Quantum use cases in banking: an evidence check

What the published evidence shows for the four quantum use cases banks are most often pitched, and how to judge each claim.

• Level: Intermediate
• 13 min read
• Applications · 2 of 3

What you’ll learn

• Describe what the evidence shows for risk simulation, portfolio optimisation, fraud detection and settlement.
• Explain why resource estimates describe specific problems on machines that don't exist yet.
• Judge a claim by the strongest kind of evidence behind it and a fair classical baseline.

Recommended background: What quantum computers are good at, How to read a quantum headline

On this page

1. Six kinds of evidence
2. Risk simulation and derivatives pricing
3. Portfolio optimisation
4. Fraud detection and machine learning
5. Payment settlement and liquidity optimisation
6. The evidence at a glance
7. How big are the machines on paper?
8. What this means in practice
9. Sources and further reading

Quantum computing is often described as transforming risk management, portfolio construction and fraud detection. This piece treats each of those statements as a claim and checks it against what has actually been published. It covers the four use cases banks are pitched most often: risk simulation and derivatives pricing, portfolio optimisation, fraud and machine learning, and payment settlement and liquidity optimisation. It's written for steering committees, risk managers and business analysts who are asked to judge proposals. How to approve and govern a pilot is covered separately in How to evaluate a quantum computing pilot.

Six kinds of evidence

"Quantum can do X" can mean very different things. It helps to ask which of these a claim rests on:

1. Theory. A mathematical proof that a quantum algorithm needs fewer steps than the best known classical one, usually as problems grow very large. It says nothing yet about real machines.
2. Resource estimate. A paper that works out what machine a specific problem would need: how many error-corrected (logical) qubits, how many operations, how fast. It's a design on paper, under stated assumptions.
3. Classical simulation. The quantum algorithm run on an ordinary computer that imitates a quantum one. Useful for testing ideas, but limited to small sizes and no evidence of speed.
4. Hardware demonstration. The algorithm run on a real quantum computer, usually on a small or simplified version of the problem. It shows feasibility, not advantage.
5. Competitive benchmark. A head-to-head test against the strongest classical method, tuned by people who know it, on the same problem, data and budget. This is where an advantage claim has to be won.
6. Production evidence. The method in routine use, making decisions the bank relies on, with results that hold up over time.

Each source in this piece is labelled by type: a peer-reviewed paper has been checked by other researchers; a preprint or working paper has not, or not yet; a company announcement is the company's own account.

Risk simulation and derivatives pricing

• The claim. Banks price complex derivatives and measure risk by running very large numbers of random scenarios, known as Monte Carlo simulation. A quantum method called amplitude estimation can, in theory, reach the same accuracy with roughly the square root of the number of samples. Ashley Montanaro proved a near-quadratic speed-up of this kind in a 2015 peer-reviewed paper.
• What's been run. IBM researchers measured value at risk for a toy Treasury-bill example on IBM quantum hardware and ran a two-asset version in classical simulation (Woerner and Egger, peer-reviewed, 2019). Woerner and Egger describe their examples as toy models. In a 2020 peer-reviewed paper, Stamatopoulos and colleagues priced several option types in simulation and tested pricing circuits on an IBM device, where noisy two-qubit operations were the main source of error and needed error mitigation.
• How big a machine. A 2021 peer-reviewed estimate by researchers at Goldman Sachs and IBM looked at two real contract types: a basket autocallable on three underlyings and a target accrual redemption forward (TARF), priced to an error of 2×10⁻³. The autocallable needed about 8,000 logical qubits and a sequence of 54 million layers of a costly operation type (the T-depth); the TARF needed about 11,500 logical qubits. To beat classical pricing, which the authors take to be about a second, the whole program would have to run in about a second. A 2024 peer-reviewed follow-up by two of the same authors cut the autocallable estimate to about 4,700 logical qubits and about a billion T gates in total. It puts the speed requirement separately, as a logical clock rate of about 45 MHz, or 45 million steps a second, so that the calculation finishes within the one second the authors assume classical pricing takes. The two numbers measure different things: many T gates run side by side within each step, so the billion gates don't execute one after another. Neither paper estimates physical qubits or error rates; both work at the logical level.
• The classical bar. Classical Monte Carlo isn't standing still. Banks run it in parallel on large clusters and graphics processors, and use variance reduction and quasi-random sampling, which can converge faster than plain Monte Carlo on many pricing problems. Some products have faster analytical or numerical methods altogether. A fair comparison uses the same contract, accuracy target and confidence level, and counts the full time, including loading the market model into the quantum computer.
• Where it stands. Theory, resource estimates and small demonstrations exist. No competitive benchmark or production use has been published. A 2021 peer-reviewed perspective from Google researchers argues that quadratic speed-ups like this one are unlikely to deliver an advantage on early generations of error-corrected machines, because of the overhead error correction adds to every operation.

