🏦 Credit Scoring and Fraud Detection with Machine Learning

About this simulation

Retail banks score millions of card transactions and credit applications a day. A gradient-boosted or neural model turns raw features — transaction amount, spending velocity, device and location signals — into a single risk score between 0 and 1. This scene renders that scoring function as a 3D risk landscape: height and colour show the model's predicted risk over two features (transaction amount and account/spending velocity). Each glowing sphere is a live transaction falling onto the landscape, being scored, and routed to an outcome gate.

How it works

• Teal → amber → red surface — the model's learned risk function; taller and redder means higher predicted fraud probability.
• Translucent plane — the decision threshold. Transactions scoring above it are declined or sent to fraud review; below it they're approved.
• Yellow band — a manual-review zone just under the threshold, where borderline cases go to a human analyst instead of an automatic decision.
• SHAP panel — for the most recently flagged transaction, shows how much each feature pushed the score up (red) or down (teal), mimicking a real explainability report used to justify a decline to a regulator or customer.

Controls

• Decision threshold — the risk score above which a transaction is declined.
• Fraud attack intensity — the share of incoming transactions drawn from a high-risk fraud pattern (large amount + high velocity).
• Velocity feature weight — how heavily the model weighs spending velocity, tilting and reshaping the risk surface.
• Manual-review band — width of the grey-zone around the threshold routed to human review instead of an automatic decision.
• Transaction throughput — how many transactions per second stream through the scorer.

Did you know?

Real fraud-detection systems score transactions in under 100 milliseconds and must stay explainable — regulations such as the EU's GDPR and US ECOA/Regulation B give consumers a right to know the main reasons behind an automated credit decision, which is exactly what SHAP-style attribution methods are used to produce.