Can LLM Embeddings Improve Time Series Forecasting? A Practical Feature Engineering Approach

The integration of Large Language Models into Time Series Forecasting has emerged as a significant area of research. However, naive approaches like embedding concatenation yield negligible gains (50.0% to 50.47% accuracy). This report evaluates 2024-2026 advances, identifying three effective alignment paradigms — reprogramming, direct vectorization, and multi-layer fusion — that produce meaningful improvements, particularly in zero-shot, data-scarce, and text-rich forecasting scenarios.

Comprehensive literature review of 2024-2026 academic publications covering LLM-based time series forecasting architectures, drawing on 21 primary sources from ArXiv, IEEE, ACM, and domain-specific journals. Includes benchmark comparisons across standard datasets (ETTh1, Weather, Traffic).