## CHAPTER 3 ADVANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARKET ACTIVITIES ## Source details **Canonical URL:** [CHAPTER 3 ADVANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARKET ACTIVITIES](https://www.imf.org/-/media/files/publications/gfsr/2024/october/english/ch3.pdf) ## Other formats - [Markdown version](/-/media/files/publications/gfsr/2024/october/english/ch3.pdf.md) - [Structured JSON version](/-/media/files/publications/gfsr/2024/october/english/ch3.pdf.json) --- ### Chapter 3 at a Glance — Key findings and context - Generative artificial intelligence (GenAI) and related breakthroughs can dramatically increase efficiency in trading, investment, and asset allocation via AI-assisted process automation and analysis of complex unstructured data. - Evidence suggests adoption effects are already beginning to be felt; adoption is likely to increase significantly in the near future (labor market and patent filings evidence). - Most current use is an extension of existing machine learning (ML) trends; more significant changes are a medium- to long-term concern. - Financial institutions have been actively using ML and related computational methods for approximately 20 years. - GenAI adoption is widespread in an “evolutionary” manner—large language models (LLMs) used as inputs into existing analytical models. - Prospect highlights: - Greatest value creation opportunities appear in publicly traded liquid asset classes: equities, government bonds, and listed derivatives. - GenAI could lower barriers to entry into less liquid asset classes (for example, corporate or sovereign bonds) by processing indentures and legal documents. ### Evidence of current scale, concentration, and talent - AI-powered ETFs account for less than $1 billion in assets under management. - Small numeric shares for AI-driven ETF share of US equity market capitalization (series corresponding to 2018–2024): 0.003, 0.001, 0.001, 0.002, 0.002. - High fixed costs of infrastructure and talent may exacerbate market concentration; development of foundation models has predominantly been based in the United States. - Job postings and talent indicators: - Share of job postings containing AI terminology for front office roles: monthly average of 4.5 percent in 2019, 6.6 percent peak in 2022, 4.9 percent in 2023. - Only a "small share of workers claims to have AI skills," but the financial services talent pool is growing and AI talent concentration within the US financial services industry exceeds the broader economy. - Patent filing trends: increasing filings referencing AI/ML in high-frequency or algorithmic trading and in asset management (2009–2023 time series referenced). ### Potential benefits to market functioning and financial stability - AI may reduce financial stability risks by enabling superior risk management, deepening market liquidity, and improving market monitoring for participants and regulators. - Specific documented improvements and research references: - Efficiency improvements and higher productivity (Boukherouaa and others 2021). - Refined portfolio investing frameworks (Park and others 2023). - Improved return forecasting (Chen, Kelly, and Xiu 2023). - Quantification of crash risks (Swinkels and Hoogteijling 2022). - SupTech and RegTech applications for supervisors and regulated institutions. - Use-case areas across investment decision, trade execution, and monitoring: client/institution profiling, security selection, asset allocation, orders placement and execution, reporting, and risk monitoring. ### Main categories of potential risks (summarized) - Increased market speed and volatility under stress if AI trading strategies become highly correlated or shut down. - Opacity and monitoring challenges as extreme AI behavior becomes hard to anticipate and AI activities migrate to NBFIs. - Increased operational risks from reliance on a few third-party AI service providers dominating computational power and LLM services. - Increased cyber and market manipulation risks, including fraud and social media disinformation. - Transmission channels to the real economy: loss of market confidence, higher borrowing costs, and potential financial system outages. ### Distributional and emerging-market considerations - AI can enable technological leapfrogging, financial development, inclusion, better credit access, and deeper local markets in emerging markets and developing economies. - High fixed costs may produce differential adoption speeds across regions, creating fragmentation risks and limiting benefits for some EMDEs. - Synthetic data can assist model training where real data are scarce but risks include misrepresentation of extremes and perpetuation of biases. --- ### Robo-Advisors’ Assets Under Management and AI-driven ETF trends - AI-driven ETF investment has grown but "remains tiny compared to the market’s size." - Time series covered: 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024, with forecasts for 2025, 2026, and 2027. - Chart axes / units noted: - Panel 1 and Panel 2: Billions of US dollars; share of US equity market cap in percent. - AI adoption in investment processes characterized as "still nascent" for trading and execution (Figure 3.5, panel 1). - Observed benefits from AI/ML adoption: improved efficiency, productivity, cost savings in algorithm design, better processing of unstructured data, and more compressed bid-ask spreads. - Supervisory and compliance applications: SupTech for anomaly detection; RegTech for anti-money laundering/know your customer processes, transaction monitoring, and fraud detection. ### Patent and labor market evidence (selected figures) - Patent filing increases for AI/ML in algorithmic trading and asset management (2009–2023 series). - Job postings with AI terms for front office roles: 4.5 percent (2019 monthly average), 6.6 percent (2022 peak), 4.9 percent (2023). - LinkedIn-based AI talent identification approach used in outreach: members with AI skills listed and/or in AI-occupations. --- ### Algorithmic trading, liquidity, volatility, and GenAI implications - Prevalence and concentration: - In the United States, algorithmic trading constitutes "about 70 percent of equities trading and more than half of futures trading." - Markets with high algorithmic trading shares tend to concentrate activity among a limited number of players. - Empirical findings: - Algorithmic trading "enhances liquidity and informational efficiency, albeit at the cost of increased short-term volatility." - Idiosyncratic jumps in individual stock returns are "less and less frequent" (data: five-minute intraday panel of 235 stocks, 2007–2024). - Idiosyncratic jumps are "more frequent when liquidity conditions are poor." - Algorithmic behavior under stress: - Algorithmic risk limits (position, price, volume, loss thresholds) can trigger simultaneous de-risking and cause cascading liquidity evaporation. - Evidence: order cancellation rates drop as implied volatility increases; hidden order rates increase with higher implied volatility. - Survey evidence of algorithm limits: position limits (14/15), price limits (13/15), volume limits (12/15) (ACM 2024). - GenAI channels and uncertainties: - GenAI could lower barriers to entry for algorithmic trading and facilitate processing of complex text-based data (for example, bond indentures). - Examples of AI-driven bond-market tools cited (Overbond, BondGPT). - Overall financial stability impact of these changes described as "highly uncertain." ### New dynamics driven by AI adoption (selected empirical points) - Procyclical trading volumes: - AI-powered ETFs show evidence of significantly higher portfolio turnover than other active or passive ETFs (Figure 3.11, panel 1). - Three sample AI-driven ETFs increased portfolio