Method · People Analytics
A model that predicts "everyone stays" is 84% accurate — and catches nobody. Accuracy rewards being right about the majority; it says nothing about whether the model finds the people who actually leave. That mismatch is why recall, not accuracy, was the metric we optimized for.
Technically hard to beat on accuracy. Catches zero leavers. Useless in practice.
Lower accuracy, but catches 6 in 10 people who actually leave. The right trade.
A false alarm costs a quiet check-in. A missed resignation costs a full replacement cycle — 50–200% of salary, per SHRM. That cost asymmetry is why a lower-precision, higher-recall model is the right trade, not a weaker one.
The 0.5 probability threshold used to flag someone as "at risk" is a dial HR can move — raise it to flag fewer people, lower it to catch more — not a fixed scientific result. And because the 294 test employees were held out entirely from training, in the same 16.1% attrition mix as the full dataset, this performance isn't inflated by an easier test set.