## tnmea2025016

## Source details

**Canonical URL:** [tnmea2025016](https://www.imf.org/-/media/files/publications/tnm/2025/english/tnmea2025016.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/tnm/2025/english/tnmea2025016.pdf.md)
- [Structured JSON version](/-/media/files/publications/tnm/2025/english/tnmea2025016.pdf.json)

---

### Abbreviations
- AEs advanced economies
- AI artificial intelligence
- ChatGPT Chat Generative Pre-Trained Transformer
- EMDEs emerging market and developing economies
- ESMA European Securities and Markets Authority
- FCA Financial Conduct Authority
- FSB Financial Stability Board
- GenAI generative artificial intelligence
- HFT high-frequency trading
- IOSCO International Organization of Securities Commissions
- IT information technology
- ML machine learning
- SEBI Securities and Exchange Board of India
- SEC Securities and Exchange Commission
- SME small and medium-sized enterprises

### I. Introduction — role of AI in capital markets
- Capital markets facilitate exchange of financial securities, enable economic growth, and drive technological innovation.
- Drivers of AI adoption:
  - larger volume of available data
  - cheaper and more easily accessible computing power and cloud storage
  - more advanced algorithms
  - greater investment in AI companies
  - availability of open-source AI platforms
- AI integration influences portfolio management, trading and investment strategies, investment advisory services (such as robo-advisors), client interfaces, and risk management.
- Historical overlap: much of what is categorized as AI in finance overlaps with long-standing statistical or econometric modeling techniques.
- Use cases include robo-advisory, asset management, trade execution, post-trade settlement, trade anomaly detection, crowdfunding.
- Client-facing AI must account for fiduciary nature of intermediary–investor relationships.
- Regulatory and supervisory imperatives:
  - Understand evolving, increasingly complex AI models.
  - Ensure regulatory compliance and data privacy.
  - Prevent overreliance on algorithms.
  - Address risks of increased market volatility.
- Concentration risks: concentration of systems and models among key players (capital market infrastructure providers or systemically important financial institutions offering multiple services).
- Heterogeneous adoption: effects will vary across countries depending on regulatory capacity and stage of capital market development.
- IMF concerns about risks include:
  - data-related risks (privacy and bias)
  - performance risks (robustness, synthetic data, explainability)
  - new cybersecurity threats (data manipulation attacks)
  - broader risks to financial stability (increased homogeneity in risk assessments and credit decisions, out-of-sample risks, interconnectedness fostering systemic risks)
- GenAI concerns: herd mentality bias, mispricing risks, hallucinations, potential solvency and liquidity challenges, and data manipulation attacks.
- Purpose of the technical note:
  - Stock take of AI uptake in securities markets and approaches to regulation and supervision.
  - Complement ongoing work by IOSCO and other standard-setting bodies and IMF initiatives.
  - Provide overview of GenAI spread, risks from rapid adoption, regulatory and supervisory developments and trends, and differences between AEs and EMDEs.
  - Identify practices authorities could adopt within supervisory frameworks, with attention to resource constraints in some EMDEs.

### II. AI Adoption in Capital Markets
- Background:
  - Use of ML in capital markets is not new; algorithmic trading exists for several decades.
  - Recent GenAI developments affect business conduct across the investment process (use of alternative data sets and forward-looking indicators by buy-side firms; sell-side use for risk assessment, pricing, forecasting).
- Drivers and firm objectives:
  - Firms increase AI investments to enhance internal operations, improve efficiency, reduce costs, and enhance client experience.
  - IOSCO (2025a) notes firms leverage GenAI and large language models primarily for internal applications (task automation, productivity enhancement through chatbots, improved risk management).
  - CFA Institute (2024b) found that 35 percent of investment professionals are already using AI tools like ChatGPT in their workflows.
- Use cases and workflow impacts (selected GenAI applications):
  - Customized content generation and interactive chatbots
  - Portfolio optimization through diffusion-based models
  - Simulation of volatility surfaces through variational autoencoders
  - Sentiment analysis through large language models
  - Risk simulation through generative adversarial networks
  - Analysis of unstructured clients’ data
  - Scenario simulations
- Current state and outlook:
  - AI use often extends existing advanced analytical tool adoption; more significant changes anticipated in the medium-to-long run.
  - Consensus: AI adoption will continue at pace and expand across business lines, bringing new risks warranting increased focus by securities market regulators.
- Adoption in EMDEs:
  - Five areas where AI adoption is material and could accelerate: asset management, wholesale trading, robo-advisory services, neo-brokerage, and crowdfunding.
- Authorities using AI:
  - Financial sector authorities use AI for automated data checks, combining multiple data sources, detecting trading anomalies; central banks use AI for nowcasting and market sentiment analysis.
  - Adoption rates vary significantly across AEs and EMDEs; resource levels are a key factor.

### Asset Management — AI adoption and potential benefits
- Current adoption:
  - Adoption in asset management is largely nascent (IMF 2024a).
  - 2024 survey shows many asset managers have yet to implement AI techniques.
  - Among adopters, emphasis is on research-related applications (big data analysis, idea generation, signal identification) rather than fully automated trade execution or decision-making.
- Expected benefits:
  - Most substantial benefits expected in client interactions and risk and compliance monitoring rather than direct investment management and trade execution.
  - Potential transformations:
    - optimizing portfolios
    - assessing and mitigating risks
    - predicting market dynamics
  - Deep learning used to identify complex patterns for asset allocation (neural networks with more than three layers characterized as deep learning algorithms).
- Scope: covers proprietary and third-party tools/service providers used by asset managers.

### 2. Estimated Efficiency Gains across the Asset Management Value Chain
- Top estimated efficiency gains (top five, range in percentage):
  - 25%
  - 29%
  - 31%
  - 32%
  - 40%
- Labels shown in source: Investment decision making; We don't use AI in investment process and alpha generation; Identifying data/signals that serve as proxies for missing information; Idea generation; Big data analysis
- Sources cited: BCG 2024; and Mercer 2024.
- AI and ML applications:
  - AI models identify nonlinear patterns and synthesize unstructured data (The Alan Turing Institute 2023).
  - Deep learning can refine portfolio policy frameworks (Simon, Weibels, and Zimmermann 2022).
  - Genetic programming can minimize model-fitting errors (Liu, Zhou, and Zhu 2020).
  - ML methods can improve forecasts of expected returns (Kelly, Malamud, and Zhou 2021) and high-frequency returns (Kolm, Turiel, and Westray 2021).
  - Nonparametric approaches can advance valuation methodologies (Filipovic, Pelger, and Ye 2022).
  - ML techniques can help quantify risk of sudden crashes (Swinkels and Hoogteijling 2022).
  - NLP enables extraction of insights from news, research reports, social media, filings.
  - CNNs and OCR could facilitate information extraction from images and documents when paired with NLP.
- GenAI potential and uses:
  - Generate synthetic data, emulate market scenarios, enhance investment approaches.
  - Use cases: hypothetical market scenarios, thematic idea generation, uncover hidden patterns or “black swan” events, apply GANs to analyze joint probability distributions and model synthetic order-book data (Luk 2023), devise in-house optimization strategies, streamline coding/testing/deployment, enhance analytics dashboards.
  - Caveat: integration of GenAI into core functions not yet fully materialized at scale; due diligence and testing needed.
- Data quality and investor outcomes:
  - Appropriate outcomes depend on quality, integrity, and reliability of underlying data.
  - Lack of sufficient data quality may increase risk of poor outcomes for investors.

### Wholesale trading and algorithmic trading
- Algorithmic trading is growing; AI advancements expected to influence development.
- AI-driven trading anticipated to thrive in liquid asset classes such as public equities.
- Globally, algorithmic trading activities in public equities are estimated to be 30 percent or higher of total transactions across regions.
- Adoption estimates in panel 1 are as of December 2024.
- In EMDEs, adoption varies significantly between and within countries.
- Algorithmic-related transactions for BSE and National Stock Exchange in panel 2 include transactions from participants identified as algorithmic, direct market access, colocation, and smart-order routing.

### Robo-advisory services
- Robo-advisors provide personalized investment guidance with minimum human intervention; typically authorized as investment advisors.
- Use AI for client profiling, investment planning, portfolio design and management, rebalancing, pattern recognition, and data extraction.
- Robo-advisory industry growth:
  - Assets under management (AUM) with robo-advisors are projected to exceed $2 trillion by 2028.
  - Estimated average growth rates (nominal AUM growth weighted by country count, from 35 AEs and 114 EMDEs):
    - United States (right scale): 418% per annum
    - AEs excluding the United States: 140% per annum
    - EMDEs: 172% per annum
  - Aggregate view: robo-advisory services have grown in EMDEs at a pace comparable to AEs; when adjusted for total gross national income, EMDEs are at similar levels with more uneven growth within EMDEs.

### Neo-broking and retail access
- Neo-brokers offer low-cost, user-friendly digital brokerage services; well positioned to adopt AI for personalized recommendations, engagement, and fraud detection.
- Order flows becoming increasingly automated may increase ecosystem vulnerability to interactions among AI-driven algorithms amplifying market volatility.
- AUM and penetration:
  - EMDEs' Growth: 223% annualized (reported annualized growth rates, 2017–22)
  - AEs' Growth: 197% annualized (reported annualized growth rates, 2017–22)
  - Penetration rate in major EMDEs increasing faster and reaching almost at par with AEs.
- Sample notes: Figures are simple averages across the 9 AEs and 10 EMDEs in the Group of Twenty; estimates exclude brokers solely providing robo-advisory services or cryptocurrency assets.

### Crowdfunding and GenAI impacts
- Two crowdfunding types: loan-based (peer-to-peer lending) and investment-based (crowd-investing).
- Platforms are fully digitalized and increasingly integrate AI for forecasting, personalization, automation, risk assessment, investor-project matching.
- GenAI explicit capabilities:
  - Automated Personalized Content Creation: campaign descriptions, fundraising updates, marketing materials, social media videos, hashtags, thank-you notes; third-party tools can generate campaign descriptions from key inputs.
  - Customized Investment Recommendations and Investors’ Matching: GenAI evaluates user preferences, risk tolerance, goals; third-party fundraising-as-a-service providers use GenAI to identify potential investors, refine targeting, create hyperpersonalized pitches, facilitate meeting scheduling; risk of encouraging fraudulent practices.
  - Personalized Investor Interaction: AI-driven chatbots and virtual assistants provide instant responses, guide investors, deliver real-time updates; third-party providers enable businesses to design, train, deploy chatbots.
- Market trends and risks:
  - Overall value of crowd-investing transactions remains low relative to total funding; EMDE growth rates high and outstripped those in AEs.
  - United Kingdom and the United States constitute more than half of overall transaction volume.
  - Average capital raising sizes remain small.
- Regulatory implications:
  - Need for increased regulatory scrutiny, particularly in EMDEs; attention areas: company eligibility, transparency and disclosure, investor protection, platform compliance, cross-border regulations.

