Amazon Data Analyst Interview Questions

“I would first remove cancelled orders, since they should not count. Then, for each customer, I would number their orders by date using ROW_NUMBER partitioned by customer. I take the rows numbered one and two and place them side by side, either with a self join or with LEAD to pull the next order date onto the first row. Then I keep customers where the gap between the dates is a month or less. Customers with a single order have no second date and drop out. I would check what happens when two orders share the same day, and test on a few known customers.”

“I would define the outcome first: how often a product page view turns into an order, and orders per active customer. Comparing before and after in the same city is risky, because seasons and promotions also change. So I would find similar cities that did not get the faster promise and compare how the gap between them moved. I would look at existing customers separately from new ones. I would also check that the promise was kept, since more orders with more late parcels is not a win. Finally, I would watch whether the lift lasts after the first weeks.”

“More orders alone does not tell me enough. I would compare the groups on orders kept after returns, and include the cost of handling a return. Then I would break the returns down. Perhaps the row suggests clothing in a size the customer sent back before. If returns gather in a few item types, the feature could ship with those excluded. I would check that both differences are larger than chance, and that the test ran long enough for returns to arrive, since they come later than orders. My recommendation would state the net effect and what I am unsure about.”

“I would first look for a reliable shared identifier, such as a barcode or manufacturer part code, and group by that. Where it is missing, I would clean the names: lower case, remove extra words like free shipping, and standardise units. Then I would match on brand, model and size together, not name alone. Fuzzy matching creates false matches, so I would check a sample by hand and measure how often it is wrong. I would keep a mapping table from listing to product, so the logic is reused. In the report, I would say what share of listings could not be matched.”

“My team believed most repeat customers came back because of discounts. While studying your project, I found most repeat orders were at full price and followed a fast first delivery. Before saying anything, I checked my query with a colleague and tried a second method, which agreed. I presented it as a question: here is what I see, what am I missing? I showed the raw numbers, not only a chart. One manager raised a fair point about a discount type I had missed. I added it and the finding held. After that the team decided your result.”

“I once presented late delivery figures to warehouse shift managers on your project. They had no interest in my method. They needed to know where to put people. I dropped most of my charts and showed one table: which hours of the day produced the most late parcels. I said it in one sentence, that parcels packed in the last hour before the truck left were the ones arriving late. I was honest that I could show the pattern but not the cause. The managers moved some staff to that hour, and afterwards your result. I learned to begin with the decision.”

“First I would ask whether the drop is real or the data is late. I would check that the overnight jobs finished and that row counts look normal. If visits are steady but orders fell, the problem is between the cart and the order. Then I would break it down: all devices or one, all payment methods or one, all categories or one. A drop in one payment method points to a provider issue. I would also check for a holiday or a release at midnight. As soon as I believe it is real, I tell the on-call team what I know and what I have ruled out.”

“I like questions where the answer changes what someone does the next day. Shopping and delivery data is full of them: why was this parcel late, why did this customer stop ordering, did this change to a page really help. The data here also covers the whole path from a search to a doorstep, so a finding can be traced to a real cause. In my last role I worked on your project and enjoyed the work most when operations staff used my numbers. I would like to learn how to run careful experiments at this size, and how to keep measures honest.”