资源描述
The use of artificial intelligence and machine learning by market intermediaries and asset managers Consultation Report The Board OF THE INTERNATIONAL ORGANIZATION OF SECURITIES COMMISSIONS CR02/2020 JUNE 2020 This paper is for public consultation purposes only. It has not been approved for any other purpose by the IOSCO Board or any of its members.ii Copies of publications are available from: The International Organization of Securities Commissions website iosco International Organization of Securities Commissions 2020. All rights reserved. Brief excerpts may be reproduced or translated provided the source is stated.iii Foreword The Board of the International Organization of Securities Commissions (IOSCO) has published this Consultation Report to assist IOSCO members in providing appropriate regulatory frameworks in the supervision of market intermediaries and asset managers that utilise AI and ML. How to Submit Comments Comments may be submitted by one of the three following methods on or before 26 October 2020. To help us process and review your comments more efficiently, please use only one method. Important: All comments will be made available publicly, unless anonymity is specifically requested. Comments will be converted to PDF format and posted on the IOSCO website. Personal identifying information will not be edited from submissions. 1. Email Send comments to consultation-02-2020iosco. The subject line of your message must indicate The use of artificial intelligence and machine learning by market intermediaries and asset managers If you attach a document, indicate the software used (e.g., WordPerfect, Microsoft WORD, ASCII text, etc) to create the attachment. Do not submit attachments as HTML, PDF, GIFG, TIFF, PIF, ZIP or EXE files. 2. Facsimile Transmission Send by facsimile transmission using the following fax number: + 34 (91) 555 93 68. 3. Paper Send 3 copies of your paper comment letter to: Alp Eroglu International Organization of Securities Commissions (IOSCO) Calle Oquendo 12 28006 Madrid Spain Your comment letter should indicate prominently that it is a Public Comment on The use of artificial intelligence and machine learning by market intermediaries and asset managersiv Contents Chapter Page 1 Executive summary 1 2 Background and scope 4 3 How firms are using AI and ML techniques 7 4 Identified risks and harms posed by the use of AI and ML 10 5 Firms response to the potential risks arising from the use of AI and ML 15 6 Proposed guidance 18 7 Conclusion and next steps 23 A1 How regulators are addressing the challenges created by AI and ML 24 A2 Guidance published by supranational bodies 351 Chapter 1 - Executive Summary Background Artificial Intelligence (AI) and Machine Learning (ML), collectively called AI and ML, are increasingly being utilised in financial services, due to a combination of increased data availability and computing power. The use of AI and ML by market intermediaries and asset managers may be altering firms business models. For example, firms may use AI and ML to support their advisory and support services, risk management, client identification and monitoring, selection of trading algorithms and portfolio management, which may also alter their risk profiles. The use of this technology by market intermediaries and asset managers may create significant efficiencies and benefits for firms and investors, including increasing execution speed and reducing the cost of investment services. However, this use may also create or amplify certain risks, which could potentially have an impact on the efficiency of financial markets and could result in consumer harm. The use of, and controls surrounding AI and ML within financial markets is therefore a current focus for regulators across the globe. IOSCO identified the use of AI and ML by market intermediaries and asset managers as a key priority. The IOSCO Board approved a mandate in April 2019 for Committee 3 on Regulation of Market Intermediaries (C3) and Committee 5 on Investment Management (C5) to examine best practices arising from the supervision of AI and ML. 1 The committees were asked to propose guidance that member jurisdictions may consider adopting to address the conduct risks associated with the development, testing and deployment of AI and ML. Potential risks identified in the consultation report IOSCO surveyed and held round table discussions with market intermediaries and conducted outreach to asset managers to identify how AI and ML are being used and the associated risks. The following areas of potential risks and harms were identified in relation to the development, testing and deployment of AI and ML: Governance and oversight; Algorithm development, testing and ongoing monitoring; Data quality and bias; Transparency and explainability; Outsourcing; and Ethical concerns. Proposed IOSCO Guidance This consultation report proposes guidance to assist IOSCO members in providing appropriate regulatory