### 1. Funds Using Alternative Data Sets
- Percentage of funds using specific sources (cumulative percentages shown):
  - Email receipt, 19%
  - Geolocation, 24%
  - Satellite, 29%
  - Web traffic, 31%
  - App usage, 33%
  - Other, 35%
  - Social/sentiment, 36%
  - Credit/debit card, 38%
  - Web data (scraping), 43%
- Additional breakdown of “Other*” (Other*, 40%):
  - Publicly available govt. data, 9%
  - News and media sentiment data, 8%
  - Employment data, 7%
  - Web-Scraped data, 7%
  - Environment, social, and governance, 7%
  - Social media, 6%
  - Real estate data, 6%
  - Consumer reviews, 5%
  - Transcription data, 5%
- Majority of these data sets are open source; Python is the most widely preferred language for handling unstructured data.
- Other data types cited: energy consumption data, e-commerce data, supply chain and logistic data, credit card transactions, weather data, app download data, court records and legal documents, satellite imagery, insider trading data, patent and intellectual property data, crypto data, geolocation from mobile foot traffic, IoT device data, flight tracking, clickstream, sensor technologies, wearables.
- Market fairness and supervisory challenges:
  - Integration of alternative data sets raises concerns about fairness and information asymmetry.
  - Institutional investors may leverage exclusive data sets, disadvantaging retail participants.
  - Widespread use of similar ML techniques increases potential for correlated strategies and herding.
  - Uncertainty over classification of aggregated consumer spending data affects legal use and market abuse implications.
  - Supervisory challenges: overseeing firms using alternative data, verifying data provenance, analyzing complex AI models, establishing cross-border data-sharing frameworks, clarifying fair access to data.
- Misconduct aimed at retail investors:
  - AI-enabled market manipulation (spoofing, layering, wash trading, quote stuffing) and AI-amplified retail misconduct (misleading advice, mis-selling).
  - Examples: 2010 Flash Crash (Navinder Singh Sarao); SEBI’s 2023 action against Nimi Enterprises.
  - Enforcement: US SEC actions on “AI washing”; enforcement against Delphia and Global Predictions (KKC 2025).
  - Retail risks: unsuitable robo-advisor recommendations, systematic demographic overexposure, “black box problem.”
  - Supervisory implication: capacity building critical for monitoring AI systems and ensuring transparency and accountability.
- Regulatory responses and international initiatives:
  - Danielsson and Uthemann (2024) proposed six criteria: data, mutability, objectives, authority, responsibility, consequences.
  - The Alan Turing Institute (2023) identified four regulatory considerations: (1) consumer protection, (2) competition, (3) market integrity, (4) operational resilience; recommended “human-in-the-loop.”
  - Bank of England (2024) emphasized training, monitoring, control, alignment with regulatory frameworks, and adversarial techniques in stress testing.
  - Standard-setting bodies:
    - FSB (2017); FSB updates November 2024 (FSB 2024) and October 2025 (FSB 2025).
    - IOSCO (2021) measures and IOSCO consultation report (2025a).
  - Jurisdictional landscape: disparity across 31 jurisdictions hosting the 25 largest stock exchanges; supervisory responses vary, with some EMDEs yet to provide guidance on robo-advisory and algo-trading.

### 4. Guidance on Algorithmic Trading — Snapshot Statistics and Examples
- Snapshot Statistics:
  - Yes, 88%
  - Not yet, 12%
  - Yes, 79%
  - Not yet, 21%
- Select examples from Advanced Economies:
  - ESMA (February 2023) found AI models enable optimization of trade execution and post-trade processes.
  - ESMA (May 2024) expects firms to comply with existing regulatory requirements when using AI.
  - EU AI Act provides requirements for model documentation, cybersecurity, robustness assessment, disclosure when AI content is generated, and human oversight for “high-risk” AI systems (Article 14).
  - Hong Kong SFC (SFC 2024) circular on generative AI: core principles governance, AI model risk management, data security, oversight of third-party providers; senior management responsible; transparency and data privacy compliance required.
- Select examples from EMDEs:
  - China: first administrative regulation on GenAI services by CAC effective August 15, 2023 (CAC 2023).
  - India: SEBI reporting requirements on AI/ML from 2019; SEBI (2024) proposed tightening governance and accountability rules; India’s algorithmic trading: over 50 percent of trades are algo-based and 80 percent of orders originate from colocation facilities.
  - United Arab Emirates: Joint “Guidelines for Financial Institutions Adopting Enabling Technologies.”
  - Thailand: algorithmic trading proportion in Stock Exchange of Thailand was 44.8 percent in October 2025; existing rules include order message limit of 25 orders per second.
  - Brazil, Mexico, South Africa: national AI plans, draft bills, and AI policy frameworks in development (various summaries provided).
- Key insights:
  - Regulatory responses diverse and evolving; frameworks must remain adaptive.
  - Chinese regulators established an algorithm registry for documentation.
  - Some regulators integrate AI into supervisory frameworks to leverage benefits.
  - EMDE focus: data challenges, market collusion/manipulation risks, and coordination outcomes; national plans prioritize AI governance and competition law reviews.
  - National frameworks increasingly hold AI users liable for outputs and encourage fair, ethical use.
  - Rise of social media and AI-driven mass content increases misinformation risk; authorities intensify efforts on investor protection and social media monitoring.

### V. Recommendations — Supervision of AI Use in Capital Markets (High-Level)
- Seven high-level recommendations complementary to IOSCO (2021) and consistent with FSB (2024) conclusions; consider incorporation into ongoing/planned regulatory or supervisory initiatives.
- Subrecommendations and key measures:

  - Enhancing AI Supervisory Capacity and Skillset
    - Build specialized supervisory teams for AI in capital markets (surveillance of AI-powered trading, automated market-making, AI-driven investment management, AI-based credit scoring).
    - Support expertise with targeted training modules (ML models in trading, market microstructure analysis, risk assessment of AI-driven market-making).
    - Establish technical partnerships with academic institutions and industry experts.
    - Develop knowledge repositories with case studies, common AI models, and documentation standards.
    - Leverage international cooperation via regional supervisory networks and ongoing education.
    - EMDE focus: address challenges evaluating risk models and trading algorithms.

  - Harnessing Technology-Enabled Supervisory Tools
    - Implement secure, validated platforms for market surveillance; adopt validated surveillance systems with basic AI pattern detection and strict data security.
    - Use open-source analytics for public data analysis; deploy secure proprietary solutions for sensitive market data complemented by real-time monitoring and alerting.
    - Consider shared regional regulatory platforms to distribute costs while maintaining data security and local control.
    - Validate surveillance systems via back-testing and regular recalibration.
    - Suggested approaches: cost-effective surveillance tailored to national market structures; regional cooperation; scalable solutions evolving with market sophistication.

  - A Proportionate Approach Toward Market Monitoring
    - Leverage surveys of AI adoption, reporting from regulated entities, and public disclosures.
    - Focus indicators on major market movements and potential risks; choose indicators based on data availability and supervisor capacity.
    - Indicators range from microsecond-level order-book changes to daily/weekly aggregate statistics for smaller markets.
    - Prioritize automated indicators that function with basic existing technology.
    - EMDE recommendations: practical monitoring for high retail participation markets; account for global AI models’ inability to capture local conditions; frameworks functioning despite infrastructure limitations.
    - Sample reference indicators (categories and frequencies preserved): Exchange/Trading Venue Metrics (Percentage of algorithmic trading order; Order-to-trade ratios; Message traffic patterns; Cancel/modify statistics — Daily or weekly); Robo-Advisory Specific (Number of active clients; AUM by strategy type; Average account size; Performance metrics — Through regulatory reporting); Trading patterns (Avg. trade size; Weekly monitoring for manipulation; Percentage mobile vs web trading; Popular stock concentration).

  - Enhanced Transparency and Disclosure Requirements
    - Clear disclosure requirements for AI-related service providers, especially robo-advisors and automated trading systems:
      - Plain-language descriptions of AI processes, key assumptions, limitations, and role of human oversight.
      - Regular reporting of performance and risk metrics suitable for retail investors.
    - Governance and accountability:
      - Clear ownership and oversight structures; rigorous validation and monitoring; transparent documentation and audit trails; accountability mechanisms.
    - Regulated entities should maintain website sections explaining AI use, with updates and risk disclaimers; provide educational materials in local languages.
    - Implement standardized disclosure templates focused on benefits, risks, fee structures, and investor rights.
    - Ensure resolution channels and right to “opt out” for users preferring human judgment.
    - Balance transparency with protection of proprietary training data and technical details to mitigate unintentional collusion risks.
    - For EMDEs: mandate retail-focused disclosures, governance frameworks, resolution channels, address information asymmetries, tailor transparency to financial literacy levels.

  - Social Media Monitoring and Market Manipulation Prevention
    - Implement basic social media sentiment monitoring on major platforms, tracking predefined keywords, hashtags, and influential accounts.
    - Use simple sentiment analysis to detect pump-and-dump schemes or misinformation; calibrate to local languages and usage patterns.
    - Establish patterns of manipulative behavior specific to markets; deploy pattern recognition to track unusual trading volumes, price movements, and correlations with social media activity; prioritize small-cap and newly listed companies.
    - Develop simplified reporting guidelines for participants to flag suspicious social media activities.
    - For EMDEs with limited alternative data or lower social media penetration: create adapted frameworks and monitoring systems operating with limited technological resources.

  - Cross-Border Supervision and Cooperation
    - Establish bilateral information-sharing with jurisdictions hosting offshore platforms targeting domestic retail investors; share suspicious trading patterns and manipulative schemes.
    - Maintain database of offshore platforms and regulatory status.
    - Prioritize tracking fund flows and trading through “official channels” while acknowledging limits monitoring “unofficial channels” (alternative trading platforms, decentralized finance).
    - Increase participation in regional supervisory colleges and informal supervisory groups on cross-border market abuse.
    - Implement initiatives gradually, starting with high-risk areas.
    - Suggested actions: create joint policies and tools to monitor algorithmic and HFT risks; develop cooperation mechanisms reflecting varying market development; share watch lists and manipulation pattern databases.

  - Measures to Address Concentration and Infrastructure Risk
    - Concern: concentration among IT infrastructure and AI software services (hardware, software, cloud services, pre-trained models).
    - Risks: single points of failure, amplified operational risk, systemic disruptions.
    - Recommended actions:
      - Coordinate in defining critical AI third-party service providers and map interdependencies.
      - Incorporate cybersecurity across design, procurement, deployment, and operations.
      - Ensure infrastructure and governance secure platforms against cybersecurity attacks or manipulation.
      - Enhance monitoring of large and concentrated market participants, including non-bank financial intermediaries.

### Financial Stability Risks and Market Structure Findings
- IMF’s October 2024 Global Financial Stability Report (IMF 2024a) identifies four key financial stability risks from AI/GenAI adoption:
  - increased market speed and volatility under stress (especially if AI trading strategies become highly correlated)
  - opacity and monitoring challenges (particularly as activities migrate to nonbank financial intermediaries)
  - operational risks from reliance on key third-party AI service providers
  - increased cyber and market manipulation risks
- Data-related challenges and market concentration:
  - Historical market data in many jurisdictions lack depth, consistency, and granularity.
  - Fragmented data across licensing, supervision, enforcement, market infrastructures, exchanges, and custodians.
  - Real-time market data access constrained by cost and technical limits.
  - Implications: models trained on incomplete data may perform poorly; limited alternative data reduces scope for novel strategies; data quality issues may amplify stress-period risks; higher costs hinder local firms.
  - Market concentration dynamics: data oligopolies and herd behavior can reinforce dominance by firms with better data and models; EMDEs may host global firms with advanced AI alongside local firms using basic tech, creating two-tier markets.

### Market Collusion, Manipulation, and Historical Episodes
- Risks of collusion-like behavior:
  - AI trading agents pretrained together can autonomously evolve collusive behaviors producing supra-competitive profits without explicit intent.
  - Algorithmic market makers may unintentionally align strategies, creating wider spreads and disadvantaging smaller participants.
- Empirical/illustrative points:
  - Thailand: foreign-driven short-selling and program trading accounted for up to 15 percent of daily trading; Thai SEC introduced short-selling limits and the “uptick rule”.
  - South Africa: foreign investors hold about 30 percent of Johannesburg Stock Exchange-listed equities and about 25 percent of government bonds.
  - Mexico: foreign-owned banks dominate OTC derivatives transactions, consolidating market power.
- Historical malfunctions:
  - Knight Capital Group’s 2012 algorithmic trading malfunction generated $440 million in losses within 45 minutes.
  - Renaissance Institutional Equities Fund experienced significant challenges in March 2020 when AI models trained on historical data struggled during unprecedented market conditions.

