Quantum computing has long been discussed in capital markets as a technology with potentially transformative applications, but one whose practical impact remains some way off. Recent work by HSBC and IBM suggests the gap between quantum research and real-world trading may be starting to narrow.
In a study using production-scale European corporate bond trading data, researchers tested whether a quantum computer could improve the ability of machine learning models to estimate the probability that a dealer’s response to a request for quote (RFQ) would result in an executed trade.
The results were striking. Conventional approaches, using commonly used models, produced an average AUC (Area Under the Curve, a standard measure of classification performance) of around 0.63, while a quantum simulation running on conventional computing infrastructure scored around 0.60. But using features generated by actual quantum hardware, the strongest configuration achieved an average score of around 0.97, representing an improvement of up to approximately 34% over models using the original trading data.
“This was really unexpected. Very unexpected,” Dr Del Rajan, Vice President, Quantum Technologies at HSBC Group, tells TradingTech Insight. “What we’re left with is an experimental study to realise that, in this specific task, quantum computing is actually quite useful on this data set.”
Transforming the data, not replacing the algo
The significance of the experiment becomes clearer when viewed in terms of the trading problem HSBC was trying to solve.
In electronic corporate bond markets, an institutional client can send an RFQ to several dealers in what is effectively a blind auction. For the dealer, accurately estimating the probability that a quoted price will result in a trade – the fill probability – is an important input into the execution strategy.
“Any improvement in execution likelihood is a competitive advantage,” says Rajan. “The key thing the algo traders said was that it helps to improve margins, find an advantage, improve risk management via higher hit rates on preferred trades, avoidance of undesired transactions, and increased future deal flow.”
Importantly, HSBC and IBM were not attempting to replace an existing trading algorithm or machine learning model with a quantum equivalent.
Instead, quantum computing was applied to feature engineering, essentially transforming the trading data into a different representation before passing it into commonly used machine learning models.
The researchers deliberately kept the rest of the process broadly unchanged, testing the transformed data against four established machine learning approaches including logistic regression, XGBoost, random forest and neural networks. This allowed them to isolate the impact of the data transformation rather than simply comparing completely different modelling approaches.
“The quantum computing was only used for the data transformation,” Rajan explains. “We kept everything else the same.”
From experiment to trading architecture
That distinction is also important when considering how quantum computing might eventually fit into a trading technology stack.
Today’s quantum machines are typically accessed remotely rather than deployed alongside conventional trading infrastructure, making latency an obvious obstacle to placing them directly within an execution path. HSBC’s approach avoids that requirement.
“The quantum computer will be used as an offline component that generates these discriminative features that represent market states,” says Rajan. “And then we would use that for our models.”
HSBC is now exploring those engineering questions as part of a larger second phase of the project, including where quantum-generated features could sit within a live trading workflow and how frequently they might need to be refreshed.
Promising results, important caveats
For all the size of the observed improvement, however, neither HSBC nor IBM is presenting the experiment as proof that quantum computing has established a general advantage in algorithmic trading.
The underlying study used real trading data covering more than one million RFQs, primarily in European corporate bonds, although the detailed quantum experiments used a representative sample within a more confined trading period.
More fundamentally, the researchers do not yet fully understand why the quantum hardware produced such a substantial improvement. The same uplift did not appear when quantum processing was simulated on conventional hardware. Researchers also tested different quantum machines and experimented with artificially introducing noise into simulations, but were unable to reproduce the hardware result through those alternative approaches. The published paper consequently describes the findings as empirical and explicitly cautions against assuming they will generalise to other datasets or market environments.
“We still need to explore it under other market conditions,” says Rajan. “There are a lot of unknowns.”
That is now a key objective for phase two: determining whether the effect persists across different datasets and more complex trading situations, while addressing the practical engineering requirements of a live environment.
For trading technology teams, therefore, quantum remains an exploratory technology rather than the next mandatory component of the execution stack. But HSBC’s experiment provides something more tangible than a theoretical use case: evidence that today’s quantum hardware can be applied to a genuine trading problem using real market data.
Rajan’s advice is correspondingly pragmatic.
“Quantum computing is still a nascent technology,” he says. “Getting educated on the use cases where it has shown some improvements, and just being educated about the technology, is a way to keep yourself abreast of these developments.”
Subscribe to our newsletter
