Insider Brief
- Oxford Quantum Circuits and Trust Base found that progress toward quantum advantage in finance will depend on strong classical comparisons, hybrid workflows and evaluations covering complete pricing and risk processes.
- Classical physics-informed neural networks offered the best balance of accuracy, speed and flexibility among the tested approaches, while hybrid quantum models remained slower despite promising results on some measures.
- OQC reported that future gains will depend on improved hardware, lower resource requirements and quantum methods that produce stable prices, sensitivities and risk calculations under real market conditions.
- Image: Photo by Adam Śmigielski on Unsplash
Research by Oxford Quantum Circuits and Trust Base suggests that finding quantum advantage in finance will require more than demonstrating that a quantum algorithm can perform a pricing calculation.
The companies examined where quantum methods could fit within financial workflows, how they compared with classical approaches and what would be required to make them operationally useful. The work covered derivative pricing, Value-at-Risk and Credit Valuation Adjustment, or CVA, according to a recent technical blog post published by OQC.
Trust Base is a digital subsidiary of Sumitomo Mitsui Trust Group. OQC, a British quantum computing company, supplied the quantum expertise for the project.
The partners compared classical Monte Carlo methods, physics-informed neural networks, hybrid quantum-classical neural networks and quantum-enhanced Monte Carlo techniques. Their findings did not establish quantum advantage, however, the project produced lessons about how financial institutions could search for applications as quantum hardware improves.
Among the methods tested, classical physics-informed neural networks currently provided the most practical combination of numerical quality, flexibility and runtime, OQC reported in the post. Hybrid models incorporating quantum circuits could perform competitively on some measures, but the cost of repeatedly executing and differentiating those circuits made the models considerably slower.
The results point toward a staged approach in which quantum processors perform selected tasks within largely classical systems. They also suggest that any claim of advantage must consider the entire financial workflow, rather than the performance of one algorithm or the accuracy of one price.
Accurate Prices Are Only the Beginning
Derivative pricing provided a useful testing ground because a model must produce more than the value of a financial contract at a selected moment.
Derivatives derive their value from another asset, interest rate or financial measure. Their prices can depend on how those underlying values move over time, whether they cross certain thresholds and how variables such as volatility change under different market conditions.
Simple products can be priced using established methods such as Black-Scholes models, finite-difference solvers and Monte Carlo simulations. The problems become more difficult with path-dependent products, whose value depends on the route an asset’s price takes; changing volatility, in which expected price swings vary over time or at different price levels; barriers, where crossing a preset level can activate or cancel a contract and nonlinear payoffs, which do not change at a constant rate as the underlying asset moves.
According to the blog post, OQC and Trust Base tested the methods on several common financial models covering stock and currency options, changing market volatility and interest-rate movements. These included the Black-Scholes, Garman-Kohlhagen, Dupire and Hull-White models. Several approaches produced the expected price for a selected example, but that result alone did not show whether they would perform consistently across wider market conditions.
The models were less consistent when tested across a wider range of market conditions, OQC reported. Targeted training could often improve delta, which measures how a derivative’s price changes as the value of its underlying asset moves.
Gamma, which measures how quickly a derivative’s sensitivity to market movements changes, proved harder to stabilize, according to the team. Even small inconsistencies in the model’s pricing estimates could produce much larger errors in this measure.
In the real world, this would be important because banks use these sensitivities to build hedging strategies and assess risk. A model that produces the correct price but unstable sensitivities would have limited value in an operational financial system. It would be like a weather forecast that gets today’s temperature right but cannot reliably predict how quickly conditions will change. Financial institutions need both to adjust hedges and manage risk as markets move.
OQC said models should consequently be evaluated across complete pricing surfaces — which means, loosely, the shape prices take as conditions change — and tested on sensitivities, exposure profiles, CVA and stress calculations. The lesson extends beyond quantum computing. A method can perform well on an isolated benchmark without producing the consistent behavior required by the larger workflow.
Finding a Place for Quantum Components
The project explored two broad routes toward quantum-enhanced financial computing.
The first involved physics-informed neural networks, or PINNs, which are models that learn how a financial product’s price should change under different market conditions by following the product’s underlying financial rules, including when and how it pays out.
