Leak-free fraud detection across three financial datasets, reported with average precision and bootstrap confidence intervals instead of accuracy. Cost-based threshold optimisation, formal hypothesis testing with effect sizes, odds-ratio interpretation, and a live interactive explorer.
How it knows: The threshold is chosen by minimising Z = 100xFN + 1xFP rather than defaulting to 0.5. On the Credit Card data the optimum lands at 0.02, catching 88 of 98 frauds for 83 false alarms - $1,083 of modelled loss against $9,800 for doing nothing.
Nicholas Smith - all work