Hybrid Quantum Algorithms for Trade Surveillance Analytics
This paper benchmarks four quantum and hybrid classifiers against classical baselines for trade surveillance, finding that while no quantum advantage was observed, a five-qubit hybrid neural network running on noisy hardware achieved competitive performance (55.4–59.8% F1) comparable to well-tuned classical models, suggesting potential for future growth with larger devices.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every year, the global financial system loses hundreds of billions of dollars to money laundering and market manipulation. To stop this, banks and regulators rely on automated systems that scan millions of transactions for suspicious activity. However, these systems are currently overwhelmed by their own success; they flag more than 95 percent of the alerts they generate as false alarms. This means human analysts spend nearly all their time investigating perfectly normal trades, leaving the truly dangerous schemes hidden in the noise. The challenge is not just finding the bad actors, but doing so without drowning the investigators in false leads. For decades, the solution has been to build better rules or use standard computer programs to spot patterns, but these methods struggle with the sheer volume of data and the clever ways criminals adapt to avoid detection.
Recently, a new approach has emerged from the world of quantum computing, a field that uses the strange laws of physics at the atomic scale to process information in ways classical computers cannot. Researchers have long wondered if these quantum machines could solve the specific problem of financial surveillance: could they spot the subtle, complex connections between accounts that traditional software misses? A team of researchers from Vanguard, the Washington Institute for STEM, Entrepreneurship, and Research, and Capgemini set out to test this idea. They did not look at real bank records, which are protected by strict privacy laws, but instead created a highly realistic, artificial dataset. This synthetic data mimicked the behavior of a "pump-and-dump" scheme, a type of market fraud where a group of traders artificially inflates a stock price before selling it off, leaving victims with worthless shares. The dataset included 1,000 trading accounts, with 20 percent acting as the fraudsters and the rest as innocent participants.
The researchers pitted four different types of quantum and hybrid quantum-classical algorithms against four standard, classical computer models. The goal was to see which system could best identify the 200 suspicious accounts hidden in the test set. The classical models included well-known tools like decision trees and neural networks, which are the current workhorses of the industry. The quantum contenders were more experimental: one used a method similar to finding the nearest neighbors in a crowd but measured by quantum similarity; another used a support vector machine, a technique that draws a line to separate groups, but built on a quantum map of the data; a third was a full quantum neural network; and the fourth was a hybrid model that used a small classical computer to control a tiny quantum circuit.
The results were clear and surprising. The best-performing model was not a quantum machine, but a classical one: a decision-tree algorithm called XGBoost. It correctly identified 61.4 percent of the suspicious accounts while keeping false alarms at a manageable level. None of the purely quantum models came close to this performance. The quantum neural network, which used 20 qubits (the basic units of quantum information), managed to find only 42 percent of the fraudsters. The quantum support vector machine performed even worse, failing to exceed 41 percent. The quantum k-nearest neighbors method did better than its classical counterpart, finding 43.5 percent of the fraudsters, but it still lagged far behind the best classical tools.
However, the story did not end with a defeat for quantum computing. The most promising result came from the hybrid model, which combined a small classical computer with a tiny five-qubit quantum circuit. This system, which used a quarter of the quantum resources required by the other quantum models, achieved a performance of 57.9 percent on a perfect simulation and 59.8 percent when run on actual quantum hardware. This placed it just behind the best classical model and ahead of six of the eight classical configurations tested. In practical terms, this hybrid model was able to flag genuine suspicious accounts with a precision that was roughly an order of magnitude better than the current industry standard, where only one in twenty alerts is real.
The study also revealed that the way data is prepared is just as important as the algorithm used to analyze it. The researchers tested two different ways of simplifying the complex trading data, and the results varied wildly depending on the method chosen. For the hybrid model and the best classical tools, one specific method of organizing the data worked significantly better than the other. This suggests that the key to unlocking the potential of these systems lies not just in building more powerful quantum computers, but in understanding how to feed them the right information.
While the study did not find a quantum advantage where a quantum machine outperformed the best classical computer, it demonstrated that a small, hybrid system running on today's noisy quantum hardware can already compete with the best traditional tools. The researchers concluded that the hybrid approach is the most effective path forward for this specific task. By using a classical computer to guide a small quantum circuit, the system learned to embed the data in a way that a fixed quantum map could not. This suggests that the future of financial surveillance may not require massive, error-free quantum supercomputers, but rather clever combinations of classical and quantum tools that can operate effectively on the imperfect machines available today. The work proves that quantum methods are no longer just theoretical; they can be trained, tested, and run on real hardware to solve real-world problems, even if they are not yet the undisputed champions.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.