Amazon Data Scientist Interview Loop
The Amazon data scientist interview starts with a recruiter call. Next is a technical phone screen, and even that has a Leadership Principles question attached. Then there's a virtual loop of four to five rounds, back to back: data manipulation, science breadth, science depth on one of your own projects, the hiring manager, and the Bar Raiser. What makes Amazon different is that the behavioral part isn't one round. It's in every round. Each interviewer scores you on assigned principles, whatever you're talking about.
What decides it
Leadership Principles and ownership. The Bar Raiser has a veto. At senior levels, stories without end-to-end ownership usually mean a down-level.
Practice by round
SQL and Python basics under time. That means aggregations, joins and filters, then lists and dictionaries. One or two Leadership Principle questions follow, and the interviewer expects a short story with a number in it.
This is the data-manipulation round. You'll use window functions and CTEs on order and customer tables, then do the same processing in Python. Amazon's SQL isn't tricky, but it is thorough. Handle nulls, duplicates and customers with no orders before anyone brings them up.
| Question | Difficulty |
|---|---|
| Manager of the Largest Department | Medium |
| Three Purchases | Medium |
| Responsible for Most Customers | Medium |
| Salary Less Than Twice The Average | Medium |
| Book Sales | Medium |
| Duplicate Training Lessons | Medium |
| Caller History | Medium |
| Car Part Price | Medium |
| Player with Longest Streak | Hard |
| Popular Posts | Hard |
| Cookbook Recipes | Hard |
| First Day Retention Rate | Hard |
| Sales Growth per Territory | Hard |
| Product Market Share | Hard |
| Multi-Status Order Tracking by Week | Hard |
The applied case in Amazon's own operations, set as mini-analyses. Delivery promise misses, split shipment rate, a Buy Again rollout and duplicate transactions each ask you to scope, measure and recommend. In the real loop, the depth round takes your own project apart three or four levels down. These are practice for defending an analysis end to end. The most common rejection here is not being able to put a number on your own impact.
| Case study | Difficulty |
|---|---|
| DCF for a Fulfillment Center | Hard |
| Data project | Difficulty |
|---|---|
| Resolving Duplicate Transactions | Medium |
| Delivery Promise Misses | Medium |
| Split Shipment Rate | Medium |
| Buy Again Rollout | Hard |
Science breadth. You get short statistics, probability and ML questions with follow-ups: bias and variance, reading a p-value, logistic-regression coefficients, what to do with few labeled instances, and which of two advertisements works better. It's broader and shallower than Google's, so range beats depth.
| Question | Difficulty |
|---|---|
| Random Bucketing | Easy |
| Interpret P-value | Easy |
| Assessing Multicollinearity | Easy |
| Colinearity in Data Analysis | Medium |
| Bias-Variance Tradeoff | Medium |
| Linear Regression & T-test | Medium |
| Broken Packages | Medium |
| Multi-dimensional Data | Medium |
| Statistical Hypothesis | Medium |
| More Effective Advertisement | Medium |
| Maximum Likelihood of Heads | Hard |
| An Instrument From a Rollout Glitch | Hard |