Goldman Sachs · Primly Community

Goldman Sachs machine learning engineer interview: what they care about and what surprised me

ml_mike · 4 replies

did the GS ML engineer loop for a role in their risk and analytics group. this is not a research-flavored ML interview. just want to set that expectation upfront.

the loop: phone screen, take-home, three technical rounds, one behavioral.

phone screen: basics. ML fundamentals, a few probability questions, brief coding problem. filter round, nothing fancy.

take-home: build a classifier on a financial dataset, evaluate it, write up your methodology. they care about how you handle imbalanced classes (the data was very imbalanced) and how you interpret results for a non-technical audience. i wrote a 2-page PDF alongside my notebook, that seemed to land well.

technical round 1, ML depth: feature engineering for tabular/time-series data, regularization tradeoffs, ensemble methods. they asked how i'd approach overfitting in a model where i can't get more data. also: model monitoring in production. how do i detect that a model has gone stale. this is serious at GS because stale models in a risk context are a real liability.

technical round 2, coding: Python, data manipulation and some algorithmic content. one problem was clearly from the actual work they do (anonymized). medium difficulty. they let me ask clarifying questions.

technical round 3, system design: design an ML pipeline for real-time credit risk scoring. latency constraints were tight. they pushed on retraining cadence, feature store, rollback strategy.

behavioral: heavier than i expected. the theme was "how do you handle a situation where your model is right but stakeholders don't trust it." came up twice in different forms.

the GS ML role is closer to ML engineering in the production + reliability sense than to research or modeling innovation. if that's what you want, great fit. if you want to push the frontier on architectures, probably not the right place.

4 replies

ds_dmitri

the imbalanced classes thing is genuinely a GS theme. makes sense given fraud and risk data. did you use SMOTE, class weights, or something else? curious what they responded to.

ml_mike

i went with class weights first and explained why (fewer assumptions), then discussed SMOTE tradeoffs. they seemed to value the reasoning over the choice. the wrong answer would have been "i just used SMOTE because that's what i do."

infra_ines

the real-time credit scoring system design is a meaty problem. latency requirements in financial risk can be sub-10ms. did they give you a specific SLA to design toward or was it open-ended?

quietquit_quincy

"how do you handle a situation where your model is right but stakeholders don't trust it" is honestly the most important ML interview question and barely any interviews ask it. the answer requires both ML knowledge and actual organizational savvy.