金工文献精译第三期:机器学习驱动下的金融对不确定性的吸收与加剧-20220211-德邦证券-19页_1mb
报告摘要
金融工程专题:机器学习驱动下的金融对不确定性的吸收与加剧
核心内容
本文探讨了机器学习(ML)在金融领域对不确定性的吸收与加剧现象。通过分析182位金融从业者,特别是45位积极应用ML技术的人员,研究揭示了ML技术在金融不确定性管理中的双重作用:一方面,它能够吸收自然不确定性,将市场风险转化为可操作的预测;另一方面,它也引入了“关键模型不确定性”,即模型的预测和决策过程难以被解释,从而增加了新的不确定性。
主要观点
- 不确定性是金融市场的固有特征:不确定性不仅是风险,也可能是可利用的资源,尤其是在大数据和ML技术日益普及的背景下。
- 机器学习作为不确定性吸收工具:ML模型通过识别数据中的相关性,帮助吸收自然不确定性,并将其转化为可管理的风险。
- 关键模型不确定性:ML模型,尤其是深度神经网络,因其高度不透明性,导致无法解释其预测和决策机制,从而产生新的不确定性。
- 期望的反身性:在金融市场中,期望影响现象的实现,而现象又反过来影响期望,这种反身性使估值和市场行为更加复杂。
- 认知标准化与模型使用:从业者通过使用工具(如ML模型)进行信息比较和解释,从而实现“认知标准化”,这有助于减少不确定性。
- 模型的可解释性与透明度:对模型的不理解会引发对模型可靠性的质疑,因此从业者倾向于使用更直观的模型以降低不确定性。
- 组织权力的变化:随着ML技术的广泛应用,组织中的权力结构正在发生变化,ML模型在某些方面取代了人类对不确定性的吸收能力。
- 奥卡姆剃刀原则的应用:为了限制模型的复杂性,许多从业者偏好简单、直观的模型,以避免不可解释的模型带来的不确定性。
关键信息
ML模型在不确定性吸收中的作用
- ML模型通过分析大量非结构化、有噪声的数据,识别其中的复杂模式和结构,从而吸收自然不确定性。
- 例如,某清算银行的量化风险团队使用无监督神经网络检测客户交易中的异常行为,帮助识别潜在风险。
关键模型不确定性的挑战
- 由于ML模型的不透明性,模型的内部逻辑难以解释,这使得模型的预测和决策过程变得不可理解,进而导致新的不确定性。
- 这种不确定性对模型的可靠性提出质疑,影响了从业者对模型的信任。
从业者对模型的态度
- 一些从业者对ML模型持怀疑态度,认为其操作方式难以理解,因此倾向于使用更直观的模型,如线性回归。
- 例如,某量化分析师表示,若不理解模型的运作逻辑,他不会将其用于投资决策。
风险提示
- 市场波动风险:市场本身存在不确定性,ML模型无法完全消除这种波动。
- 数据可用性风险:ML模型依赖大量数据,而数据的获取和质量可能影响模型的有效性。
- 模型失效风险:模型可能在样本外表现不佳,尤其是当模型过度拟合训练数据时。
- 国内外市场差异风险:不同市场的结构和行为可能影响ML模型的适用性。
结论
本文指出,虽然ML技术在金融不确定性管理方面具有重要作用,但它也带来了新的挑战。关键模型不确定性成为研究重点,表明模型的不透明性可能加剧金融系统中的不确定性。因此,ML技术在金融领域的应用需要进一步研究,尤其是在模型解释性和组织权力配置方面。
参考文献
- Alpaydin, E. (2016). Machine learning: The new AI. The MIT Press.
- Arnoldi, J. (2006). Frames and screens: The reduction of uncertainty in electronic derivatives trading. Economy and Society.
- Arnott, R. D., Harvey, C. R., & Markowitz, H. (2018). A backtesting protocol in the era of machine learning. SSRN.
- Basili, M. (2001). Knightian uncertainty in financial markets: An assessment. Economic Notes.
- Beckert, J. (2016). Imagined futures: Fictional expectations and capitalist dynamics. Harvard University Press.
- Borch, C. (2016). High-frequency trading, algorithmic finance and the flash crash. Economy and Society.
- Carruthers, B. G. (2013). From uncertainty toward risk: The case of credit ratings. Socio-Economic Review.
- Esposito, E. (2011). The future of futures: The time of money in financing and society. Edward Elgar Publishing.
- Hansen, K. B. (2020a). Model talk: Calculative cultures in quantitative finance. Science, Technology, & Human Values.
- Hansen, K. B. (2020b). The virtue of simplicity: On machine learning models in algorithmic trading. Big Data & Society.
- March, J. G., & Simon, H. A. (1993). Organizations. Blackwell Publishers.
- Thévenot, L. (2002). Conventions of co-ordination and the framing of uncertainty. In E. Fullbrook (Ed.), Intersubjectivity in economics: Agents and structures.
- Varian, H. R. (2014). Big data: New tricks for econometrics. Journal of Economic Perspectives.
- Weigand, A. (2019). Machine learning in empirical asset pricing. Financial Markets and Portfolio Management.