turnover during March 2020 market turmoil (Figure 3.11, panel 2). - Faster market reactions to textual releases: - After LLM introduction, initial market reactions to Federal Open Market Committee minutes (up to 45 seconds) "tend to reflect its eventual impact more accurately" than before LLMs. Event selection: index price change surpassed 0.2 percent at the 15-minute mark; both subsamples contain 15 qualifying datapoints. - Risks of collusion, manipulation, and concentration: - Theoretical models indicate possible tacit algorithmic collusion and "winner takes all" dynamics. - Simulations (Fan, Pelger, and Yu forthcoming) illustrate price gaps depending on the number of uninformed agents and the presence of an informed reinforcement learning agent. ### Table 3.1 — Potential Positive and Negative Scenarios (verbatim categories) - Market liquidity - Negative Scenario: "AI magnifies existing risks related to algorithmic trading by facilitating its growth. AI could 'democratize' and expand algorithmic trading activity to a broader set of assets and geographic areas. This could exacerbate risks related to sudden liquidity withdrawal under stressed conditions." - Positive Scenario: "AI increases the stability of algorithmic trading under stressed conditions. AI-driven algorithms could operate in a wider set of market conditions than traditional algorithms, with lower flash-crash risk, and reduced liquidity-withdrawal under stress." - Leverage - Negative Scenario: "AI-driven strategies boost short-term leverage. As arbitrage opportunities are exploited more efficiently by more advanced algorithms, remaining opportunities might require higher leverage to deliver similar returns." - Positive Scenario: "AI improves the management of leverage and related risks. AI could facilitate more frequent and automated management of leveraged positions, based on more inputs, and mitigate operational lags." - Interconnectedness - Negative Scenario: "AI increases interconnectedness. AI could proliferate algorithmic trading to other asset classes, geographic regions, and trading venues, and also operate in between different market segments; that is, in a multi-asset and multitrading venue approach. Increased interconnectedness leads to higher correlations between capital market segments, facilitating spillovers and transmission of stress." - Positive Scenario: "Market access, efficiency, and liquidity improve for some market segments, including emerging markets." ### Market concentration and IT vendor risks (key figures) - Vendor/IT concentration indicators: - IT infrastructure outage hours: 2023 (205.3 hours) and 2022 (133.5 hours). - Vendor concentration and overdependence on a limited number of AI model providers cited as a systemic risk. --- ### Regulatory responses, supervisory posture, and recommended actions - Expected regulatory responses and stakeholder preferences: - Provide clarity and guidance on model risk management; emphasize stress testing for extreme scenarios; require transparency and clearer disclosures; issue guidance on consumer-facing AI and accountability frameworks. - Stakeholders prefer balanced regulation that ensures responsible use without stifling innovation; capital market supervisors encouraged to provide guidelines and best practices rather than strict rulemaking. - Need to address bias in AI models and to upskill supervisors to integrate AI/ML into supervision and market surveillance. - Financial stability challenges summarized: - Continued AI growth could thin margins, incentivize leverage, produce herd-like behavior in stress, and cause flash sales or fire-sales in adverse conditions. - Migration of activity to NBFIs can increase systemic opacity and make holistic monitoring more difficult. - Reliance on a few third-party AI service providers could create a single-point-of-failure analogous to failures of key financial market utilities. - Cyber, manipulation, data poisoning, and model hallucination are highlighted as material concerns. - Adoption could exacerbate spillovers of advanced economy shocks to EMDEs. ### Regulatory and supervisory developments (examples and guidance) - Ongoing initiatives and frameworks: - IOSCO two-year project on AI risks with potential policy guidance expected by the first quarter of 2025 (IOSCO 2024). - FSB guidance on third-party risk and cyber incidents (FSB 2020, 2023a). - Basel Committee frameworks on data governance and operational/cyber risk (BCBS 2013; BIS 2023). - NIST AI framework (NIST 2024). - IOSCO work on algorithmic trading and market volatility (IOSCO 2018). - Recommended supervisory posture: - Remain vigilant on AI deployment; update supervisory skills and tools; process more granular data in real time; proactively question whether existing frameworks adapt to novel AI forms. - Re-parameterize circuit breakers and margining practices where necessary; test algorithms in controlled environments; identify good practices to reduce procyclicality of margin models. ### Data coverage and sample notes - Analysis covers stock exchange operators with market capitalization of listed companies exceeding $1 trillion as of March 2024. - Sample: 24 jurisdictions (11 members of the European Economic Area) and 26 financial market authorities (11 from the European Economic Area and 3 from the United States). --- ### Operational resilience, third-party providers, and SupTech adoption - Third-party AI service providers: - Authorities should coordinate regulation and supervision of AI service providers and map relationships between critical AI service providers and essential IT infrastructure providers. - Definition of critical service providers should capture systemic use of common AI models. - Protocols should cover avoidance, protection, response, and recovery across AI system life cycle stages. - SupTech adoption and information advantages: - GenAI enhances information retrieval, content creation, code generation, debugging, and legacy code optimization—accelerating fraud detection and market monitoring. - SupTech adoption (2023): 79 percent of advanced economies and 54 percent of emerging market and developing economies had adopted SupTech tools, compared to 50 percent and 31 percent, respectively, in 2022 (Cambridge SupTech Lab 2023). - Cyber risk indicators and incidents: - IMF measure of potential maximum annual financial firm losses from cyber incidents increased from $300 million to $2.2 billion since 2017. - Reported 2024 case: finance worker in Hong Kong SAR allegedly tricked by AI-generated deepfake, leading to a $25 million payout to fraudsters. - Cyber incidents have spillovers to markets (example: 2023 ransomware attack on Industrial and Commercial Bank of China reportedly affected US Treasury market conditions). --- ### Glossary — Selected AI and market-related terms (verbatim definitions highlighted) - Algorithmic trading (AT): Trading in financial instruments whereby an algorithm independently executes trading decisions. Strategies vary in complexity and latency; reinforcement learning allows algorithms to learn dynamically. - Artificial intelligence (AI): Broad definition including well-established predictive analytics; chapter focuses on more sophisticated AI, including generative AI and complex nongenerative applications. - Deep learning: Form of ML using layered artificial neural networks for supervised, unsupervised, or reinforcement learning. - Foundation models: Models trained with deep learning on massive data (text, images); typically pretrained and shared for refinement. - Generative AI: AI that generates new content, often powered by foundation models such as LLMs. - High-frequency trading (HFT): A type of algorithmic trading; not all algorithmic trading is HFT. - Large language models (LLMs): AI systems designed to learn grammar, syntax, and semantics to generate coherent language. - Reinforcement learning (RL): ML paradigm where an agent learns by taking actions and receiving feedback in the form of rewards or penalties. - Robo-advisors: Digital interfaces with algorithms (and possibly ML) providing automated financial recommendations or portfolio management with limited human intervention. - SupTech: FinTech applications used by regulatory, supervisory, and oversight authorities. - Synthetic data: Artificially generated information designed to mimic real-world data in statistical properties and structure. *Source: Chapter 3, "ADVANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARKET ACTIVITIES," Global Financial Stability Report, International Monetary Fund, October 2024.