Morgan Stanley · Primly Community

Morgan Stanley machine learning engineer interview, what the rounds look like and what they're really testing

ml_mike · 5 replies

Interviewed for an ML engineer role at Morgan Stanley earlier this year. The team was on the risk modeling side, not the quant trading desk, so your experience may vary if you're targeting a different group. But here's what I saw.

Screening

First call was with a tech recruiter. Nothing interesting. They confirmed I had Python and ML experience and asked about my current role.

Second call was with the hiring manager. This was actually substantive. She asked about specific models I'd deployed, how I think about production reliability for ML systems, and what I know about monitoring model drift. This wasn't a warmup, it was a real filter. I'd say prep for this like a real technical conversation, not a soft screen.

Onsite (virtual, 3 rounds)

ML fundamentals: core ML concepts, bias-variance tradeoff, regularization, how you'd choose between models for a specific classification problem. They asked me to walk through how I'd approach training a credit default prediction model. They didn't need me to know the specific MS models, they wanted to see my reasoning process.

Coding: Python, LeetCode-medium range, data manipulation. One problem was essentially pandas/numpy style (not algorithmic), one was more traditional. They seem to be fine with either approach as long as you're fluent in Python.

MLOps/system design: this one surprised me. They cared a lot about how I'd operationalize a model. Feature stores, batch vs. real-time inference, monitoring for data drift, rollback strategy when a model degrades. In finance the production reliability bar is high, you can't just retrain and hope for the best.

What I noticed

They don't care much about deep learning or transformer architectures for most ML eng roles. This isn't a GenAI team (at least not the one I interviewed for). They care about classical ML, clean feature engineering, and models you can explain to a risk committee. Interpretability over accuracy is a real cultural preference.

Level was VP (their equivalent of senior IC). Offer I declined was around $240k total comp in NYC, mostly base-heavy.

5 replies

ds_dmitri

The interpretability preference makes total sense in a regulated environment. You can't put a neural net in a credit decisioning system without an explainability layer that satisfies auditors. That's just the domain. Did they ask about SHAP or LIME specifically or more conceptual?

ml_mike

Conceptual mostly. They asked me how I'd explain a model's decision to a business stakeholder who isn't technical. I mentioned feature importance and decision boundaries. SHAP came up when I brought it in, they seemed familiar but didn't require it.

finance_faye

The credit default model walkthrough question is a classic for any finance ML role. If you're prepping for MS ML interviews, make sure you can talk through: target variable definition, handling class imbalance, feature selection rationale, and how you'd validate without data leakage. Those are the real failure modes.

quietquit_quincy

240k TC declined is interesting. Was that a comp mismatch or a role/culture thing? Asking because I'm looking at similar roles and trying to calibrate if the bank premium is real.

ml_mike

Mostly comp. I had a competing offer from a tech company with higher equity. The MS offer was base-heavy which felt safer but the upside was lower. If you value stability over lottery tickets, the bank comp is actually fine. Just don't expect startup equity math.