(AAAI'26) (In Submission) T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusion

If you find our work useful in your research. Please consider giving a star ⭐ and citation 📚:

@article{chowdhury2025t3time,
 title={T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusion},
 author={Chowdhury, Abdul Monaf and Akter, Rabeya and Arib, Safaeid Hossain},
 journal={arXiv preprint arXiv:2508.04251},
 year={2025}
}

Abstract

Multivariate time series forecasting (MTSF) seeks to model temporal dynamics among variables to predict future trends. Transformer-based models and large language models (LLMs) have shown promise due to their ability to capture long-range dependencies and patterns. However, current methods often rely on rigid inductive biases, ignore inter variable interactions, or apply static fusion strategies that limit adaptability across forecast horizons. These limitations create bottlenecks in capturing nuanced, horizon-specific relationships in time-series data. To solve this problem, we propose T3Time, a novel trimodal framework consisting of time, spectral, and prompt branches, where the dedicated frequency encoding branch captures the periodic structures along with a gating mechanism that learns prioritization between temporal and spectral features based on the prediction horizon. We also proposed a mechanism which adaptively aggregates multiple cross-modal alignment heads by dynamically weighting the importance of each head based on the features. Extensive experiments on benchmark datasets demonstrate that our model consistently outperforms state-of-the-art baselines, achieving an average reduction of 3.37% in MSE and 2.08% in MAE. Furthermore, it shows strong generalization in few-shot learning settings: with 5% training data, we see a reduction in MSE and MAE by 4.13% and 1.91%, respectively; and with 10% data, by 3.70% and 1.98% on average. Code is available at: https://github.com/monaf-chowdhury/T3Time

Dependencies

• Python 3.11
• PyTorch 2.1.2
• cuda 12.1
• torchvision 0.8.0

> conda env create -f env_windows.yaml
> conda install anaconda::h5py=3.12.1
# > conda install conda-forge::transformers=4.51.3
> pip install transformers==4.51.3

Datasets

Datasets can be obtained from TimesNet and TFB.

Usages

• Last token embedding storage

bash Store_{data_name}.sh

• Train and inference

bash {data_name}.sh

Contact Us

For inquiries or further assistance, contact us at monafabdul15@gmail.com or open an issue on this repository.