* ### Chapter 3 at a Glance ### Chapter 3 at a Glance ### Key findings on AI adoption and effects in capital markets - Generative artificial intelligence (GenAI) and related breakthroughs have the potential to dramatically increase the efficiency of capital markets—trading, investment, and asset allocation—through artificial intelligence–assisted process automation and analysis of complex unstructured data. - Evidence suggests these effects are already beginning to be felt. - New evidence from labor markets and patent filings suggests that the adoption of artificial intelligence (AI) in capital markets is likely to increase significantly in the near future. - Analyses of pricing patterns and trading dynamics already show changes in some markets consistent with the adoption of these new technologies. - Most current use of AI appears to be an extension of existing trends in the use of machine learning and other advanced analytical tools; more significant changes are a medium- to long-term concern. - Financial institutions have been actively using ML and other AI-related computation methods for approximately 20 years, and these methods are now well integrated into their investment processes. - GenAI is already seeing widespread “evolutionary” adoption—use cases that build upon existing analytical methods and investment strategies—with many market participants using large language models as inputs into existing analytical models. - GenAI could lower barriers to entry for quantitative investors into less liquid asset classes (such as corporate or sovereign bonds) that require extensive analysis of indentures and other legal documents. - Prospects for creating value through AI appear to be most promising in publicly traded liquid asset classes; equities, government bonds, and listed derivatives are highlighted as primary areas. ### Evidence of current scale and concentration - AI-powered ETFs—where AI is used to construct and adjust an ETF’s portfolio—still account for a very small share of the market, with less than $1 billion in assets under management. - Cost per unit of calculation has declined, but “notable” models used in leading GenAI applications have become much more complex, leading to much higher overall costs. - The high fixed costs of infrastructure and talent may exacerbate market concentration among a few private sector developers; development of foundation models has predominantly been based in the United States. - References to data sources: Epoch AI; Stanford University’s Ecosystem Graphs; IMF staff calculations. ### Potential benefits to market functioning and financial stability - AI may reduce financial stability risks by enabling superior risk management, deepening market liquidity, and improving market monitoring by both participants and regulators. - Specific productivity and capability improvements noted: - Efficiency improvements and higher productivity (Boukherouaa and others 2021). - Refined portfolio investing frameworks (Park and others 2023). - Improved return forecasting (Chen, Kelly, and Xiu 2023). - Quantification of crash risks (Swinkels and Hoogteijling 2022). - SupTech and RegTech applications for supervisors and regulated institutions. - Use cases across investment decision, trade execution, and monitoring include: client/institution profiling, security selection, asset allocation, orders placement and execution, reporting, and risk monitoring (see Figure 3.1 use-case matrix). ### Main categories of potential risks identified - Increased market speed and volatility under stress, especially if AI trading strategies become highly correlated or shut down in response to an unforeseen event. - Opacity and monitoring challenges as extreme behavior of AI systems becomes increasingly difficult to anticipate and as AI activities migrate to nonbank financial intermediaries (NBFIs). - Increased operational risks due to reliance on a few key third-party AI service providers that dominate computational power and large language model services. - Increased cyber and market manipulation risks, particularly in generating fraud and social media disinformation. - These risks could transmit stress to the real economy through loss of market confidence, higher borrowing costs, and potentially significant financial system outages. ### Distributional and emerging-market considerations - For emerging markets, AI is widely seen as a positive development—enabling technological leapfrogging, increased financial development, and inclusion through increased access to credit and deepening of local financial markets. - However, high fixed costs could lead to different speeds of adoption across regions, creating fragmentation risks and potentially limiting benefits for some emerging market and developing economies. ### Current market practices and attitudes - Many market participants view human oversight (“human in the loop”) as essential; autonomous AI generating and executing trades without human oversight is widely seen as a nonstarter. - IMF staff outreach indicates equities and derivatives are the most likely areas for AI adoption in the investment process, followed by fixed income and foreign exchange. - Buy-side firms employ AI/ML for productivity enhancement, exploration of new asset classes, and extraction of signals from alternative and unstructured data to support investment decisions. ### Policy recommendations for authorities - Undertake the calibration of circuit breakers and a review of margining practices in light of potentially rapid AI-driven price moves. - Enhance monitoring and data collection of the activity of large traders, including nonbank financial intermediaries. - Address dependency on data, models, and technological infrastructure by requesting a risk mapping from regulated entities (that is, data on the internal and external interconnections and interdependencies that are necessary to deliver the institutions’ critical services). - Adopt a coordinated approach for the definition of critical AI third-party service providers and continue to strive for resilience in capital markets by enhancing cyberattack protocols. - Adopt measures that ensure continued market integrity, efficiency, and resilience of over-the-counter markets when AI use proliferates. *Source: Chapter 3, "ADVANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARKET ACTIVITIES," Global Financial Stability Report, International Monetary Fund, October 2024.