From Traditional Methods to Large Language Models: A Comparative Study of Three Stock Return Prediction Models

Abstract

Stock return forecasting has shifted direction several times over the past thirty years, starting with linear factor models, moving through traditional machine learning, and lately turning to large language models that extract signals from text. Each appro ach has its own empirical literature, yet the three have rarely been compared within a single analytical frame. This paper takes on that comparison. Using the Fama -French model to represent the linear factor tradition, LASSO with random forests and neural nets for machine learning, and LLM-based text methods for the newest wave, we examine them across their theoretical foundations, the data they consume, how they model returns, and how they perform empirically. The findings are noteworthy. LLMs have, for th e first time, brought unstructured text into the forecasting pipeline, and their predictive edge survives after standard factors are accounted for. But considerable unknowns remain. Long horizon robustness is unproven, hallucinations pose real risks, and cross-market portability is largely untested. The paper ends by sketching an integrated asset pricing framework, a conceptual starting point for braiding the three traditions together.