### Use of Alternative and Unstructured Data: Usage and Risks
- Usage statistics:
  - CFA 2024b: 64 percent of investment professionals now use alternative data across regions.
  - 48 percent of this unstructured data comes from open sources.
- Common alternative data types: satellite imagery, social media data, web-scraped data, transcription data, anonymized and aggregated retail data.
- Risks:
  - Insider dealing and market misconduct when processing breaches privacy/confidentiality.
  - GenAI plus social engineering can amplify disinformation.
  - Adversarial manipulation (e.g., adversarial patches) can mislead AI models and induce herd behavior.
  - Reliance on open-source data and tools increases systemic vulnerabilities when models/data are concentrated.

### BOX 1. Market Collusion versus AI-Powered Coordination Outcomes
- Market Collusion:
  - Requires intent, direct communication, and legal breach; example: Trod Ltd/GB Eye Ltd coordinated pricing using automated pricing algorithms (CMA 2018).
- AI-Powered Coordination:
  - Unintentional coordination arises when independent AI systems interact and converge on outcomes resembling collusion without explicit intent.
  - Experiments (Calvano and others 2018) show AI pricing algorithms can autonomously punish deviations and return to higher price levels, leading to partially collusive equilibria.
  - Dou, Goldstein, and Ji 2025: behavior stems from optimization processes without deliberate programming for collusion.
- Regulatory implication:
  - Intent-based liability frameworks do not capture emergent, unintentional coordination; traditional market abuse tools may not detect subtle emergent patterns.
  - Markets using similar or identical pricing algorithms, especially via third-party providers, are vulnerable to unintended coordination.

### VI. Navigating the Path Forward
- AI adoption presents opportunities (expanded access via robo-advisory and crowdfunding; enhanced trading efficiency via algorithmic systems) and new risks requiring oversight and mitigation.
- Recommendations account for resource constraints and local market conditions, aiming to:
  - Bridge technological divides between international and local participants.
  - Account for limited supervisory resources and expertise.
  - Address AI adoption risks in less developed markets.
  - Create frameworks that evolve with market sophistication.
  - Enable effective supervision despite infrastructure limitations.
- For EMDEs:
  - No one-size-fits-all solution; balance AI innovation with resource constraints.
  - International cooperation and knowledge sharing essential.
  - Proportional monitoring frameworks must reflect specific market conditions and supervisory capabilities.
  - Recognize heterogeneous AI adoption: some EMDEs have advanced uptake requiring robust frameworks; others should build on existing approaches before adopting advanced tools.
  - Core principle: maintain appropriate safeguards for financial stability and investor protection.

*Source: IMF technical note content (tnmea2025016 - Appendix 1. Key Concepts 42; tnmea2025016 - 2. Estimated Efficiency Gains; tnmea2025016 - 1. AUM Held and Executed by Neo-Brokers; tnmea2025016 - 3. Average Funding Per; tnmea2025016 - 4. Guidance on Algorithmic Trading).*

### Appendix 1. Key Concepts 42

### Appendix 1. Key Concepts 42

### Abbreviations
- AEs advanced economies
- AI artificial intelligence
- ChatGPT Chat Generative Pre-Trained Transformer
- EMDEs emerging market and developing economies
- ESMA European Securities and Markets Authority
- FCA Financial Conduct Authority
- FSB Financial Stability Board
- GenAI generative artificial intelligence
- HFT high-frequency trading
- IOSCO International Organization of Securities Commissions
- IT information technology
- ML machine learning
- SEBI Securities and Exchange Board of India
- SEC Securities and Exchange Commission
- SME small and medium-sized enterprises

### I. Introduction — role of AI in capital markets
- Capital markets facilitate exchange of financial securities, enable economic growth, and drive technological innovation.
- AI can analyze vast amounts of data, identify patterns, and make rapid decisions; drivers include:
  - larger volume of available data
  - cheaper and more easily accessible computing power and cloud storage
  - more advanced algorithms
  - greater investment in AI companies
  - availability of open-source AI platforms
- AI integration is influencing many aspects of capital markets, including:
  - portfolio management
  - trading and investment strategies
  - investment advisory services (such as robo-advisors)
  - client interfaces
  - risk management
- Historical note: much of what is categorized as AI in finance overlaps with long-standing statistical or econometric modeling techniques.
- AI use cases in capital markets are extensive across intermediaries and customer interactions, including:
  - robo-advisory
  - asset management
  - trade execution
  - post-trade settlement
  - trade anomaly detection
  - crowdfunding
- Client-facing AI (e.g., chatbots) must account for the fiduciary nature of intermediary–investor relationships typical in capital markets.

- Regulatory and supervisory imperatives:
  - Understand the role of evolving, increasingly complex AI models in capital markets.
  - Ensure regulatory compliance and data privacy.
  - Prevent overreliance on algorithms.
  - Address risks of increased market volatility.

- Concentration risks:
  - Increased AI adoption may lead to concentration of systems and models among key players, such as capital market infrastructure providers or systemically important financial institutions offering multiple financial services.1,3

- Heterogeneous adoption:
  - AI adoption remains highly heterogeneous across jurisdictions; direction, timing, and magnitude of AI’s effect on financial systems are uncertain and will vary across countries depending on regulatory capacity and stage of capital market development.

- Definitions and risk framing:
  - No single unanimous definition of AI; examples of adopted definitions include OECD (2024), FSB (2024), and IOSCO (2021) definitions cited in source text.1
  - IMF concerns about risks include:
    - data-related risks (privacy and bias)
    - performance risks (robustness, synthetic data, explainability)
    - new cybersecurity threats (data manipulation attacks)
    - broader risks to financial stability (including increased homogeneity in risk assessments and credit decisions, out-of-sample risks, interconnectedness fostering systemic risks)5
  - Generative AI (GenAI) introduces concerns such as herd mentality bias, mispricing risks, hallucinations, potential solvency and liquidity challenges, and data manipulation attacks.

- Purpose of the technical note:
  - Stock take of AI uptake in securities markets and approaches to regulation and supervision.
  - Complement ongoing work by IOSCO and other standard-setting bodies and IMF initiatives.4
  - Provide an overview of GenAI spread, risks from rapid adoption, regulatory and supervisory developments and trends, and differences between AEs and EMDEs.
  - Identify practices authorities could adopt within supervisory frameworks, with attention to resource constraints in some EMDEs.

### II. AI Adoption in Capital Markets
- Background:
  - Use of ML in capital markets is not new; algorithmic trading exists for several decades.
  - Recent GenAI developments are starting to affect business conduct throughout the investment process, including use of alternative data sets and forward-looking indicators by buy-side firms.
  - Sell-side institutions use AI for risk assessment, pricing, and forecasting.

- Drivers and firm objectives:
  - Across financial services, firms are increasing AI investments to:
    - enhance internal operations
    - improve efficiency
    - boost productivity by reducing costs
    - enhance client experience
  - IOSCO consultation report (IOSCO 2025a) notes firms leverage GenAI and large language models primarily for internal applications (task automation, productivity enhancement through chatbots, improved risk management) rather than customer-facing purposes.
  - CFA Institute (2024b) found that 35 percent of investment professionals are already using AI tools like ChatGPT in their workflows.

- Use cases and workflow impacts (Figure 1 highlights GenAI applications across investment process):
  - GenAI applications listed include:
    - Customized content generation and interactive chatbots
    - Portfolio optimization through diffusion-based models
    - Simulation of volatility surfaces through variational autoencoders
    - Sentiment analysis through large language models
    - Risk simulation through generative adversarial networks
    - Analysis of unstructured clients’ data
    - Scenario simulations
  - Workflow stages impacted: Market, Portfolio Construction/Rebalancing, Trading/Execution, Risk & Compliance, Investment and Risk Policy Statement, Reporting.

- Current state and outlook:
  - Outreach with buy-side and sell-side institutions (IMF 2024a) indicates current AI use in capital market activities often extends existing advanced analytical tool adoption; more significant changes anticipated in the medium-to-long run.
  - Consensus: AI adoption will continue at pace and expand across business lines and activities, bringing new risks that warrant increased focus by securities market regulators.

- Adoption in EMDEs:
  - For EMDEs, five areas where AI adoption is material and could accelerate: asset management, wholesale trading, robo-advisory services, neo-brokerage, and crowdfunding.6
  - Specific AI-related trends—such as the use of alternative and unstructured data sets—pose regulatory challenges (see Section III referenced in source).

- Authorities using AI:
  - Financial sector authorities increasingly integrate AI into their own activities (IMF 2024a), using AI for automated data checks, combining multiple data sources, and detecting trading anomalies.
  - Central banks have used AI to strengthen nowcasting techniques and evaluate market sentiment from unstructured sources (IMF 2023).
  - Adoption rates in capital markets activities, including regulatory use, vary significantly across AEs and EMDEs; divergent resource levels are a key factor.

### Asset Management — AI adoption and potential benefits
- Current adoption:
  - Adoption of AI in asset management is largely nascent (IMF 2024a).
  - A 2024 survey shows a significant number of asset managers have yet to implement AI techniques (Figure 2, panel 1).
  - Among adopters, emphasis is on research-related applications (big data analysis, idea generation, signal identification) rather than fully automated AI-driven trade execution or decision-making.

- Expected benefits:
  - Most substantial benefits are expected in client interactions and risk and compliance monitoring rather than direct improvements to investment management and trade execution (Figure 2, panel 2).
  - AI has potential to transform asset and portfolio management by:
    - optimizing portfolios
    - assessing and mitigating risks
    - predicting market dynamics
  - AI-driven tools assist managers to analyze historical data, associated economic indicators, and geopolitical events to optimize portfolios and proactively respond to market dynamics.
  - Deep learning (a subfield of ML) is used to identify complex patterns for precise asset allocation; note: neural networks with more than three layers are characterized as deep learning algorithms.8

- Scope of covered tools:
  - Section covers both proprietary and third-party tools/service providers used by asset managers for discharging their functions.7

*Source: IMF technical note content (Appendix 1. Key Concepts 42).*

### 2. Estimated Efficiency Gains across

### 2. Estimated Efficiency Gains across the Asset Management Value Chain

### Top estimated efficiency gains (top five, range in percentage)
- 25%
- 29%
- 31%
- 32%
- 40%
- Labels shown in source (as presented): Investment decision making; We don't use AI in investment process and alpha generation; Identifying data/signals that serve as proxies for missing information; Idea generation; Big data analysis

Sources cited for these estimates: BCG 2024; and Mercer 2024.

### AI and machine learning applications in asset management
- AI models can identify nonlinear patterns and synthesize unstructured data from diverse sources (The Alan Turing Institute 2023).
- Deep learning techniques can refine portfolio policy frameworks (Simon, Weibels, and Zimmermann 2022).
- Genetic programming can minimize model-fitting errors (Liu, Zhou, and Zhu 2020).
- ML methods can improve forecasts of expected returns (Kelly, Malamud, and Zhou 2021) and high-frequency returns (Kolm, Turiel, and Westray 2021).
- Nonparametric approaches can advance valuation methodologies (Filipovic, Pelger, and Ye 2022).
- ML techniques can help quantify risk of sudden crashes by revealing complex nonlinear relationships (Swinkels and Hoogteijling 2022).
- Natural language processing enables extraction of insights from news articles, research reports, social media posts, and company filings to inform market sentiment and risk assessment.
- Convolutional neural networks and optical character recognition could facilitate information extraction from images and documents when paired with NLP.

