10 Point72 Quantitative Researcher (New Grad) Interview Questions (2026)

Outline

STAR. Pick a real project — thesis, internship, hackathon. Walk through: problem definition, methodology, the hardest technical or interpersonal challenge, how you addressed it, the outcome. Point72 values researchers who can articulate the hard parts honestly.

Outline

Specific reasons: multi-strategy with both discretionary and systematic sides, Cubist Systematic offers research-heavy track, the Point72 Academy program offers structured new-grad training. Avoid the 'high pay' lead. Reference Point72's growth story and the founder's role in shaping the firm.

Outline

Sliding window with hash map of last-seen indices. Expand right; if current character was seen and its index is within window, contract from left. Track max length. O(n) time, O(min(n,k)) space. Code cleanly.

Outline

Single pass. Track min_seen. For each price, compute price - min_seen and update max_profit. O(n) time, O(1) space. Walk through example. Edge cases: monotonically decreasing (return 0), single element (return 0).

Outline

No. A trivial classifier predicting always-negative achieves 99% accuracy. Use better metrics: precision, recall, F1, AUC-ROC. For imbalanced data, accuracy is misleading. Discuss the implications: under-sampling, over-sampling (SMOTE), class weights, threshold tuning. State the diagnosis cleanly.

Outline

Bagging: train multiple models on bootstrap samples in parallel, average predictions. Reduces variance. Example: Random Forest. Boosting: train models sequentially, each focusing on errors of the previous. Reduces bias. Example: AdaBoost, Gradient Boosting, XGBoost. Discuss tradeoffs: bagging is robust to overfitting (parallel); boosting can overfit but achieves lower bias.

Outline

Mean = (a+b)/2 by symmetry. Variance = (b-a)^2 / 12. Derive: E[X] = ∫_a^b x · 1/(b-a) dx = (a+b)/2. E[X^2] = ∫_a^b x^2 · 1/(b-a) dx = (a^2+ab+b^2)/3. Var = E[X^2] - E[X]^2 = (b-a)^2/12. State the derivation if asked.

Outline

Cholesky decomposition. Let L be lower triangular with L·L^T = Σ. Generate z = vector of independent standard normals. Then x = μ + L·z has the desired multivariate normal distribution. State the algorithm. Mention that if Σ is not positive-definite (numerical issues), use eigenvalue decomposition or nearest-PSD projection.

Outline

CLT: sum of n i.i.d. random variables with finite mean μ and finite variance σ^2 has distribution approaching N(nμ, nσ^2) as n grows. Equivalently, sample mean is approximately N(μ, σ^2/n). Apply when: large n, finite variance, near-iid samples. Fails when: heavy-tailed distributions (infinite variance), strong dependence. Discuss why it matters in finance (return distributions, sample statistics).

Outline

Overfitting: model fits training data noise rather than signal. Detect: training loss low, test/validation loss high. Symptoms: huge gap between train and val curves. Cures: regularization (L1/L2), dropout, early stopping, more data, simpler model, cross-validation. State the diagnostic framework: always evaluate on held-out data.