Mfon Akpan
The predictive abilities of generative artificial intelligence (AI) are changing the landscape for analytic workflows across sectors. Nevertheless, its capacity and implications for use cases in high-stakes, non-stationary environments — like financial markets — have been empirically under-researched (Tan et al., 2023; Lin & Marques, 2024). This research paper examines generative AI’s zero-shot forecasting capabilities using two large language model (LLM) architectures, OpenAI’s GPT-4o, and Anthropic’s Claude 3.5 Sonnet, as they forecast stock prices. Specifically, the paper evaluates the LLMs’ predictive powers in terms of actual closing prices for a portfolio of in-use equities across sectors on February 3, 2025. A rigorous quantitative approach is used throughout the analyses. In the results section, standardized metrics including mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), correlation scores, and R-squared are calculated to assess predictive accuracy and directional bias. Results show that Claude 3.5 Sonnet outperformed GPT-4o on all accuracy metrics, and also showed better accuracy in forecasting actual movement in the stock market, confirming the study hypothesis and demonstrating performance can vary significantly between LLM architectures (Xu et al., 2024). Further analysis of sector-based performance can be undertaken. The study concludes that while the LLM Claude 3.5 Sonnet does yield encouraging strategic implications for use in investment analytics, there still exist significant challenges in relation to interpretability, model calibration, and model sensitivity to rapidly changing market dynamics.