DRW Interview ProcessTier 2

The AI / ML screen is an auto-graded coding task: given an imbalanced binary dataset and a baseline model, write train and test functions that beat the baseline — only training-side rebalancing (resampling or class weights) reliably moves the score. A full practice replica of this task is in the pipeline and will appear here when it ships.

Topic weights tell you where to spend the week. Of the 87 DRW-tagged problems, coding (24), expectation (12), and probability (11) are the largest blocks, and that is the split the OA uses: the SDE track is pure coding, and the quant track is expectation and probability with a linear-algebra and calculus edge (matrix unknowns, an inverse-matrix derivative, a Gamma-function area). Finance (8), game theory (7), and market microstructure (6) are a quarter of the DRW bank, but they belong to the trading-intuition and onsite rounds, not the OA. The AI / ML task is a coding exercise, not a quiz: with an imbalanced binary dataset and a fixed baseline, only training-side rebalancing (resampling or class weights) reliably beats the baseline, so fix the data before tuning the model. The two reported DRW machine-learning interview problems, Live-Coding a CNN Binary Image Classifier in PyTorch and Designing an ML System for Stock Return Prediction, show how DRW frames ML past the screen.

Two consequences for how you prepare. First, the OA tests a narrower slice than the interviews: iterated market-making estimation drills, the pirate game, AKQ poker, and Penney's game are signature DRW interview material, but none of it appears in the OA, so park those until you have passed. Second, the OA is the only stage with no human in the loop. The numbers rotate per sitting, the SDE problems are auto-graded with no hidden-test feedback, and a clean approach that lands on the wrong number earns nothing. In the live rounds DRW's marking rubrics reward seeing the key insight fast and reasoning cleanly under follow-ups; in the OA only the final answer counts. Once you clear the screen, the 27 candidate-reported DRW interview questions (each with a worked solution) cover the video and onsite rounds, and the DRW quant salary page shows what the offer at the end of the funnel looks like.

Day 1, diagnose. Take the interactive DRW OA once, cold, under the 45-minute clock, and note which families cost the most time. Then open the three free DRW-reported problems: Gambler's Ruin on a Fair Coin, Conditional Marble Draw with Three Colors, and Futures Exchange Data Feed Delay. Day 2, probability. Work the conditional-probability and Bayes problems in the probability hub; the OA's Bayes items reduce to posterior odds equals prior odds times the likelihood ratio. Add two-state Markov chains: write the flow-balance equation and solve for the stationary share. Day 3, expectation. The expectation hub covers dice EV games, linearity of expectation, and the tower rule (condition on the first event, then average). Optimal-stopping and Sharpe-type items reduce to the mean and standard deviation of a two-point payoff. Day 4, linear algebra and moments. From the linear algebra hub: trace and determinant of 2x2 and 3x3 matrices, eigenvalues from the trace plus a null vector, the PSD condition for a correlation matrix (determinant at least zero), and the trace of an inverse as e2 over det. Then covariance algebra: Var(aX+bY), and E[XY] as the product of means plus the covariance. Day 5, speed and exact arithmetic. The OA is fill-in, so carry exact fractions (15/29, 8/9, and 3 sqrt(5)/4 are the reported shape of answers) and convert to decimals only if the field demands it. Spend 20 minutes on fraction arithmetic from the mental math hub, then re-time the Day 1 set. Day 6, SDE track (quant candidates: buffer day). Software candidates: the coding hub plus the three recurring DRW OA problems above. With no hidden-test feedback, write your own edge cases before submitting: empty input, a single element, all-equal values, maximum size. Quant candidates: the brain-teaser hub is for the later video round, not the OA, so redo the families you missed on Day 1. Day 7, full replica. Sit the 2027 QR OA replica question for question, then the rotating-pool timed set. Aim to finish with time left to re-check arithmetic; if a family still costs more than 10 minutes, repeat Day 3 or Day 4 before your real sitting.