附录
图1:所有被采访者
| Type | Number |
|---|---|
| Trading firms (High-frequency trading firms in particular) | 61 |
| Asset management firms | 12 |
| Hedge fund management firms | 19 |
| Banks | 19 |
| Broker, broker-dealer firms | 10 |
| Exchanges and other trading venues | 23 |
| Regulators | 10 |
| Data, technology, and analytics vendors | 15 |
| Other | 13 |
| Total | 182 |
资料来源:Hansen and Borch(2021),德邦研究所
图2:机器学习子样本采访列表
| Interview ID | Type of firm | Role | Location | Date |
|---|---|---|---|---|
| C002 | Investment bank | Developer | London | November 6, 2017 |
| C005 | Hedge fund | Machine learning researcher | London | November 7, 2017 |
| C006 | Algorithms trading firm | Machine learning engineer | London | November 22, 2017 |
| D006 | Hedge fund | Sr. research scientist | New York | December 12, 2017 |
| D012 | Hedge fund | Head of computer trading | New York | December 6, 2017 |
| D020 | Algorithms trading firm | Trading operations specialist | Chicago | September 27, 2017 |
| D021 | Algorithms trading firm | Software developer | Chicago | September 26, 2017 |
| D024 | Hedge fund | Trading algorithm engineer | Chicago | October 25, 2017 |
| D029 | Algorithms trading firm | Algorithms trading lead | Chicago | October 20, 2017 |
| D032 | Algorithms trading firm | Fund manager | Chicago | October 16, 2017 |
| D033 | Hedge fund | Chief scientist and CTO | San Francisco | January 22, 2018 |
| D038 | Algorithms trading firm | Quantitative trading analyst | Chicago | January 24, 2018 |
| BC001 | Algorithms trading firm | Founder and CEO (two persons) | London | August 30, 2018 |
| BC003 | Algorithms trading firm | Sr. Software Engineer | London | August 30, 2018 |
| BC004 | Algorithms trading firm | Delivery managers (two persons) | London | August 30, 2018 |
| BC005 | Algorithms trading firm | Head of market risk | London | August 31, 2018 |
| BC006 | Algorithms trading firm | Delivery manager, software engineer, and compliance officer (three persons) | London | August 31, 2018 |
| BC007 | Algorithms trading firm | CTO | London | August 31, 2018 and November 28, 2018 |
| BC008 | Algorithms trading firm | Infrastructure engineer | London | August 31, 2018 |
| BC009 | Algorithms trading firm | CEO | London | February 28, 2019 |
| BC010 | Algorithms trading firm | CEO and CTO (two persons) | London | February 28, 2019 |
| BC011 | Algorithms trading firm | Production team | London | March 1, 2019 |
| BC012 | Algorithms trading firm | Leadership team | London | March 1, 2019 |
| BC015 | Algorithms trading firm | CRO | London | August 29, 2019 |
| BC016 | Algorithms trading firm | Trader | London | August 29, 2019 |
| BC017 | Algorithms trading firm | CEO and CTO (same two persons as BC010) | London | August 29, 2019 |
| K007 | Pension fund | Quantitative portfolio manager | London | January 30, 2018 |
| K009 | Clearing bank | Head of quant risk team and machine learning quant (two persons) | Amsterdam | April 12, 2018 |
| K012 | Analytics vendor | Machine learning quant | New York | May 29, 2018 |
| K013 | Consultant | Quant trader and machine learning specialist | Spain | May 31, 2018 |
| K017 | Hedge fund | Quant analyst | Paris | June 19, 2018 |
| K018 | Hedge fund | Researcher | London | June 25, 2018 |
| K019 | Analytics vendor | Head of research | London | June 26, 2018 |
| K024 | Brokerage firm | Global head of product management and head of EMEA | London | September 5, 2018 |
| K026 | Hedge fund | Director of investment strategies | London | September 6, 2018 |
| K027 | Technology vendor | CSO | London | September 20, 2018 |
| K029 | Hedge fund | Deputy head of research | London | October 11, 2018 |
| K031 | Hedge fund | Quantitative researcher | London | November 2, 2018 |
| K038 | Analytics vendor | CEO | London | November 22, 2018 |
| K039 | Hedge fund | Senior quantitative analyst | London | December 20, 2018 |
| K040 | Hedge fund | Fund manager | London | March 5, 2019 |
| K041 | Investment bank | E-trading risk quant | London | March 28, 2019 |
| BK001 | Brokerage firm | Quantitative researcher and machine learning quant (two persons) | London | September 26, 2018 |
| BK002 | Brokerage firm | Head of quantitative trading | London | June 6, 2019 |
| G002 | Algorithms trading firm | Algorithms trading lead | Chicago | May 23, 2018 |
资料来源:Hansen and Borch(2021),德邦研究所
投资要点
- ML模型在吸收自然不确定性方面具有优势,但也引入了关键模型不确定性。
- 从业者对模型的不透明性存在担忧,倾向于使用更直观的模型以增强可解释性。
- 期望的反身性使金融市场的不确定性更加复杂和动态。
- 随着自动化和数据驱动的金融模式发展,不确定性不仅限于未来,也影响过去和现在。
- ML技术的广泛应用正在改变组织的权力结构,使其更多地依赖于算法而非人类判断。
风险提示
- 市场波动风险
- 数据可用性风险
- 模型失效风险
- 国内外市场差异风险
展开完整摘要
试读结束,高清完整版pdf/doc/ppt,请点下载