CC BY
57AI-BASED TOOLS IN EUROPEAN SECURITIES MARKETS: AN OVERVIEW OF ADOPTION, USE CASES, FINDINGS, AND IMPLICATIONS
ALEXANDER APOSTOLOV Ph.D - Economic Research Institute at BAS, Sofia, Bulgaria
Abstract: This article provides an overview of the current application of artificial intelligence (AI) in European Union (EU) securities markets and assesses the degree of adoption of AI-based tools. It highlights that while an increasing number of asset managers leverage AI in investment strategies, risk management, and compliance, only a few of them have developed fully AI-based investment processes and publicly promote the use of AI. Similarly, in other parts of the market, such as credit rating agencies, proxy advisory firms, and financial market participants, AI is used to enhance information sourcing and data analysis. However, the overall adoption of AI is not leading to a fast and disruptive overhaul of business processes. The report concludes that while AI is increasingly being adopted to support and optimize certain activities, there are risks associated with its widespread use, such as the concentration of systems and models among a few "big players." Therefore, further attention and monitoring are required to ensure that AI developments and the related potential risks are well understood and taken into account. The report is based on an analysis of information collected by the European Securities and Markets Authority (ESMA) through several channels between April and November 2022, including interviews with securities market participants in the EU and written questionnaires sent to entities directly supervised by ESMA.
Keywords: Artificial Intelligence, Exchange-traded financial assets, Non-bank financial institutions
JEL classification: C45, D83, G17, G17, G23
Introduction
In recent years, the financial sector has been confronted with the increasing availability of vast amounts of data from a variety of sources, accompanied by a steady increase in computing power. These conditions have created a favorable environment for the use of artificial intelligence (AI) and machine learning (ML) technologies. AI and ML are transformative technologies that can be used to harness data through advanced statistical techniques and computer science. Policymakers, regulators, and supervisors around the world are paying increasing attention to how AI techniques are applied in the financial services sector and are assessing the potential risks associated with them. However, detailed data on the recent use of AI in European financial markets is limited. This paper therefore aims to explore the common applications of AI currently used by entities operating in different sectors of the EU securities markets. The article is aligned with the Digital Finance Strategy adopted by the European Commission in September 2020, which highlights the responsible use of AI tools in the financial sector. Even then, the Commission recognized that digital innovation in the financial sector has inherent risks that need to be monitored to ensure consumer protection, financial stability, and an orderly digital transformation. Since the advent of ChatGPT and the wave of innovations and competing products based on AI, the situation has become even more complicated. This article aims to provide insight into the interaction between industry practices and the regulatory framework.
1. Integrating natural language processing into portfolio management investment
strategies in asset management companies: An AI-based approach
Portfolio managers use a range of techniques that fall under the category of artificial intelligence. AI tools can be employed by quantitative funds as part of systematic investing strategies and by discretionary portfolio managers to improve fundamental analysis (e.g., to optimize portfolio construction by estimating the structure of dependence among financial assets). There are considerable differences in how these strategies are incorporated into the investing process and how
much AI may influence the resulting investment decisions. AI seems to be mostly employed as a tool to carry out specialized activities that make use of vast volumes of data. In reality1, a rising number of funds are adopting AI in some capacity due to its efficiency in extracting information from a variety of massive numerical and textual databases with minimum human supervision. The increasing incorporation of alternative and unstructured datasets among the information sources important to the identification of investment possibilities is a distinguishing feature of AI2 in comparison to more conventional approaches for technical or fundamental analysis. In this context, NLP techniques, which extract economically valuable information from a variety of text sources, appear to be particularly well-liked by AI users. NLP is frequently used to find important news stories and create sentiment indicators. The identification and evaluation of environmental, social, and governance (ESG) disclosures is one application that industry professionals frequently bring up. Market research reveals that investment professionals create real-time ESG assessments based on business communications, such as corporate social responsibility reports, using natural language processing (NLP). Several portfolio construction experts are promoting solutions that meet the evolving business needs of institutional investors with regard to ESG. According to comments from industry professionals, AI does not appear