Machine Learning in Asset Management: “AI doesn’t want anything – it only optimizes a measurable number”

leitwolf: Ever since ChatGPT, artificial intelligence (AI) has been on everyone's lips. How do large language models work, and what makes them successful?

Maximilian Kasy : The first step involves compiling large text datasets. This now essentially encompasses the entire internet, including Wikipedia databases, the scientific article database "arXiv," and the coding platform "GitHub." Based on all this data, so-called "foundation models" are trained, which statistically predict the most likely next word.

Fine-tuning takes place in subsequent steps. In recent months, this has focused in particular on reinforcement learning, which involves specific tasks, such as those in mathematics or programming, with correct solutions. The model is incentivized to learn to predict this solution independently. While media attention is currently focused almost exclusively on such language models, AI is much more.

Marvin Labod : 90 percent of our colleagues also think of language models when it comes to AI. However, we also use simpler, more proven models, for example, in volatility estimation. But one thing is also clear: The development in the acceptance, user-friendliness, and accessibility of such language models is enormous, given how instinctively students who come to us from universities today work with language models or solve problems, use them as research assistants, or employ them in coding. For us, these are all very important efficiency tools.

A major limitation of these models has long been cited as their inability to draw logical conclusions or to relate cause and effect. Instead, only correlations play a role. What's behind this?

Kasy : Behind this lie old debates in AI. Alternatives to machine learning, i.e., to statistics, have long emphasized logical reasoning, but with limited success. The reinforcement learning method seems to work very well for areas with clearly "correct" answers.

What's less clear, however, is whether the method helps in areas where multiple answers are possible. What's important to remember: The largest application of AI is still online advertising! Algorithms constantly estimate the causal effect of various decisions on advertising revenue. Causality isn't a problem in this sense.

AI models can be prone to "hallucinations" when applied to applications beyond the scope of their training data. How can this problem be reduced or eliminated?

Kasy : I don't think there are any fundamental solutions. The most helpful approach seems to me to be "tool use": Chatbots can search for sources on the internet or run code in Python, for example, to obtain externally verified answers. A good metaphor here is the Emmental cheese with holes in it: Even the information that language models can correctly answer has holes in it.

What do you think about the discussion surrounding the much-discussed term "Artificial General Intelligence (AGI)"? It is supposed to be largely similar to human intelligence and solve tasks like humans, only perhaps faster, better...

Kasy : We humans have a strong tendency to project human characteristics onto the non-human environment and attribute intentions and consciousness to it. This is even stronger with AGI. But the model itself doesn't "want" anything. I think we understand what AI does much better if we understand it as optimization based on statistical data. Put simply: Some measurable number should become as large as possible. Then we need to ask: What exactly should be optimized? And based on which data?

Labod : Exactly. The reward function, or how the model is trained, what it is trained on—that is very fundamental to the answers the model will later give us. What interests us as users—and my research is primarily concerned with this—is: What can the new models do and where and how can I use them? Theoretical aspects also play a role. This includes the question of how exactly and at which levels the reward systems work, which nodes are triggered, and what happens when certain nodes in the model are parameterized differently than before.

There are also voices warning against excessive hype surrounding AI. While there are many applications that are economically promising and have the potential to be disruptive, others are even further away than one might think, because AI is repeatedly faced with challenges for which it is not trained. What do you think about this?

Kasy: Machine learning has a relatively clear "production function." Its basis is statistical prediction; more data means better forecasts, and more complex application areas lead to poorer predictions. Thus, the success of machine learning always depends on the application area, or rather, the ratio of complexity to the (potentially) available data. The internet provides a vast amount of data, but there are also areas where there are fundamental limitations, for example, with macroeconomic information about financial crises or data on rare diseases.

Where do you see the limits of the responsible use of AI?

Labod : We need informed users. We need a say for the people affected by AI decisions. At the very least, we need much more understanding and awareness of the fact that different players have very different target functions in their AI models. Consequently, it's important to use a variety of useful tools when forming opinions or using AI, and always maintain a critical eye on the tools used, their limitations, and the companies behind them.

Kasy : AI is much more than language models. In many areas, AI is about automated decision-making systems, about optimizing measurable goals. Precisely which goals are optimized is decided by those who control the (respective) AI. The AI, in turn, is controlled by those who control the inputs—that is, data, computer chips, and expertise. We must place control over the goals in the hands of those affected by decisions—and create democratic structures for this. The informed user is a prerequisite for this, but it cannot be left to the individual: We need a broad societal debate on this topic.

About the interviewees:

Maximilian Kasy is a professor of economics at the University of Oxford. His current research interests include the development of foundations for statistics and artificial intelligence in a social context.

Marvin Labod is Head of Quantitative Analysis at Lupus Alpha and, as a portfolio manager in the Derivative Solutions division, is responsible for capital preservation concepts, overlay mandates and derivative volatility strategies.

This interview was kindly provided by Lupus Alpha and is taken from the current issue of leitwolf magazine: (leitwolf-magazin.de)