TimeCatcher

TimeCatcher: Official implementation of "TimeCatcher: A Variational Framework for Volatility-Aware Forecasting of Non-Stationary Time Series". This efficient MLP-based framework integrates a VAE encoder and a volatility-aware module to significantly improve long-term forecasting, especially in high-volatility scenarios.

Getting Started

Follow these steps to reproduce our experiments:

1. Create a new Conda environment with Python 3.10: conda create -n TimeCatcher python=3.10

2. Install dependencies pip install -r requirements.txt

3. Prepare datasets

• Download pre-processed datasets from:
  • Google Drive
  • Baidu Drive
  • Kaggle Datasets
• Create a dataset directory: mkdir ./dataset
• Place all downloaded files in ./dataset

4. Run experiments
   Execute the provided training script: bash ./scripts/TimeCatcher_ecl.sh Additional task scripts available in ./scripts/

Note: All experiments were conducted on a single NVIDIA RTX 4090 GPU