Reverso

Efficient time-series foundation models for zero-shot forecasting.

Paper • GitHub • Hugging Face

By combining long convolutions with linear RNN layers, Reverso matches the performance of transformer-based models that are over 100x larger.

Key Results

Evaluated on Gift-Eval, a comprehensive time-series forecasting benchmark spanning 97 tasks within 23 datasets across 7 domains.

For reference, Xihe-Max (1.5B params) achieves 0.711 and TimesFM-2.5 (200M params) achieves 0.705 on the same benchmark.

Installation

pip install -r requirements.txt
pip install --no-build-isolation git+https://github.com/HazyResearch/flash-fft-conv.git#subdirectory=csrc/flashfftconv
pip install --no-build-isolation git+https://github.com/HazyResearch/flash-fft-conv.git
pip install -e .

Requirements

• Python >= 3.11
• PyTorch 2.6.0
• CUDA-compatible GPU
• FlashFFTConv
• flash-linear-attention

Model Architecture

Reverso uses a hybrid architecture that interleaves:

1. Long convolution layers (FlashFFTConv) with gated short convolutions
2. DeltaNet layers for modeling sequential dependencies
3. MLP layers for channel mixing
4. Attention-based decoder head for producing the final forecast

Input sequences are normalized to [0, 1] and processed point-wise (no patching). The model predicts 48 time steps at a time and rolls out autoregressively for longer horizons.

The modeling code is in reverso/.

Quick Start

import torch
from reverso import load_model, forecast

model, cfg = load_model(
    "checkpoints/reverso_small/checkpoint.pth",
    "checkpoints/reverso_small/args.json",
    device="cuda",
)

context = torch.full((1, 2048, 1), 5.0, device="cuda")  # (batch, seq_len, 1)
predictions = forecast(
    model, context,
    prediction_length=96,
    seq_len=cfg.seq_len,
    output_token_len=cfg.output_token_len,
)
print(predictions.shape)  # (1, 96, 1)

Examples

Install the example dependencies first:

pip install -r example/requirements.txt

Forecast Demo

Run Reverso on synthetic signals (constant, linear, sine, sawtooth, square):

python example/forecast_demo.py --signal all

Use --signal sine to run a single signal, or --list to see all options.

Gift-Eval Benchmark

To reproduce the benchmark results, first follow the Gift-Eval setup instructions to install the package and download the data. Then run:

python example/eval_gift.py \
    --checkpoint checkpoints/reverso_small/checkpoint.pth \
    --output-dir results/ \
    --force-flip-invariance

Note: Dependencies within Gift-Eval may conflict with those in Reverso. If you encounter issues, try upgrading huggingface_hub:

pip install --upgrade huggingface_hub

Note: While running this benchmark, it is recommended to use flip invariance, but this requires two forward passes of the model. The inference speed is also not fully optimized and could be further sped up.

Available Checkpoints

Citation

@misc{fu2026reversoefficienttimeseries,
      title={Reverso: Efficient Time Series Foundation Models for Zero-shot Forecasting},
      author={Xinghong Fu and Yanhong Li and Georgios Papaioannou and Yoon Kim},
      year={2026},
      eprint={2602.17634},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.17634},
}

License

This project is licensed under the MIT License.