Most ML projects stop at "my model has 97% recall." This one asks: ok, but what does that actually cost? Predicting who will leave is only half the problem. Knowing what each wrong prediction costs your business is the other half.
Employee attrition costs companies between 50–200% of an annual salary per lost employee. For a Swiss IT profile, that's roughly CHF 70,000 per person — recruiting, onboarding, lost productivity, and institutional knowledge walking out the door.
The standard approach is to train a model and pick the one with the best accuracy or F1. But that ignores something important: in this problem, a false negative (missing someone who's about to leave) is not the same cost as a false positive (intervening with someone who was staying anyway). The default threshold of 0.5 assumes they're equal. They're not.
Instead of optimizing for recall or F1, I optimized for total business cost. The threshold that minimizes cost isn't 0.5 — it's 0.35 for the Swiss IT market.
The IBM dataset uses US salaries. I recalibrated all cost assumptions using real Swiss IT market data (jobs.ch, SECO 2025) with a scaling factor of 1.287.
The "best" model depends on your company. Retaining senior talent? Use threshold 0.12. Budget constraints? Use 0.50. Mixed profiles? 0.25.
The model predicts if someone will leave, not when. Overtime is the top predictor but may reflect job type rather than direct cause. Documented openly.
One threshold doesn't fit all companies. I defined three scenarios based on different business priorities:
Threshold 0.12 · Recall 88% · Detects 30/34 employees. Best for senior profiles where losing anyone is expensive.
Threshold 0.50 · Precision 79% · Only 4 unnecessary interventions. Best when budget for retention is limited.
Threshold 0.25 · Detects 22/34 · Moderate interventions. Best for mixed profiles and general use.
The most interesting insight from this project isn't technical — it's that GridSearchCV with hyperparameter tuning (LogReg, 97% recall, cost 1,782k CHF) was dramatically worse than simple threshold tuning on a base LDA model (88% recall, cost 1,008k CHF). Better metrics don't mean better business decisions.
I also learned that adapting a model to a specific market context requires more than changing currency symbols. The cost ratio between false negatives and false positives fundamentally changes what the optimal threshold is — from 0.12 with SHRM's 6.5:1 ratio to 0.35 with Switzerland's 3.4:1 ratio.