* ### 1. Robo-Advisors’ Assets Under Management ### 1. Robo-Advisors’ Assets Under Management ### Robo-Advisors AUM and AI-Driven ETFs - AI-driven exchange-traded fund (ETF) investment has grown, but it "remains tiny compared to the market’s size." - Time series presented cover years: 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024, with forecasts for 2025, 2026, and 2027 (light blue bars are forecasts for 2025, 2026, and 2027). - Chart axes / units: - Panel 1: Billions of US dollars; share of US equity market cap in percent. - Panel 2: Billions of US dollars; share of US equity market cap in percent. - Small numeric values shown for AI-driven ETF share of US equity market capitalization in panel 2 include: 0.003, 0.001, 0.001, 0.002, 0.002 (series corresponding to 2018–2024 respectively as shown). - The share of assets under management as a percentage of the US equity market capitalization is based on the MSCI US Equity Index market capitalization. ### AI Adoption in Investment Processes and Markets - Adoption of AI in core investment processes such as trading and execution is described as "still nascent" (Figure 3.5, panel 1). - Use cases gaining traction: - AI-powered exploration of alternative and text data to uncover causal relationships in markets previously unknown, potentially leading to new investment strategies. - Adoption of traditional AI/ML applications to increase robustness and accuracy of existing models, especially forecasting. - Empirical outreach findings: - Within a three- to five-year horizon, participants expect greater integration of sophisticated AI in investment and trading decisions. - Participants expect a "human in the loop" approach to persist in the near term (three to five years), especially for large capital allocation decisions. - Complete autonomy is "not anticipated soon"; models will continue to operate within predefined rules. - Some participants mentioned potential for agent-to-agent trading and development of complete AI-driven workflows in trading. - Evidence on autonomous trading: - Sophisticated AI has not yet been implemented widely to build autonomous AI trading agents (Authority for Consumers and Markets 2024, p. 18). - AI is more frequently used to generate a signal that is then used as an input in an existing analytical system where a human trader may make the trading decision. - Benefits observed from AI/ML adoption include improved efficiency and productivity, cost savings in designing trading algorithms, better processing of unstructured data, and more compressed bid-ask spreads. - Supervisory and compliance use: - Financial supervisors use AI-driven SupTech tools to monitor markets and detect anomalies in large data sets. - Banks use RegTech tools to manage regulatory compliance and enhance "anti-money laundering/know your customer" processes, including automating tasks for higher accuracy in clients’ data, transaction monitoring, and fraud detection. ### Evidence from Patent Filings and Labor Markets - Patent filing trends: - Number of filings that reference AI/ML terminologies in the context of high-frequency or algorithmic trading has increased (Figure 3.6, panel 1). - Recent filings lean toward improving operational efficiency of brokerage or trading platforms and developing systems that compute trading signals with low latency and high throughput. - AI/ML-related filings have driven a surge in patents in asset management (Figure 3.6, panel 2). - Asset management filings detail use of ML techniques to: - Enhance efficiency of cash flow and liquidity management. - Automate asset class rebalancing. - Improve valuation and forecasting methods. - Determine capital requirements tailored to individual needs. - Interpret unstructured data and process information from alternative data sources. - Access and manage alternative asset classes (methodologies for trading emissions and managing digital assets). - Evaluate cryptographically signed transactions. - Patent chart scales / units: - Panel 1: Number of patents; percent of patents (time series from 2009 to 2023 shown). - Panel 2: Number of patents; percent of patents (time series from 2009 to 2023 shown). - Labor market evidence: - Only a "small share of workers claims to have AI skills" (Figure 3.7, panel 1), but the talent pool in financial services is growing. - ML, natural language processing, and deep learning are among the top 30 competencies listed in US quantitative researcher and analyst profiles. - Job postings: - Share of job postings containing AI terminology for front office roles rose from a monthly average of 4.5 percent of front office roles in 2019 to 4.9 percent in 2023 and a peak of 6.6 percent in 2022. - A LinkedIn member is considered AI talent if they have explicitly added AI skills to their profile and/or are occupied in an AI-occupation representative. - The share of job postings for front office roles and the financial services industry requiring AI skills has outpaced the overall share of AI-related job postings for the broader US economy. - AI talent concentration within the US financial services industry exceeds the broader economy. ### Implications for Emerging Markets and Synthetic Data - Potential benefits for emerging markets and developing economies: - Improvements in access to financial services, credit scoring, loan origination, robo-advising, and portfolio construction. - GenAI-enabled parsing of fragmented and unstructured data could reduce investment barriers and improve liquidity of some emerging market assets. - Synthetic (AI-generated) data may assist in training investment models where real data are scarce. - Caveats on synthetic data: - Risk of unintended over- or under-representation of certain values of real-world data distribution, undermining extreme event performance of AI systems. - Potential biases perpetuated by synthetic data when generation processes fail to account for specificities and requirements of second-order applications. - Divergence risk: - Some market participants indicate potential differentiation between large and less-significant emerging markets based on the extent of AI implementation. - Overall, the risk of fragmentation between advanced economies and emerging market and developing economies "seems to be limited," with advances in AI potentially supporting greater financial inclusion. - Automation could affect lower-skilled jobs in some countries. ### Market Structure and Dynamics: Algorithmic Trading and NBFIs - Nonbank financial institutions (NBFIs): - NBFIs now hold "over half of all financial market assets globally" and may grow more important with AI adoption. - NBFIs are generally more agile and subject to fewer constraints for AI adoption compared with some larger banks that face legacy infrastructure and stricter model governance/accountability and model explainability requirements. - Algorithmic trading prevalence and concentration: - In the United States, algorithmic trading constitutes "about 70 percent of equities trading and more than half of futures trading" (Figure 3.8, panel 1). - Other jurisdictions lag in algorithmic equities trading share but "could catch up briskly" (Figure 3.8, panel 2). - Markets with relatively high share of algorithmic trading activity tend to see concentration of activity among a limited number of players (Figure 3.8, panel 3). - High fixed costs associated with internal development or deployment of sophisticated AI benefit larger trading firms and may lead smaller players to rely on third-party cloud and AI software providers, amplifying outsourcing, market concentration, and vendor lock-in risks (Figure 3.9). - Evolution of strategies: - Trading strategies have evolved from relatively simple trading rules to more complex algorithms and are poised to use more sophisticated AI. - AI’s ability to process large amounts of high-frequency and unstructured data in short amounts of time is expected to enable further automation of trading decisions and provide competitive advantages. *Sources: Bloomberg Finance L.P.; Statista Digital Market Insights; Authority for Consumers and Markets; Mercer Investments (2024); World Intellectual Property Organization, PATENTSCOPE; IMF staff calculations and outreach summarized in Chapter 3.