are free.

Frequently asked questions

What machine learning topics come up in quant interviews?

The core set is bias-variance tradeoff, regularization (ridge vs lasso), cross-validation and its failure modes on time-ordered data, feature leakage, tree ensembles vs linear models, and evaluation metrics like information coefficient and out-of-sample R². Firms like Two Sigma, D.E. Shaw, and G-Research focus less on model zoo trivia and more on whether you understand generalization on low signal-to-noise financial data.

Why is standard k-fold cross-validation wrong for financial data?

K-fold assumes observations are i.i.d., but financial labels are usually computed over forward return horizons that overlap fold boundaries, so information leaks from test folds into training folds. The accepted fix is purged k-fold with an embargo: drop training samples whose label windows overlap the test fold and add a buffer period after it. Interviewers expect you to explain both the failure and the fix.

What is feature leakage and how do interviewers test it?

Feature leakage is any pathway by which information unavailable at prediction time enters training, such as normalizing features using full-sample statistics, using restated fundamentals, or building a universe from stocks that survived to the present. Interviewers typically describe a plausible research pipeline and ask you to find the flaw. Strong candidates name specific, subtle leaks rather than just saying "don't use future data."

Do I need deep learning for quant researcher interviews?

Rarely. Most desks want you to explain why heavily parameterized models are a poor default for daily-frequency financial data, where a few thousand noisy observations cannot support millions of parameters. Knowing when regularized linear models or shallow gradient-boosted trees are the honest choice signals more experience than reciting transformer architecture details.

Practice the real thing

QuantVault has 2,800+ quant interview problems with full solutions, intuition, and hints, firm-by-firm interview funnels, and an auto-graded coding judge. Start free.

Browse problems Firm interview guides