Portfolio optimisation

• The claim. Choosing the best mix of assets under real constraints, such as limits on the number of holdings, minimum trade sizes and transaction costs, is a hard optimisation problem. Quantum annealers, hybrid quantum services and gate-based algorithms are all proposed as faster or better solvers.
• What's been run. A 2022 peer-reviewed study by Multiverse Computing and BBVA ran dynamic portfolio optimisation on real price data using a D-Wave annealer and its hybrid service, an IBM quantum computer and a quantum-inspired classical method. The largest case had 8 assets over 53 time steps. The authors found no single best method: results depended on whether they measured profit or risk-adjusted return, and the classical solver they used was sometimes better than the quantum and quantum-inspired ones. Broader reviews of quantum optimisation, discussed in the pilot article, reach similar conclusions.
• How big a machine. A 2023 peer-reviewed analysis worked through every step of a fault-tolerant quantum algorithm for portfolio optimisation (a quantum interior point method). For 100 assets at the precision studied, it estimated about 8 million logical qubits and around 7×10²⁹ costly operations, spread over 2×10²⁴ sequential layers. It also assumed a large quantum memory (QRAM) that doesn't exist yet. The authors note that the same 100-asset problem can be solved in seconds on a laptop, and conclude that fundamental improvements to the algorithm are needed for any practical advantage.
• The classical bar. The textbook version of the problem is solved efficiently by standard software. Realistic constraints make it harder, but commercial mixed-integer solvers and well-tuned heuristics are widely used for these constrained versions. A fair comparison sets the quantum method against those solvers on the same constraints, and judges solution quality, whether every constraint is met, run time, total cost, and how the portfolio performs out of sample.
• Where it stands. Small hardware demonstrations and resource estimates exist. No demonstrated advantage over leading classical solvers, and no production use, has been published.

Fraud detection and machine learning

• The claim. Quantum machine learning could find patterns that classical models miss, or speed up the linear algebra inside some learning methods. Some early algorithms promised very large speed-ups, but they depend on loading data into a quantum computer quickly, which no machine can yet do at scale, a caveat Scott Aaronson set out in a 2015 commentary. In 2019, Ewin Tang showed in a peer-reviewed paper that a well-known quantum recommendation algorithm could be matched by a classical one that is only polynomially slower, removing its claimed exponential advantage.
• What's been run. A 2022 peer-reviewed study by CERN, IBM and IRIS Analytics tested quantum support vector machines on card-payment fraud. The full data set held about 2.4 million transactions, but the quantum model was trained and tested on a reduced sample of 1,500 and 1,000 records with seven features, and the reported results came from a simulated quantum computer. Classical and quantum models performed similarly; combining them improved accuracy by about one percentage point. A preprint first posted in 2022 by Pasqal, Multiverse Computing and Crédit Agricole CIB predicted credit rating downgrades on a neutral-atom processor with up to 60 qubits and reported performance competitive with a random forest model; the bank's 2023 announcement described the results as "as accurate as the results in production", meaning the bank's existing models. HSBC and IBM's 2025 bond-trading trial is discussed in the pilot article.
• How big a machine. We found no published resource estimate for a fraud or credit model at banking scale, so there is no supportable figure to give.
• The classical bar. Gradient-boosted trees, such as XGBoost, and network-based methods that follow money between accounts are the established tools, trained on millions of records. A fair comparison uses the same data and features for both sides, not a reduced set for the quantum model only, tests on data the model hasn't seen, and counts cost and response time, since fraud decisions often have to be made while a payment is in progress.
• Where it stands. Experiments exist, on small or reduced data sets, with results similar to classical models. No competitive advantage or production use has been published.

