Quote:
Originally Posted by ragnarkar
https://www.cnbc.com/2019/04/03/ibm-...will-quit.html
I think this will also be used to find people to let go. After all, all else being equal, you'll want to retain the one who isn't intendingvon leaving. No more getting fired just for looking for another job: any signs that you want to leave and you could be out the door.
|
I wonder what the default is. We had a machine learning competition like this for our national conference last year that I won. The default for the company was actually 82.1% stayed. So if you just said everyone would stay you were 82.1% accurate. We used ROC AUC as the metric, which is an ordering of the probability where a 50% ROC AUC had an ordered probability that had no order (i.e. those that didn't term had probabilities of terming similar to those that did term). A 1 is perfect. I won with an .84 ROC AUC. But even then my actual accuracy was only 85%, so a delta on the baseline of just 3%.
As far as flight risk models go specifically (i.e. only caring about people that are about to term, not trying to predict it historically) there's actually some pretty good data out there on things as simple as Linkedin changes/updates. If you update your LinkedIn you're more likely to be considering leaving. Sometimes there are tells in your work product to if it's being tracked well enough. I had a buddy that built a flight risk model for financial advisors at a firm he worked for and he was insanely accurate because all the FAs that were leaving stopped scheduling meetings with clients and typically had 1-2 months of very low earnings compared to their historical averages. It turned out they were waiting to schedule appointments with their clients until they jumped ship and wanted to kick the new gig off right with some good months.
I could see IBM having a 90% baseline (i.e. in any given 3 year timeframe only 10% of employees term), so all they need to do is beat that by 5%. Not quite as mindblowing as the headline makes it out to be honestly.