to be fundamentally altering the way that portfolio managers make investments. It's known that only a small number of funds have created fully end-to-end AI-based investing processes. Some industry insiders claim that asset managers that employ sophisticated AI techniques are often specialized hedge funds led by analysts with solid expertise in ML who bank on AI as a selling point. When AI (and especially ML) is applied to systematic investing, it might be challenging to determine if a certain method is more effective than others. As these models have been proven to perform best, neural networks are probably a common strategy, although "ensemble" approaches that integrate a variety of different machine learning techniques have been demonstrated to yield better predictions than any single ML methodology (see, for example, Borghi and De Rossi, 2020). In conclusion, adding machine learning to the investment process shouldn't be viewed as a simple, automated solution to boost fund performance. According to the CFA Institute (2021), in
1 It is difficult to precisely quantify how many asset managers employ AI. Recent research indicates that 32% of respondents in the entire financial services industry use natural language text understanding (Zhang et al., 2022, p. 162), whereas 22% of EMEA finance professionals use big data analysis and ML techniques "to conduct market research that leads to investment decisions" (CFA Institute, 2020). In a global survey conducted in March 2019 (CFA Institute, 2019), 31% of portfolio managers reported using at least one of the ten listed AI techniques for creating trading algorithms, and 10% reported using AI/ML techniques in the investment strategy "to find a nonlinear relationship or estimate" (a narrower interpretation that is likely to apply primarily to systematic strategies). For technologically adept entities, such as quantitative investment funds, these numbers are likely to be an underestimate. CFA Institute (2019) argues that discretionary managers who conduct fundamental analysis also use unstructured and alternative data. 25% of a distinct sample of equity and credit analysts reported using AI or ML for industry and company analysis, while 56% reported using unstructured and/or alternative data. Linciano et al. (2022) discovered that, of the eight asset management firms comprising 60% of the Italian market, seven use AI systems in some aspects of their business, and three have fully incorporated AI systems to optimize the investment process. It should be noted that, as FSB (2017) and OECD (2021) emphasize, there is not always a common definition or understanding of what is included within the concept of an AI system, and firms are typically reluctant to disclose specific details about their investment process.
2 Recent years have seen the development of distinct, but largely compatible, definitions of AI. The European Commission defines artificial intelligence (AI) as software that "can generate outputs such as content, predictions, recommendations, or decisions affecting the environments they interact with" and is developed using one of a number of techniques, including machine learning (ML), logic- and knowledge-based approaches, and statistical methods. The European Council (2022) employs a similar definition, adding that an AI system "is designed to operate with elements of autonomy" and "infers how to accomplish a given set of objectives." AI is defined by IOSCO (2021) as "the science and engineering of creating intelligent machines, or more simply, the study of methods to make computers emulate human problemsolving decisions." AI is defined by the OECD in 2021 as "machine-based systems with varying levels of autonomy that can make predictions, recommendations, or decisions for a given set of human-defined objectives." The FSB (2017) characterizes it as "the theory and development of computer systems capable of performing duties that have traditionally required human intelligence."
order for an ML-based investing approach to produce meaningful outcomes, it is essential to hire individuals with specialized technical knowledge. Data is also essential for the effective use of AI. Financial data time series are frequently mentioned as being short compared to cross sections or, in the case of high-frequency data, as having a poor signal-to-noise ratio. These differ from the data that is normally utilized in successful ML applications outside of banking due to these properties. Several of the AI models now in use are better suited for short-term frameworks than long-term decision making due to structural breakdowns in time series or "regime changes," which may restrict the use of this data to produce economic projections. Notwithstanding these obstacles, research is being conducted at several leading organizations using AI in the investing process with the goal of creating ML models that take into account shifting regulatory environments in the financial markets. AI has the potential to form the foundation of portfolio risk management models in addition to assisting fundamental and technical analysis. By monitoring portfolio managers' actions, automating the execution of quality reports, and determining the risk to market liquidity, certain hedge funds and asset managers are automating risk management and compliance procedures (see IOSCO, 2021). Early warning systems can potentially employ AI approaches to anticipate market volatility and financial disasters (see Bartram et al., 2020).
