FinBa: Surprise-Aware Stock Prediction with LLMs and World Models

Supervisor: Turgud Valiyev

Author: Georg Unterlechner

Abstract

Stock price prediction remains challenging due to market volatility, com- plex temporal dependencies, and the unpredictable effects of financial news. Traditional machine learning (ML), deep learning (DL), and large language models (LLMs) have been widely explored for financial fore- casting. However, effectively capturing how unexpected events influ- ence future market behavior remains an important challenge, particu- larly when distinguishing events that significantly affect the stock mar- ket from routine news. To address this, we propose FinBa, a surprise- aware financial forecasting framework that combines state space models (SSMs), LLM-based event understanding, and predictive world model- ing. FinBa first identifies financial events relevant to future stock move- ments. Then, it applies a surprise-aware world model to learn how these events may change future market representations by comparing expected and event-conditioned market states to estimate future stock movements. We evaluate FinBa against traditional ML, DL, SSM, and LLM-based methods. Furthermore, we investigate the contributions of event selection, world modeling, and surprise-aware learning. The study aims to determine whether integrating event understanding with predic- tive market-state modeling improves forecasting accuracy and robustness.