* ### 1. Share of Algorithmic Trading Activity by ### ch3 - 1. Share of Algorithmic Trading Activity by ### Algorithmic trading: liquidity, efficiency, and volatility - Research findings: - Algorithmic trading "enhances liquidity and informational efficiency, albeit at the cost of increased short-term volatility" (Hendershott, Jones, and Menkveld 2011; Hendershott and Riordan 2012; Boehmer, Fong, and Wu 2021). - Algorithmic trading can "increase volatility following macroeconomic news and can disincentivize informed traders from participating in the market" (Scholtus, van Dijk, and Frijns 2014; Yadav 2015). - In the US Treasury market, digitalization "has dramatically improved liquidity on aggregate, but this may have come at the cost of rare but extreme bouts of illiquidity under stress" (Bouveret and others 2015). - Market liquidity is affected by the extent to which high-frequency traders are present (Adrian, Fleming, and Vogt 2017). - Empirical evidence on return jumps and liquidity: - A decomposition of high-frequency US stock returns shows that idiosyncratic jumps in individual stock returns are "less and less frequent" (Figure 3.10, panel 1). - Idiosyncratic jumps are "more frequent when liquidity conditions are poor" (Figure 3.10, panel 2). - Data: five-minute trading hour intraday data covering a balanced panel of 235 stocks between 2007 and 2024 (Figure 3.10, panel 1). ### Algorithmic behavior under stress and risk limits - Mechanisms and observed patterns: - Algorithmic trading strategies are often programmed to de-risk or shut down during periods of high volatility; such algorithmic risk limits can include "restrictions on the total volume of trades, maximum loss thresholds, or limits on exposures to specific assets or markets." - Under stress, simultaneous triggering of limits can cause cascading effects, feedback loops, and "the sudden evaporation of liquidity provided by algorithmic trading." - Evidence of reduced liquidity under stress in US equity markets: - Order cancellation rates drop significantly as implied volatility increases (Figure 3.10, panel 3). - Hidden order rates increase under higher implied volatility—consistent with "flighty liquidity-under-stress." - Survey evidence: algorithms are subject to position limits (14/15), price limits (13/15), volume limits (12/15), and other limits (ACM 2024) in a Dutch Authority for Consumers and Markets energy-trader survey. ### GenAI and the proliferation of algorithmic trading - Capabilities and channels: - "GenAI could lower barriers to entry for algorithmic trading, as it facilitates coding, testing, and automation of trading in less technologically sophisticated trading venues." - GenAI can "facilitate the processing of complex text-based data (such as bond indentures) to enable more standardized risk analysis" and support pricing tools for liquidity in asset classes that "do not naturally lend themselves to automated trading" (for example, corporate bonds). - Examples of AI-driven tools in bond markets include Overbond and BondGPT (as cited in the source). - Uncertainty: "On balance, the impact of these changes from a financial stability perspective is highly uncertain." ### New dynamics that could be driven by further AI adoption - Procyclical trading volumes: - AI can process vast amounts of information and spur larger and more frequent portfolio adjustments, increasing trading volumes. - Evidence: "Portfolio turnover for AI-powered ETFs provides evidence" — ETFs with AI-driven strategies "have experienced significantly higher turnover than other active or passive ETFs" (Figure 3.11, panel 1). - Three sample AI-driven ETFs increased their portfolio turnover during the March 2020 market turmoil (Figure 3.11, panel 2). - Faster market reactions to textual releases: - Intraday market data suggest that after the introduction of large language models, the initial market reaction following the release of Federal Open Market Committee minutes (up to 45 seconds) "tends to reflect its eventual impact more accurately than in the period before the introduction of these technologies" (Figure 3.11, panel 3). - Event selection note: events included if the index price change surpassed 0.2 percent at the 15-minute mark; the periods before and after LLMs are separated by the publication of Vaswani and others (2017). Both subsamples contain 15 qualifying datapoints. - Risks of collusion, manipulation, and market structure outcomes: - Theoretical models show possible tacit algorithmic collusion (Dou, Goldstein, and Ji 2024) and "winner takes all" dynamics (Baron and others 2017). - Manipulation is more likely if one algorithm has an information or latency advantage and when the market has fewer players (Figure 3.11, panel 4). - Simulation note: Panel 4 shows simulated scenarios from Fan, Pelger, and Yu (forthcoming) with one informed reinforcement learning agent holding one-eighth of total market buying power and varying numbers of uninformed agents; scenarios show price gaps between market price and fundamental value that depend on the number of uninformed agents. ### Table 3.1 — Potential Positive and Negative Scenarios (verbatim categories and assessments) - Market liquidity - Negative Scenario: "AI magnifies existing risks related to algorithmic trading by facilitating its growth. AI could 'democratize' and expand algorithmic trading activity to a broader set of assets and geographic areas. This could exacerbate risks related to sudden liquidity withdrawal under stressed conditions." - Positive Scenario: "AI increases the stability of algorithmic trading under stressed conditions. AI-driven algorithms could operate in a wider set of market conditions than traditional algorithms, with lower flash-crash risk, and reduced liquidity-withdrawal under stress." - Leverage - Negative Scenario: "AI-driven strategies boost short-term leverage. As arbitrage opportunities are exploited more efficiently by more advanced algorithms, remaining opportunities might require higher leverage to deliver similar returns." - Positive Scenario: "AI improves the management of leverage and related risks. AI could facilitate more frequent and automated management of leveraged positions, based on more inputs, and mitigate operational lags." - Interconnectedness - Negative Scenario: "AI increases interconnectedness. AI could proliferate algorithmic trading to other asset classes, geographic regions, and trading venues, and also operate in between different market segments; that is, in a multi-asset and multitrading venue approach. Increased interconnectedness leads to higher correlations between capital market segments, facilitating spillovers and transmission of stress." - Positive Scenario: "Market access, efficiency, and liquidity improve for some market segments, including emerging markets." ### Market concentration, herding, and other financial stability implications - Outreach and perceived risks: - Market participants cited herding and market concentration as key financial stability risks from wider AI adoption, especially where strategies "were to become largely derived from open-source AI and trained on similar data sourced from the same set of vendors." - Vendor concentration is a potential systemic risk: "overdependence on a limited number of AI model providers and data vendors could lead to mass disruptions to trading and investment were one or some of these vendors to fail." - Participants also cited risks of market manipulation through deepfakes or misinformation and market fragility, including drying up of market liquidity. - IT concentration and outages (figure-level indicators preserved): - IT infrastructure and AI software services market show concentration in market share (Figure 3.9, panel 1). - There was longer outage time for IT infrastructure in 2023 than in 2022: 2023 (205.3 hours) and 2022 (133.5 hours) (Figure 3.9, panel 2). *Source: IMF staff and figures as presented in the chapter excerpt.