AI equity forecasting benchmark reveals LoRA-adapted TimesFM lacks directional skill

A new research paper challenges the effectiveness of large language models like TimesFM for equity forecasting, particularly when using LoRA adapters. The study introduces a base-rate-honest benchmark to expose how seemingly high directional accuracy can be misleading, often achieved by simple "always-up" rules in rising markets. The findings indicate that pooled LoRA adapters show no directional skill over these naive baselines and can even perform worse than zero-shot TimesFM, with fine-tuning only offering marginal improvements in point-forecast error. AI

IMPACT Challenges the efficacy of current LLMs for financial forecasting, suggesting a need for more robust evaluation methods beyond simple accuracy metrics.

RANK_REASON The cluster contains a research paper detailing a new benchmark and findings related to AI model performance in financial forecasting.

Read on arXiv cs.LG →

• alphaXiv
• Benjamini–Hochberg procedure
• CatalyzeX
• DagsHub
• Diebold-Mariano
• Gotit.pub
• Hugging Face
• LoRA
• McNemar
• Nasdaq-100
• ScienceCast
• S&P 500
• TimesFM

• paper
• model release

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →