[ICML'25] Sample-level Adaptive Model Fusion for Time Series Forecasting

TimeFuse is an ensemble time series forecasting framework that adaptively fuses multiple models at the sample level to exploit their complementary strengths. We extract the meta-features of the input time series and use them to learn a fusion model that dynamically combines the predictions of a diverse set of base models. The framework is model-aganostic, allowing for the integration of various forecasting models, including deep learning and traditional statistical methods.

Our implementation is based on the thuml/Time-Series-Library to take advantage of its extensive collection of forecasting models and searched training configs.

🧳 Usage

1. Install dependencies:

   pip install -r requirements.txt

2. Prepare the datasets:

   Simply download https://drive.google.com/file/d/1H40tc6cO1lWVcU6WBJ0sg_jwOGh_XBDE/view?usp=sharing and put all files in the root directory of the project. The dataset should be in the dataset/ directory. We also provide some pretrained model checkpoints in the checkpoints/ directory, and extracted meta-training data in the meta_data/ directory. You can use them to quickly run the example notebook run_timefuse.ipynb with the default config without training the base models from scratch.

3. Run an TimeFuse experiment:

   We provide a step-by-step guide to run TimeFuse experiment in the run_timefuse.ipynb notebook, please check the notebook for more details.

   The general steps are:

   1. Load experiment configs: Load the default config from run_config.json or create your own config file.
   2. Train base models: Train the base models with loaded configs, if the checkpoints does not exist.
   3. Extract meta-features: Extract meta-training data and store in the meta_data/ folder. Given a training time series pair (X_in, X_out) and k base models, we extract the following meta-training data:
      • x_meta: the meta-features of the input time series X_in
      • y_model_preds: the predictions of the k base models (i.e., their preditions of X_out)
      • y_true: the ground truth X_out
   4. Train the fusion model: Train the fusion model with the meta-training data.
   5. Evaluate the fusion model: Compare TimeFuse and all base models on the test set.

🗂️ Project Structure

• args.py, load_configs.py, meta_feature.py, timefuse.py: Core scripts for configuration, meta-feature extraction, and fusion logic.
• run.py: Main entry point for running experiments.
• run_base_model_train.ipynb: Example notebook for base model training.
• run_config.json: Example configuration file.
• data_provider/: Data loading and preprocessing utilities.
• dataset/: Datasets for experiments.
• exp/: Experiment modules.
• layers/: Model layers and components.
• meta_data/: Meta-data for fusion.
• models/: Model definitions.
• scripts/: Helper scripts.
• utils/: Utility functions.
• checkpoints/: Saved model checkpoints.