### Generative AI (GenAI) potential and uses
- GenAI can generate synthetic data, apply deep learning, emulate market scenarios, and enhance investment approaches.
- Use cases include:
  - Generating hypothetical market scenarios and simulating effects on portfolio performance and risk parameters.
  - Generating thematic ideas via processing unstructured data sets for sentiment analysis or keyword searches.
  - Uncovering hidden patterns or “black swan” events using general-purpose AI and more specialized generative architectures.
  - Applying generative adversarial networks to analyze joint probability distributions of asset prices, tune strategy hyperparameters, and model synthetic order-book data (Luk 2023).
  - Devising in-house portfolio optimization strategies and reducing reliance on historical data.
  - Streamlining coding, testing, validation, and deployment of trading and risk-management algorithms.
  - Enhancing analytics on dashboards for portfolio performance and market dynamics.

- Caveats:
  - Integration of GenAI into core functions has not yet fully materialized at scale.
  - Asset managers need sufficient due diligence and testing to ensure recommendations and synthetic data reliability.

### Data quality and investor outcomes
- Appropriate outcomes from data-driven activities depend largely on the quality, integrity, and reliability of underlying data.
- Lack of data of sufficient quality may increase the risk of poor outcomes for investors from firms’ use of AI.

### Wholesale trading and algorithmic trading
- Algorithmic trading is growing and becoming significant in some capital markets; development is expected to be influenced by AI advancements (IMF 2024a).
- AI-driven trading is anticipated to thrive in liquid asset classes such as public equities.
- Globally, algorithmic trading activities in public equities are estimated to be 30 percent or higher of total transactions across regions.
- Adoption estimates in panel 1 are as of December 2024.
- In EMDEs, adoption of algorithmic trading varies significantly, both between and within countries.
- Algorithmic-related transactions for BSE and National Stock Exchange in panel 2 include transactions from participants identified as algorithmic, direct market access, colocation, and smart-order routing.

### Robo-advisory services
- Robo-advisors provide personalized investment guidance with minimum human intervention and are typically authorized as investment advisors.
- Robo-advisors use AI for client profiling, investment planning, portfolio design and management, rebalancing, pattern recognition, and data extraction.
- Robo-advisory industry growth and impact:
  - Assets under management (AUM) with robo-advisors are projected to exceed $2 trillion by 2028.
  - Estimated average growth rates (nominal AUM growth weighted by country count, from 35 AEs and 114 EMDEs):
    - United States (right scale): 418% per annum
    - AEs excluding the United States: 140% per annum
    - EMDEs: 172% per annum
  - Adoption in AEs ex-US has been relatively modest by comparison, but AUM has grown significantly in recent years.
  - At an aggregate level, robo-advisory services have grown in EMDEs at a pace comparable to AEs, and when adjusted for total gross national income, EMDEs are at similar levels—albeit with more uneven growth within EMDEs.

### Neo-broking and retail access
- Neo-broking platforms offer low-cost, user-friendly digital brokerage services and are well positioned to adopt AI techniques for personalized recommendations, user engagement, and fraud detection.
- Order flows becoming increasingly automated may increase vulnerability of the trading ecosystem to interactions among AI-driven algorithms that could amplify market volatility.
- Neo-broking AUM and penetration:
  - AUM held and executed by neo-brokers has been steadily increasing, with EMDEs’ growth appearing to be on par with AEs from 2017 to 2022.
  - Penetration rate in major EMDEs is increasing faster than and reaching almost at par with that of AEs.

### Regulatory and infrastructure considerations
- IOSCO (2025b) highlighted concerns about neo-brokers: potential conflicts of interest, need for robust information technology infrastructure, fee disclosures, advertising practices, client consent for ancillary services, effects of noncommission revenues on order execution, and IT system robustness.
- Market development initiatives that have spurred algorithmic trading include electronification of trading, facilitation of high-frequency trading, listing of securities on public exchanges, and bundling into exchange-traded funds.

*Regulatory Considerations Regarding Accelerated Use of AI in Securities Markets (Technical Notes and Manuals).*

### 1. AUM Held and Executed by Neo-Brokers

### 1. AUM Held and Executed by Neo-Brokers

### Growth and User Penetration
- The increase in AUM held and executed by neo-brokers is driven by growth in the use of such services.
- Users’ dependence on neo-brokers in some EMDEs is increasingly more significant compared with AEs.
- Reported annualized growth rates, 2017–22:
  - EMDEs' Growth: 223% annualized
  - AEs' Growth: 197% annualized
- Sample and scope notes:
  - Sample data sets are sovereigns in the Group of Twenty.
  - Figures in panel 2 are estimated penetration rates based on a simple average across the 9 AEs and 10 EMDEs in the Group of Twenty.
  - Estimates of AUM are those held and executed by on-line brokers and platforms, including social trading platforms, but exclude brokers that solely provide robo-advisory services or cryptocurrency assets.
  - Figures include only entities that offer either exclusive business-to-customer (B2C) services or a combination of B2C and business-to-business (B2B) services, excluding those that cater solely to B2B clients.
- Definitions and acronyms preserved:
  - AEs = advanced economies
  - AUM = assets under management
  - CAGR = cumulative annualized growth rate
  - EMDEs = emerging market and developing economies
  - GNI = gross national income

### AUM Per User and Exposure
- The recent growth in assets-per-user on these platforms appears to be more pronounced in some EMDEs than in AEs, suggesting the demand gap these service providers seek to fill.
- Individual investors’ exposure to neo-brokers in some EMDEs is much larger than in AEs.
- The trend could continue, particularly if AI techniques are able to unlock value to meet individual investors’ needs through digital platforms with on-demand customized user experiences.
- Note on AUM per user presentation: AUM per user is shown relative to GNI per capita in the source figures.

### Crowdfunding and AI/GenAI Impacts
- Two primary types of crowdfunding:
  - Loan-based (peer-to-peer lending): individuals lend money to borrowers who promise to repay with interest; focused on debt financing and interest income.
  - Investment-based (crowd-investing): individuals invest in a start-up or business in exchange for equity and partial ownership.
- Platforms:
  - Crowdfunding platforms typically act as intermediaries, are fully digitalized, and have increasingly integrated AI techniques to enhance forecasting, personalization, automation, risk assessment, and investor-project matching.
  - Platforms facilitate issuance of securities (equity and debt), typically in smaller amounts compared with conventional banking facilities.
- GenAI effects on crowdfunding (explicit enumerated capabilities):
  - Automated Personalized Content Creation:
    - GenAI enhances creation of campaign descriptions, fundraising updates, marketing materials, social media videos, hashtags, and thank-you notes.
    - Third-party tools can generate campaign descriptions from key inputs such as funding goals and objectives.
  - Customized Investment Recommendations and Investors’ Matching:
    - GenAI evaluates user preferences, risk tolerance, and goals to generate personalized investment recommendations and improve matches between funders and businesses.
    - Third-party fundraising-as-a-service providers use GenAI to identify potential investors from public information, refine targeting, create hyperpersonalized pitches (including customized pitch decks), and facilitate meeting scheduling.
    - Example risk: GenAI could encourage fraudulent practices, misrepresentation, or false advertising.
  - Personalized Investor Interaction:
    - AI-driven chatbots and virtual assistants provide instant responses, guide investors through processes, deliver real-time updates, and can be trained on campaign information to handle inquiries.
    - Third-party providers enable businesses to design, train, and deploy their own chatbots on platforms.

### Market Trends, Risks, and Regulatory Considerations
- Growth patterns:
  - Although overall value of crowd-investing transactions remains low relative to total funding in the economy, EMDE growth rates have been high and consistently outstripped those in AEs.
  - Overall capital raised is still driven by AEs, with the United Kingdom and the United States constituting more than half of the overall transaction volume.
  - Average capital raising sizes remain small, reflecting the nascent stage and size of companies.
- Regulatory implications:
  - The growth of crowdfunding activities reflects the critical need for increased regulatory scrutiny, particularly in EMDEs, where rapid expansion could outpace authorities’ ability to adapt regulatory and supervisory frameworks.
  - Areas requiring attention include company eligibility, transparency and disclosure, investor protection, platform compliance, and cross-border regulations to ensure safe and effective expansion.

*Sources: OECD; Statista Digital Markets Insights; World Bank Databank; IMF World Economic Outlook Database; and IMF staff calculations.*

### 3. Average Funding Per

### 3. Average Funding Per

### AI Adoption and Capital Markets: Snapshot and Use Cases
- Table 1 summarizes AI use cases across five capital market segments: Asset Management, Wholesale Trading, Robo-Advisory, Neo-Broking, and Crowdfunding.
- Key service/features and growth trends noted:
  - Asset Management: portfolio optimization, risk assessment, market analysis, client-servicing; growing interest in unstructured data and client customization; adoption gaining momentum in AEs, EMDEs in early stages.
  - Wholesale Trading: advanced algorithmic trading, AI-driven price optimization, liquidity management; the United States leads in algo-trading adoption; EMDEs experiencing continued growth.
  - Robo-Advisory: personalized guidance, automated client profiling, lower fees; global AUM expected to exceed $2 trillion by 2028; EMDEs show higher growth rates despite lower volume in AEs.
  - Neo-Broking: fully digitalized retail experience; use of unstructured and alternative data for recommendations and fraud detection; growing use in some EMDEs.
  - Crowdfunding: lending-based growth serving SMEs; fraud detection, chatbots, GenAI for content; EMDEs’ growth rates exceed AEs’.
- Note: Sample data includes 34 AEs and 53 EMDEs. Figures in panel 2 are simple averages across equal-weighted sovereigns in AEs and EMDEs. AEs = advanced economies; EMDEs = emerging market and developing economies.

### Financial Stability and Risk Categories (IMF Findings)
- IMF’s October 2024 Global Financial Stability Report (IMF 2024a) identifies four key financial stability risks from AI/GenAI adoption:
  - increased market speed and volatility under stress (especially if AI trading strategies become highly correlated);
  - opacity and monitoring challenges (particularly as activities migrate to nonbank financial intermediaries);
  - operational risks from reliance on key third-party AI service providers;
  - increased cyber and market manipulation risks.
- The IMF (2024a) frames current AI adoption in capital markets as largely evolutionary, building on existing analytical methods and investment strategies.
- IMF (2023) highlights GenAI-specific risks: data privacy, embedded bias, model robustness issues (particularly “hallucination”), and synthetic data challenges.

### Supervisory Challenges and Opportunities
- Opportunities:
  - AI can improve price discovery, deepen liquidity, enhance risk management, and strengthen market surveillance.
  - Examples include AI-enhanced pricing in derivatives markets, dynamic counterparty risk assessment, optimized collateral management, and automated market making that could reduce bid-ask spreads.
  - Successful infrastructure adoption requires competitive capital markets and an independent, well-resourced securities markets authority.
- Challenges for authorities (especially in EMDEs):
  - Need for expertise in data science, AI, market surveillance, and risk management to assess AI-driven trading risks.
  - Difficulty evaluating risk models and trading algorithms used by market participants.
  - Constrained capacity to manage increased market interconnectedness and volatility spillovers.
  - Heightened cybersecurity concerns due to AI system vulnerabilities.

