Hybrid Quantum Algorithms for Trade Surveillance Analytics

AI summary

Overview

Research area: Quantum machine learning applied to financial compliance — specifically anti-money laundering (AML) and trade-surveillance analytics, using a synthetic pump-and-dump detection benchmark.

Technical level: Intermediate. The paper is readable without a quantum computing background, but the methods section assumes familiarity with quantum feature maps, fidelity/kernel methods, ansätze and NISQ hardware concepts.

Scope (one sentence): A like-for-like benchmark of four quantum and hybrid quantum–classical classifier families against classical baselines on one shared synthetic dataset, split and pair of feature representations, with one hybrid model executed on real superconducting hardware.

What This Paper Is About

Trade surveillance and AML monitoring still run largely on static rules, and by industry estimates more than 95% of the alerts they raise are false positives, each one reviewed by a human analyst. The authors ask whether quantum and hybrid quantum–classical classifiers can do better on a pump-and-dump detection task, and they test that question by holding the data, the train/test split and the feature representations fixed across every model so that quantum and classical approaches can be compared fairly. The goal is not to claim a speed-up but to establish where today's small, noisy quantum models actually stand relative to well-tuned classical ones.

Key Contributions

1. A controlled four-way comparison. Four quantum/hybrid classifier families — a fidelity-based quantum k-nearest-neighbours classifier (QKNN), a quantum support vector machine (QSVM) on a one-body projected quantum kernel, a 20-qubit quantum neural network (QNN), and a hybrid quantum neural network (HQNN) — are scored against logistic regression, XGBoost, one- and three-hidden-layer neural networks and a classical k-NN control, all on the same 800 training and 200 test accounts with the same ANOVA and PCA representations.

2. An execution-level batching scheme for QKNN. Multiple independent fidelity computations are packed into one reusable circuit, with batch size derived as B = ⌊H/q⌋ for a device with H qubits and q = ⌈N/2⌉ qubits per encoded data point. The authors show this reduces the number of circuit launches from M_test × M to M_test × ⌈M/B⌉ without changing the similarity measure or classification rule.

3. A family of QKNN variants. The paper formalises fidelity-weighted voting, mean fidelity-weighted voting (to counter class imbalance), anomaly-amplified voting that recursively adds confidently flagged suspicious samples to the voting memory, and higher-order quantum neighbourhood propagation through powers of the fidelity matrix.

4. A hardware demonstration and a synthetic data generator. The trained HQNN — in which a small classical feed-forward network generates every rotation angle of a five-qubit circuit — is executed on the ibm_fez superconducting processor, and the paper documents a configurable synthetic pump-and-dump generator with three-phase price lifecycles, role-based collusion and deliberate decoy accounts.

Main Findings

• Classical XGBoost wins outright. On ANOVA features it reaches a suspicious-class F1 of 61.4%, the highest score in the study, and the paper states plainly that no quantum advantage is observed on this benchmark.

• The HQNN is the strongest quantum-hybrid model. It reaches 54.5–57.9% F1 on an ideal simulator and 55.4–59.8% F1 on ibm_fez, ahead of six of the eight classical configurations.

• Small qubit count, competitive performance. The HQNN uses a five-qubit circuit — a quarter of the qubits of the 20-qubit QNN — and outperforms that QNN's 42.0% F1.

• Fewer false alarms at comparable detection. Compared with the 20-qubit QNN, the HQNN raises fewer than half as many false alarms at a similar detection rate. At a 15% false-positive rate, one in every two of its alerts is a genuinely suspicious account.

• QKNN beats its classical control. The quantum k-NN reaches 43.5% F1 on the ideal simulator and 44.4% under simulated device noise, above the classical k-NN control at 35.6%.

• QSVM is the weakest quantum entry. It does not exceed 41% F1.

• Noise tolerance is not the bottleneck. The QKNN's score under simulated device noise is slightly higher than its ideal-simulator score, and the HQNN performs as well or better on ibm_fez than on the ideal simulator.

• Granularity caveat. With only 40 suspicious test accounts, a single additional detected account moves recall by 2.5 percentage points and a single false alarm moves the false-positive rate by 0.625 points; the authors flag that differences of a few points should be read with this in mind.

Methodology in Plain English

The authors built a synthetic dataset rather than using real bank records, because privacy rules make genuine AML data essentially unavailable for research. The generator simulates pump-and-dump schemes in three stages: a quiet accumulation period with a slow price rise, an aggressive pump producing an exponential run-up, and a dump producing an exponential decay. Events last from one to fourteen days, including one special one-day "flash pump". Suspicious accounts are given distinct roles — orchestrators, helpers and victims — and everyone, including the suspicious accounts, does a large share of ordinary trading in unrelated securities to camouflage the manipulation. A fraction of suspicious accounts are made to look almost benign, and a fraction of benign accounts are made to look suspicious, deliberately narrowing the gap that a detector could exploit.

From the transaction records the authors derive 33 behavioural features over intraday windows (2, 4, 6 and 12 hours) and daily windows (1, 2 and 5 days). These cover trade-size anomalies against a 21-day lookback, wash-trading round trips, price impact, profit-and-loss ratios, "buy followed by multiple sells" patterns, inter-transaction timing statistics, low-priced-stock concentration, and co-trading network measures. Missing values are imputed by type (medians for ratio and temporal features, zero for z-scores and amount features), everything is standardised, and the 33 features are compressed to 20 by either ANOVA F-test selection or PCA. Each reduced feature is then rescaled to [0, π] on the training data so it can serve directly as a rotation angle in a quantum circuit.

