Since the eighteenth century, mathematical modelling has attempted to represent physical reality and has more recently tackled the economic and financial domains. The golden rule of modelling has always been parsimony: try, if possible, to use as few equations and as few parameters as possible. Moreover, the estimation of parameters from a dataset using statistical techniques is all the better the smaller the number of these parameters, which implicitly validates the rule. Does this rule also apply to artificial intelligence?
Two recent articles on financial topics come to the opposite conclusions. The first, “The virtue of complexity” by Yale researcher Bryan Kelly, claims that the abundance of parameters improves modeling, while the second, “Less is more? Biases and overfitting in machine learning return predictions” by Clint Howard, an analyst at Robeco, tends to confirm the golden rule. The first attempts to show theoretically and empirically that by adding more nodes and layers in a neural network, one of AI’s flagship tools, the quality of predictions improves. In particular, on an example the author takes, these networks predict the monthly evolution of the Dow Jones from 1926 to the present much better than the models of classical econometrics and obtain Sharpe ratios for portfolio management strategies based on these forecasts that are twice as high as those of conventional linear models.
The second article attempts to show, on the other hand, that abundance is not always synonymous with improvement. The researcher is interested in forecasting the U.S. equities returns over the since 1957. He compares the results obtained by using the full dataset to those obtained by classifying companies according to their size (small, medium and large). The very simple experiment is to look at the results of the models depending on whether all the information is used or not. In this example, the information ratios (measuring the performance of the models) are higher regardless of the models (classical or AI) with companies divided into several groups… so with less (and more homogeneous) data.
AI is disrupting habits and beliefs, and finance is no exception to this challenge. Several researchers have tried to theoretically show why “more is better”, like the first paper. But the demonstrations are based on assumptions that often reduce the scope of the conclusion. The example taken by the second article, admittedly on a different dimension (sample size, shows that the opposite is sometimes true! The purported ability of AI to differentiate data and overcome artifacts, so well illustrated in the case of image analysis, is put at fault on a banal example…
Many articles also use the textual analysis capabilities of AI and contradict common misconceptions, particularly in terms of market efficiency. The ability to predict market developments in the short term has been repeatedly proven in recent articles. The combination of quantitative tools such as neural networks and textual analysis promises another qualitative leap in this area. Moreover, the increase in available data, for example with the recent reporting obligations for non-financial data, is in line with the use of more complex tools than the models used over the past fifty years, which had revolutionized economic and financial disciplines.
However, the reliability and stability of mathematical or computational models remains and will remain debatable. Are we sure that the world we live in is predictable? As JM Jancovici likes to say, laws of nature cannot be modified by the National Assembly, but finance laws can. How could markets abstract themselves from the political and social context that surrounds them? In addition, AI, like classical modeling, works all the better if the right predictive variables have been well chosen!
“The virtue of complexity”, Bryan Kelly, Inquire Europe seminar, Cologne, September 2024
Less is more? Biases and overfitting in machine learning return predictions”, Clint Howard, Inquire Europe seminar, Cologne, September 2024
This article is written by Jean-François Boulier, Member of the Board and Chairman of the Inquire Europe Prize Committee