Affirm Data Scientist Interview Questions

1.Tell me about a model or analysis you built that stakeholders kept misusing or misinterpreting. How did you figure out what was going wrong, and what did you do about it?

2.Walk me through a time you noticed something unexpected in your data — a metric spike, an anomaly, a model drift — and had to trace it back across multiple systems or teams to find the real cause. What did you find, and how long did it take?

3.Describe a model or feature you shipped where you spotted a non-obvious risk — something that wasn't in the PRD or your manager's ask — and had to decide whether to raise it or just ship. What did you do?

4.Tell me about the last time you pushed back on a senior data scientist, PM, or executive who wanted to use a metric, model, or methodology you thought was wrong. How did you make the case, and how did it land?

5.Tell me about a time you inherited or audited someone else's model or pipeline and discovered it was behaving in a way that no one on the team could fully explain. How did you handle that?

6.Walk me through a time when your model's predictions looked right on your end, but something was clearly wrong in production — and tracing the issue required you to go well outside your own code. What was the gap and how did you close it?

7.You're computing repayment rates for Affirm's Pay-Over-Time product and you notice the numbers look different depending on whether you query Snowflake or pull from the PostgreSQL transactional database. How do you figure out which one is right and what do you do about it?

8.Say you're asked to evaluate whether a new feature in Adaptive Checkout — like showing a longer repayment term to certain users — actually improved conversion. What experiment design would you use and what are the main validity threats you'd want to address?

9.Walk me through how you'd build a training dataset for a model that predicts whether an Affirm borrower will miss their next installment payment. What are the most important data engineering decisions you'd need to get right?

10.You've trained a gradient boosting model to predict loan default for Affirm's underwriting pipeline. Offline AUC looks great, but after a few weeks in production the approval rate has dropped significantly with no corresponding change in default rates. What's your diagnosis and how do you investigate?

11.Affirm is considering adding a new repayment behavior signal — number of times a user has changed their auto-pay bank account in the last 90 days — as a feature in the underwriting model. Walk me through how you'd evaluate whether this feature is worth adding, and what concerns you'd raise before shipping it.

12.You're asked to build a real-time fraud scoring system for the Affirm Card that needs to score every swipe within 200ms. You'll have access to a feature store, a trained model, and Kafka event streams. Design the inference architecture and tell me where the hardest failure modes are.

13.Design a pipeline that computes a single, authoritative credit utilization score for each Affirm user across all their active Pay-Over-Time loans. How do you ensure the score is consistent, fresh enough to be useful, and doesn't cause double-counting across loan types?

14.We want to build a monitoring system that alerts us whenever a key metric in the Affirm Money Account — say, daily deposit volume — moves in an unexpected direction. Walk me through how you'd design it, from data ingestion to alert routing.

15.Design an offline A/B experiment analysis system for Adaptive Checkout that can handle the fact that the same user might see multiple checkout variants across different merchants and over time. How do you structure the data and analysis to avoid unit-of-analysis violations?

16.Design a system that computes and serves merchant-level default rate benchmarks so that Affirm's credit risk team can see, in near real-time, whether a specific merchant's borrower cohort is performing worse than comparable merchants. What are the hardest data modeling problems here?

17.Design a feature store for Affirm's underwriting models that serves both real-time loan decisions and offline training jobs. The tricky part: some features are computed from the borrower's full payment history, which can span hundreds of transactions going back years. Walk me through your architecture.

18.Design a system that detects when Affirm's underwriting model has silently become biased against a legally protected class — say, approvals for a demographic segment have dropped 15% over 90 days without any intentional model change. How do you build it, and what do you do when it fires?

19.Tell me about a time you had to explain a model's behavior or a data finding to someone who wasn't technical — a PM, a credit risk officer, someone like that. How did you decide what to simplify and what to preserve?

20.Tell me about a time you joined a project mid-stream — someone else had already designed the approach or set the direction — and you disagreed with a core assumption. What did you do?

21.Describe a time you had to work with a cross-functional partner — an engineer, a product manager, a risk analyst — who had a fundamentally different definition of success than you did. How did you get to alignment?

22.Tell me about a time you owned an analysis or a model that turned out to be wrong in a meaningful way — not just imprecise, but actually guiding the team in the wrong direction. How did you handle telling people?

23.You've been the most senior data scientist on a project for a few months and a new engineer or analyst joins the team. How do you approach onboarding them into your work — especially the parts that are poorly documented or that you've been carrying in your head?

24.Tell me about a time you had to decide whether to use a simple, explainable model versus a more complex one that performed better on the metrics. What was the context, and how did you make the call?

25.Estimate the annual revenue impact if Affirm's average loan approval rate increased by 3 percentage points. Walk me through how you'd structure the calculation.

26.Affirm's 7-day repayment reminder email has a 40% open rate. How would you estimate what the open rate 'should' be, and how would you decide if 40% is actually a problem?

27.Affirm's approval rate for Adaptive Checkout is 10 points lower on mobile web than on native iOS, for the same merchant and similar-looking users. How do you figure out if this is a real underwriting difference or a measurement artifact?

28.You're building a dashboard to track the health of Affirm's Pay-Over-Time product for a single merchant partner. What are the five most important metrics you'd put on it, and how would you define each one precisely?

29.Affirm is thinking about launching a referral program where existing users get $10 off their next purchase for each friend they refer who completes a loan. How would you measure whether this program is actually profitable, and what are the trickiest parts of that calculation?

30.The data science team is asked to choose a single 'north star' metric for the Affirm Card that the entire product team will optimize against. What would you propose, how would you define it, and what's the strongest argument against your choice?