### Data-Related Challenges and Market Concentration
- Data limitations:
  - Historical market data in many jurisdictions lack depth, consistency, and granularity, especially in over-the-counter markets and less-liquid securities.
  - Fragmented data across licensing, supervision, enforcement, market infrastructures, exchanges, and custodians; limited integration capabilities.
  - Real-time market data access constrained by cost and technical limits, making sophisticated reporting difficult for smaller participants.
- Implications:
  - Models trained on incomplete data may perform poorly or generate unreliable signals.
  - Limited alternative data reduces scope for novel trading strategies.
  - Data quality issues may amplify risks during market stress periods.
  - Higher costs and technical barriers hinder local firms from implementing AI solutions.
- Market concentration dynamics:
  - Data oligopolies enable firms with superior nonpublic data to develop more effective AI models, increasing concentration.
  - Herd behavior, procyclicality, and program trading volatility can reinforce dominance by firms with better data and models.
  - EMDEs may host global financial firms with advanced AI alongside local firms using basic technology, creating a two-tier market and potential regulatory arbitrage.

### Market Collusion, Manipulation, and Historical Episodes
- Risks of collusion-like behavior:
  - AI trading agents pretrained together can autonomously evolve collusive behaviors, producing supra-competitive profits through strategic underreaction to market information.
  - Algorithmic market makers may unintentionally align strategies, creating wider spreads and disadvantaging smaller participants.
- Empirical and illustrative points:
  - Thailand: foreign-driven short-selling and program trading accounted for up to 15 percent of daily trading; Thai SEC introduced short-selling limits and the “uptick rule”.
  - South Africa: foreign investors hold about 30 percent of Johannesburg Stock Exchange-listed equities and about 25 percent of government bonds, creating systemic vulnerabilities.
  - Mexico: foreign-owned banks dominate over-the-counter derivatives transactions, consolidating market power.
- Historical market malfunctions:
  - Knight Capital Group’s 2012 algorithmic trading malfunction generated millions of erroneous orders, resulting in $440 million in losses within 45 minutes.
  - Renaissance Institutional Equities Fund experienced significant challenges in March 2020 when AI models trained on historical data struggled during unprecedented market conditions.

### Use of Alternative and Unstructured Data: Opportunities and Risks
- Usage statistics:
  - A global study (CFA 2024b) found that 64 percent of investment professionals now use alternative data across the Americas, Europe, the Middle East and Africa, and Asia Pacific regions.
  - 48 percent of this unstructured data comes from open sources.
- Common alternative data types: satellite imagery, social media data, web-scraped data, transcription data, anonymized and aggregated retail data.
- Risks:
  - Use of unstructured data raises insider dealing and market misconduct risks when data processing breaches privacy or confidentiality obligations.
  - GenAI combined with social engineering can amplify disinformation and exploit human vulnerabilities.
  - Adversarial manipulation of unstructured data (e.g., adversarial patches) can mislead AI models, inducing incorrect outputs and herd behavior across funds using similar data and models.
  - Widespread reliance on open-source data and processing tools promotes accessibility and innovation but increases systemic vulnerabilities when models and data are highly concentrated.

### Regulatory and Policy Responses Highlighted
- IMF recommendations and examples:
  - Recalibrate circuit breakers and enhance monitoring of large traders.
  - Address technological dependencies and strengthen market integrity measures.
  - Enhance international cooperation, cross-sectoral collaboration, and information sharing to monitor AI adoption implications (FSB 2024).
  - Thailand’s 2024 market-conduct reforms as a template: short-selling limits, the “uptick rule”, order message limits, mandatory strategy reporting, enhanced audit trails.
- Supervisory capacity building:
  - Prioritize building fundamental data capabilities and integration frameworks before pursuing advanced AI applications.
  - Adopt a phased and pragmatic approach aligned with supervisory priorities and capacity to ensure AI supports regulatory goals without overstraining resources.
  - Consider safeguards to mitigate market concentration, data oligopolies, and potential regulatory arbitrage by global firms operating in EMDE jurisdictions.

*Source: IMF staff compilations (tnmea2025016 - 3. Average Funding Per).*

### 1. Funds Using Alternative Data Sets

### 1. Funds Using Alternative Data Sets

### Types of alternative data used
- Percentage of funds using a specific source (cumulative percentages shown in source figure):
  - Email receipt, 19%
  - Geolocation, 24%
  - Satellite, 29%
  - Web traffic, 31%
  - App usage, 33%
  - Other, 35%
  - Social/sentiment, 36%
  - Credit/debit card, 38%
  - Web data (scraping), 43%
- Additional breakdown of “Other*” (Other*, 40%):
  - Publicly available govt. data, 9%
  - News and media sentiment data, 8%
  - Employment data, 7%
  - Web-Scraped data, 7%
  - Environment, social, and governance, 7%
  - Social media, 6%
  - Real estate data, 6%
  - Consumer reviews, 5%
  - Transcription data, 5%
- Note: The majority of these data sets are open source, and Python, also open source, is the most widely preferred language for handling unstructured data.
- Other data types cited in the source include: energy consumption data, e-commerce data, supply chain and logistic data, credit card transactions, weather data, app download data, court records and legal documents, satellite imagery, insider trading data, patent and intellectual property data, crypto data, geolocation data from mobile foot traffic, data from IoT devices, flight tracking data, clickstream data, sensor technologies, and wearables data.

### Market fairness, information asymmetry, and supervisory challenges
- Integration of alternative data sets (for example, consumer spending or transaction data) into investment decisions raises concerns about market fairness and information asymmetry.
- Institutional investors often leverage exclusive data sets to gain a competitive edge, disadvantaging retail participants.
- Widespread use of similar machine learning techniques (for example, natural language processing) on alternative and unstructured data increases the potential for correlated investment strategies and herding behavior, amplifying systemic risks.
- Uncertainty exists over classification of aggregated consumer spending data as public information or inside information; classification affects legal use in trading and potential market abuse implications.
- Supervisory challenges identified:
  - Difficulties in overseeing firms (for example, peer-to-peer lenders) using alternative data, which has led to predatory lending practices affecting small businesses and individuals in areas with limited traditional credit data.
  - Verifying data quality and provenance.
  - Analyzing complex AI models and their emergent behaviors.
  - Establishing effective cross-border data-sharing frameworks.
  - Need to clarify what constitutes fair access to data while balancing innovation and protection of market participants and public confidence.

### Misconduct targeted toward retail investors
- Two primary dimensions of misconduct in securities markets:
  - AI-enabled market manipulation (examples include spoofing, layering, wash trading, quote stuffing) that artificially influences market prices or trading volumes.
    - Case examples cited in the source: the 2010 Flash Crash (Navinder Singh Sarao), and the Securities and Exchange Board of India’s 2023 action against Nimi Enterprises.
  - Retail investor misconduct (misleading financial advice, mis-selling of products, exploitation of behavioral biases and financial literacy gaps), which AI can amplify through hyperpersonalized strategies.
- Examples and enforcement activity:
  - AI tools have been implicated in manipulative cases (UK FCA versus Da Vinci Invest case (2015); Navinder Sarao case (2016); Paul Axel Walter case (2017) cited as algorithmic manipulation instances).
  - The US SEC has brought enforcement actions against individuals, issuers, investment firms, and investment professionals involved in “AI washing” (false claims about AI use).
  - Recent enforcement actions against firms like Delphia and Global Predictions have resulted in significant penalties (KKC 2025).
- Specific retail investor risks from AI-driven platforms:
  - Robo-advisors and automated investment platforms may provide unsuitable recommendations by failing to properly account for individual circumstances and risk tolerance.
  - Potential systematic overexposure of certain demographic groups to high-risk investments or failure to adjust portfolios during market volatility because of rigid algorithmic rules.
  - The “black box problem”—AI models that are highly complex and may be incapable of comprehension by human users—makes detection and accountability for misconduct difficult.
- Practical supervisory implication:
  - Capacity building is critical for supervisors to monitor AI systems, ensure transparency in algorithmic decision making, and adapt regulatory and supervisory tools to AI-specific risks (data dependencies, model biases, emergent behaviors).

### Regulatory responses and international initiatives
- Recent analytical frameworks and recommendations:
  - Danielsson and Uthemann (2024) proposed six criteria for judging suitability of AI use by the private sector: data, mutability, objectives, authority, responsibility, consequences; identified channels through which AI can destabilize the system (malicious use, misinformed use and overreliance, misalignment and control avoidance, resource concentration).
  - The Alan Turing Institute (2023) identified four regulatory considerations arising from AI adoption: (1) consumer protection, (2) competition, (3) market integrity, and (4) operational resilience; recommended “human-in-the-loop” to encourage AI explainability.
  - Bank of England (2024) emphasized training, monitoring, and control (including risk and stop-loss limits); alignment of AI models with regulatory frameworks on an ongoing basis; and use of adversarial techniques in stress testing AI systems.
- Standard-setting bodies:
  - Financial Stability Board (FSB):
    - FSB (2017) identified risks to financial stability including third-party dependencies, emergence of new systemically important players outside the regulatory perimeter, and lack of interpretability/auditability of AI/ML methods.
    - FSB issued an update in November 2024 (FSB 2024) taking stock of recent advancements and AI-related vulnerabilities; followed by a further report in October 2025 (FSB 2025) offering advice on monitoring AI adoption in the financial sector.
  - International Organization of Securities Commissions (IOSCO):
    - IOSCO (2021) covered risks including governance and oversight; algorithm development, testing, and ongoing monitoring of data quality and bias; transparency and explainability; outsourcing; and ethical concerns.
    - IOSCO set out six measures for authorities (including designation of senior management responsible for AI/ML processes and continuous testing/validation/monitoring of AI/ML algorithms).
    - IOSCO published a consultation report (2025a) aimed at creating shared understanding of emerging AI technologies in financial products and services and is considering additional guidance as appropriate.
- Jurisdictional and supervisory landscape:
  - A review across a sample of 31 jurisdictions hosting the 25 largest stock exchanges shows disparity between national policy responses to AI and supervisory responses; supervisory responses—especially in EMDEs—do not uniformly reflect the rapid growth of key AI use cases in domestic capital markets.
  - Market regulatory authorities have generally implemented IOSCO (2021) measures as a first step, adapting or complementing those measures where desirable, with initiatives ranging from clarificatory guidance and outreach to principle-based approaches and enforcement.
  - Observed regulatory gaps include some EMDEs yet to provide guidance on robo-advisory services and a similar trend for algo-trading despite its long history.

*Sources: Alternativedata.org; CFA Institute’s July 2023 Generative AI, Unstructured/Alt data and open-source survey; and text contained in the source PDF.*

### 4. Guidance on Algorithmic Trading

### 4. Guidance on Algorithmic Trading

### Snapshot Statistics
- Yes, 88%
- Not yet, 12%
- Yes, 79%
- Not yet, 21%

Sources: UN Institute for Disarmament Research AI Policy Portal; UN data protection and privacy legislation worldwide; and IMF staff analysis and review of AI-related initiatives taken by financial sector authorities.
Note: The analysis covers 31 jurisdictions that are host to stock exchanges with the highest domestic market capitalization as of April 2024 (of the 31 jurisdictions, 11 are members of the EEA, and there are 33 financial market authorities, of which 13 are from the EEA and 3 from the US). AI = artificial intelligence; EEA = European Economic Area; EMDEs = emerging market and developing economies.