* ### 2. How Do You Expect Regulatory Authorities to Respond to the Risks ### 2. How Do You Expect Regulatory Authorities to Respond to the Risks of Generative AI? ### Expected regulatory responses and stakeholder preferences - Regulatory authorities are expected to: - Provide clarity and guidance on model risk management. - Emphasize stress testing for extreme scenarios. - Provide transparency and clearer disclosures. - Offer guidance on industry-specific regulatory structures to avoid violation of existing regulations. - Issue guidelines on AI use in consumer-facing applications and accountability frameworks. - Stakeholder preferences and emphases: - Buy-side and sell-side entities, academia, and market infrastructure providers emphasized the need for balanced regulation that ensures responsible use of AI without stifling innovation while ensuring adequate consumer protection. - Consensus that capital market supervisors should focus on providing guidelines and best practices rather than strict rulemaking, given the rapidly evolving nature of AI technology in financial markets. - Need to address bias in AI models and ensure supervisors improve AI preparedness through continuous upskilling and integrate AI/ML (including sophisticated AI) in supervision and market surveillance functions. - Overall sentiment: regulatory approach should be flexible and adaptable to keep pace with rapid advancements in AI technology in the financial sector. - Definitions and notes: - Infrastructure refers to market infrastructure firms. - Other industry types include AI vendors, academia, and rating agencies. - AI = artificial intelligence; EMDE = emerging market and developing economies. ### Financial stability challenges (current and prospective) - The use of AI in capital markets is nascent; risks appear contained currently, but rapid adoption is likely and may drive transformative impacts that create several financial stability challenges. - Increased market speed and volatility under stress: - Continued growth of AI-enhanced algorithmic trading strategies could enhance market liquidity and bring efficiency gains, manifesting in more prompt price adjustments and thinner margins for traders. - Thinner margins could incentivize increased use of leverage across the financial system and result in more amplification between falling asset prices, volatility, and deleveraging in periods of stress. - AI models may herd and produce similar decisions, especially during stress periods, resulting in procyclical financial stability risks. During normal times, AI models may uncover new trading opportunities, leading to more diverse investment strategies that would be positive for financial market resilience. During adverse shocks, however, models could simultaneously rebalance portfolios toward safe assets, creating a self-fulfilling spiral of fire sales. - Novel adverse events—such as the COVID-19 pandemic in 2020—may drive AI model outcomes that are difficult to comprehend, or models may simply shutdown, requiring humans to make decisions on and process a voluminous number of trades. - Vulnerability could be heightened if AI trading algorithms collude with each other, resulting in a winner-dominated market that could be more easily upended by adverse shocks. - More opacity and monitoring challenges: - AI may spur further migration of activities to NBFIs. Since the global financial crisis, trading and investment activity, especially capital market activities, have steadily migrated out of the banking sector and into NBFIs. - Some NBFIs have built extensive expertise and technology to exploit AI advances. Regulatory requirements for banks regarding explainability and transparency of internal models—compared to comparatively lighter requirements for NBFIs—give NBFIs a competitive advantage, raising systemic opacity. - AI models could generate portfolios across different asset classes, geographic regions, and trading venues, creating correlations and interconnectedness not currently relevant, undermining regulators’ ability to monitor financial risks holistically. - Likely emergence of new forms of risks (for example, potential complex interactions between autonomous AI agents not visible at the level of individual institutions or at the regulatory level). - Increased operational risks from reliance on a few third-party AI service providers: - AI models and related IT services currently reside with a handful of key providers with dominant computational power and established large language models. - If capital markets activities become too reliant on these models, the failure of these providers may lead to market stress akin to the failure of key financial market utilities such as clearing houses. - Increased market manipulation and cyber risks: - Fraud, disinformation, and deepfakes will likely become more sophisticated as AI advances and could be used by bad actors to manipulate financial markets and asset prices. - Data integrity and confidentiality could be compromised, leading to AI models producing suboptimal trading and investment decisions. - Additional concerns raised by participants: - Liquidity, excess volatility, and flash crashes arising from fast-paced decision making and ineffectiveness of guardrails due to poor guardrail design, increasing AI complexity, or malicious intervention. - Cyberattacks on financial intermediaries and market utilities, and large-scale data poisoning as potential sources of systemic risk. - Lack of model explainability and model hallucination could be detrimental to trust in markets. - High costs associated with fine-tuning sophisticated models using large data sets could create an unlevel playing field favoring large firms. - Customer fraud, unauthorized use, and data access could pose risks and compliance issues, leading to reputational damage. - Adoption of AI could exacerbate spillovers of advanced economy shocks to EMDEs, particularly if AI models are more sensitive to price fluctuations and managed against a basket of various asset classes. ### Regulatory and supervisory developments - Existing frameworks and orientations: - Capital markets are already subject to regulation and supervision; institutions are responsible for AI systems they deploy, whether internally developed or externally sourced. - Existing regulatory and supervisory frameworks for capital markets are largely technology-neutral and are applicable to AI systems. - Ongoing work by financial sector authorities explores application of existing prudential frameworks and the need for additional frameworks to cover risks specific to AI use, with current focus more on conduct issues such as ethics, fairness, and transparency. - Recommendations for financial sector authorities: - Remain vigilant on AI deployment by capital market participants and be prepared to respond to an acceleration in the pace of adoption. - Update skills and supervisory tools to monitor more complex investment strategies and process more granular data in real time. - Proactively question whether extant regulatory frameworks adapt to novel forms of AI with a comprehensive view of emerging risks. - Standard-setter and supervisory activity examples (as noted in the chapter): - The FSB issued guidance for managing third-party risk and cyber incidents (FSB 2020, 2023a). - The Basel Committee frameworks include principles and recommendations on data governance and operational and cyber risk management (BCBS 2013; BIS 2023). - The US National Institute of Standards and Technology (NIST) has issued a relevant AI framework (NIST 2024). - IOSCO has addressed algorithmic trading and market volatility (IOSCO 2018) and AI risks in market intermediaries. - Existing