CATF: A Manager-Worker Framework for Context-Aware Multivariate Time-Series Forecasting

CATF is a flexible framework built upon a manager–worker architecture for multivariate time-series forecasting. It enables specialized learning across different context patterns and improves predictive performance over standard baselines. This repository provides implementations of standard baseline models for multivariate time-series forecasting, as well as their enhanced versions using our proposed CATF framework. We include all our core implementation in cap folder, which stands for context-aware prediction.

Figure 1: Manager Worker Architecture.

Dataset Setup

Download the full dataset from the link. And put the long_term_forecast folder into dataset folder.

It includes 7 different datasets:

• Electricity Transformer Temperature (ETT) benchmark dataset (in ETT-small folder), which has 4 individual datasets: ETTh1.csv, ETTh2.csv, ETTm1.csv, ETTm2.csv.
• Exchange rate dataset (in exchange_rate folder): exchange_rate.csv.
• Weather dataset (in weather folder): weather.csv.
• Illness dataset (in illness folder): national_illness.csv.

Environment Setup

Create and activate a virtual environment:

python3 -m venv venv && source venv/bin/activate

Install all required dependencies:

pip install -r requirement.txt

Training

Run a Baseline Model:

To train a standard baseline model on a specific dataset:

python experiments/train_et_model.py —config cap/configs/baseline/<dataset>/config_<dataset_lower_case>_<model>.yaml

Replace <dataset> and <model> with the abbreviations of the values. <dataset> includes: ETTh1, ETTh2, ETTm1, ETTm2, exchange, illness, weather; <model> includes: Informer (<info>), Autoformer (<auto>), EFDFormer (<fed>), TimesNet (<times>), i-Transformer (<it>)

Run CATF-<baseline>:

To train CATF-enhanced variants of the baselines:

python experiments/train_et_catf.py --config cap/configs/cap/<dataset>/et_cap_<model>.yaml

Run multiple times (with GPU selection and output saving):

python experiments/run_multiple_times.py --command "export CUDA_VISIBLE_DEVICES=<GPU_Number> && python experiments/train_et_catf.py --config cap/configs/cap/<dataset>/et_cap_<model>.yaml" --times <number of experiments> --save-output

Example:

python experiments/run_multiple_times.py --command "export CUDA_VISIBLE_DEVICES=0 && python experiments/train_et_catf.py --config cap/configs/cap/ETTh1/et_cap_times.yaml" --times 10 --save-output

Results

CATF vs Baseline

We compare CATF-enhanced models (CATF-Baselines) with their original counterparts across multiple benchmark datasets.

Figure 2: CATF vs. baseline models across multiple datasets.

CATF-TimesNet vs. Recent SOTA Models

Figure 3: CATF-TimesNet vs. recent state-of-the-art models.

Repository Structure

CATF/
├── cap/                        # Core package containing all modules (CAP stands for Context-Aware Prediction)
│   ├── configs/                # YAML configuration files (for baselines and CATF models)
│   ├── data/                   # Data loading and preprocessing utilities
│   ├── models/                 # Model architectures (baselines and catf)
│   │   └── catf.py             # CATF-specific model definitions
│   ├── training/               # Training logic and trainer classes
│   │   └── catf_trainer.py     # CATF training loop
│
├── run_multiple_times.py       # Script to run training multiple times with logging
├── train_et_catf.py            # Main training script for CATF
├── train_et_model.py           # Script for training baseline models
├── requirement.txt             # Python package dependencies
├── README.md                   # Project documentation (you are here)