2. The market share and performance of investment funds employing artificial intelligence
and machine learning in the EU
Although traditional investment funds are increasingly interested in utilizing innovative AI tools, their actual use appears to be hampered by technological and knowledge barriers, particularly among smaller asset managers, as well as conflicting client feedback. In an industry where artificial intelligence is not yet widely accepted, the perceived risks of a "black box" and the difficulty explaining negative outcomes may deter some investors. To cast additional light on this topic, we use ESMA's data to determine how many funds disclose to investors that they employ AI or ML tools. This was accomplished by analyzing the universe of open-ended investment funds covered by two financial data suppliers. Specifically, we screened approximately 145,000 financial documents issued by investment funds domiciled in the EU (including prospectuses, key investor information documents (KIIDs), shareholder reports, factsheets, etc.), covering at least 22,000 funds, using text-mining techniques. Then, we examined all documents containing the terms "artificial intelligence" or "machine learning" and identified the funds that mentioned incorporating AI or ML into their investment process. To supplement this search, we inspected all funds recorded in either the Morningstar Direct or Refinitiv Lipper databases whose names contain the terms "artificial intelligence" or "machine learning" or their abbreviations, and - for funds not already included in our sample - similarly identified those entities that stated that AI or ML underpin their investment process (as opposed to, for example, funds that invest predominantly in companies developing AI technologies). The results of this exercise indicated that the majority of investment funds do not advertise their use of artificial intelligence (AI): we identified 65 funds (offered by 40 distinct fund management firms) that indicated using AI (or, more specifically, ML) in their investment strategies. Sixty percent of these companies offer share classes to retail investors. In the past five years, the number of existing funds has increased fivefold to 54 entities as of 3Q22, with 29 equity funds, 11 alternative asset funds, 10 mixed asset funds, one bond fund, and one real estate fund. Additionally, eleven funds have been liquidated. Despite the increase in the number of funds disclosing the use of artificial intelligence, their market share remains negligible: less than 0.2% of the number of undertakings for collective investment in transferable securities (UCITS) funds in the EU and around 0.03% of UCITS' assets under management. In five of the last eight quarters, there have been net outflows from funds employing artificial intelligence. In contrast, investor outflows from other EU investment funds have only been negative in the two most recent periods. In a subsequent analysis, we identified all funds that mention NLP and found only nine additional entities that reported employing similar investment methods. Overall, these numbers appear low, particularly considering that the definition of artificial intelligence can be expansive and there is no legally binding definition
ОФ "Международный научно-исследовательский центр "Endless Light in Science"
that would prevent fund management firms from advertising its use (for example, to avoid the risk of making false claims and attracting regulatory oversight). The decision to omit references to artificial intelligence may be explained by the fact that, as stated in the preceding section, many funds use AI to execute limited steps of the decision-making process that have no immediate impact on the investment strategy and investment policy, which are typically the focus of investor disclosure documents. Nonetheless, some companies' discretion may be influenced by the market conditions they confront. Consistent with anecdotal evidence from market participants that some investors may associate AI with a lack of transparency or accountability, investment firms may be wary of the reputational ramifications of promoting the use of AI explicitly. Consequently, our numbers may be an underestimate of the number of funds that could credibly assert that AI aids their decision-making processes. However, they are likely to acquire the majority of those specialized funds that adopt a systematic investment strategy and rely heavily on machine learning-based models. This is more likely to be the focal point of their marketing strategy for these funds. In fact, 35 of the 65 identified funds include AI and/or ML in their names, making these concepts central to their value proposition. These vehicles may have been created by asset management firms with the intention of catering to a clientele that prefers innovative and sophisticated investment instruments. To further evaluate the effectiveness of AI-powered funds, we investigated whether these techniques enabled them to outperform their competitors. The average returns and alphas of funds using AI during the three years preceding October 2022 were not substantially different from those of funds whose investment strategies did not mention the use of AI. In addition, we conducted additional data analyses that controlled for other characteristics in order to compare the performance of AI-powered funds to that of similarly situated funds. There was no statistically significant distinction between the two categories. Lastly, we compared the expense ratios of AI-powered funds to those of their counterparts and found no significant difference in either direction. Therefore, we were unable to discover evidence that AI enables lower fees by containing costs or that it is used as a selling point to charge investors excessive fees. In accordance with the apparent absence of a strong demand intrinsic to AI or any "hype" surrounding its use that explicitly advertises it, the technology may not have yet consistently translated into superior returns for fund investors.3 The use of AI and ML in investment management offers the possibility of efficient investment decisions and, if implemented on a larger scale, the possibility of a reduction in fund operating expenses over time. Nevertheless, based on the main findings, the technology may not yet have consistently resulted in superior outcomes for fund investors for those funds that promote it explicitly.
