Hybrid Quantum Algorithms for Trade Surveillance Analytics (arXiv (Cornell University))

Summary

• Rule-based AML monitoring systems raise alerts of which, by industry estimates, more than 95% are false positives.p.1
• Money laundering is estimated to take two to five percent of global GDP each year, or $800 billion to $2 trillion annually.p.2
• All models share one stratified split of 1000 accounts, 800 for training and 200 for testing, with 20% suspicious.p.1
• The held-out test set has 200 accounts, of which 160 are benign and 40 are suspicious.p.7
• The hybrid quantum neural network uses a classical network with 948 trainable weights and biases feeding a five-qubit circuit.p.20
• For an 80/20 imbalance the class weights are 2.5 and 0.625, penalizing missed suspicious accounts more heavily.p.22

Key insights

1. The paper benchmarks four families of quantum and hybrid classifiers against classical baselines on a synthetic pump-and-dump detection task.p.1
2. False positive rates of rule-based systems often exceed 95%, burying compliance teams in alerts.p.2
3. The training set has 640 benign and 160 suspicious accounts.p.7
4. Each representation uses 20 features, chosen by ANOVA selection or by PCA.p.7
5. A classical k-nearest-neighbours control with k = 5 serves as the like-for-like comparison for QKNN.p.8
6. The HQNN maps 20 input features through two hidden layers of 16 ReLU units to a five-qubit circuit.p.20
7. Synthetic data gives full control over fraud scenarios and certain ground-truth labels, enabling reproducible benchmarking.p.5