frameworks to supervise market intermediaries and asset managers that utilise AI and ML. 1 Board Priorities - IOSCO work program for 2019, March 25, 2019, available at: iosco/library/pubdocs/pdf/IOSCOPD625.pdf2 The proposed guidance consists of six measures that reflect expected standards of conduct by market intermediaries and asset managers using AI and ML. Although the guidance is not binding, IOSCO members are encouraged to consider these proposals carefully in the context of their legal and regulatory frameworks. IOSCO members and firms should also consider the proportionality of any response when considering these proposals. The use of AI and ML will likely increase as the technology advances, and it is plausible that the regulatory framework will need to evolve in tandem to address the associated emerging risks. Therefore, this report, including the definitions and guidance, may need to be reviewed in the future to remain up to date. Measure 1: Regulators should consider requiring firms to have designated senior management responsible for the oversight of the development, testing, deployment, monitoring and controls of AI and ML. This includes requiring firms to have a documented internal governance framework, with clear lines of accountability. Senior Management should designate an appropriately senior individual (or groups of individuals), with the relevant skill set and knowledge to sign off on initial deployment and substantial updates of the technology. Measure 2: Regulators should require firms to adequately test and monitor the algorithms to validate the results of an AI and ML technique on a continuous basis. The testing should be conducted in an environment that is segregated from the live environment prior to deployment to ensure that AI and ML: (a) behave as expected in stressed and unstressed market conditions; (b) operate in a way that complies with regulatory obligations. Measure 3: Regulators should require firms to have the adequate skills, expertise and experience to develop, test, deploy, monitor and oversee the controls over the AI and ML that the firm utilises. Compliance and risk management functions should be able to understand and challenge the algorithms that are produced and conduct due diligence on any third-party provider, including on the level of knowledge, expertise and experience present. Measure 4: Regulators should require firms to understand their reliance and manage their relationship with third party providers, including monitoring their performance and conducting oversight. To ensure adequate accountability, firms should have a clear service level agreement and contract in place clarifying the scope of the outsourced functions and the responsibility of the service provider. This agreement should contain clear performance indicators and should also clearly determine sanctions for poor performance. Measure 5: Regulators should consider what level of disclosure of the use of AI and ML is required by firms, including: (a) Regulators should consider requiring firms to disclose meaningful information to customers and clients around their use of AI and ML that impact client outcomes. (b) Regulators should consider what type of information they may require from firms using AI and ML to ensure they can have appropriate oversight of those firms.3 Measure 6: Regulators should consider requiring firms to have appropriate controls in place to ensure that the data that the performance of the AI and ML is dependent on is of sufficient quality to prevent biases and sufficiently broad for a well-founded application of AI and ML.4 Chapter 2 - Background and Scope Previous IOSCO work in this area IOSCO has undertaken several workstreams on the use of AI and ML in financial markets, including: Committee on Emerging Risks (CER): The CER undertook a mandate on the use of novel technologies deployed by regulators to increase the efficiency and effectiveness of supervisory and oversight programs and published a report in February 2017. 2 CER examined the regulatory use of tools such as big data analytics and data visualisation technologies; AI and ML, and deep learning technologies; and distributed ledger technologies. Committee on Regulation of Secondary Markets (C2): C2 published a report in April 2013 on Technological Challenges to Effective Market Surveillance Issues and Regulatory Tools. 