README.md

Introduction

Official implementation of BALM-TSF (CIKM 2025).

BALM-TSF is a lightweight framework that alleviates modality imbalance between text and time series via statistical + learnable prompts and a balanced alignment module (scaling + contrastive learning).

Modality Imbalance Issue:

Model Structure:

Long-term Forecasting Results

Efficiency Analysis

Requirements

• Python 3.11 (recommended via Miniconda)

• Install dependencies:

  pip install -r requirements.txt
  

Datasets

Download the pre-processed datasets from Google Drive, and place the extracted contents under ./dataset. This is a public dataset sharelink from Time-LLM.

Quick Start

1. Download and unzip datasets into ./dataset.
2. Tune the model using the provided scripts.
3. The training and test results will be visualized and recorded via Wandb.

Long-term Forecasting

bash ./scripts_long_term/BALM_ETTh1_GPT2.sh

Few-shot Forecasting

bash ./scripts_few_shot/BALM_ETTh1_GPT2.sh

Acknowledgements

We appreciate Time-Series-Library and Time-LLM for code references and datasets.

Citation

If you find this repository useful, please consider citing our work:

@article{zhou2025balmtsf,
  title={BALM-TSF: Balanced Multimodal Alignment for LLM-Based Time Series Forecasting},
  author={Zhou, Shiqiao and Schöner, Holger and Lyu, Huanbo and Fouché, Edouard and Wang, Shuo},
  journal={arXiv preprint arXiv:2509.00622},
  year={2025}
}