3. Cloud-based systems and ESG-related data: Key tools for AI-driven asset management The rise of AI-native software companies that provide institutional investors with services in one or more domains, such as portfolio management, risk management, and compliance, is a major trend in the European asset management sector. Tools for data anomaly detection, automatic reporting, and the automated creation of legal documents such as fund prospectuses, PRIIPs KIDs, and the European ESG Template are all examples of AI-based compliance solutions. These use cases serve as illustrations of the revolutionary potential of RegTech (the use of technology to improve compliance and regulatory procedures). Many companies in the compliance space recognized the
3 In recent years, numerous active exchange-traded funds (ETFs) claiming to employ AI-driven strategies have been introduced in the United States. Bartram et al. (2021) identified 11 active ETFs whose investment process is powered by AI/ML by performing a systematic search on all active ETFs traded on the US market, albeit with a limited scope that included only fund descriptions and the Reuters/Refinitiv newsfeed. These funds contain less than one percent of the total assets managed by active ETFs, but have experienced significant growth. After taking into account the portfolio's exposure to a standard set of style determinants, they do not outperform active ETFs by a significant margin. Previously, Rabener (2019) estimated that AI-powered ETFs had not attained a significant scale by the end of 2019, with total assets under management estimated to be around $100 million. Boyde (2021) identified six ETFs enabled by AI from two management firms with a total of USD 270 million in assets more recently.
regulatory roadmap as the primary driver influencing demand for their services, but they also saw an autonomous push toward corporate transformation and digitalization. Moreover, technologies designed to provide data and signals for customers to utilize as input for quantitative or discretionary investing strategies often center on ESG-related data. Cloud-based systems are often used to provide the services. It's interesting to note that these companies seem to make their usage of AI their key differentiator. This approach contrasts with the comparatively limited use of AI as a branding tool among investment funds, despite it being difficult to measure the performance and market share of these enterprises. The majority of the companies in the poll claimed to provide customized solutions, giving customers varied levels of control. For instance, depending on the use case and the client's requirements, the data utilized is either given by the client or acquired internally. Some businesses provide their customers with the required product in addition to analysis or tools designed expressly for analyzing it. In fact, there is general agreement among company representatives that the success of their enterprises depends on offering "explainable" tools - that is, tools that their customers can understand regardless of how sophisticated the models used are. As some firm executives noted, one reason for this is that clients are more willing to accept the results of AI when it is presented as making "recommendations" rather than "decisions," because they are frequently wary of the role that AI plays in the asset management industry, which they often associate with excessive automation of the decision-making process. Executives from the company frequently showed knowledge of the value of giving their customers access to and control over the models, data, and procedures used. On the one hand, these third-party providers perform a useful role by giving less knowledgeable market players access to more effective instruments, which should eventually increase market efficiency and provide economic surplus on a systemic level. On the other hand, it can be argued that outsourcing important aspects of a company's core operations carries some risks, particularly if these tasks are given to organizations with potentially opaque systems. This would make it more difficult for clients to monitor and manage the models, data, and procedures used. These dangers may, however, be reduced by appropriate and effective outsourcing procedures and are not fundamentally specific to the employment of AI.
4. AI applications in trading: An overview of the value chain
AI is widely used by various entities involved in the trading process across the value chain. The trade lifecycle is typically divided into three phases: pre-trade analysis, trade execution, and posttrading. In the pre-trade analysis phase, AI models can aid asset managers and other investors in assessing the characteristics of financial assets to identify investment opportunities before executing a trade. In contrast, high-frequency traders rely on algorithmic trading strategies that both accept and execute investment decisions. AI can also be utilized in specific trade execution algorithms to optimize the costs involved in executing a trade, especially for large orders, by minimizing its market impact. Lastly, AI models can optimize post-trade processing by efficiently allocating liquidity during the settlement cycle. This article will explore use cases of AI successively in pre-trade analysis and investment decision algorithms, trade execution, and post-trade processing.
