Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing few-shot time series forecasting with LLM-guided diffusion

Haonan Shi, Dehua Shuai, Liming Wang, Xiyang Liu, Long Tian

TL;DR

This work tackles few-shot multivariate time series forecasting by introducing LTSM-DIFF, a framework that uses a fine-tuned LLM as a temporal memory to extract rich representations from history and condition a diffusion-based forecaster. The encoder employs a GPT-2 backbone with LoRA to produce $\boldsymbol{x}_0 \in \mathbb{R}^{T \times m}$, while a UViT diffusion module models the joint distribution of future trajectories with forward noising on both the condition and the target and a denoising process during inference. The training objective combines autoregressive learning with diffusion loss: $\mathcal{L}=\mathcal{L}_{LLM}+\lambda\mathcal{L}_{diff}$, enabling robust few-shot transfer demonstrated on seven real-world datasets, including strong performance when fine-tuning with as little as 1% of target data. Overall, the approach establishes a new paradigm for cross-domain knowledge transfer from language models to time-series forecasting, delivering both high accuracy and uncertainty quantification in data-scarce settings.

Abstract

Time series forecasting in specialized domains is often constrained by limited data availability, where conventional models typically require large-scale datasets to effectively capture underlying temporal dynamics. To tackle this few-shot challenge, we propose LTSM-DIFF (Large-scale Temporal Sequential Memory with Diffusion), a novel learning framework that integrates the expressive power of large language models with the generative capability of diffusion models. Specifically, the LTSM module is fine-tuned and employed as a temporal memory mechanism, extracting rich sequential representations even under data-scarce conditions. These representations are then utilized as conditional guidance for a joint probability diffusion process, enabling refined modeling of complex temporal patterns. This design allows knowledge transfer from the language domain to time series tasks, substantially enhancing both generalization and robustness. Extensive experiments across diverse benchmarks demonstrate that LTSM-DIFF consistently achieves state-of-the-art performance in data-rich scenarios, while also delivering significant improvements in few-shot forecasting. Our work establishes a new paradigm for time series analysis under data scarcity.

Enhancing few-shot time series forecasting with LLM-guided diffusion

TL;DR

This work tackles few-shot multivariate time series forecasting by introducing LTSM-DIFF, a framework that uses a fine-tuned LLM as a temporal memory to extract rich representations from history and condition a diffusion-based forecaster. The encoder employs a GPT-2 backbone with LoRA to produce , while a UViT diffusion module models the joint distribution of future trajectories with forward noising on both the condition and the target and a denoising process during inference. The training objective combines autoregressive learning with diffusion loss: , enabling robust few-shot transfer demonstrated on seven real-world datasets, including strong performance when fine-tuning with as little as 1% of target data. Overall, the approach establishes a new paradigm for cross-domain knowledge transfer from language models to time-series forecasting, delivering both high accuracy and uncertainty quantification in data-scarce settings.

Abstract

Time series forecasting in specialized domains is often constrained by limited data availability, where conventional models typically require large-scale datasets to effectively capture underlying temporal dynamics. To tackle this few-shot challenge, we propose LTSM-DIFF (Large-scale Temporal Sequential Memory with Diffusion), a novel learning framework that integrates the expressive power of large language models with the generative capability of diffusion models. Specifically, the LTSM module is fine-tuned and employed as a temporal memory mechanism, extracting rich sequential representations even under data-scarce conditions. These representations are then utilized as conditional guidance for a joint probability diffusion process, enabling refined modeling of complex temporal patterns. This design allows knowledge transfer from the language domain to time series tasks, substantially enhancing both generalization and robustness. Extensive experiments across diverse benchmarks demonstrate that LTSM-DIFF consistently achieves state-of-the-art performance in data-rich scenarios, while also delivering significant improvements in few-shot forecasting. Our work establishes a new paradigm for time series analysis under data scarcity.
Paper Structure (16 sections, 9 equations, 2 figures, 4 tables)

This paper contains 16 sections, 9 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time Series Forecasting with LLMs
  4. Time Series Forecasting with Diffusion Models
  5. Method
  6. Problem Definition
  7. Temporal Representation with a Foundation Model
  8. Conditional Forecasting with a Diffusion Model
  9. Joint Training Objective
  10. Experiment
  11. Experimental Setup
  12. Full-data Comparison
  13. Few-shot Transfer
  14. Ablation Study
  15. Visualization of the Denoising Process
  16. ...and 1 more sections

Figures (2)

  • Figure 1: Overall framework of our proposed method. The encoder based on a foundation model extracts temporal representations, while the UViT-based diffusion module performs conditional forecasting.
  • Figure 2: Visualization of the denoising process on an ETTm2 sample: black lines represent historical input and ground truth, the red line is the initial LTSM prediction, and the green shaded band shows the diffusion uncertainty range.