Payment settlement and liquidity optimisation

• The claim. In a high-value payment system, each payment settles in full and in real time, so banks need liquidity on hand. Changing the order in which queued payments settle can reduce how much they need. Finding the best order is a hard combinatorial problem, which makes it a natural candidate for quantum optimisation. The studies below reorder queued payments in a gross settlement system. That's different from netting, where obligations are offset against each other before settlement: neither study demonstrates optimisation of bilateral or multilateral netting, and the 2026 paper uses multilateral netting only as a theoretical upper bound on the savings available.
• What's been run. Bank of Canada researchers used D-Wave's hybrid solver, which combines a quantum annealer with classical computing, on historical data from Canada's former high-value system, LVTS, for 30 sampled days from 2015 to 2017 (peer-reviewed, 2024). Reordering batches of 70 payments saved an average of about C$240 million of liquidity a day, for a delay of about 90 seconds. The paper doesn't claim a quantum advantage, and its classical comparison was limited to one exact solver that struggled with the problem. A 2026 working paper by Bank of Italy and Bank of Canada researchers applied the same approach to the Italian part of the euro area's TARGET2 system, using 35 days from 2021. For batches of 70 payments, the hybrid quantum solver saved about €23 million a day, and a classical method, simulated annealing, given about the same time, saved about the same. Using simulated annealing on larger batches of 700 payments raised average savings to about €269 million a day, at a cost of about 25 minutes' delay per batch; the paper notes that current quantum hardware can't handle batches that large.
• How big a machine. No published resource estimate exists for this problem; the limit reported so far is the batch size today's hardware can take.
• The classical bar. High-value payment systems already include liquidity-saving mechanisms, and they vary in sophistication: some simply hold payments in a queue, while others can settle offsetting queued payments together, as one of the settlement mechanisms in Canada's Lynx system does. A fair classical baseline includes the mechanism the system actually uses, alongside general-purpose methods such as the simulated annealing that matched the hybrid quantum solver in the 2026 paper. It runs every method on the same batches and time budget, and weighs liquidity saved against the delay to each payment.
• Where it stands. This is the most banking-specific evidence, from central bank researchers using real payment data, and the only use case here with a published comparison against a classical method given the same time budget. That comparison found no advantage for the quantum solver. Both studies are research; neither describes a production deployment.

The evidence at a glance

Use case	Strongest evidence published	Published resource estimate	Advantage over the best classical method	Production
Risk simulation and derivatives pricing	Theory, resource estimates, small hardware demonstrations	About 4,700 to 11,500 logical qubits for specific contracts, finishing in about a second; the 2024 estimate needs about a billion T gates in total at a logical clock rate of about 45 MHz	Not shown	None published
Portfolio optimisation	Small hardware demonstrations with limited classical comparison	About 8 million logical qubits for 100 assets, plus quantum memory that doesn't exist	Not shown	None published
Fraud and machine learning	Experiments on reduced data, partly in simulation	None found	Not shown; results similar to classical	None published
Payment settlement and liquidity optimisation	Central bank studies on real payment data, one head-to-head comparison with a classical method	None found	Not shown; a classical heuristic matched it	None published

How big are the machines on paper?

For comparison, the error-corrected demonstrations described in Quantum threat indicators report dozens of logical qubits, not thousands, and none has yet run the long calculations these estimates assume.

What this means in practice

• For steering committees. Ask which of the six kinds of evidence a proposal rests on. A hardware demonstration or a vendor case study is not a competitive benchmark. "Transforming" is a claim about production results, and none has been published for these use cases. A proposal that presents itself as learning and readiness is consistent with the evidence; one that promises a near-term edge needs a benchmark to back it.
• For risk managers. Treat resource estimates as moving indicators, not forecasts. Between 2021 and 2024, better algorithms cut the autocallable estimate from about 8,000 to about 4,700 logical qubits; a similar drop in a use case you care about is worth recording. Any model built in a pilot that could influence decisions falls under model risk management like any other.
• For business analysts. Write down the classical baseline before the work starts: which method, who tunes it, the same data and constraints, and how full cost and elapsed time are counted. The payment settlement studies show why: the 2026 paper is informative because its authors ran a classical heuristic alongside the quantum solver.