Rather than calculate values across a fixed numerical grid, a PINN learns a continuous representation covering time and other market variables. Once trained, the model can be used repeatedly to estimate prices, sensitivities and exposures.
OQC and Trust Base also investigated quantum-compressed physics-informed neural networks, known as QPINNs. These hybrid models incorporate a parameterized quantum circuit that functions as a trainable feature map inside the wider neural network.
The objective was not to replace the entire pricing system with a quantum computer. The partners instead examined whether quantum-generated features could help a model represent a pricing surface more efficiently or reduce the number of trainable parameters required.
QPINNs could be competitive in parameter count and sometimes improved individual price or exposure calculations, according to OQC. Those results did not make the quantum models faster. Repeatedly evaluating the quantum circuit and calculating its derivatives substantially increased the total runtime.
That difference illustrates one of the central benchmarking problems in quantum machine learning. A model that uses fewer parameters is not necessarily more computationally efficient once a range of other factors — including circuit execution, gradient calculations, classical optimization and communication between quantum and classical systems — are included.
Under the project’s particular implementations and evaluation conditions, OQC recommended classical PINNs as the most practical of the tested methods. The company said they were comparatively fast, adaptable across different financial products and suitable for repeated calculations such as expected positive exposure and CVA.
OQC characterized QPINNs as a research platform for studying hybrid models and future quantum-classical interfaces, rather than a production-ready replacement for conventional solvers.
Financial Knowledge Improved the Models
The project also found that training strategy and financial constraints could matter more than increasing the size of a model.
Some portions of a pricing problem are harder to learn than others. These can include regions close to a derivative’s strike price, barriers that activate or terminate a contract, exercise boundaries, maturity dates and the edges of the modeled domain.
Adaptive sampling — which means directing more of the model’s training toward the areas where it is making the largest errors — concentrated more of the training process in those difficult regions. That approach had a material effect on model performance because it directed computational resources toward areas where the models had the most difficulty producing prices that followed the financial rules and limits built into the problem, according to the post.
The finding suggests that incorporating knowledge about a financial problem can produce greater improvements than indiscriminately adding model parameters. Quantum-generated features do not remove the need to understand the product being priced or the numerical behavior of the underlying equation.
It also reinforces the need for strong classical comparisons. Quantum methods should be tested against optimized Monte Carlo simulations, finite-difference techniques and classical neural networks. Otherwise, an apparent quantum improvement could result from comparing a specialized quantum method with a weak or poorly tuned classical alternative.
For banks, the relevant question is not simply whether a quantum model can solve a pricing equation. The test is whether the quantum component adds enough accuracy, speed or efficiency to an already capable classical system to justify its additional cost and complexity.
Hardware Will Help Determine the Advantage
OQC and Trust Base examined a second possible route through quantum amplitude estimation, a family of algorithms for estimating expected values.
Many financial calculations use Monte Carlo simulation to estimate an average across numerous possible future outcomes. Quantum amplitude estimation offers a theoretical route to achieving a desired precision with fewer samples, which could eventually produce an advantage for large pricing or risk problems.
OQC reported that practical performance depended heavily on circuit construction, compilation, qubit placement and hardware noise. In some experiments, less costly circuit choices produced smoother empirical behavior than more exact but deeper circuits.
The result illustrates why an algorithm’s theoretical scaling cannot by itself establish a practical advantage. A deeper circuit may lose its mathematical benefit if noise accumulates before the computation finishes.
OQC also presented fault-tolerant resource estimates to show how different hardware assumptions could affect the required system. Raising the assumed error-correction threshold from 1% to 5% could reduce the physical-qubit requirement for one 32-qubit configuration from approximately 404,000 to 130,000, according to the company.
Achieving that higher threshold could depend on hardware approaches such as dual-rail or erasure qubits. It is therefore an assumption about the capabilities of a future machine, rather than an improvement that can simply be applied to existing quantum processors.
Even under more favorable assumptions, reaching the required accuracy of about one part in 1,000 to one part in 10,000 ( to ) demanded more Grover-style repetitions than the researchers could include in their current classical resource estimates, OQC said.
OQC said future techniques, including quantum signal processing, could potentially reduce T-gate requirements by as much as 16 times and logical-qubit requirements by four times. Those figures are projections about possible resource reductions rather than performance gains demonstrated on a production financial workload.
