python 3.12.13, scikit-learn 1.9.1, pandas 3.0.6
rows 1000, bad 300 (30%), features 18
cost-optimal threshold for calibrated probabilities: 1/(1+5) = 0.167

Test cost per applicant over 20 stratified 70/30 splits (lower is better)
  strategy                         mean    min    max  bad caught  good rejected
  default 0.5                     0.866  0.740  0.990         48%            13%
  cost theory 0.167               0.545  0.450  0.643         86%            48%
  TunedThresholdClassifierCV      0.547  0.450  0.640         89%            55%
  all rejected                    0.700  0.700  0.700        100%           100%
  tuned on the test set (leak)    0.507  0.410  0.573 

Thresholds chosen per split
  TunedThresholdClassifierCV   mean 0.138, range 0.090 to 0.191
  tuned on the test set        mean 0.149, range 0.070 to 0.250

Average cost curve: at 0.50 0.866, at 0.17 0.545, lowest 0.536 at 0.15