Fraud model, pick your threshold

Scores of the last 15 % of the ULB credit-card dataset (42,722 transactions, 52 frauds), produced by a LightGBM and a logistic regression that were trained only on the 70 % before it. Move the cutoff and see what an ops team would get.

loading test scores (about 1 MB)

model threshold 0.200

recall

–

precision

–

alerts / 10k txns

–

frauds missed

–

of 52 in the window

LightGBM logistic regression · the dot is your current threshold

Model card Dataset Training script GitHub built by hari-krishna-ai