Two Sigma Quant Research Intern Onsite Interview Experience and Insights

This post was last edited by Anonymous on 2025-10-09 11:19 OA: This has been posted many times on the forum. One question was linear interpolation, another was pandas data processing + scikit-learn li

This post was last edited by Anonymous on 2025-10-09 11:19 OA: This has been posted many times on the forum. One question was linear interpolation, another was pandas data processing + scikit-learn linear regression, and the third was a handwritten batch univariate linear regression. Just optimize using the OLS formula. The first round was Data Analysis. Build a prediction model to predict housing prices given historical data. You need to discuss how to define features and outcomes, how to choose the model, and how to choose the evaluation metric. The interviewer will keep asking follow-up questions, including how to handle missing values, non-linear relationships, feature selection, data drift, variance drift, data leakage, why this particular non-linear model may work, how to handle shocks like COVID, etc. It ended when they said there wasn't enough time. I previously attended CMU (17-445), did research and wrote similar systems, and reviewed Alex Xu's book before the exam and mocked various similar problems with AI. I luckily passed this round. The second round was Domain Interview + Coding. The Domain Interview was terrible. It was supposed to be about CS. The interview focused on PhD-related knowledge, but the interviewer suddenly switched to math halfway through, which I hadn't prepared for. I encountered a Markov chain problem, and while I offered an approximation, the interviewer insisted on an analytical solution. I then wrote a cubic equation and had to solve it on the spot, but I ran out of time. In coding, I spent an hour writing two problems: one a variation of problem 694 (which only required handling corner cases), and the other a variation of problem 3387, which asked for the path with no repeating nodes and the maximum product of edge weights on a directed complete graph. This problem was NP-hard; I spent 10 minutes writing a brute-force approach, then got stuck on an incorrect bitmask dynamics solution, finishing within five minutes. Just as I finished, I realized that since all weights were greater than 0, taking the logarithm of the weights would fix the initial dynamics solution. I received a rejection letter the day after the interview.