PRISM

Hierarchical multiscale forecasting for long, multivariate time series.

PRISM builds a learnable multiresolution tree that decomposes signals into band-limited fragments, routes them with soft gates, mixes sibling bands, and fuses forecasts at the root. It is end-to-end, interpretable, and strong on long horizons with lower memory.

Highlights

• Multiscale decomposition with learnable filters
• Temperature-controlled routing for band assignment
• Band-mixing block to capture cross-scale dependencies
• Squeeze-and-gate aggregation at the root
• Fast inference on long horizons

Quick Start

1. Create your environment and install deps.
2. Put datasets under ./temp/datasets as in the scripts.
3. Run a quick script.

# From repo root
cd PRISM
bash scripts/ETTh1_quick.sh

Installation

This setup is tested for Python 3.10 and recent NVIDIA GPUs (e.g., 3090/4090).

git clone git@github.com:nerdslab/PRISM.git
cd PRISM

python3.10 -m venv tsdiff310
source tsdiff310/bin/activate  # Linux/Mac
# or: tsdiff310\Scripts\activate  # Windows

pip install gluonts==0.12.8
pip install aeon torch einops omegaconf lightning hydra-core matplotlib seaborn
pip install wandb==0.18.7 torchvision==0.20.1
pip install dtaidistance PyWavelets opt_einsum
pip install -U rich
pip install pytorch-lightning==1.9.5
pip install imageio patool sktime pykeops orjson torchtyping

Data Layout

Each script expects a dataset folder and a CSV file under ./temp/datasets. Match the script paths exactly.

Examples:

• ./temp/datasets/ETT-small/ETTh1.csv
• ./temp/datasets/electricity/electricity.csv
• ./temp/datasets/traffic/traffic.csv

Tip: open a script in scripts/ to see the exact root_path and data_path.

Run Training

Pick one of the quick scripts:

bash scripts/ETTh1_quick.sh
bash scripts/ETTh2_quick.sh
bash scripts/ETTm1_quick.sh
bash scripts/ETTm2_quick.sh
bash scripts/electricity_quick.sh
bash scripts/exchange_quick.sh
bash scripts/traffic_quick.sh
bash scripts/weather_quick.sh

Run with the CLI

Use run_prism.py directly if you want full control:

python run_prism.py \
  --data ETTh1 \
  --root_path ./temp/datasets/ETT-small \
  --data_path ETTh1.csv \
  --target OT \
  --features M \
  --seq_len 336 \
  --pred_len 96 \
  --enc_hidden 336 \
  --dec_hidden 336 \
  --dropout 0.25 \
  --tree_depth 2 \
  --num_components 5 \
  --batch_size 512 \
  --train_epochs 120 \
  --lr 1e-4 \
  --patience 15 \
  --loss mae \
  --lradj 1 \
  --use_amp \
  --decomp_kind haar \
  --use_multi_gpu \
  --devices 0,1,2,3 \
  --seed 14 \
  --overlap 8

Key Hyperparameters

• --tree_depth: depth of the multiresolution tree (typ. 2-3)
• --num_components: number of band fragments K (typ. 5-8)
• --overlap: overlap between fragments (typ. 8-24)

Outputs

Logs are saved here:

./temp/logs/exp_prism/{dataset}_cont{context_length}_{features}/

Each log filename includes hyperparameters for easy tracking.

Benchmarks

PRISM is evaluated on:

• ETT (ETTh1, ETTh2, ETTm1, ETTm2)
• Electricity
• Exchange Rate
• Traffic
• Weather

Citation

If you use PRISM, please cite:

@article{prism2025,
  title={PRISM: A Hierarchical Multiscale Approach for Time Series Forecasting},
  author={Chen, Zihao and Andre, Alexandre and Ma, Wenrui and Knight, Ian and Shuvaev, Sergey and Dyer, Eva},
  journal={...},
  year={2025}
}

Acknowledgements

• Time Series Library (TSLib)
• GluonTS
• PyWavelets
• D-PAD training pipeline inspiration