4.1. Pre-trade market analysis
Prior to executing a transaction, investors utilize AI models to analyze asset price signals and identify investment opportunities. This pre-trade analysis can either be evaluated by a human decision-maker (as is typically the case in investment funds) or be incorporated into algorithmic trading strategies designed to make and execute investment decisions. Typically, high-frequency traders engaged in market making and arbitrage, as well as proprietary traders, quantitative hedge funds, and other buy-side investors, use these investment decision algorithms. According to FMSB (2020), the majority of algorithmic trading conducted by banks and significant non-bank market makers is still based on relatively transparent rules-based models. According to other recent accounts, many large proprietary trading firms have incorporated machine learning (ML) models into their trading algorithms, albeit primarily in the form of supervised learning with ongoing research into reinforcement learning. Where it is deployed, ML is predominantly used to trade liquid instruments,
ОФ "Международный научно-исследовательский центр "Endless Light in Science"
such as equities, futures, and foreign exchange instruments, for which abundant and timely data is available. AI can support trading algorithms with the goal of minimizing the market impact of transactions; the use of AI for this purpose is discussed in detail in the section on trade execution that follows. Regarding the use of AI in securities pricing algorithms, models that optimize hedging and quotation decisions are examples. Some brokers use machine learning (ML) models fed with client-related historical data (such as the historical "hit ratio," which defines their relationship with the client), to automate their response to client's requests for quotes, optimizing their price and the likelihood of the client's acceptance. The pricing of securities lending transactions is an additional application of AI that holds promise. Each day, securities lenders must respond to thousands of inquiries regarding the available inventory of securities for short sale. To remain competitive, they must promptly respond to price inquiries despite substantial uncertainty regarding the demand and supply of securities. A growing number of lenders are employing artificial intelligence to address two problems: determining optimal securities lending prices and predicting which securities will become "hard-to-borrow" (HTB) securities. Some utilize models such as random forest and polynomial regression for the first objective, supplying them with a large number of variables that reflect, among other things, market capitalization, utilization, duration, and convexity. In order to address the second issue, some lenders use supervised clustering algorithms, such as the k-nearest neighbors, to predict the HTB status of a security based on the similarity of its features to those of other securities whose HTB statuses are already known. In addition, some lenders are investigating the possibility of using NLP techniques to process inquiries from prospective borrowers, thereby outsourcing the beginning of the negotiation process.
4.2. Trade execution
In the trade execution phase, AI finds some of its most promising applications: in completing an order, a broker endeavors to minimize the most significant transaction cost, which is the market impact. It has become crucial for investment banks and other brokers operating low-margin businesses to accurately predict market impacts. However, this quantity is notoriously difficult to estimate, particularly for less liquid securities, for which analogous historical trade data is scarce (see FSB, 2017). In addition, these market effects are nonlinear, meaning that they tend to be disproportionately greater for meta orders (i.e., large trading orders that are typically divided into multiple "child orders" and filled over the course of several business days). Some brokers and significant buy-side investors, such as pension funds and hedge funds, have developed machine learning (ML) algorithms to divide and execute metaorders optimally across various trading venues and periods in order to minimize their market impact and, consequently, transaction costs. This optimization task lends itself well to reinforcement learning, which can be used to determine the optimal size and execution time of a metaorder's offspring orders. However, a significant obstacle for these models is the dearth of specific data on metaorders, which are only held by the entity fulfilling the order. This has prompted brokers to develop models that are trained on a limited data set and, as a result, have limited utility. There are ongoing efforts to combine data, despite the fact that these initiatives are subject to privacy concerns. Some asset managers are considering transforming the aggregated data using techniques such as principal component analysis or synthetic data. However, some of these techniques may reduce the explainability of the models, making it more difficult to determine the effect of each variable on the outcome. Lastly, market participants appear to rely increasingly on nonparametric models, which may capture the nonlinear effects of large transactions on the market more accurately. Deep learning models such as neural networks and Bayesian neural networks outperform parametric models such as the I-Star model, according to industry practitioners.
4.3. Post-trade processing
Post-trade processing comprises the reporting, clearing, and settlement of a transaction. In this context, ML methods are used by some central securities depositories (CSDs) and brokers to predict the probability of a trade not being settled given the resources allocated to it so as to optimally distribute said resources (namely liquidity). The "failed" or "successful" label assigned to each past
ОФ "Международный научно-исследовательский центр "Endless Light in Science"
trade of a client provides an ideal supervised learning set-up. Models are thus calibrated on past, labeled, client-specific data in a supervised learning fashion, whereas new clients, for whom past data is unavailable, can be assigned to clusters of existing clients through unsupervised learning techniques. Against the backdrop of cash penalties imposed on settlement failures envisaged under the CSDR, research on ML models to improve post-trade processes is likely to bring substantial advantages to market participants and increase settlement efficiency. Notwithstanding the nascent applications of ML in settlement activities, feedback received from central clearing counterparties (CCPs) and CSDs surveyed by ESMA suggests that, at least for the time being, most of these entities are not widely using AI. Indeed, investment in AI seems rather limited currently concerning CCPs, most of which argue that there is still limited additional value in the adoption of AI. The majority of CSD representatives noted that although they are largely still using technology infrastructures that were created over a long period of time, they have plans to increase their use of AI in the near future. Finally, some data reporting service providers (DRSPs) and trade repositories (TRs) have either deployed or started to develop AI solutions (based on ML models or NLP) for anomaly detection, data verification, data quality checks, and automated data extraction from unstructured documents. For these purposes, DRSPs and TRs tend to turn to cloud services offered by third-party providers. When asked about the main advantages of using AI, these entities stated that they expect improvements in terms of efficiency and accuracy, as well as facilitating decision-making.
