Machine Learning & Black Swans

Black Swans were defined by Nassim Taleb in 2007 as events characterised by their rarity and unpredictability, the severity of their consequences, and the tendency to rationalise them after the fact. In finance, these events have the characteristic of being able to be exogenous: natural events, pandemic, war, but also of being able to be generated by the financial markets themselves in the form of speculative bubbles. Economic and financial markets have thus experienced several global events of this kind, for example the real-estate bubble of 2008, often classified in the category of Black Swans. The unpredictability of these bubbles rests on the difficulty of determining whether they really exist or not, of measuring their evolution and their capacity to burst, and above all their potential impact on the markets and the economy.

The irruption of ChatGPT-4 in 2023 constituted a turning point in the understanding of the possibilities of AI and in the acceleration of the tools used until then. In market finance, the rapid development of new devices made it possible to generalise and extend algorithms that were, for some, already advanced (high-frequency trading, robot-advisory, web scraping), while also multiplying the number of players active in the field. Whereas the emphasis was placed first on the massive processing of documents, synthesis and rewriting by Large Language Models (LLM), intelligence is now increasingly used in investment itself, applying in particular the principles of machine learning.

Machine-learning algorithms are designed to learn on the basis of vast volumes of financial data and to identify trends and formulate predictions or decisions without human intervention or with limited human intervention. However, impressive as it may be, AI's capacity to predict or mitigate the impact of particular events is not assured. This is linked both to the by-definition unpredictable nature of the Black Swan, but also to the characteristics of learning.

Indeed, the general objective of learning is to minimise entropy so that the predicted distribution is as close as possible to the observed distribution, thereby tending to treat separately or to ignore outliers, whether aberrant or difficult to explain. Learning is also dependent on the access to and length of historical data, on their availability according to the domains, but also on the hypothesis that past data, massively used, make it possible to determine trends having a certain predictive power. Finally, an important part of the fine-tuning is to avoid "overfitting", that is, operating by favouring the training data to the detriment of predictive capacity. This naturally leads to setting aside or minimising the particular events that do not fall within the average.

The characteristics of learning, whether supervised or unsupervised, raise the question of the generation of a particular bubble and of the self-production of a Black Swan of a new kind through the multiplication of automated players whose diversity would be spurious. According to the principles established by Shannon, the number and multiplication of players does not mean that there is more information, nor that it is differentiated. This is illustrated by the paradox of the game invented by Shannon in 1948, which creates a discrete universal predictor profiting from the biases of the information provided (in this case an experimenter who generates numbers that are in reality not random). The facilitated access to advanced algorithms causes a multiplication of players, but these risk processing information in a similar manner. It is likely that these players, by making similar hypotheses and by using essentially the same data and the same algorithms, obtain similar results and reinforce the same predictions while limiting the possibilities of arbitrage. A parallel can be drawn with high-frequency trading, which made it possible to generate significant outperformance as long as the number of players was limited.

One may then ask whether the use of learning models makes it possible to achieve greater market efficiency, by multiplying the players, or on the contrary increases the risk of inefficiency by reproducing artificially cloned players. The paradox would then be that the maturity of AI-driven investment ends up reproducing the biases held against index investments: excessive concentration, fear of bubble creation, herd behaviour. However, it does not a priori bear the same major disadvantages associated with these index funds: lack of outperformance, lack of flexibility and the impossibility of being able to profit from opportunities on the market.

Thierry Vessereau joined BCGE Asset Management in July 2019 as an institutional manager on the passive and active funds of the Synchrony range. Previously, he worked from 2005 to 2019 at the bank Pictet & Cie as project and development manager, then as an investment specialist in the Indexation department of Pictet Asset Management. Thierry Vessereau holds a master's degree in computer science and a doctorate in economics.