Deep Learning in Finance: From Implementation to Regulation

Abstract

Despite important theoretical questions that remain to be solved, Artificial Intelligence and Deep Learning are being increasingly used in the Finance and Insurance sector. Beyond straightforward data analytics, decision models are being implemented with Deep Learning. These algorithms cannot be used blindly. The understanding of the underlying problem is key. Humans, engineers or mathematicians, are essential. One trendy application is the use of Deep Learning (specifically GANs) to generate datasets. In finance, data are often scarce and having the possibility to generate new data (similar to an original dataset) can be decisive. In many applications, explainability of Artificial Intelligence is critical to protect consumers. Explainability is not a one-size-fits-all concept, and several degrees of explainability may have to be reached. Explainability to non-specialists is an additional challenge. Biais in the learning data is critical to assess because biases will be reproduced by the algorithm, and lead to unexplained discriminations. The role of regulatory agencies will be crucial to protect consumers while allowing innovation. There is currently no unified regulatory framework. The European Commission's Artificial Intelligence Act (draft proposal in April 2021) lists prohibited artificial intelligence practices and defines high-risk application areas for which they identify requirements ( risk management system, data governance, technical documentation and record keeping, transparency, human oversight, accuracy, robustness and cybersecurity).

Keywords: AI, deep learning, regulation, explainability

JEL Classification: G21,G22, G28