No more free lunch: The increasing popularity of machine learning and financial market efficiency

Abstract

In this paper, we show that the increasing popularity of machine learning improves market efficiency. By analysing the performance of a set of popular machine learning-based investment strategies, we find that profits from these strategies experience significant declines since the wide adoption of machine learning techniques, especially for profits based on the more preferred method of neural networks. These declines mainly come from long legs. Using the ‘machine learning’ Google search index as a proxy for machine learning-based trading intensity, we find that returns from the neural networks-based long–short and long-only strategies are weaker following high levels of machine learning intensity, while no relation is found between machine learning intensity and the short-only neural networks-based strategy.

Keywords:

• Machine learning
• market efficiency
• mispricing
• neural networks
• arbitraging activities

Acknowledgements

We thank Shiyang Huang, Zhigang Qiu, Hanwen Sun, Chi-Yang Tsou, Yanyi Wang, Hong Xiang, Weinan Zheng, and participants at the 2021 International Forum on Development of FinTech for helpful comments and suggestions. Substantial parts of this research were done when Jian Feng was at Renmin University of China. All errors are our own.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 For the survey results, see “The Machine Learning Trends Transforming Finance” published on 17 April 2019, https://www.refinitiv.com/perspectives/ai-digitalization/the-machine-learning-trends-transforming-finance/.

2 Google search indices are available starting from 2004.

3 For detailed reviews of this literature, see Welch and Goyal (2008), Koijen and Van Nieuwerburgh (2011), and Rapach and Zhou (2013).

4 Our sample starts in 1975 because we need sufficient quarterly accounting variables from Compustat to compute firm characteristics.

5 These operations are typically referred to as activation functions. They play a key role in NN by adding non-linearity into the model while maintaining the feasibility of calculating gradients and optimisation. Commonly used non-linear operations are Sigmoid, Tanh, and ReLU, among others.