Artificial intelligence (AI) is reshaping modern finance, supporting applications such as stock market forecasting, portfolio management, and investment advice. However, researchers from Pusan National University and their international collaborators argue that accurate market predictions do not always translate into better investment decisions. Instead, they suggest that financial AI should be assessed by its ability to improve real-world decision-making rather than prediction accuracy alone.

To address this challenge, Professor Yoontae Hwang of Pusan National University and Professor Stefan Zohren of the University of Oxford developed the Signature-Informed Transformer (SIT), a decision-focused AI framework. Rather than concentrating solely on predicting future prices, the model learns from how markets evolve and how different assets influence one another, enabling it to optimise investment decisions while accounting for risk. Published in the Proceedings of the 43rd International Conference on Machine Learning on 30 April 2026, the study lists Professor Hwang as first author.

The researchers evaluated the SIT framework using equity market data from the United States and China. Compared with conventional forecasting-based methods, the decision-focused approach delivered stronger risk-adjusted returns and more consistent wealth accumulation. According to Professor Hwang, the findings suggest that future financial AI systems should prioritise decision quality over prediction accuracy to achieve better investment outcomes.

In a second study, the research team investigated whether the reported success of financial AI can be reliably trusted. Analysing 164 studies on large language models (LLMs) in finance published between 2023 and 2025, they identified several recurring sources of bias that could overstate model performance. These included the unintended use of future information, survivor bias resulting from the exclusion of failed companies, unrealistic evaluation settings, and the omission of practical considerations such as transaction costs. Published in the Proceedings of the 43rd International Conference on Machine Learning on 1 May 2026, the study lists Professor Hwang as co-first author.

To improve research quality, the team introduced a Structural Validity Framework, a practical checklist designed to help researchers evaluate whether financial AI systems are tested under realistic conditions and whether their reported performance is likely to generalise beyond laboratory settings. The framework encourages more transparent and rigorous evaluation practices that better reflect real-world financial markets.

Together, the two studies highlight a common principle: AI should be designed to support meaningful financial decisions and evaluated using realistic benchmarks. Looking ahead, the researchers envision AI-powered “flight simulators” for financial markets, enabling institutions and regulators to test investment strategies, financial products, and market shocks in virtual environments before they affect real investors. Such advances could ultimately promote more transparent financial advice and more trustworthy AI systems.

More information: Yoontae Hwang et al, Signature-Informed Transformer for Asset Allocation, Proceedings of the 43rd International Conference on Machine Learning. DOI: 10.48550/arXiv.2510.03129

Journal information: Proceedings of the 43rd International Conference on Machine Learning Provided by Pusan National University