The 1000 accounts are split 80:20, stratified, at the account level, with a fixed seed: 800 for training (640 benign, 160 suspicious) and 200 for testing (160 benign, 40 suspicious). Classical hyperparameters are tuned by five-fold stratified cross-validation inside the training set (seed 42), and decision thresholds are chosen by sweeping out-of-fold scores to maximise pooled out-of-fold F1 — never using test data. XGBoost and the two neural networks are additionally refitted under nine further random seeds to check sensitivity to training randomness.

On the quantum side, the four classifier families are built as described in the contributions above, with the QKNN and QSVM run both on an ideal simulator and under the noise models of IBM fake backends, and the HQNN's trained circuit executed on real ibm_fez hardware.

Why This Matters

The paper delivers something the quantum machine learning literature often lacks: a fair fight. By fixing the data, the split, the feature representations, the threshold-selection protocol and the evaluation metric across four quantum families and eight classical configurations, it turns "could quantum help here?" into a measurable question rather than a speculative one. The headline result — no advantage yet, but a five-qubit hybrid model on today's noisy hardware landing among the classical pack — gives the field a calibrated reference point.

Real-world applications:

• AML transaction monitoring, where false-positive rates above 95% drive enormous manual review cost; a model that raises fewer false alarms at the same detection rate directly reduces investigator workload.
• Trade surveillance for market abuse, covering pump-and-dump, spoofing and insider dealing, where manipulators deliberately mimic legitimate strategies to stay below static thresholds.
• Alert triage and prioritisation, where a hybrid scorer can rank alerts so analysts see genuinely suspicious accounts first.
• Compliance reporting under AMLD6 and FinCEN priorities, where regulators increasingly hold institutions accountable for missing complex, cross-asset, cross-jurisdiction schemes.

Industry relevance is reinforced by the author affiliations: two authors are at Vanguard and two at Capgemini, alongside the Washington Institute for STEM, Entrepreneurship, and Research (WISER) — a sign that quantum-enhanced compliance tooling is being evaluated inside financial and consulting organisations, not just in academia.

Future Directions

• Richer circuits and larger devices. The authors explicitly name richer circuits, larger devices and better training as the routes by which quantum approaches could pull ahead; the HQNN's five-qubit result is framed as a floor, not a ceiling.

• Scaling the benchmark. The dataset is 1000 synthetic accounts, chosen partly to fit what the quantum pipelines can process. Whether the relative ordering holds on larger, relational, real-world transaction graphs is untested.

• Making the quantum side practical. The QKNN's cost is dominated by pairwise fidelity evaluations, and kernel methods scale poorly with sample count; the batching scheme addresses launch overhead but not the underlying quadratic number of comparisons.

• Robustness to adversarial drift. The paper's own motivation section describes concept drift and deliberate adversarial adaptation, but the benchmark is a static synthetic snapshot. Testing whether quantum feature maps confer any robustness as manipulators adapt is an open question.

Target Audience

This paper suits quantum machine learning researchers who want a rigorously controlled, application-grounded benchmark rather than an isolated algorithmic result; machine learning engineers and data scientists in financial compliance evaluating whether quantum methods are worth prototyping; and technical decision-makers in banks, asset managers and consultancies who need a realistic read on the current capability gap. It is also useful to readers new to quantum ML who want a concrete worked example of how QKNNs, QSVMs, QNNs and hybrid models are constructed and evaluated, since the methods section spells out the encodings, voting schemes and circuit designs from first principles.

Authors’ abstract

Anti-money laundering (AML) and trade surveillance remain largely manual: rule-based monitoring raises alerts of which, by estimate, over 95% are false positives, each reviewed by a human analyst. We benchmark four quantum and hybrid classifiers against classical baselines on a synthetic pump-and-dump detection task: a fidelity-based quantum k-nearest-neighbours classifier (QKNN) with clustering-based routing; a quantum support vector machine (QSVM) on a one-body projected quantum kernel; a 20-qubit quantum neural network (QNN) with a Pauli-ZZ feature map and RealAmplitudes ansatz; and a hybrid quantum neural network (HQNN) in which a small classical network generates every rotation angle of a five-qubit circuit. Baselines are logistic regression, XGBoost, one- and three-hidden-layer neural networks, and classical k-nearest neighbours. All models share one stratified split of 1000 accounts (800 train, 200 test, 20% suspicious) and two 20-feature representations (ANOVA and PCA); QKNN and QSVM are also tested under simulated noise, and the HQNN is run on the ibm_fez processor. The best model is classical XGBoost on ANOVA features, at a suspicious-class F1 of 61.4%. The HQNN is the strongest quantum-hybrid model, at 54.5-57.9% F1 on an ideal simulator and 55.4-59.8% on ibm_fez, ahead of six of eight classical configurations; at a 15% false-positive rate, half its alerts are genuinely suspicious. Versus the QNN (42.0% F1), it raises under half the false alarms at similar recall with a quarter of the qubits. The QKNN reaches 43.5% F1 (44.4% under noise), above classical KNN (35.6%); the QSVM does not exceed 41%. We observe no quantum advantage, but a five-qubit hybrid on today's noisy hardware already matches well-tuned classical models, with room to grow via richer circuits and larger devices. We contribute a like-for-like comparison with classical models on shared data, splits and features.