Wells Fargo · Primly Community

Wells Fargo machine learning engineer interview, what it actually tests vs what you'd expect

ml_mike · 6 replies

Went through the Wells Fargo machine learning engineer interview process earlier this year. Role was on their risk and fraud detection team in San Francisco. Sharing because most ML interview content online is either FAANG-focused or generic. Banks are different.

What I expected: heavy leetcode grind, vague ML system design, maybe some statistics.

What I actually got: substantially more applied and domain-specific than I anticipated.

Process overview: Recruiter screen. Standard. ML coding screen (60 min). They gave me a real dataset (anonymized transaction records) and asked me to build and evaluate a simple fraud classifier in Python. Jupyter notebook environment. They cared about: how I handled class imbalance, which metric I optimized (they pushed back when I defaulted to accuracy, which was the right pushback), and how I would explain the model to a compliance team. That last part was the actual test. XGBoost with SHAP values came up in my answer and they asked follow-up questions about feature importance interpretation. ML system design (45 min). How would you design a real-time fraud scoring system that has to make a decision in under 200ms, handle concept drift, and log everything for regulatory audit? This is the kind of problem that's very specific to financial services. I talked through feature stores, model serving infrastructure, shadow mode testing, and the challenge of retraining on labeled data that takes weeks to arrive (chargebacks are slow). They seemed to appreciate that I'd thought about the label delay problem. Behavioral panel. Two rounds. They really wanted STAR-method answers and were explicit about it. They asked about working with stakeholders who didn't understand model limitations, and handling a model failure in production.

What mattered most: domain knowledge around financial services ML. If you've worked on fraud, credit risk, AML, or anything adjacent, highlight it aggressively. If you haven't, spend time reading about the problems. The general ML skills are table stakes. What differentiates candidates is whether they understand why precision/recall tradeoffs in fraud detection have real business consequences.

Leveling landed me at what they call Senior Data Scientist II (the ML eng titles at WF are a bit unusual). Base around $165k in SF, bonus target 15%, RSUs over 3 years. Not FAANG total comp, but the role had real scope.

6 replies

marketer_mei

The label delay problem is real and most candidates absolutely miss it. Fraud ML in banking is not like a standard supervised learning classroom problem. You're often training on data where the ground truth won't be confirmed for 60-90 days. Modeling on transactions where the chargeback window hasn't closed yet means you're training on noise. Nice that you flagged it.

sec_sasha

SHAP for regulatory explainability is almost required in any model-at-a-bank conversation now. SR 11-7 compliance (the Fed's guidance on model risk management) basically mandates that you can explain why your model made a decision. If you can't explain it to a model risk officer, it doesn't matter how accurate it is.

ml_mike

Yep. SR 11-7 came up explicitly in my conversation. I didn't know the exact citation going in but I knew the concept. Definitely worth reading the Wikipedia summary at minimum before any bank ML interview.

content_cole

Adding to the reading list: also look at OCC model risk guidance. If you can casually reference 'model validation' as a separate team from model development, it signals you understand how the bank actually operates.

visa_vik

How long did the whole process take from recruiter screen to offer? I'm under the OPT clock and trying to figure out which processes are realistic to pursue.

ml_mike

About 5 weeks total, which felt slow but they were communicative. They mentioned that ML roles at regulated institutions have extra approval steps before extending an offer. If you're on OPT, factor that in. Their H1B sponsorship was confirmed early in the recruiter conversation.