Quantitative analysts — “quants” — must compete fiercely in developing and deploying models that enable portfolio managers to better assess the value of financial assets, particularly derivatives, and to meticulously manage their operations and risks through continuous portfolio rebalancing. Yet these mathematics and finance specialists have generally received no training or introduction to artificial intelligence techniques — knowledge that could genuinely take their performance to the next level.
AI: the future of intelligent finance
What gains could be unlocked by leveraging cutting-edge AI technologies? The advantages of artificial intelligence techniques in the world of finance are already well known. Why, then, aren’t they being exploited to a greater extent?
Indeed, AI and Machine Learning can truly become the cornerstones of our financial intelligence systems. They are used in business process optimisation and decision-support applications, as well as in predictive analytics for risk assessment, for identifying intraday liquidity peaks, for detecting anomalies in cash flows, to name just a few. All of these AI systems involve massive data exploration and the continuous improvement of models. This is a process of ongoing refinement with vast potential yet to be explored. All the more so as evidence of the relevance of AI in finance is already there: hedge funds that have used artificial intelligence have reported returns of around triple those of conventional systems.
But what happens when an event disrupts the market in a way that could not have been anticipated? Is a fully autonomous programme capable of responding appropriately to a situation that was not foreseen at the time of its design?
Contrary to what one might assume, AI systems are far from incapable of adaptation.
The COVID-19 health crisis brought brutal economic growth to a halt across world economies. Despite this, European portfolios managed using AI limited the damage.
This is clear proof that not only is AI capable of adapting to unforeseen market developments, but that it can do so better than a conventional system.
Nevertheless, while this technology is undeniably powerful, it has not yet reached full maturity. Moving beyond Proof of Concept prototypes built by an R&D team requires sustained effort to enable confident adoption across the entire organisation and to guarantee the relevance of results.
There is never an immediate, easy win.
To begin glimpsing the full potential of AI, repeated testing and diligent research are necessary. Investment in people and infrastructure, constant commitment and patience are required to exploit the market and outperform. This research brings together teams of researchers, data scientists and investment experts to ensure that the technology is used — without being overexploited — to its full potential.
It is demanding, methodical work that must be carried out with intelligence — human intelligence, at that.
Knowing and exploring the potential of AI: this is a project that promises to be fruitful. But what are its concrete applications in quantitative finance? How can AI be put to work on one’s projects?
With any new technology comes the need to be able and to know how to use it. This requires carefully selecting the people and competencies you surround yourself with. Here as elsewhere, nothing is more effective than a multidisciplinary team. One might think to invest in a few data scientists or to train quants in AI. But the latter risk encountering a number of obstacles: the long learning curve, pressure from management to produce quick conclusions, and the complexity of making the right technological choices — all of which frequently leads to starting the R&D process with even less experienced interns. This process of AI exploration is observed across most players in finance, and it often leads to a shared conclusion: AI requires heavy investment and organisational transformation, AI is not suited to our specific challenges, AI is not for us.
At the organisational level, implementing AI in medium and large well-established structures is often more difficult, as it is a domain that straddles very different teams (quant and IT), which are often in separate reporting lines, and with a legacy system that frequently hinders the rapid adaptation of the organisation.
At the human level, the training of quants and managers does not yet lead them to an applied familiarity with artificial intelligence. It is high time for university curricula to adapt to today’s possibilities and tools such as MLOps and high-performance software development, as AI is an indispensable tool in the toolkit of a modern quant.
Then there is another challenge specific to the world of quantitative finance that must integrate AI: choosing the most relevant application for your business — the one that will make you money, optimise and accelerate processes, or reduce risks. But there will be obstacles to overcome.
The first obstacle is the lack of relevant data for the model. Certain financial products — such as illiquid products, OTC products and internal NMRF models within the FRTB framework — are difficult to price due to insufficient data to build reliable financial models. For these types of use cases, AI is a serious candidate for recreating robust risk indicators and pricing models, despite the scarcity of training data. Solutions exist to compensate for the lack of data, such as GANs (Generative Adversarial Networks), and simpler AI modelling techniques that work with small amounts of data, such as regressive models, Bayesian networks, and decision trees.
The second obstacle is the modelling of complex relationships between financial factors. For example, in the context of integrating ESG data into portfolio optimisation strategies and green investment, the data received from various providers is of variable quality, not readily usable by portfolio managers, and requires processing and refinement to extract useful information. A data scientist spends 80% of their time cleaning data. Even when this work is done perfectly, classical statistical metrics such as correlation fail to model complex relationships and exhibit biases and spurious correlations between data points. If, for example, a manager wishes to identify signals with an improved signal-to-noise ratio, it is possible to offer AI-based solutions using causal AI to improve signal quality. Causal AI is a new dimension of AI that enables the modelling of causal relationships between data using graphical models and allows counterfactual questions to be posed. Causal AI helps to improve investment decisions and alpha generation as well as portfolio optimisation.
The third obstacle is accelerating the computational performance of complex financial models and reducing infrastructure costs. Analytical solutions exist for pricing structured products (equity-based, bonds, ETF, rough volatility models, etc.) and risk calculation (xVA, Expected Shortfall, etc.), which are typically based on nested Monte Carlo calculations that are very costly in terms of time and hardware resources.
Furthermore, in high-frequency trading, the challenges are often the volume of data rather than the quantitative model itself. For example, 1 tick of data can require processing several tens of gigabytes of data in a matter of milliseconds. Investment hypotheses and strategies must be revalidated with each data change (in real time), which is not at all the challenge in traditional low-frequency portfolio management. This means the entire backtesting chain — Pricing/Modelling/Calibration/Filtering — must be re-run. And the more complex or exotic products a portfolio contains, the longer the processing takes.
AI can also massively accelerate analytical processing: this is the domain of Deep Pricing or Deep Hedging. ML/DL models are trained to approximate analytical computational functions. These models can calculate prices or volatility at speeds at least 100 times faster than analytical models, and will consume far fewer computational resources (CPUs). The key, however, is to guarantee robust and reliable models.
A residual lack of trust in the handling of financial models by AI may persist, and this must be taken into account. It is essential to be able to offer an explainable and transparent AI model, to which the portfolio manager, risk manager or data scientist has contributed deeply and directly. To achieve this, it is preferable that the model in question comply with certain constraints. In its design, it must combine quantitative logic with qualitative logic, integrate the researcher into the AI model design process, and ensure that the strategy is clearly explained to investors and regulatory authorities.
A world of “intelligent finance” making meaningful use of AI is within reach and full of potential.
Its complexity — mirroring that of finance itself — requires methodical preparation and the assembly of a solid, capable team ready to take on the challenge.
The specificities of the domain must be taken into account and are crucial to the success of the project. Effectively integrating all competencies in-house can only be achieved after establishing a collaboration with trusted, specialised players across the entire value chain (business consulting, infrastructure, specific data, and AI technology itself). This will require surrounding oneself with business and technology partners, data scientists, architects, and software engineers in order to quickly validate feasibility and the anticipated return on investment.
A multidisciplinary team — with the richness and complementarity of diverse skills — is therefore the key to successfully integrating AI within one’s organisation.