Pusan National University Research Shifts Financial AI Focus From Prediction to Decision

Researchers at Pusan National University have proposed a fundamental shift in how artificial intelligence is applied to investment decisions. In two studies presented at the 43rd International Conference on Machine Learning (ICML), a team led by Professor Yoontae Hwang argued that financial AI must be optimized for decision quality rather than simple prediction accuracy to be effective in real-world markets.

To implement this shift, the researchers introduced the Signature-Informed Transformer (SIT), a new model designed for asset allocation. Unlike traditional models that attempt to forecast specific end prices, the SIT model focuses on how market prices evolve and the complex ways different assets influence one another. When tested across major equity markets in the U.S. and China, the SIT model demonstrated more robust wealth accumulation and stronger risk-adjusted performance than traditional forecasting-based approaches.

The Prediction Gap

This research addresses a critical gap in traditional financial AI, which typically prioritizes the accuracy of market predictions. The Pusan National University team notes that high prediction accuracy does not automatically translate into better investment outcomes. They liken this discrepancy to a weather application that correctly predicts the temperature but fails to warn the user to carry an umbrella—the data is accurate, but the resulting action is not optimized for the user's needs.

Professor Yoontae Hwang emphasized this distinction, stating, "Our findings indicate that future financial AI systems may need to shift their focus from maximizing prediction accuracy to optimizing decision quality."

Correcting LLM Biases

Beyond the SIT model, the researchers tackled the reliability of Large Language Models (LLMs) in finance. By reviewing 164 financial LLM studies published between 2023 and 2025, the team identified recurring systemic biases that may inflate performance claims. These include survivor bias—the exclusion of failed companies from datasets—and look-ahead bias, where future information is unintentionally used to inform past decisions.

Professor Hwang noted that the team observed several other issues, including unrealistic evaluation settings and the omission of practical constraints such as transaction costs. To combat these errors, the researchers proposed a Structural Validity Framework, designed to identify these biases and ensure that financial AI evaluations are grounded in reality.

Industry Implications

This shift from "prediction" to "decision" could lead to the development of more reliable automated asset allocation tools. By prioritizing the quality of the decision over the precision of the forecast, the research provides a pathway toward AI that is more resilient to market volatility.

Furthermore, the introduction of the Structural Validity Framework provides a necessary corrective for the industry. By exposing the flaws in current LLM research, the framework promotes greater transparency and trustworthiness in automated financial advice, ensuring that performance claims are not merely the result of flawed evaluation settings.

Future Outlook

As the industry integrates these findings, the focus will likely move toward creating more transparent validation processes for financial AI. The Pusan National University research sets a new benchmark for how models are tested, suggesting that the next generation of financial tools will be judged not by how well they guess the future, but by how effectively they manage risk and wealth.