3 The report made recommendations to help market authorities address the technological difficulties facing effective market surveillance. IOSCO Fintech Network: The IOSCO Fintech Network was established in May 2018 to facilitate the sharing of knowledge and experiences among IOSCO members. The IOSCO Fintech Network considered the ethical implications of the use of AI and ML technologies. IOSCO Mandate Building on the previous IOSCO work, the proposed guidance seeks to address the potential risks and harms that may be caused by the use of AI and ML by market intermediaries and asset managers and looks to help ensure that market intermediaries and asset managers have: appropriate governance, controls and oversight frameworks over the development, testing, use and performance monitoring of AI and ML; staff with adequate knowledge, skills and experience to implement, oversee, and challenge the outcomes of the AI and ML; robust, consistent and clearly defined development and testing processes to enable firms to identify potential issues prior to full deployment of AI and ML; and appropriate transparency and disclosures to their investors, regulators and other relevant stakeholders. 2 IOSCO Research Report on Financial Technologies (Fintech), February 2017, available at: iosco/library/pubdocs/pdf/IOSCOPD554.pdf 3 Technological Challenges to Effective Market Surveillance Issues and Regulatory Tools, August 2012, available at: iosco/library/pubdocs/pdf/IOSCOPD389.pdf5 Defining the terms Al and ML for this consultation report AI can be understood as a combination of mass data, sufficient computing resources and machine learning. ML is a sub-set of AI which can be defined as a method of designing a sequence of actions to solve a problem, which optimise automatically through experience with or without human intervention. Artificial Intelligence The term “Artificial Intelligence”, first coined by data scientist John McCarthy 4 in 1956, can be understood as a combination of mass data, sufficient computing resources and ML, which can accomplish simple, repetitive tasks, or can be more sophisticated and, to some degree, self- learn and perform autonomously, based on a system that mimics human cognitive skills or human capabilities. However, the prospect of a computer having such a level of intelligence, also called “strong artificial intelligence” is not expected in the foreseeable future. AI in the financial services industry is still in its relative infancy and is poised to become more common, and with that will come legal, ethical, economic and regulatory challenges. Machine Learning The term “Machine Learning” is a specific subset and application of AI, which focuses on the development of computer programs that analyse and look for patterns in large quantities of data, with the aim of building knowledge to make better future decisions. ML algorithms differ from traditional algorithms in their ability to harness inductive reasoning, i.e., ML algorithms learn and develop using past data trends to predict future outcomes. Inductive reasoning is often used to predict future outcomes, with the accuracy of the hypothesis improving as more observations are made and the quality of the data improves. This however cannot be guaranteed due to potential data shortcomings. A successful ML algorithm will therefore learn and evolve over time and will possibly make recommendations that were not explicitly envisaged when it was created. 5 There are various categories of ML algorithms. These categories are based on the level of human intervention required in feeding back, and/or interpreting data from the algorithm. “Deep Learning”, a form of computationally intensive ML that learns association and statistical patterns often from a very large dataset, can also include any of these categories: 6 4 What is AI? available at: jmc.stanford.edu/artificial-intelligence/index.html 5 A famous example of this is the “move 37” in the game of Go: when Googles AlphaGo5 algorithm was pitted against professional Go player Lee Sedol in March 2016, making a move on the 37th turn that was previously unimaginable. This algorithm used “deep learning”, a form of ML technique that efficiently learns associations and statistical patterns from a very large dataset. 6 Deep learning is a method that analyses data in multiple layers, starting with learning about simple concepts then learning more complex concepts. Deep learning can be used for all three categories of machine learning algorithms.6 Supervised learning: the algorithm is fed an initial set of data that has been labelled. Based on this training set, the algorithm will learn classification rules and predict the labels for the remaining observations in the data set. Reinforcement learning: the algorithm is fed an initial set of data that has not been labelled and is asked to identify clusters of observations underpinned by similar characteristics. As it chooses an action for the data points, it receives feedback that helps it learn. 7 Unsupervised learning: the algorithm detects patterns in the data by identifying clusters of observations underpinned by similar characteristics it uncovers the structure of the data on its own. 7 R Sutton, A Barto, Reinforcement Learning: an introduction, MIT Press, 1998.7 Chapter 3 How firms are using AI and ML techniques AI and ML use by market intermediaries and asset managers Market intermediaries and asset managers use of AI and ML is growing, as their understanding of the technology and its utility evolves. The rise
展开阅读全文