5. Data quality and model risk in the era of AI-driven securities markets The increasing prevalence of AI in the financial system is typically associated with a variety of potential risks. The following are some of the most significant hazards associated with AI in the context of securities markets: - explicability; - concentration, interdependence, and systemic risk; -algorithmic bias; - operational risk; - data quality and model risk. The majority of these hazards are not inherent to AI-branded models or algorithms. However, they can be amplified through the use of AI, as AI systems typically operate at a greater scale, complexity, and degree of automation than conventional statistical tools. Explainability is arguably one of the most distinctive AI risk factors, especially for certain ML models. The inability to articulate an AI model has the potential to hinder model performance and risk management. A growing adoption of AI in the securities markets poses systemic concentration risks. As making considerable advances in the development of AI systems is resource-intensive, a number of observers believe that entry barriers may emerge and lead to outsourcing by the few large asset managers who can invest in technology, data, infrastructure, and talent. However, it should be noted that concentration and interconnectedness risks arising from the dominance of certain providers apply to the broader digital financial services sector, as highlighted in the European Supervisory Authorities' Advice on Digital Finance (ESAs, 2022). Prenio and Yong (2021) warn that an excessive reliance on third-party service providers may also result in commercial capture and dependency risk. According to a number of studies, the concentration of AI tools among a small number of systemically significant providers may increase systemic risk, particularly in the context of algorithmic trading,4 by inducing herding behavior, convergence of investment strategies, and uncontrolled chain reactions that exacerbate volatility during shocks. In a similar vein, FSB (2017) observes that correlated risks posed by numerous financial market participants employing
4 Article 4(1)(39) of MiFID II defines algorithmic trading as "trading in financial instruments where a computer algorithm automatically determines individual parameters of orders, such as whether to initiate the order, the timing, price, or quantity of the order, or how to manage the order after it has been submitted, with limited or no human intervention." WEF (2019) argues that off-the-shelf algorithms may converge towards a single view of the market, thereby fueling asset bubbles and amplifying market disruptions. Tomanek (2021) argues that the UCITS Directive and the AIFMD, unlike the applicable provisions of MiFID II, do not require funds to implement instruments such as automated volatility interrupts or emergency interrupt schemes, which could effectively limit or prevent chain reactions. Consult Directive 2009/65/EC of the European Parliament and of the Council of July 13, 2009 (UCITS Directive) and Directive 2011/61/EU of the European Parliament and of the Council of June 8, 2011 (AIFMD) for further information. Also see OECD (2021).
comparable ML models may threaten financial stability. This risk could become significant if more successful algorithmic trading strategies are adopted, although there is currently no evidence that AI is driving this trend.5 Despite the fact that these risks have not yet materialized, it is interesting to note that the current European investment fund regime does not comprehensively address issues such as market concentration of the type described above, potential systemic risks arising from the use of AI in algorithmic trading, and algorithmic bias and overfitting (Tomanek, 2021). In actuality, algorithmic bias is a common concern when AI influences financial decision-making. The term refers to the systematic behavior of an algorithm that produces outcomes that may be deemed unjust - for example, because they penalize certain individuals based on their biological characteristics, and that may differ from the algorithm's intended function. Algorithmic bias can result from the algorithm's design or the way data is collected and utilized. Compared to applications in banking and insurance, where the use of clients' personal information is inherent to activities such as lending, credit extension, and consumer finance, the risk of algorithmic bias in an AI model leading to discriminatory outcomes in asset management and securities markets may be lower. However, certain types of bias have the potential to distort the results of an asset allocation model, leading to suboptimal outcomes or threatening the market's integrity. For instance, an AI algorithm may overweight the equities of companies with certain characteristics that were historically correlated with outperformance but whose use is now discouraged, such as the ethnicity or gender of the chief executive officer. Lastly, there is prevalent concern that the quality of the datasets used in the learning phase can have a material influence on the results and performance of AI and ML applications (see IOSCO, 2021). Experts in the field frequently emphasize the importance of data as a prerequisite for utilizing AI. Briefly, AI relies on data as its "fuel": the success of AI tools is highly dependent on data quality, and low-quality, chaotic data can easily lead to unreliable models. Despite the fact that AI is not yet pervasive in the securities markets, the risks outlined above warrant additional surveillance given the widespread interest in and focus on the topic among market participants. Nonetheless, governance and supervision of the existing processes are likely to be effective in mitigating a substantial portion of these risks.