### Select Examples from Advanced Economies
- ESMA (February 2023) found AI models enable optimization of trade execution and post-trade processes, reducing market impact of large orders and reducing settlement failures.
- ESMA (May 2024) expects firms to comply with existing regulatory requirements when using AI and noted inherent risks including algorithmic biases, data quality issues, and privacy and security concerns.
- The EU AI Act provides requirements for model documentation, cybersecurity, robustness assessment, disclosure when AI content is generated, and requires human oversight for “high-risk” AI systems (Article 14), including provisions for human intervention.
- Hong Kong SFC (SFC 2024) issued a circular on responsible use of generative AI language models:
  - Emphasizes efficiency gains in client interactions, research, and operations and highlights risks: inaccuracies, biases, cybersecurity threats, and overreliance.
  - Core principles: governance, AI model risk management, data security, oversight of third-party providers.
  - Senior management responsible for policies and controls across AI life cycle; high-risk use cases (e.g., investment recommendations) require heightened scrutiny and human oversight.
  - Licensed corporations must maintain transparency with clients and comply with data privacy and intellectual property laws.

### Select Examples from EMDEs
- General: EMDE policymakers address AI risks via data protection, cybersecurity, consumer protection, and risk management; responses vary from monitoring and outreach to targeted guidance.
- China:
  - First administrative regulation on management of GenAI services issued by the Cyberspace Administration of China came into effect on August 15, 2023 (CAC 2023).
  - Measures apply to companies providing GenAI services to the public within China, regardless of incorporation, to promote regulated development, safeguard national security and social public interests, and protect lawful rights and interests.
- India:
  - SEBI issued circulars on reporting requirements for AI and ML applications as early as 2019 (SEBI 2019).
  - SEBI (2024) proposed tightening governance and accountability rules for AI in the private sector; proposed that every SEBI-regulated entity using AI in securities markets be solely responsible for consequences of such use, including privacy, security, and data integrity.
  - India’s algorithmic trading ecosystem: over 50 percent of trades are algo-based and 80 percent of orders originate from colocation facilities.
  - SEBI’s regulatory evolution includes: 2016 colocation reforms; 2018 managed colocation services and free tick-by-tick feeds; 2021 retail algorithmic trading regulations requiring exchange approvals for algorithms and classifying all API-Generated orders as algorithmic orders.
- United Arab Emirates:
  - Joint “Guidelines for Financial Institutions Adopting Enabling Technologies” set out cross-sector principles for technologies including AI, APIs, Big Data Analytics, Biometrics, Cloud Computing, and Distributed Ledger Technology.
- Thailand:
  - Thailand SEC studied high-frequency trading; proportion of algorithmic trading in the Stock Exchange of Thailand was 44.8 percent in October 2025.
  - Existing rules include intermediary systems to review and verify customer orders and an order message limit set at 25 orders per second.
  - SEC used GenAI in a social media–based sentiment analysis project addressing “finfluencers” and potential issues: lack of transparency, hidden marketing, inadequate disclosures, conflicts of interest.

- Brazil, Mexico, South Africa:
  - Brazil announced an AI investment plan to reduce reliance on imported AI tools; discussions on a comprehensive AI bill similar to EU AI Act.
  - Mexico adopted a National AI Agenda; a draft bill (submitted April 2024) would require every AI system to obtain regulatory authorization before distribution in Mexico.
  - South Africa released an AI Policy Framework in August 2024 balancing benefits with ethical, social, and economic considerations.

### Key Insights and Implications
- Regulatory responses are diverse and evolving; frameworks must remain adaptive due to rapid AI development.
- Examples illustrate measured, wide-ranging approaches addressing both general AI risks and jurisdiction-specific issues.
- Chinese regulators established an algorithm registry to enhance documentation and understanding of AI models.
- Some regulators (e.g., Thailand) are integrating AI techniques into supervisory frameworks to deepen understanding and leverage benefits.
- EMDE focus: data challenges, risks of market collusion/manipulation, and bad coordination outcomes. National plans prioritize AI governance and competition law reviews.
- National frameworks increasingly hold AI users liable for outputs and encourage fair, ethical use.
- Rise of social media and AI-driven mass content increases risk of misinformation; authorities are intensifying efforts to dispel rumors and tighten cybersecurity through multiagency collaboration.
- Retail investor protection, investor education, and social media monitoring are prioritized to prevent misinformation harms.

### V. Recommendations — Supervision of AI Use in Capital Markets (High-Level)
- The paper provides seven high-level recommendations for market regulatory authorities, complementary to IOSCO (2021) and consistent with FSB (2024) conclusions.
- Authorities should consider incorporating recommendations into ongoing/planned regulatory or supervisory initiatives (risk-based supervision, IT infrastructure enhancements).
- Sandboxes and SupTech tools require feasibility assessment and may need additional resources.

Subrecommendations and Key Measures

- Enhancing AI Supervisory Capacity and Skillset
  - Build specialized supervisory teams for AI in capital markets (surveillance of AI-powered trading, automated market-making, AI-driven investment management, AI-based credit scoring).
  - Support expertise development with targeted training modules: machine learning models in trading, market microstructure analysis, risk assessment of AI-driven market-making.
  - Establish technical partnerships with academic institutions and industry experts.
  - Develop comprehensive knowledge repositories with case studies, common AI models, and documentation standards.
  - Leverage international cooperation via regional supervisory networks and ongoing education.
  - For EMDEs: addresses challenges evaluating risk models and trading algorithms.

- Harnessing Technology-Enabled Supervisory Tools
  - Implement secure, validated platforms for market surveillance; adopt validated surveillance systems with basic AI pattern detection and strict data security.
  - Use open-source analytics for public data analysis (market sentiment, company disclosures, published market statistics).
  - Deploy secure proprietary solutions for sensitive market data (order-book monitoring, regulatory reporting) complemented by secure internal real-time monitoring and alerting.
  - Consider shared regional regulatory platforms to distribute costs while maintaining data security and local control.
  - Validate surveillance systems via rigorous back-testing and regular recalibration (alert conversion rates, system response times).
  - Suggested approaches:
    - Implement cost-effective surveillance systems tailored to national market structures.
    - Establish regional cooperation for shared technology platforms with secure data handling.
    - Create scalable solutions evolving with market sophistication.

- A Proportionate Approach Toward Market Monitoring
  - Leverage periodic/ad hoc surveys of AI adoption, reporting from regulated entities, and public disclosures.
  - Focus monitoring indicators on major market movements and potential risks; choose indicators based on data availability and supervisor processing capacity.
  - Indicators can range from microsecond-level order-book changes in large markets to daily/weekly aggregate statistics and basic pattern recognition in smaller markets.
  - Monitoring framework should account for market liquidity, retail participation, and vulnerability to external shocks.
  - Prioritize automated indicators that function with basic existing technology.
  - For EMDEs, recommendations include:
    - Practical monitoring for markets with high retail participation.
    - Addressing global AI models' inability to account for local conditions.
    - Frameworks that function despite infrastructure limitations.
  - Sample reference indicators (categories and frequencies preserved in source) include:
    - Exchange/Trading Venue Metrics: Percentage of algorithmic trading order; Order-to-trade ratios; Message traffic patterns; Cancel/modify statistics — Daily or weekly.
    - Robo-Advisory Specific: Number of active clients; AUM by strategy type; Average account size; Performance metrics — Through regulatory reporting.
    - Trading patterns: Avg. trade size; Weekly monitoring for manipulation; Percentage mobile vs web trading; Popular stock concentration.
    - Risk indicators and others as presented in Table 2 (frequencies: monthly, quarterly, semiannual as per category).

- Enhanced Transparency and Disclosure Requirements
  - Establish clear disclosure requirements for AI-related service providers, especially robo-advisors and automated trading systems:
    - Plain-language descriptions of AI processes, key assumptions, limitations, and role of human oversight.
    - Regular reporting of performance and risk metrics in terms suitable for retail investors.
  - Governance and accountability:
    - Clear ownership and oversight structures; roles and responsibilities across AI life cycle.
    - Rigorous validation and monitoring to ensure model performance, fairness, and compliance.
    - Transparent documentation and audit trails for traceability and explainability.
    - Accountability mechanisms for errors, biases, or unintended consequences.
  - Regulated entities should maintain website sections explaining AI use, with updates and risk disclaimers; provide educational materials in local languages.
  - Implement standardized disclosure templates focused on benefits, risks, fee structures, and investor rights.
  - Ensure resolution channels and right to “opt out” for users preferring human judgment; support timely human intervention for disputes.
  - Balance transparency with protection of proprietary training data and technical details to mitigate unintentional collusion risks; disclosures should focus on outcomes, assumptions, and risks.
  - Supplement monitoring with traditional supervisory tools: outreach, structured information gathering, and engagement with industry associations.
  - For EMDEs, consider:
    - Mandating retail-focused disclosures.
    - Providing governance/accountability frameworks.
    - Ensuring resolution channels and human intervention.
    - Addressing information asymmetries between global and local firms.
    - Tailoring transparency to varying financial literacy levels.

- Social Media Monitoring and Market Manipulation Prevention
  - Implement basic social media sentiment monitoring on major platforms, tracking predefined keywords, hashtags, and influential accounts.
  - Use simple sentiment analysis to detect unusual patterns indicating pump-and-dump schemes or misinformation campaigns; calibrate tools to local languages and usage patterns.
  - Establish patterns of manipulative behavior specific to markets (coordinated social campaigns, cross-border schemes, multiple small accounts).
  - Deploy basic pattern recognition to track unusual trading volumes, price movements, and correlations with social media activity; prioritize small-cap and newly listed companies.
  - Develop simplified reporting guidelines for market participants to flag suspicious social media activities.
  - For EMDEs with limited alternative data or lower social media penetration:
    - Create frameworks adapted to local social media usage.
    - Address cross-border manipulation risks.
    - Develop monitoring systems that operate with limited technological resources.

- Cross-Border Supervision and Cooperation
  - Establish bilateral information-sharing arrangements with jurisdictions hosting offshore platforms targeting domestic retail investors; share suspicious trading patterns and known manipulative schemes.
  - Maintain database of offshore platforms and regulatory status in major jurisdictions.
  - Prioritize tracking fund flows and trading patterns through “official channels” (regulated exchanges, centralized platforms) while acknowledging limits of monitoring “unofficial channels” (alternative trading platforms, decentralized finance).
  - Increase participation in regional supervisory colleges and informal supervisory groups on cross-border market abuse cases.
  - Implement initiatives gradually, starting with high-risk areas and expanding as resources and capabilities allow.
  - Suggested actions:
    - Create joint policies and tools to monitor and mitigate algorithmic and HFT risks across jurisdictions.
    - Develop cooperation mechanisms that reflect varying market development levels.
    - Share watch lists, databases of manipulation patterns, and coordinate cross-border responses.

- Measures to Address Concentration and Infrastructure Risk
  - Concern: concentration risk among IT infrastructure and AI software services, including hardware (accelerated computing chips), software, cloud services, and pre-trained models.
  - Risks: single points of failure, amplified operational risk, systemic disruptions.
  - Recommended actions:
    - Coordinate in defining critical AI third-party service providers and map interdependencies between critical AI and essential IT infrastructure providers.
    - Incorporate cybersecurity needs across design, procurement, deployment, and operations stages.
    - Ensure infrastructure and governance processes secure platforms against cybersecurity attacks or manipulation.
    - Enhance monitoring of large and concentrated market participants, including non-bank financial intermediaries.

### VI. Navigating the Path Forward
- AI adoption in capital markets presents opportunities (expanded access via robo-advisory, crowdfunding; enhanced trading efficiency via algorithmic systems) and new risks requiring oversight and mitigation.
- Recommendations address supervisory challenges while accounting for resource constraints and local market conditions; aim to provide practical, implementable solutions to:
  - Bridge technological divides between international and local participants.
  - Account for limited supervisory resources and expertise.
  - Address AI adoption risks in less developed markets.
  - Create frameworks that evolve with market sophistication.
  - Enable effective supervision despite infrastructure limitations.
- For EMDEs:
  - No one-size-fits-all solution; balance AI innovation with resource constraints.
  - International cooperation and knowledge sharing are essential.
  - Proportional monitoring frameworks must reflect specific market conditions and supervisory capabilities.
  - Recognize heterogeneous AI adoption: some EMDEs have advanced AI uptake requiring robust frameworks; others should build on existing approaches before adopting advanced tools.
  - Core principle: maintain appropriate safeguards for financial stability and investor protection.

