Research from the Department of Finance examines the limits of LLMs to improve accuracy for financial markets and ensure AI helps rather than hinders as it shapes the economy.
The problem
Artificial intelligence is reshaping financial markets, yet its limitations – when it comes to its cognitive abilities and decision-making processes – are underexplored. Like humans, AI systems have biases that affect how financial markets work, so we need to understand what those are and how they play out to safeguard the economy. We need to know how AI "thinks" and makes decisions to get the full picture of how modern financial markets operate.
The research
Contributing to this effort, Dr Antoine Didisheim and Dr Attila Balogh document ‘information overload’ in Large Language Models (LLMs) applied to financial analysis. Performing two tests (in corporate finance and asset pricing) they show that beyond a certain threshold, excess context – or ‘more’ information – actually degrades accuracy of the LLM. This ‘information overload’ is due to the absence of cognitive biases, since the ‘rational’ choice is to give as much information as possible. The findings show a fundamental limitation of AI-driven finance: more data is not always better. If we want accurate outputs, how much information is given to LLMs matters.
The impact
By understanding this limitation, practitioners can be more strategic approach, ensuring AI is helpful rather than creating new problems or distorting the market. The research also asserts the importance of studying AI in finance. As it becomes increasingly integrated into financial markets, and therefore more important in shaping our economy, a comprehensive understanding of how AI works, including potential flaws and biases, is essential.
Department: Finance
Area: AI in finance
Researchers
Sustainable Development Goals
We align our research activity with the United Nations' Sustainable Development Goals (SDGs).