Conclusion
Numerous promising AI use cases exist in the securities markets, although the actual degree of implementation varies between sectors (depending on the business case for the use of innovative AI tools) and within sectors (depending on a number of factors, such as company resources and business models). NLP is gaining popularity across all industries where entities can gain an advantage by parsing large quantities of text to identify specific unstructured information that may not otherwise be accessible, such as ESG policies, or conduct sentiment analysis. In asset management, an increasing number of market participants use AI in the investment process, for risk management, and for compliance activities, either by developing AI-based tools in-house or by acquiring them from external providers. Nonetheless, we find that investment funds that disclose and promote the use of AI in their investment strategies are still uncommon, indicating an incremental adoption of the
5 Although the use of complex, opaque, or rapidly-updating AI models may exacerbate the potential negative effects of algorithmic trading, these risks do not inherently originate from the use of AI. MiFID II already contains provisions that address the hazards associated with algorithmic trading. The European Securities and Markets Authority (ESMA) has routinely monitored these risks and engaged in dialogue with relevant parties on EU financial markets. Concerning potential novel risks, the ESMA's assessment report on algorithmic trading states that "the majority of market participants were unable to identify risks and effects on market structures other than those already specified in MiFID II that would merit additional regulatory attention." In addition, the evaluation states, "Only the largest market participants can keep up with the costly investments required by the current technological "arms race," according to a minority of respondents." This not only decreases competition, but also concentrates hazards among a limited number of firms (including CCPs). (See ESMA, 2021, paragraph 30).
technology in the industry. AI is also delivering tangible benefits in the trading life cycle, from the execution of trading orders to post-trade procedures, despite the fact that it is not yet pervasive. In execution, machine learning (particularly reinforcement learning) enables brokers and large institutional investors to minimize the market impact of large orders by determining how to optimally divide them across venues and time periods. Some CSDs use supervised learning in post-trade processes to predict the probability of settlement failures and optimally allocate liquidity. However, ESMA surveys indicate that the majority of CCPs and CSDs are not currently utilizing AI models. In other segments of the market, CRAs and proxy-advisory firms are predominantly experimenting with AI tools for information procurement, while only a few entities appear to be experimenting with models that support core aspects of their businesses. Although market participants are progressively utilizing AI to support certain activities and optimize specific business phases, this does not appear to be resulting in a rapid and disruptive overhaul of business processes. This is attributable to a number of factors, including technological limitations, client preferences, and regulatory uncertainty. Regarding the interplay between industry practices and the current regulatory framework, market participants polled by ESMA did not identify any significant regulatory barriers to the deployment of AI-based technology. However, they welcomed the prospect of a defined framework for the effective and dependable use of AI to help reduce the skepticism that still exists among many market participants regarding its adoption. In this context, the hazards associated with the use of AI in the securities markets are substantial but appear to be manageable. Nevertheless, AI has the potential to make crucial business and decision-making processes substantially speedier, more complex, and less transparent, which are all central concerns of regulation and oversight. Appropriate governance frameworks assuring the accountability and responsibility of both AI providers and end users are required to mitigate the risks arising from the intricacy of some AI systems and the frequently enormous volumes of data employed. If AI-based investment and trading models become increasingly successful in the future, additional hazards may emerge, such as the concentration of AI systems in the hands of a few "big players." Despite the fact that complexity and a lack of transparency may not be inherent characteristics of artificial intelligence, they may impede the adoption of innovative tools due to the need to maintain effective human supervision and upskill management. Some companies appear to be restricting or forgoing the use of AI and ML algorithms due to operational concerns, such as AI's compatibility with their legacy technology. Companies that implement AI and ML typically rely on their existing governance and supervision structures and do not employ compliance specialists to challenge and supervise the development of ML algorithms (IOSCO, 2021).
