Use of survival analysis for trigger mining

Question

I am tasked with analyzing data to find "triggers" to an event. Specifically, this is transaction data from a bank (e.g., checking account daily balances, daily over draft fees, daily number of checks cleared etc) and the event of interest is the checking account being closed by the customer. It sounds like I am really needing to do a feature selection (important features might be something like "having three over drafts in 6 months").

I was thinking about cox regression (possibly with time varying covariates) and some variable selection (the paper here was an inspiration). In this way significant variables could be considered triggers. The business will be using these triggers in monitoring software (if a trigger event is detected), flag the customer as an attrition risk, and then do something to keep them from leaving.

Couple questions:

1. Does this seem like the correct approach or are there others?

2. It seems like it will be important to construct variables to determine the best triggers. For example, is the best trigger:

   a. Number of overdrafts in 3 months?
   b. Number of overdrafts in last 7 days?
   c. 3 or more overdrafts in the past 2 months
   d. (...)

Any suggestions on how to determine these triggers?

Answer 1

FWIW after 2 months, the survival method you describe does seem like a productive approach. Not sure I'm qualified to give it the mantle of being "correct." And, just as with skewed predictors we often try transforming via a square root, a log, and so on in order to find the strongest relationship with an outcome, it makes sense that you would try each of your ideas about overdrafts. It's interesting: on the one hand, a variable like "...last 7 days" might be too narrowly distributed to predict well (might involve overly rare events), but on the other, perhaps that narrowness will be overriden by its "a propos-ness." Just make sure you test the different versions separately so that they don't cannibalize one another.