American Express Data Analyst Interview Questions, Process, and Tips

• Q1. How would you decide if a person can be a customer of a private jet company
  View Answers (1)

Three Data model questions will be given to solve within 24 hours.

• Q1. Q1. Case study for Number of Green T-shirts sold in US?
  View Answers (1)
• Q2. Q2. SQL joins and Python simple questions
  Add Answer

You are building a model to predict if an order will be returned for a furniture ecommerce site.
1. steps to be followed.
2. What all features would you select based on your business sense.

• Q1. Why this job interests you?
  Add Answer
• Q2. Tell us your past experience
  Add Answer

• Q1. Explain why and when Analysis of Variance (ANOVA) is used. {Question from CV, since I had mentioned the use of it before}
  Add Answer
• Q2. ,You want to create an index that comprises of a bunch of components. How would you choose to aggregate all the individual components? Is it context specific?
  View Answers (1)

Question:
Suppose you are trying to detect if a particular credit card transaction is fraudulent or not. The credit score of the individual to which the card belongs to had a very healthy credit score. All bills were paid in time and average transaction amount was not that high ($800). The individual had not been out of the country in the last couple of decades. Here is a list of transactions:

1) Gold jwelleries worth $5000
2) Groceries worth $35
3) Second hand car worth $8,000
4) Burgers worth $10

Which transaction looks fraudulent to you?
There is no specific answer. They just want to see how you think through the problem. One can potentially make use of data in order to deal with this problem. From that, one can estimate the probability of each of these transactions being fraudulent. Econometrically, one can develop a potential binary logit model. That would involve identifying certain features that belong to individuals like the one considered above and use these features to come up with an estimate of the probability of the transaction being a fraud.
Not just that, this also needs to include not individual specific features but external features as well. For example, the first transaction might not be as fraudulent as it looks like, because in heavily regulated markets, the risk associated with reselling the gold or exchanging it for money might be high enough to disincentivise the fraudster from buying gold. Thus regulation might also be a valid feature, and different from features describing an individuals characteristics.
Ofcourse problems of overfitting would arise ifan excessive number of features are used. Various means of finding the optimal Degrees of Freedom can be employed.
Obviously one can do better with more complicated decisioning algorihms that involve machine learning models as well.
Eventually one needs to estimate at what threshold of probability will the trasaction be declared fraudulent.

Mettle test on quant and machine learning

• Q1. How will you handle an office emergency
  Add Answer
• Q2. What are your strengths
  Add Answer

Round duration - 90 Minutes
Round difficulty - Medium

5-9pm

• Q1. Maximum In Sliding Windows Of Size K Given an array/list of integers of length ‘N’, there is a sliding window of size ‘K’ which moves from the beginning of the ...read more
  View Answers (4)

Round duration - 45 Minutes
Round difficulty - Medium

Timing was Morning-noon
Interviewer was super friendly. 
He even helped me at places I got Stuck

• Q1. DBMS Question

  A quite Complex Joins Problem.

  View Answers (1)

Round duration - 15 Minutes
Round difficulty - Easy

Easiest HR round ever.
Just 15 mins after the technical interview round

• Q1. Basic HR questions What do you know about amex. Why amex. Willing to relocate. What excites you about this role. Why did you apply for this role in the first pl...read more
  View Answers (1)