REFERENCES:
1. Alfonsi, A., Fruth, A. and Schied, A. (2010), "Optimal execution strategies in limit order books with general shape functions", Quantitative Finance, Vol. 10, No 2, pp. 143-157.
2. Artificial Intelligence Public-Private Forum (2022), Final Report - Artificial Intelligence Public-Private Forum, Bank of England and Financial Conduct Authority
3. Bartram, S., Branke, J. and Motahari, M. (2020), 'Artificial intelligence in asset management', CFA Institute Research Foundation, Charlottesville, VA
4. Bartram, S., Branke, J., De Rossi, G. and Motahari, M. (2021), "Machine learning for active portfolio management", Journal of Financial Data Science, Vol. 3, No 3, pp. 9-30.
5. Borghi, R. and De Rossi, G. (2020), "The artificial intelligence approach to picking stocks", in Jurczenko, E. (ed.), Machine Learning for Asset Management, Hoboken, NJ.
6. Briere, M, Lehalle, C. A., Nefedova, T. and Raboun, A. (2019), "Modelling transaction costs when trades may be crowded: A Bayesian network using partially observable orders imbalance", Université Paris-Dauphine Research Paper No 3420665
7. Bucci, F., Mastromatteo, I., Eisler, Z., Lillo, F., Bouchaud, J. P. and Lehalle, C.A. (2018),
"Co-impact: Crowding effects in institutional trading activity", working paper
8. Carpenter, M. and Poon, S. (2018), "Lessons learned from AI prototype designed for corporate AGM voting decisions", working paper
9. CFA Institute (2021), T-shaped Teams: Organizing to adopt AI and big data at investment firms, Charlottesville, VA
10. Dupont, L., Fliche, O. and Yang, S. (2020), "Governance of Artificial Intelligence in Finance, discussion document", ACPR, Fintech-Innovation Hub, Paris.
11. EBA (European Banking Authority) (2021), "EBA discussion paper on machine learning for IRB models", EBA/DP/2021/04, Paris La Défense
12. EFAMA (European Fund and Asset Management Association) (2022), "Trends in the European investment fund industry in the fourth quarter of 2021 & results for the full year of 2021", Quarterly Statistical Release, No 88, Brussels
13. EIOPA (European Insurance and Occupational Pensions Authority), Consultative Expert Group on Digital Ethics in insurance (2021), "Artificial Intelligence Governance Principles: Towards ethical and trustworthy artificial intelligence in the European insurance sector", Publications Office of the European Union, Luxembourg
14. ESMA (European Securities and Markets Authority) (2021), "MiFID II Review Report -MiFID II / MiFIR review report on algorithmic trading"
15. European Commission (2020), Communication from the Commission to the European Parliament, "The European Economic and Social Committee and the Committee of the Regions on a digital finance strategy for the EU", C0M(2020) 591 final
16. European Commission (2021), "Proposal for a Regulation laying down harmonised rules on artificial intelligence", C0M(2021) 206 final
17. European Council (2022), "Proposal for a Regulation of the European Parliament and of the Council laying down harmonized rules on artificial intelligence (Artificial Intelligence Act) and amending certain Union legislative acts"
18. FSB (Financial Stability Board) (2017), "Artificial intelligence and machine learning in financial services - Market developments and financial stability implications"
19. IOSCO (International Organization of Securities Commissions) (2021), The use of artificial intelligence and machine learning by market intermediaries and asset managers - Final report, FR06/21, Madrid
20. Joint Committee of the European Supervisory Authorities (2018), "Joint Committee Final Report on Big Data", JC/2018/04
21. Linardatos, P., Papastefanopoulos, V. and Kotsiantis, S. (2021), "Explainable AI: A review of machine learning interpretability methods", Entropy, Vol. 23, No 1, 18.
22. OECD (Organisation for Economic Co-operation and Development) (2021), "Artificial intelligence, machine learning and big data in finance: Opportunities, challenges, and implications for policy makers"
23. WEF (World Economic Forum) (2019), "Navigating Uncharted Waters: A roadmap to responsible innovation with AI in financial services", The Future of Financial Service
24. Zhang, D., Maslej, N., Brynjolfsson, E., Etchemendy, J., Lyons, T., Manyika, J., Ngo, H., Niebles, J. C., Sellitto, M., Sakhaee, E., Shoham, Y., Clark, J. and Perrault, R. (2022), "Artificial Intelligence Index Report 2022", Stanford University, Stanford, CA
