AI in financial markets: from trade surveillance to pre-trade revolution

This content was originally published by The AI Journal

Artificial intelligence (AI) has quickly become an unavoidable topic in conversations across financial markets. The power of AI and machine learning (ML) models to revolutionise the way markets operate is undeniable: from facilitating trade automation to spotting cases of market abuse and insider trading, the potential to use this new technology is no longer confined to the realm of possibility. AI and ML-powered solutions are fast becoming a reality.

Part of the furniture

A recent report from the Bank of England shows that, within UK financial services, 75 percent of firms already use AI. A further 10 percent plan to do so in the next three years. Moreover, 55 percent of all AI use cases involve some degree of automated decision-making. The same report showed that a third of all AI use cases in UK markets are third-party implementations, a sharp increase from the 17 percent reported in a 2022 survey.

The report’s findings reflect a global trend that AI and automation are increasingly becoming a part of the furniture for firms operating within the financial system. As vendors and solutions providers become more confident in AI’s capabilities and the value of its integration, AI and ML models are likely to become an established part of solutions and wider offerings. There are three notable areas where ION has tested AI and ML models being productive across financial markets:

• Enhancing trade surveillance capabilities to ensure market abuse is highlighted and addressed;
• Streamlining pre-trade calculations to provide faster risk evaluation; and
• Supporting more precise instrument pricing to achieve more accurate risk modelling.

Trade surveillance: a helping hand, or a helping eye?

AI and ML models can process high volumes of data efficiently, which is important in streamlining and improving trade surveillance and risk management processes. While many market players have already adopted this technology successfully, it is important to factor in AI and ML’s ever-evolving capabilities. Compliance teams have found success in using AI for analytics and pattern recognition, as it enables them to classify alerts effectively and quickly. It also helps identify subtle and complex market abuse tactics that may elude traditional surveillance methods.

Furthermore, AI-driven trade surveillance technologies are scalable and have low latency capabilities. These features have proved invaluable for large organisations that handle high and increasing trade volumes. Automated surveillance systems can handle growing amounts of data and transactions without compromising efficiency, while low latency ensures that alerts are generated and analysed in real time. The traditional ‘bulk’ analysis method, which sees compliance teams analysing sets of alerts at a time, often causes delays with a greater chance of inaccuracies. AI-powered systems reduce the staff necessary and allow for real-time analysis, enabling more timely responses to potential violations.

Regarding regulation, market participants are already exploring ways that AI can be used. New technology could help firms and regulators align on constantly evolving approaches. In March, the US Securities and Exchange Commission (SEC) acknowledged the potential of AI to manage risk across financial services during a roundtable in Washington DC. The panel discussed the inherent risks and opportunities in using AI in market innovation, bringing together regulators, market participants, legal experts, and technology providers. With the growing market volatility in recent months, AI and ML have been highlighted as potential solutions to help navigate the associated risks.

Transparency and oversight still pose significant barriers to the adoption of AI systems. But, regulators also recognise the potential of AI to enhance surveillance capabilities, mostly due to its more thorough and efficient monitoring of market activities. In the face of growing unpredictability, there is no denying that the ability to adapt quickly and effectively will be key in equipping stakeholders to navigate these challenges. AI and ML technology offer an obvious solution, and confidence is building with the greater transparency and explainability that has come with new approaches to using this technology in trade surveillance. Given the growing number of regulators advocating for this to be explored, the markets may not have to wait long until this becomes a reality.

A pre-trade revolution?

But the uses of AI and ML in financial markets don’t stop there. With AI models becoming more sophisticated, opportunities to apply them to pre-trade calculations will also become increasingly viable. The capability of these models to process large volumes of data at high speed makes them ideal for such uses. They can be used to calculate margins, support the selection of pre-execution counterparties in a Request for Quotation (RFQ), or decide when to execute orders, informing traders’ and algos’ decision-making and risk control. Integrating AI and ML into these models will not be a straightforward process, but could revolutionise pre-trade workflows.

When using AI and neural networks specifically, it is important that they are trained in the right way. From a risk perspective, if pre-trade approximations undershoot significantly, they could leave firms exposed to insufficient coverage. Solutions such as CME’s SPAN 2 calculations and SPAN 2 approximations have worked to address these concerns, but continue to face issues when combining speed and accuracy. From our recent integrations of AI within ION’s pre-trade solutions, we have learned that it is possible to overcome these barriers using neural networks: they can be trained to err on the side of overshooting, while still aiming to be as close as possible with approximation calculations.

AI can be effective in training models based on historical information. For example, they use past decisions to inform the selection of counterparties for RFQs, and in identifying prime execution timing to reduce market impact and maximise profit. Going beyond this, the trained machine can also be integrated with trading tools and algos to improve the effectiveness of the decision-making process. At ION, we have successfully integrated supervised learning techniques with our algos to optimize the execution of large orders, reducing the complexity of configuration. This approach improves scalability, increasing the number of large orders that a single trader can manage simultaneously. By nature, AI-based tools absorb information to learn and improve continuously. This ability means that they can achieve results comparable to the most experienced of traders.

Though time-consuming, training and testing are crucial for the successful and effective integration of AI and ML processes. Solution providers and market participants alike must ensure that models can process high volumes of complex data within a multitude of market scenarios before financial systems become dependent upon them.

Accurate pricing models

AI models have proved useful in increasing the accuracy of pricing financial instruments, especially when modelling illiquid assets.

In the case of implied volatility (IV), this is particularly interesting. Neural networks excel at finding hidden patterns in data and can be used to improve the accuracy of risk calculations. This is important when managing implied volatility, as deviations from market data within calculations can lead to significant pricing errors. Traditional implied volatility models, notably the Surface Stochastic Volatility Inspired (SSVI) model, often fail to capture real market behaviour. The simplicity of such models’ parameters means they struggle to accommodate complex market data. Imagine trying to paint a detailed landscape with only four colours: models like SSVI get the broad strokes right, but miss the finer details.

This is where ML models come in. Ackerer et al’s 2019 report introduced a new approach combining neural networks with the SSVI model. This hybrid corrects SSVI’s predictions where they deviate from market data, while preserving key advantages of these models, such as arbitrage-free conditions. The capability of neural networks to process complex market data has been gradually developed in recent years. The ION team ran its own empirical study over 234 different volatility surfaces across multiple underlyings and dates: across all of these scenarios, the mean absolute percentage error obtained from the neural network did not exceed 4 percent.

Looking ahead

If AI and ML can retain these levels of accuracy throughout testing, it seems inevitable that their integration in pre-trade calculations and solutions will become increasingly common over the years to come. It is equally likely that models will continue to enhance trade surveillance and compliance efforts in the face of changing regulations and a growing trend of market volatility and unpredictability.

The opportunities posed by technology’s new capabilities will be central in optimising pre- and post-trade processes. However, rigorous testing and training of these models is paramount if market players are to maintain forward progress, rather than compromising existing systems by running before they can walk.