TSGym is a novel automated MTSF (multivariate time series forecasting) solution, performing fine-grained component selection and automated model construction. Based on meta-learning method, TSGym enables the creation of more effective solutions tailored to diverse time series data or forecasting task.
🚩2025.5: The code for all experiments in our paper is open-sourced on GitHub. We plan to release all code scripts and full details of the paper.
✅TSGym decouples existing SOTA methods according to the standard Pipeline of MTSF modeling.
✅TSGym significantly expands the diversity of the modeling pipeline and structures each pipeline step according to distinct Design Dimensions (see the following Table).
✅TSGym automatically constructs DL-based time-series forecasting model by specified Design Choices via meta-learning, achieving significantly better performance compared to current SOTA solutions in both short-/long-term forecasting tasks.
TSGym's ability to deconstruct deep time series models facilitates large-scale, component-level (i.e., design choice) evaluations. For detailed results, please see the full paper.
Compared with other state-of-the-art forecasters, TSGym demonstrates superior capability in both short-term and long-term MTSF tasks.
- einops==0.8.0
- local-attention==1.9.14
- matplotlib==3.7.0
- numpy==1.23.5
- pandas==1.5.3
- patool==1.12
- reformer-pytorch==1.4.4
- scikit-learn==1.2.2
- scipy==1.10.1
- sktime==0.16.1
- sympy==1.11.1
- tqdm==4.64.1