Empirical Evaluation of Large Language Models for Asset‑Return Prediction

In an era of exploding financial‐market information and rapid algorithmic iteration, traditional asset‑return forecasting models struggle to exploit unstructured text. Using cross‑asset data—equities, Treasuries and commodity futures—from 2004 to 2024, we build an integrated prediction framework that fuses semantic factors extracted by Large Language Models (LLMs) with price‑volume and macro‑numerical factors. We benchmark it against Logit, Random Forest, LightGBM and bidirectional LSTM. A comprehensive evaluation with weighted F₁, ROC‑AUC, Information Ratio and Sharpe Ratio shows that (i) LLM‑based semantic factors significantly improve directional accuracy (F₁ + 20.5 %, ROC‑AUC + 11.9 %); (ii) after a 3 bp transaction cost, the LLM‑driven long–short portfolio achieves annualised information and Sharpe ratios of 0.96 and 1.17, markedly outperforming all baselines; (iii) robustness checks confirm this edge across high‑volatility regimes, asset classes and text‑lag scenarios; and (iv) the combination of SHAP and attention visualisation traces keyword‑level contributions, enhancing interpretability. Our results provide reproducible, quantifiable evidence for large‑scale LLM deployment in quantitative investing and point to future work on model compression, slippage estimation and multimodal extension.