frameworks by the Financial Stability Board (FSB), Bank for International Settlements, and national regulators build on principles of technology-neutral, results-based, and proportional regulation and supervision (Monetary Authority of Singapore 2018; Hong Kong Monetary Authority 2019). - Supervisory posture: - Financial supervisors are cautious and respond through targeted outreach and clarification of existing standards. - It is recommended that financial sector authorities update their skills and supervisory tools to monitor more complex investment strategies and process more granular data in real time, and to proactively assess whether existing frameworks adapt to novel AI forms with a comprehensive view of emerging risks. ### Data points and sample coverage noted in the chapter - Analysis covers jurisdictions with stock exchange operators with a market capitalization of listed companies exceeding $1 trillion as of March 2024. - Sample consists of 24 jurisdictions, of which 11 are members of the European Economic Area. - Sample includes 26 financial market authorities, of which 11 are from the European Economic Area and 3 are from the United States. *Source: ch3 - 2. How Do You Expect Regulatory Authorities to Respond to the Risks of Generative AI?; IMF, October 2024 Global Financial Stability Report.* ### CHAPTER 3 AdvANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARkET ACTIvITIES ### CHAPTER 3 AdvANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARkET ACTIvITIES ### Regulatory landscape and supervisory posture - IOSCO is conducting a two-year project to assess risks and challenges associated with the use of AI, with potential policy guidance expected by the first quarter of 2025 (IOSCO 2024). - An IMF review of actions taken by 26 authorities in large capital markets finds governments have begun to formulate comprehensive AI strategies and act in the areas of data protection, governance, and cybercrime. - Some jurisdictions are considering dedicated AI legislation to ensure robust governance. - Supervisory authorities have so far focused primarily on clarification and outreach, rather than on enforcement. ### Best practices and engagement mechanisms - Public/private forums to develop overarching principles (Office of the Superintendent of Financial Institutions 2023). - Partnering with industry to build a risk framework (Monetary Authority of Singapore 2024). - Conducting surveys on the applicability of existing frameworks (US National Archives 2021; Institute for Workplace Equality 2022; Bank of England 2024). - Requesting notification by banks prior to adoption of certain technologies or arrangements with third parties (BCBS 2024). - Periodic upskilling and upgrading of supervisory staff to identify AI-specific issues such as models designed to “game the regulation” and detect algorithmic coordination. - Cross-sectoral thematic reviews to reveal potential herding or material interconnectedness (Securities and Exchange Board of India 2019). ### Supervisory applications, SupTech adoption, and information advantages - AI can automate data quality checks to ensure completeness, correctness, and consistency. - AI can combine multiple data sources lacking unique identifiers and help detect anomalies in trading patterns (di Castri and others 2019). - GenAI enhances information retrieval, content creation, and code generation, debugging, explanation, and legacy code optimization—accelerating fraud detection, market monitoring, and data management tasks. - Adoption of SupTech tools (2023): 79 percent of advanced economies and 54 percent of emerging market and developing economies had adopted SupTech tools, compared to 50 percent and 31 percent, respectively, in 2022 (Cambridge SupTech Lab 2023). ### Market speed, volatility, and margining under stress - Authorities and trading venues should assess whether new or modified volatility response mechanisms are necessary for crash events potentially originated in AI-driven trading. - Existing circuit breakers may need re-parameterization; poorly designed circuit breakers may exacerbate volatility and interfere with market efficiency and price discovery (Vereckey 2023). - Testing algorithms in controlled environments is recommended to assess behavior in extreme circumstances. - Margining and buffers: further international work is needed to: - foster market participants’ preparedness for large variation margin calls during market stress; - identify good practices for variation margin collection and distribution by the central counterparty; - understand the degree and nature of central counterparty margin models’ responsiveness to volatility and other market stresses; - review initial margin levels in non-stress times, including the effectiveness of tools to reduce procyclicality of margin models (BCBS, CPMI, and IOSCO 2022). ### Opacity, interdependencies, and nonbank financial intermediation (NBFI) - Financial institutions should regularly map interdependencies between data, models, and technological infrastructure supporting AI models, including shared or interdependent data sources, common architectures, and reliance on a small number of providers. - Regulatory frameworks require assessing cumulative effects of models but do not mandate a joint assessment of data dependencies; an updated view of interdependencies enables proactive risk management. - Financial sector authorities should strengthen oversight and regulation of NBFIs by requiring them to identify themselves and disclose AI-relevant information. - Authorities could monitor “large traders” by unique identification and reporting of activities to their registered broker-dealer to enable monitoring. - Measures to enhance NBFI resilience: improved risk management, strengthened liquidity buffers, incentives for market-makers to enhance liquidity, improved incentives for central clearing, and margin requirements for non-centrally cleared derivatives. Backstops could include central bank liquidity provision to market-making banks or indirectly support non-bank dealers by easing market funding conditions (CGFS 2014). ### Third-party AI service providers and operational resilience - Authorities should undertake coordinated regulation and supervision of AI service providers and map relationships between critical AI service providers and essential IT infrastructure providers. - Comparable and interoperable regulatory approaches to critical service providers facilitate compliance and coordination among financial sector authorities (FSB 2023b). - The definition of critical service providers should be broad enough to capture systemic use of common AI models (Bank of England 2024). - Financial sector authorities should require protocols to avoid, protect against, respond to, and recover from attacks across AI system life cycle stages (design, development/procurement, deployment, operations) (National Cyber Security Centre 2023). ### Manipulation and cyber risk: incidents, trends, and examples - Cyberattacks have increased, with the finance and insurance sector share rising over the past decade. - One IMF measure of potential maximum annual financial firm losses from cyber incidents increased from $300 million to $2.2 billion since 2017 (see Chapter 3 of the April 2024 Global Financial Stability Report). - Generative AI can be used by bad actors to produce deepfakes and manipulate audio/video to impersonate key individuals, facilitate fraudulent transactions, manipulate stock prices, or erode trust—potentially triggering selloffs or deposit runs. - A 2024 case: a finance worker at a multinational firm in Hong Kong SAR was reportedly tricked by AI-generated deepfake video and audio, allegedly leading to a $25 million payout to fraudsters. - Although a dedicated AI