01
Who leaves next, and why?
1,470 employees in IBM’s HR data. 237 of them left.
02
Accuracy lies here.
Predict that nobody leaves and you are 83.9% right. So I compared models by AUC.
03
My first result was a mirage.
A synthetic copy of the data gave AUC 0.565, close to a coin flip. On the real data: 0.80.
04
The simplest model won.
Logistic regression: AUC 0.80, ahead of three more complex models.
05
Overtime nearly triples the risk.
30.5% of people on overtime left. Without overtime, 10.4%.
06
Sales reps leave most.
39.8% of sales representatives left. 2.5% of research directors.
07
Pay matters.
Leavers earned a median 3,202 a month. Stayers, 5,204.
08
IBM does this for real.
In 2019 IBM said its AI spots leavers with 95% accuracy, saving about $300 million.
Sources: CNBC, April 2019 ↗
09
Next: catch more of them.
The model finds 36% of leavers. Raising that is the next job.