PayPal Data Scientist Interview Experience — Two Brutal Phone Rounds on SQL, Python, and Airport Causal Inference

Overall impression: it was really hard.

First there was a 15-minute recruiter call. They asked briefly about my current job, why I was looking, and my visa/status situation, then said they'd send an email with the interview process.

The phone interviews were two rounds total, each 1 hour.

Round 1: coding (including SQL and Python), plus a case study. The SQL question was built around drivers and riders — calculating top orders, cancellation rate, and so on. There were a lot of filter conditions, so I had to be careful about when to use WHERE versus ON. Python had two questions. The first: I'm rolling dice against an opponent, and the opponent is the banker. If I win, the opponent pays me n dollars; if the opponent wins or it's a tie, I pay the opponent m dollars. Assuming the dice are fair, should I play this game?

Round 2: XP/Stats/Causal Inference. The interviewer was from the airport team, and the questions centered on supply and demand for airport orders and on reducing the cancellation rate. It went extremely deep and detailed. For an airport, you can't easily run a geo test or a switchback test, and DID also can't guarantee it accounts for time-variant effects. Synthetic control is kind of ideal in theory but rarely used in practice, and there are also network effects. This kind of airport case is just special to begin with, because drivers at the airport are in a queue — they have to wait in line. If a driver or rider cancels an order, it seriously affects the driver's experience, because the driver either has to leave the airport after waiting a long time, or loop back around to the queue they were originally waiting in. For riders, airport riders are also a special group — on the way to the pickup location, can they find the correct pickup location, can they choose the correct pickup location, are they a one-time user, is it a business trip or a personal trip — all of that has to be considered. The whole case is very different from a regular in-city case. The interviewer asked: what methods would you use to separately measure driver and rider happiness/satisfaction, what angles would you consider, which of those are data-driven, and what other hypotheses do you have — for example, guessing from a psychological angle, even without data to back it up, you could talk about that too.