University Honors

This study evaluates the effectiveness of large language models known as LLMs in numeric financial time-series forecasting by comparing three distinct forecasting workflow architectures, including prompt-driven, autoregressive generative workflows, fixed-horizon rolling-origin forecasting, and a structured, block-based forecasting pipeline with periodic resetting. The autoregressive generative workflows represent the LLM-based aspect of the overall experiment. While LLMs have shown strong performance in sequence modeling tasks, how well they perform in the area of financial forecasting remains unclear. All workflows were evaluated under a shared out-of-sample period from December 2023 through December 2024, with identical accuracy, bias, and directional metrics, using S&P 500 daily closing prices. The autoregressive workflows are deployed through the use of prompt-driven LLM generation, where each predicted value is reused as input for the next prediction without periodic adjustment. The fixed-horizon rolling-origin workflow uses classical statistical models that are re-estimated at each forecast origin using a progressively increasing historical data window. Lastly, the block-based workflow creates forecasts over fixed horizons and then replaces the predicted values with the realized observations before refitting. The results of the experiment reveal that structured workflows, including rolling-origin and block-based design, achieve strong accuracy and relatively stable bias. The block-based workflow performs notably the best overall due to its near-zero bias and high stability because of its emphasis on error containment and periodic re-anchoring. The autoregressive workflows, in contrast, underperform across all accuracy and bias metrics, showing strong tendencies to accumulate error, drift, and instability over long horizons. The findings from this experiment imply that forecasting performance is driven more by workflow and structural constraints as opposed to model complexity. Because of the nature of LLMs, they are able to capture short-term patterns or directional signals, however, they arenít suited for sustained numeric forecasting without explicit stabilizing mechanisms and error control. Overall, the study reinforces the overall relevance of classical forecasting principles and emphasizes the central role that forecasting workflow architectures play in achieving reliable financial forecasts.