Bloomberg · Primly Community

Bloomberg machine learning engineer interview: not what you expect if you're coming from pure AI companies

ml_mike · 4 replies

Did the Bloomberg MLE loop this year. 8 years in ML (NLP background, more recently working on recsys), so I've been through a lot of these loops. Bloomberg's is distinct and honestly more honest than most, so I want to write it up properly.

First: Bloomberg MLE roles are not research-adjacent. This is applied ML in a financial data context. The work is: models that extract information from text (earnings calls, news, regulatory filings), models that detect anomalies in market data, classification systems for financial instruments. If you want to work on foundational models or novel architectures, wrong place. If you want to ship ML systems that have real reliability requirements and run in a production financial data pipeline, good fit.

The technical loop. Three technical rounds: ML fundamentals. They went deep. Explain bias-variance tradeoff not as a definition but in terms of model selection for a specific problem. When do you use gradient boosting vs a neural network and why. How would you debug a model that performs well on your eval set but poorly in production. Calibration came up. They clearly want people who understand what's happening, not just people who can call sklearn fit. ML system design. Design a news sentiment classification pipeline for financial instruments. This is a real product Bloomberg ships. I had to think through: data collection and labeling, model selection (they pushed on why I chose what I chose), serving architecture, latency requirements, monitoring and drift detection. The monitoring piece got a lot of attention. Coding. Python. Not pure algorithms, more data manipulation and implementation. One question was essentially: implement logistic regression from scratch. Know your numpy, know your math.

Behavioral. Very standard Bloomberg behavioral. Incident stories, ownership stories.

They did NOT ask me to implement a transformer from scratch or derive backprop on a whiteboard. But they did probe on whether I actually understand gradient descent or just apply it. The bar feels like: senior applied ML engineer, not research scientist.

Offer comp for a senior MLE in NYC was around 240-265k total depending on how you slice the equity.

4 replies

ds_dmitri

The monitoring and drift detection emphasis is interesting and tracks with what I've seen from other financial services ML roles. Production reliability matters way more there than at a startup where you can just retrain and redeploy. Did they ask specifically about distribution shift detection methods?

corp_refugee

240-265k total for senior MLE in NYC is below Google/Meta for the same level but Bloomberg is not the same environment. The terminal business generates extremely reliable revenue and layoff risk is genuinely lower. Different tradeoff.

pivot_pat

The "implement logistic regression from scratch" question is something I need to practice. I use sklearn and torch all day but I haven't implemented the underlying math since grad school. Any resources you found useful for refreshing on that?

ml_mike

Honestly just work through it once on paper first. Write out the log loss, compute the gradient manually, then implement the update step in numpy. It's not that many lines of code. The interviewer cares less about the code and more about whether you can explain what's happening at each step.