Regulators and model-risk managers have been kept busy in recent years by rapid advances in machine-learning techniques that power banks’ modeling of everything from interest rates to customer fraud. If quantum computing continues its rapid march towards commercial viability, they might be about to get a whole lot busier.

Speaking yesterday to a panel of experts at Risk Live, hosted by WatersTechnology sister publication Risk.net, Lee Braine, managing director of research and engineering at Barclays, asked whether banks and other financial firms operating in “a highly regulated industry” could face particular governance and explainability issues with the adoption of quantum computing. This is already happening with models powered by machine learning, whose outputs can be difficult to trace back to data inputs in order to satisfy scrutiny.

“Is it similar to machine learning, in this regard, where we have [specialized] model risk management, for example?” Braine asked. “How would we go about addressing requirements for financial services? There are so many topics here—fairness, bias—but explainability I think will be key. What’s the path forward for topics such as explainability? Because even those with PhDs in this topic find it difficult to understand,” he added.

The lack of ready explainability around machine learning-based models—for instance, understanding how a model that relies on a particular approach to achieve its methodological speedups or reach its own intuitive conclusions based solely on the dataset that feeds it—is already proving a drag on adoption of newer ML techniques at some banks.

Model-risk managers may soon need to understand how such models have been additionally boosted by the power of quantum computing, said Stefan Woerner, quantum applications research and software lead at IBM Quantum, a developer of quantum computers.

“If we now move to a quantum model, that doesn’t help. That rather makes it more complex to understand what’s going on. The research needs to also investigate explainable quantum machine learning,” he said during the panel conversation.

Woerner suggested quantum models based on Monte Carlo simulations, for example, might qualify more easily for regulatory approval, as their classical-based counterparts are already compliant and approved—but added explainability could be less straightforward for models that rely on quantum-powered machine-learning approaches.

The problem could loom sooner than many think, he added: quantum advantage could be demonstrated in machine learning within the next five years, the firm’s own research suggests, with IBM last year publishing a proof demonstrating an artificial classification problem could be solved with a quantum speed-up.

What’s the path forward for topics such as explainability? Because even those with PhDs in this topic find it difficult to understand

Lee Braine, Barclays

Quantum advantage is defined as a quantum machine doing something a classical computer cannot. Super-fast quantum computers should be able to achieve that by their exploitation of quantum mechanics. A bit, or unit of information, in a classical computer can be in one of two states: it can store either a one or a zero. But a quantum bit, or qubit, can exist in both states simultaneously, meaning the computers can theoretically handle a far higher number of calculations.

Embryonic quantum computers are already being rented by banks to perform tasks such as complex, multivariate analysis, while high-frequency traders are using quantum-inspired shortcuts to exploit arbitrage opportunities.

But the advances will mean regulators also need to get up to speed. Rory McLaren, technology strategist at Deutsche Börse, said during the panel that a partnership is required between the industry and regulators to explain to supervisors—in a “non-quantum physicist manner”—the advances and applications quantum computing is being used for.

Deutsche Börse unveiled a quantum computing initiative in March, which it hopes could see the technology used to calculate proprietary risks faced by the exchange within the next three years.

In some cases, mathematical proofs can be presented to regulators, said McLaren. “In others, we will have to explain the process that we’ve gone through in terms of explainability to give the regulator a flavor on how they want to control and manage that.”

That will enable regulators “to get an early insight, so they can plan how to incorporate it into regulations,” he said.

Often, regulations are defined in terms of outcomes given from a known state of technology, he said. “When we introduce a new way, the actual facets of the regulation or the guidelines might not make sense when applied to that new probabilistic nature.”

Barclays’ Braine asked panelists whether the “non-determinism” inherent in a lot of quantum computing was a particular challenge. A non-deterministic algorithm can—even for the same input—exhibit different behaviors on subsequent runs.

Zapata Computing chief executive Christopher Savoie argued that while quantum computers exist in “a kind of limbo state” as they gear up to do calculations, their bits do ultimately “collapse to either a zero or one,” making the outputs of the hardware mathematically explainable to regulators.

Explainability instead becomes an issue when heuristics are used in machine learning, he says—but that can happen with either classical or quantum computing methods, “as this involves some level of guessing as to what the answer is.” Heuristic techniques use short-cuts to solve problems by finding an approximate solution that is not guaranteed to be optimal.

US-based software developer Zapata is applying quantum software to the US Federal Reserve’s annual stress test of large banks, the Comprehensive Capital Analysis and Review, and to credit valuation adjustment calculations for banks.

Savoie adds that, on a basic level, it is important to show regulators a workflow system that explains where data was obtained and how it was cleaned “in very non-quantum language.”