Risk-Aware Fraud Decision System

Risk × Uncertainty × Novelty - Transaction X-Ray

Devansh Kumar devantaris
Checking API…
Dataset Kaggle Credit Card Fraud — 284,807 transactions, 0.17% fraud rate
Feature Space PCA-transformed V1–V28, Amount, Time (31 dimensions)
Model Calibrated XGBoost Bootstrap Ensemble (5 members)
Uncertainty Bootstrap standard deviation across ensemble predictions
Novelty Isolation Forest trained exclusively on legitimate transactions
Decision Policy Cost-optimized routing engine with asymmetric loss functions

This system is designed for financial risk environments where decisions must account for asymmetric loss and uncertainty.

This system does not only predict fraud. It decides what action to take based on three independent axes: how risky a transaction looks, how confident the models are, and whether the transaction pattern has been seen before.

It balances risk, uncertainty, and cost to simulate how banking systems design transaction routing policies.

Approve — Transaction proceeds automatically.
Step-Up Authentication — Request additional verification from the cardholder.
Abstain — System is not confident. Hold for manual review.
Escalate Investigation — Route to human fraud analyst for review.
Block — Decline the transaction automatically.

Intended For

  • Banking risk teams evaluating transaction routing policies
  • Payment processors designing fraud detection pipelines
  • Fraud analytics teams studying decision-theoretic approaches
  • Decision engineering research and academic study

This is not a consumer fraud-checking tool. It simulates institutional decision routing under uncertainty.

  1. Generate a transaction scenario using the random generator or preset buttons.
  2. Observe the three detection layers — risk assessment, predictive uncertainty, and novelty detection.
  3. Review the routing decision and its triggering rule.
  4. Interpret the cost impact — expected loss, review cost, and net utility.
Risk Score Calibrated fraud probability from the XGBoost bootstrap ensemble. Higher values indicate stronger fraud signals.
Uncertainty Standard deviation of predictions across 5 bootstrap models. High values indicate disagreement between models.
Novelty Score Deviation from legitimate transaction behavior as measured by Isolation Forest. Negative scores indicate novel patterns.
Final Decision Derived from Risk × Uncertainty × Cost policy. The routing engine applies threshold-based rules to select the optimal action.
◇

Transaction X-Ray

Generate or select a transaction to see it pass through
the three detection layers.

Ensemble Risk → Uncertainty → Novelty → Decision