What you’ve learned

• Published evidence doesn't show a quantum advantage, or production use, in any of the four use cases banks are most often pitched.
• Resource estimates are for specific problems on machines that don't exist yet; read their assumptions, not just their qubit counts.
• Judge every claim by the strongest kind of evidence behind it, and insist on a fair classical baseline.

Check your understanding optional

Has any of these four banking use cases shown a quantum advantage in production?

Show the answer

No. Published evidence doesn't show a quantum advantage, or production use, in any of the four.

Written by Muhammad Atif · Reviewed 17 October 2026 · How sources are chosen and checked · Report a correction

Sources and further reading

Tags show what kind of source each one is. A standard or government guidance is an official document; a peer-reviewed paper has been checked by other experts; a preprint has not been peer-reviewed yet; an experiment reports a real-world demonstration; a company announcement is the company's own account; vendor documentation is a supplier's own technical guide; expert commentary is a specialist's published analysis or opinion; a news report is a journalist's account. Dates and figures were checked against these sources on 17 October 2026. Spotted an error? Email hello@plainquantum.com and it will be corrected, with a note.

1. Peer-reviewed paperQuantum speedup of Monte Carlo methodsAshley Montanaro, Proceedings of the Royal Society A, 2015
2. Peer-reviewed paperQuantum risk analysisStefan Woerner and Daniel J. Egger, npj Quantum Information, 2019
3. Peer-reviewed paperOption pricing using quantum computersStamatopoulos et al., Quantum, 2020
4. Peer-reviewed paperA threshold for quantum advantage in derivative pricingChakrabarti et al. (Goldman Sachs, IBM), Quantum, 2021
5. Peer-reviewed paperDerivative pricing using quantum signal processingNikitas Stamatopoulos and William J. Zeng, Quantum, 2024
6. Peer-reviewed paperFocus beyond quadratic speedups for error-corrected quantum advantage (perspective)Babbush et al. (Google), PRX Quantum, 2021
7. BookMonte Carlo Methods in Financial EngineeringPaul Glasserman, Springer, 2003
8. Peer-reviewed paperDynamic portfolio optimization with real datasets using quantum processors and quantum-inspired tensor networksMugel et al. (Multiverse Computing, BBVA), Physical Review Research, 2022
9. Peer-reviewed paperEnd-to-end resource analysis for quantum interior-point methods and portfolio optimizationDalzell et al., PRX Quantum, 2023
10. Peer-reviewed paperChallenges and opportunities in quantum optimization (Quantum Optimization Working Group)Abbas et al., Nature Reviews Physics, 2024
11. Expert commentaryRead the fine printScott Aaronson, Nature Physics, 2015
12. Peer-reviewed paperA quantum-inspired classical algorithm for recommendation systemsEwin Tang, ACM Symposium on Theory of Computing (STOC), 2019
13. Peer-reviewed paperMixed quantum-classical method for fraud detection with quantum feature selectionGrossi et al. (CERN, IBM, IRIS Analytics), IEEE Transactions on Quantum Engineering, 2022
14. PreprintFinancial risk management on a neutral atom quantum processorLeclerc et al. (Pasqal, Multiverse Computing, Crédit Agricole CIB), arXiv, 2022, revised 2024
15. Company announcementQuantum computing: two real-world experiments conducted by Crédit Agricole CIB with Pasqal and Multiverse ComputingCrédit Agricole press release, 2023
16. Company announcementHSBC demonstrates world's first-known quantum-enabled algorithmic trading with IBMHSBC media release, 2025
17. Peer-reviewed paperImproving the efficiency of payments systems using quantum computingMcMahon et al. (Bank of Canada), Management Science, 2024
18. Central bank working paperLiquidity optimization in gross settlement systems with quantum reordering: application to TARGET2 (Staff Working Paper 2026-32)Astuti et al. (Bank of Italy, Bank of Canada), 2026
19. Central bank discussion paperPayment coordination and liquidity efficiency in the new Canadian wholesale payments system (Staff Discussion Paper 2022-3)Francisco Rivadeneyra and Nellie Zhang, Bank of Canada, 2022
20. News reportIs quantum computing suitable for instant payments? (interview with Danica Marsden, Bank of Canada)SIX PAY Magazine, 2023
21. Peer-reviewed paperQuantum computing for finance (review)Herman et al., Nature Reviews Physics, 2023

← Applications