FraudLens

the threshold is a business decision
You are the fraud manager. Drag the line, pay the price — every number below updates from a real model's test-set predictions (56,962 transactions, 75 frauds). Educational simulation.

Draw the line

Lower threshold = stricter guard: more frauds caught, more customers wrongly blocked.
lenient strict 0.50
–frauds caught (of 75)
–£ slipping through (missed fraud)
–customers wrongly flagged
Cost defaults are illustrative, not research claims — set them to your business. “Amount-aware” prices each missed fraud at its actual transaction value.

The price of your line

Total cost across every possible threshold. Green pulse = the cheapest point.

Where it fails

Every dot is a real fraud: score vs £ amount. Left of your line = missed. Notice the misses skew high-value.

Why did the model say that?

Pick a transaction — the bars show which features pushed it toward fraud (red) or legit (green). V-features are anonymised, so we see which feature mattered, not a human reason.

Is the blind spot the model — or the data?

This model misses high-value frauds. The same cost-based pipeline, rerun on a dataset with customer-behaviour signals, catches them.

ULB (anonymised data)

~0%

of high-value frauds caught · PR-AUC 0.79

Sparkov (rich data, +1 behaviour feature)

99.8%

of high-value frauds caught · PR-AUC 0.90

Same algorithm, same approach — only the data changed. The blind spot is a data limitation, not a model one.
dataset comparison — same pipeline, richer data

How honest is this?

illustrative costs 75-fraud test fold synthetic proof dataset precomputed & auditable