cyberattack database does not yet exist, AI incidents are being tracked by multiple databases; most AI incidents have occurred in advanced economies even accounting for GDP differences, largely reflecting AI usage rates. - Fake or genuine social media activity can amplify news and contribute to panic, possibly through manipulation (reports suggest First Republic Bank was targeted by an online manipulation campaign). - Cyber incidents can have spillovers to markets; example cited: a 2023 ransomware attack on the Industrial and Commercial Bank of China reportedly affected US Treasury market conditions. ### IMF outreach, definitions, and market intelligence - IMF staff conducted outreach with buy-side and sell-side firms, AI vendors, asset managers, academia, rating agencies, and market infrastructure firms; detailed responses received from 27 stakeholders directly involved in AI topics and business. - IMF staff regulatory outreach included 10 capital market supervisors of advanced and emerging markets. - IMF adopted definitions for outreach: - “AI/ML models”: well-established predictive analytics, including neural networks, clustering algorithms, natural language processing, decision trees, and so on. - “Sophisticated AI models”: latest innovations, such as deep learning, reinforcement learning, and large language models, including generative AI capable of generating text, codes, images, and other content. - Participants in IMF’s Market Intelligence Outreach (percent): Asset managers 28, Bank/dealers 28, Market infrastructure 12, Academia 12, Others 20. *Italicized source: CHAPTER 3 AdvANCES IN ARTIFICIAL INTELLIGENCE: IMPLICATIONS FOR CAPITAL MARkET ACTIvITIES (PDF).* ### Box 3.2 (continued) ### Box 3.2 (continued) ### Glossary — AI and market-related terms - Algorithmic trading (AT) - Trading in financial instruments whereby an algorithm independently executes trading decisions.26 Algorithmic trading can be used for trade execution, market-making, or in other proprietary trading strategies. Algorithmic trading strategies vary in complexity and latency; in its simplest guise, algorithmic trading may involve the use of basic trading rules or instruction to feed portions of an order into the market at preset intervals to minimize market impact cost. More complex applications can involve multi-asset trading strategies based on advanced machine learning models. Reinforcement learning allows algorithms to learn dynamically from evolving trading patterns, as well as the actions of other algorithms. - Artificial intelligence (AI)27 - The theory and development of computer systems able to perform tasks that traditionally have required human intelligence. The definition of AI is very broad and would include many simple applications that would generally not be described as AI in the public discourse. For example, simple linear regression would fall under this broad definition, even though most would not classify this as AI. The focus of this chapter is on more sophisticated AI, which includes not only generative AI but also more complex nongenerative applications such as clustering algorithms, neural networks, gradient-boosted decision trees, support vector machines, etc. - Deep learning - A form of machine learning that uses algorithms that work in “layers” inspired by the structure and function of the brain. Deep learning algorithms, whose structures are called artificial neural networks, can be used for supervised, unsupervised, or reinforcement learning (itself a form of machine learning). - FinTech28 - Technology-enabled innovation in financial services that could result in new business models, applications, processes, or products with an associated material effect on the provision of financial services. - Foundation models29 - An umbrella term referring to a diversity of models that are usually trained by applying deep learning to massive quantities of data, such as text and images. Because the expertise, time, and computing power involved in training foundation models from scratch are typically prohibitive for most nonspecialist firms, these models are usually pretrained and shared with end users for further refinement. - Generative AI - AI that generates new content, such as text, images, and videos, often based on user prompts. Generative AI is powered by foundation models, such as large language models. ### Market-structure and AI-application terms - High-frequency trading (HFT) - HFT is frequently equated to algorithmic trading. However, whereas HFT is a type of algorithmic trading, not all forms of algorithmic trading can be described as high frequency. Algorithmic trading predates HFT and has been extensively used as a tool to determine some or all aspects of trade execution like timing, price, quantity, and venue. Many intermediaries use algorithmic trading for their own proprietary trading or offer it to their clients. It has also become a standard feature in many buy-side firms, mainly with the purpose of devising execution strategies that minimize price impact or to rebalance large portfolios of securities as market conditions change. A number of common features and trading characteristics related to HFT are identified by IOSCO.30 - Large language models (LLMs) - Large language models are AI systems designed to learn grammar, syntax, and semantics of one or more languages to generate coherent and context-relevant language. - Machine learning (ML) - A method of designing a sequence of actions to solve a problem that optimizes automatically through experience and with limited or no human intervention. - Market-making31 - The provision of liquidity for clients in financial instruments, whereby a trader sets firm bid-offer quotes and thereby provides liquidity for a specific product or a particular product class. This is designed to avoid temporary imbalances between supply and demand for certain products. - Proprietary trading - Describes a trading unit which is separate from the rest of an organization’s trading activities and is not involved in client business. It generates profits exclusively from taking positions. This trading unit has no client contact and is not involved in the broker market. - RegTech32 - Any range of applications of FinTech for regulatory and compliance requirements and reporting by regulated financial institutions. This can also refer to firms that offer such applications. - Reinforcement learning (RL) - A type of machine learning paradigm where an agent learns to make decisions by taking actions in an environment to achieve some goal. The learning process is driven by the feedback the agent receives from the environment in the form of rewards or penalties. - Robo-advisors or automated advice - Applications that combine digital interfaces and algorithms, and can also include machine learning, to provide services ranging from automated financial recommendations to contract brokering to portfolio management to their clients, without or with very limited human intervention. Such advisors may be standalone firms and platforms or can be in-house applications of incumbent financial institutions. - SupTech - Any application of FinTech used by regulatory, supervisory, and oversight authorities. - Synthetic data - Artificially generated information that is designed to mimic real-world data in terms of statistical properties and structure. Unlike real data, which are directly collected from real-world events or interactions, synthetic data are created through algorithms and simulation models. *International Monetary Fund | Global Financial Stability Report: Steadying the Course: Uncertainty, Artificial Intelligence, and Financial Stability — Box 3.2 (continued)* --- _Source: https://www.imf.org/-/media/files/publications/gfsr/2024/october/english/ch3.pdf_