Reproduce “LoFT-LLM: Low-Frequency Time-Series Forecasting with Large Language Models” on OpenResearch

About this paper

Jiacheng You, Jingcheng Yang, Yuhang Xie, Zhongxuan Wu, Xiucheng Li, Feng Li, Pengjie Wang, Jian Xu and 2 more

LoFT-LLM, developed by researchers at Harbin Institute of Technology (Shenzhen) and collaborators, presents a three-phase pipeline for low-frequency time-series forecasting with large language models. The method effectively addresses data scarcity and high-frequency noise by integrating frequency-aware learning with semantic calibration, achieving average MAE reductions of 26.53% on financial data and 15.42% on energy data compared to baseline models.

Key insights

• Combining frequency-aware signal processing with LLM-based semantic reasoning creates a synergistic approach that effectively addresses data scarcity and high-frequency noise in time-series forecasting.
• Explicitly focusing on low-frequency components through specialized modules and frequency-domain loss functions provides a more stable and robust foundation for predictions, especially when data is limited.
• Leveraging the in-context learning capabilities of LLMs via structured prompts enables the effective integration of heterogeneous auxiliary context and domain knowledge, enhancing forecasting accuracy and offering a pathway to improved interpretability.

Summary by alphaXiv.