*Regulatory Considerations Regarding Accelerated Use of AI in Securities Markets (Technical Notes and Manuals).*

### BOX 1. Market Collusion versus AI-Powered Coordination Outcomes

### BOX 1. Market Collusion versus AI-Powered Coordination Outcomes

### Market Collusion
- Collusion involves explicit intent and coordination between firms or market participants to manipulate prices, trading patterns, or market conditions for mutual benefit.
- Required elements of collusion:
  - Intent: Deliberate and premeditated actions to suppress competition.
  - Direct Communication: Active agreements or information exchanges (for example, meetings, emails).
  - Legal Breach: Collusion is illegal under most antitrust and competition laws because of its harmful effects on competition and consumer welfare.
- Illustrative example:
  - Trod Ltd/GB Eye Ltd case investigated by the UK Competition and Markets Authority (CMA): firms explicitly coordinated their prices using automated pricing algorithms. The presence of intent and direct communication made this a textbook case of illegal collusion (CMA 2018).

### AI-Powered Coordination of Actions and Outcomes
- Unintentional coordination arises when multiple independent AI systems interact in a shared market environment; systems programmed to optimize individual profits can inadvertently converge on outcomes that resemble collusion—without any explicit intent or communication.
- Mechanism:
  - AI systems interact repeatedly with the market, adjusting strategies based on observed outcomes. Over time they may ‘learn’ that actions like maintaining higher price levels or mimicking profitable patterns yield better returns.
- Empirical observations referenced:
  - Experiments reported by Calvano and others (2018) show that AI pricing algorithms can autonomously punish price deviations by temporarily cutting prices and gradually returning to higher price levels, leading to partially collusive equilibria.
  - Dou, Goldstein, and Ji 2025 note that this observed behavior stems from optimization processes within AI systems, without any deliberate programming for collusion or explicit agreements.
- Key distinction emphasized:
  - This form of coordination is emergent and unintentional.

### Intentional Versus Unintentional Coordination
- Definitions and contrast:
  - Intentional Collusion: Requires deliberate agreements, direct communication, and shared intent among participants to manipulate market outcomes.
  - Unintentional Coordination: Emerges naturally as independent agents optimize their objectives in a shared environment, leading to outcomes that mimic collusion but lack intent or explicit agreements.
- Regulatory implication:
  - Although intentional collusion is clearly illegal, unintentional coordination presents a regulatory gray area. It does not involve intent or communication but can still harm market efficiency and consumer welfare by creating artificially high prices or reduced competition.

### Supervisory and Regulatory Challenges
- Limitations of Intent-Based Tests:
  - Conventional liability frameworks rely on proving intent and causation, which are not applicable to autonomous systems that lack explicit programming for collusion.
- Market Monitoring Gaps:
  - Traditional market abuse detection tools are designed to identify clear and deliberate manipulation. They struggle to capture more subtle, emergent patterns of coordination that arise from the iterative interactions of AI systems.
- Algorithmic Similarity Risk:
  - Markets where competitors use similar or identical pricing algorithms, particularly through third-party providers, are especially vulnerable to unintended coordination, amplifying risks of market-wide anti-competitive outcomes.

*Sources: Ezrachi and Stucke 2015; Calvano and others 2018; CMA 2018; Azzutti, Ringe, and Stiehl 2021; Harrington 2024; Dou, Goldstein, and Ji 2025.*

### References

### References

### Cited works on AI and finance applications
- Ai, Wei, Junhui Jin, Qiazhu Mei, Jingyi Qiu, Teng Ye, and Jingnan Zheng. 2024. “Using Artificial Intelligence to Unlock Crowdfunding Success for Small Businesses.” https://ssrn.com/abstract=4806426.
- Boston Consulting Group (BCG). 2024. “AI and the Next Wave of Transformation.” Global Asset Management Report, 22nd Edition. https://www.bcg.com/publications/2024/ai-next-wave-of-transformation.
- Cambridge SupTech Lab. 2024. “State of SupTech Report 2024.” University of Cambridge. https://www.cambridgesuptechlab.org/SOS.
- Deloitte. 2016. “The Expansion of Robo-Advisory in Wealth Management.” https://www.scribd.com/document/422637475/Deloitte-Robo-Safe.
- Luk, Martin. 2023. “Generative AI: Overview, Economic Impact, and Applications in Asset Management.” ht tps://doi.org/10.2139/ssrn.4574814.
- Mercer. 2024. “Mercer Investments’ AI Integration in Investment Management 2024 Global Manager Survey.” https://www.mercer.com/insights/investments/portfolio-strategies/ai-in-investment-management-survey/.
- Northern Trust. 2023. “Investment Management in 2030: How Will Generative Artificial Intelligence Transform the Way Portfolio Managers Invest?” https://www.northerntrust.com/united-states/insights-research/asset-servicing/a-suite/insights-hub/investment-management-in-2030-generative-artificial-intelligence.
- Swinkels, Laurens, and Tobias Hoogteijling. 2022. “Forecasting Stock Crash Risk with Machine Learning.” https://www.robeco.com/en-us/insights/2022/06/forecasting-stock-crash-risk-with-machine-learning.

### Technical machine learning, robustness, and adversarial research
- Brown, Tom B., Dandelion Mané, Aurko Roy, Martín Abadi, and Justin Gilmer. 2017. “Adversarial Patch.” https://doi.org/10.48550/arXiv.1712.09665.
- Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. 2015. “Explaining and Harnessing Adversarial Examples.” ht t ps://doi.org /10.4 8550/arXi v.1412.6572.
- Harrington, Joseph E., Jr. 2024. “The Challenges of Third Party Pricing Algorithms for Competition Law.” May 11. https://doi.org/10.2139/ssrn.4824953
- Filipovic, Damir, Markus Pelger, and Ye Ye. 2022. “Stripping the Discount Curve—A Robust Machine Learning Approach.” Swiss Finance Institute Research Paper 22-24. ht tps://doi.org/10.2139/ssrn.4058150.
- Kelly, Bryan T., Semyon Malamud, and Kangying Zhou. 2021. “The Virtue of Complexity in Return Prediction.” Swiss Finance Institute Research Paper 21-90, October. ht tps://doi.org/10.2139/ssrn.3984925.
- Kolm, Petter N., Jeremy Turiel, and Nicholas Westray. 2021. “Deep Order Flow Imbalance: Extracting Alpha at Multiple Horizons from the Limit Order Book.” ht tps://doi.org/10.2139/ssrn.3900141.
- Simon, Frederik, Sebastian Weibels, and Tom Zimmermann. 2022. “Deep Parametric Portfolio Policies.” ht tps://doi.org/10.2139/ssrn.4150292.
- Liu, Yang, Guofu Zhou, Yingzi Zhu. 2020. “Maximizing the Sharpe Ratio: A Genetic Programming Approach.” ht tps://doi.org/10.2139/ssrn.3726609.
- Wah, Elaine, Mason Wright, and Michael P. Wellman. 2017. “Welfare Effects of Market Making in Continuous Double Auctions.” Journal of Artificial Intelligence Research 59. h t t p s : //d o i .o r g /10 .1613 /jair.5360.
- Budi, Indra, Betty Purwandari, and Ryan R. Suryono. 2021. “Detection of Fintech P2P Lending Issues in Indonesia.” Heliyon 7 (4): e06782. https://doi.org/10.1016/j.heliyon.2021.e06782.
- Gregoriades, Andreas, and Christos Themistocleous. 2025. “Improving Crowdfunding Decisions Using Explainable Artificial Intelligence.” Sustainability 17 (4): 1361. https://doi.org/10.3390/su17041361.
- Rudin, C. 2019. “Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead.” Nature of Machine Intelligence 1:206 –15. ht tps://doi.org/10.1038/s42256-019-0048-x.

### Market structure, algorithmic trading, liquidity, and market manipulation studies
- Bank for International Settlement (BIS). 2016. “Electronic Trading in Fixed Income Markets.” https://www.bis.org/publ/mktc07.pdf.
- Beketov, Mikhail, Kevin Lehmann, and Manuel Wittke. 2018. “Robo Advisors: Quantitative Methods Inside the Robots.” Journal of Asset Management 19:363 –70. ht tps://doi.org/10.1057/s41260 - 018 - 0 092-9.
- Calvano, Emilio, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello. 2018. “Artificial Intelligence, Algorithmic Pricing and Collusion.” CEPR Discussion Paper 13405, Centre for Economic Policy Research, London, UK.
- Calvano, Emilio, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello. 2020. “Artificial Intelligence, Algorithmic Pricing, and Collusion.” American Economic Review 110 (10): 3267–97.
- Kolm, Petter N., Jeremy Turiel, and Nicholas Westray. 2021. “Deep Order Flow Imbalance: Extracting Alpha at Multiple Horizons from the Limit Order Book.” ht tps://doi.org/10.2139/ssrn.3900141.
- Ramos, Henrique, and Marcelo Perlin. 2019. “Liquidity and Algorithmic Trading in Brazil.” ht tps://doi.org /10. 2139/s s r n. 3413130.
- Bloomberg. 2020. “Renaissance Says Quant Models Misfired during March Mayhem.” https://www.bloomberg.com/news/articles/2020-04-17/renaissance-says-quant-models-misfired-during-march-mayhem.
- Reuters. 2022. “Indonesia Seeks to Crack Down on “Unhealthy” Peer-to-Peer Lenders.” https://www.reuters.com/markets/rates-bonds/indonesia-seeks-crack-down-unhealthy-peer-to-peer-lenders-2022-07-22/.
- UK Financial Conduct Authority (UK FCA). 2017. “Final Notice to Paul Axel Walter—PAW01128.” ht tps://www.fca.org.uk/publication/final-notices/paul-axel-walter-2017.pdf.
- US Department of Justice (US DOJ). 2016. “Futures Trader Pleads Guilty to Illegally Manipulating the Futures Market in Connection with 2010 ‘Flash Crash.’” https://www.justice.gov/archives/opa/pr/futures-trader-pleads-guilty-illegally-manipulating-futures-market-connection-2010-flash.

### Regulatory, supervisory, and policy reports and guidance
- Central Bank of the United Arab Emirates (CBUAE). 2021. “Guidelines for Financial Institutions Adopting Enabling Technologies.” https://rulebook.centralbank.ae/en/rulebook/guidelines-financial-institutions-adopting-enabling-technologies.
- CFA Institute Research and Policy Center (CFA). 2024a. “Net-Zero Investing: Harnessing the Power of Unstructured Data.” https://rpc.cfainstitute.org/research/reports/2024/net-zero-investing-harnessing-the-power-of-unstructured-data.
- CFA Institute Research and Policy Center (CFA). 2024b. “Unstructured Data and AI: Fine-Tuning LLMs to Enhance the Investment Process.” https://rpc.cfainstitute.org/sites/default/files/-/media/documents/article/industry-research/unstructured-data-and-ai.pdf.
- Cyberspace Administration of China (CAC). 2023. “Interim Measures for the Administration of Generative Artificial Intelligence Services.” ht tps://www.cac.gov.cn/2023-07/13/c _1690898327029107.htm.
- European Commission. 2024. “AI Act enters into force.” https://commission.europa.eu/news-and-media/news/ai-act-enters-force-2024-08-01_en.
- European Parliament. 2020. “The Ethics of Artificial Intelligence: Issues and Initiatives.” https://www.europarl.europa.eu/thinktank/en/document/EPRS_STU(2020)634452.
- European Securities and Markets Authority (ESMA). 2023. “Artificial Intelligence in EU Securities Markets.” ESMA TRV Risk Analysis ESMA50-164-6247. https://www.esma.europa.eu/sites/default/files/library/ESMA50-164-6247-AI_in_securities_markets.pdf.
- European Securities and Markets Authority (ESMA). 2024. “On the Use of Artificial Intelligence (AI) in the Provision of Retail Investment Services.” Public statement issued on May 30. https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf.
- Financial Industry Regulatory Authority (FINRA). 2024. “Insider Trading Detection: FINRA’s Vital Role in Ensuring Market Integrity.” https://www.finra.org/media-center/finra-unscripted/insider-trading-detection-program-update.
- Financial Sector Conduct Authority of South Africa (South Africa FSCA). 2025. “Regulatory Strategy for 2025–2028”. https://www.fsca.co.za/Documents/FSCA%20Regulatory%20Strategy%202025-2028.pdf.
- Financial Stability Board (FSB). 2017. “Artificial Intelligence and Machine Learning in Financial Services: Market Developments and Financial Stability Implications.” https://www.fsb.org/2017/11/artificial-intelligence-and-machine-learning-in-financial-service/.
- Financial Stability Board (FSB). 2020. “Peer Review of Mexico”. Review Report. https://www.fsb.org/uploads/P190320.pdf.
- Financial Stability Board (FSB). 2024. “The Financial Stability Implications of Artificial Intelligence”. Reports to the G20. https://www.fsb.org/2024/11/the-financial-stability-implications-of-artificial-intelligence.
- Financial Stability Board (FSB). 2025. “Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector.” https://www.fsb.org/uploads/P101025.pdf.
- International Monetary Fund (IMF). 2021. “Powering the Digital Economy: Opportunities and Risks of Artificial Intelligence in Finance.” IMF Working Paper 2021/166, International Monetary Fund, Washington, DC.
- International Monetary Fund (IMF). 2022. “South Africa: Financial Sector Assessment Program Technical note on Systemic Liquidity Management.” IMF Country Report 22/183, Washington, DC.
- International Monetary Fund (IMF). 2023. “Generative Artificial Intelligence in Finance: Risk Considerations.” IMF Fintech Note 2023/006, International Monetary Fund, Washington, DC.
- International Monetary Fund (IMF). 2024a. “Advances in Artificial Intelligence: Implications for Capital Market Activities.” Global Financial Stability Report, International Monetary Fund, Washington, DC, Oc tober.
- International Monetary Fund (IMF). 2024b. “Crisis Amplifier? How to Prevent AI from Worsening the Next Economic Downturn.” Remarks by Gita Gopinath at the AI for Good Global Summit on May 30. ht tps://www.imf.org/en/News/Articles/2024/05/30/sp053024-crisis-amplifier-how-to-prevent-ai-from-worsening-the-next-economic-downturn.
- International Organization of Securities Commissions (IOSCO). 2021. “The Use of Artificial Intelligence and Machine Learning by Market Intermediaries and Asset Managers.” Final Report FR/06/2021. ht tps://www.iosco.org/library/pubdocs/pdf/IOSCOPD684.pdf.
- International Organization of Securities Commissions (IOSCO). 2022. “Report on Retail Distribution and Digitalisation.” Final Report FR/12/2022. https://www.iosco.org/library/pubdocs/pdf/IOSCOPD715.pdf.
- International Organization of Securities Commissions (IOSCO). 2023. “Retail Market Conduct Task Force Final Report.” Final Report FR/05/2023. https://www.iosco.org/library/pubdocs/pdf/IOSCOPD730.pdf.
- International Organization of Securities Commissions (IOSCO). 2025a. “Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges.” Consultation Report 2025/017. https://www.iosco.org/library/pubdocs/pdf/IOSCOPD788.pdf.
- International Organization of Securities Commissions (IOSCO). 2025b. “Neo-Brokers.” Final Report FR/18/2025. https://www.iosco.org/library/pubdocs/pdf/IOSCOPD810.pdf.
- Prenio, Jermy, Andrei Pustelnikov, and John Yeo. 2024. “Building a More Diverse Suptech Ecosystem: Findings from Surveys of Financial Authorities and Suptech Vendors”. FSI Briefs 23, Financial Stability Institute of the Bank for International Settlements, Basel, Switzerland.
- Securities and Exchange Board of India (SEBI). 2016. “Discussion Paper on ‘Strengthening of the Regulatory Framework for Algorithmic Trading & Co-location.’” Reports for public comments on August 5. ht tps://www.sebi.gov.in/reports/reports/aug-2016/discussion-paper-on-strengthening-of-the-regulatory-framework-for-algorithmic-trading-and-co-location-_32940.html.
- Securities and Exchange Board of India (SEBI). 2019. “Reporting for Artificial Intelligence (AI) and Machine Learning (ML) Applications and Systems Offered and Used by Market Intermediaries.” Circular to stock brokers and depository participants, recognized stock exchanges and depositories on January 4. https://www.sebi.gov.in/legal/circulars/jan-2019/reporting-for-artificial-intelligence-ai-and-machine-learning-ml-applications-and-systems-offered-and-used-by-market-intermediaries_41546.html.
- Securities and Exchange Board of India (SEBI). 2021. “Algorithmic Trading by Retail Investors.” Consultation paper on December 9. https://www.sebi.gov.in/reports-and-statistics/reports/dec-2021/consultation-paper-on-algorithmic-trading-by-retail-investors_54515.html.
- Securities and Exchange Board of India (SEBI). 2024. “Proposed Amendments with Respect to Assigning Responsibility for the Use of Artificial Intelligence Tools by Market Infrastructure Institutions, Registered Intermediaries and Other Persons Regulated by SEBI.” Consultation paper for public comments published on November 13. https://www.sebi.gov.in/reports-and-statistics/reports/nov-2024/proposed-amendments-with-respect-to-assigning-responsibility-for-the-use-of-artificial-intelligence-tools-by-market-infrastructure-institutions-registered-intermediaries-and-other-persons-regulated-b-_88470.html.
- Securities and Exchange Commission of Thailand (Thailand SEC). 2024. “The Rise of Finfluencers: Mapping the Landscape on Thailand’s Capital Market.” Presentation at the SEC Capital Market Symposium 2024. https://www.sec.or.th/TH/Documents/SEC_Symposium_2024/SEC_Symposium_2024-08.pdf.
- Securities and Futures Commission of Hong Kong Special Administrative Region (SFC). 2024. “Circular to Licensed Corporations—Use of Generative AI Language Models.” Issued on November 12. Securities & Futures Commission of Hong Kong.
- UK Finance. 2022. “Fair Use of AI.” UK Finance White Paper. https://www.ukfinance.org.uk/system/files/2022-06/AI%20fairness%20in%20financial%20services_FINAL.pdf.
- Lin, Kaitao, and Pedro Gurrola-Pérez. 2024. “An Analysis of Market Manipulation Definitions around the World.” https://wp.lancs.ac.uk/ffmm2024/files/2024/09/FFMM-2024-018-An-Analysis-of-Market-Manipulation-Definitions.pdf.
- Blackstone Chambers. 2015. “FCA v Da Vinci Invest Limited and others.” https://www.blackstonechambers.com/news/case-fca_v_da_vinci.
- Clifford Chance. 2019. “Big Data and Artificial Intelligence—Evolving Market Misconduct Risks.” https://www.cliffordchance.com/content/dam/cliffordchance/briefings/2019/03/big-data-and-artificialintelli-gence-evolving-market-misconduct-risks-1.pdf.
- CMA (UK Competition & Markets Authority). 2018. “Pricing Algorithms: Economic Working Paper on the Use of Algorithms to Facilitate Collusion and Personalised Pricing.” Working Paper, UK Competition & Markets Authority, London, UK.
- Ezrachi, Ariel, and Maurice E. Stucke. 2015. “Artificial Intelligence & Collusion: When Computers Inhibit Competition.” Oxford Legal Studies Research Paper 18/2015, University of Oxford, Oxford, UK.
- Calvano, Emilio, Giacomo Calzolari, Vincenzo Denicolò, and Sergio Pastorello. 2018 and 2020 entries (see above) on AI, algorithmic pricing, and collusion.
- Dou, Winston Wei, Itay Goldstein, and Yan Ji. 2025. “AI-Powered Trading, Algorithmic Collusion, and Price Efficiency.” ht tps://doi.org/10.2139/ssrn.4452704.
- Danielsson, Jon, and Andreas Uthemann. 2024. “On the Use of Artificial Intelligence in Financial Regulations and the Impact on Financial Stability.” ht tps://doi.org/10.2139/ssrn.4604628.
- Prenio, Jermy, Andrei Pustelnikov, and John Yeo. 2024. “Building a More Diverse Suptech Ecosystem: Findings from Surveys of Financial Authorities and Suptech Vendors”. FSI Briefs 23, Financial Stability Institute of the Bank for International Settlements, Basel, Switzerland.

### Institutional, standards, and explanatory references on AI
- The Alan Turing Institute. 2023. “The AI Revolution: Opportunities and Challenges for the Finance Sector.” https://arxiv.org/abs/2308.16538.
- Organization for Economic Co-operation and Development (OECD). 2024. “Explanatory Memorandum on the Updated OECD Definition of an AI System.” OECD Artificial Intelligence Paper 8. https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/03/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system_3c815e51/623da898-en.pdf.
- International Business Machine Corporation (IBM). 2023. “AI Vs. Machine Learning Vs. Deep Learning Vs. Neural Networks: What’s the Difference?” https://www.ibm.com/think/topics/ai-vs-machine-learning-vs-deep-learning-vs-neural-networks.
- Frankenfield, J. 2023. “Artificial Intelligence: What It Is and How It Is Used.” Investopedia. https://www.investopedia.com/terms/a/artificial-intelligence-ai.asp.
- UK Finance. 2022. “Fair Use of AI.” UK Finance White Paper. https://www.ukfinance.org.uk/system/files/2022-06/AI%20fairness%20in%20financial%20services_FINAL.pdf.

### Legal, market conduct, and case-focused citations
- Blackstone Chambers. 2015. “FCA v Da Vinci Invest Limited and others.” https://www.blackstonechambers.com/news/case-fca_v_da_vinci.
- Kohn, Kohn & Colapinto LLP (KKC). 2025. “AI in Finance: The SEC’s Rules and Whistleblowing.” ht tps://kkc.com/frequently-asked-questions/ai-finance-sec-rules-whistleblowing/.
- Nikkei Asia. 2024. “Thai SEC Sets Curbs on Foreign-Dominated Short-Selling.” https://asia.nikkei.com/Business/Markets/Thai-SEC-sets-curbs-on-foreign-dominated-short-selling.
- Securities and Exchange Board of India (SEBI) entries (2016, 2019, 2021, 2024) on algorithmic trading, reporting for AI/ML, and proposed amendments (see above).
- Securities and Futures Commission of Hong Kong Special Administrative Region (SFC). 2024. “Circular to Licensed Corporations—Use of Generative AI Language Models.” Issued on November 12. Securities & Futures Commission of Hong Kong.
- Financial Industry Regulatory Authority (FINRA). 2024. “Insider Trading Detection: FINRA’s Vital Role in Ensuring Market Integrity.” https://www.finra.org/media-center/finra-unscripted/insider-trading-detection-program-update.

*tnmea2025016 - References*

---


_Source: https://www.imf.org/-/media/files/publications/tnm/2025/english/tnmea2025016.pdf_
