Aurora

DecisionIntelligence/Aurora

Aurora is a time series forecasting model from DecisionIntelligence. Use it for the time series forecasting task on the model card, and read the license before you ship it in a product. The card lists the license as mit.

<div align="center" <img alt="Aurora Logo" src="https://cdn-uploads.huggingface.co/production/uploads/66276727368ec2a0b933772c/ytpsIAr98keUvNouoOVmb.png" width="30%" /

Downloads · 30 days

8.3K

41% of all-time downloads

All-time downloads

20.2K

Public

Parameters

211M

844 MB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors844 MB · 100%

At a glance

Task: Time Series Forecasting
License: mit
Model type: aurora
Access: Public
Created: Jan 27, 2026
Updated: Apr 29, 2026
SHA: c495b02c

Model card

From the Hugging Face model README

<div align="center"> <img alt="Aurora Logo" src="https://cdn-uploads.huggingface.co/production/uploads/66276727368ec2a0b933772c/ytpsIAr98keUvNouoOVmb.png" width="30%" /> <h1>Aurora: Towards Universal Generative Multimodal Time Series Forecasting</h1> <p> The official code repo of our ICLR 2026 paper: <a href="https://arxiv.org/pdf/2509.22295"> Aurora: Towards Universal Generative Multimodal Time Series Forecasting </a> </p> <p> <a href="https://arxiv.org/pdf/2509.22295"> <img alt="ICLR 2026" src="https://img.shields.io/badge/ICLR%202026-Aurora-orange" /> </a> <a href="https://www.python.org/"> <img alt="Python" src="https://img.shields.io/badge/Python-3.10%2B-blue" /> </a> <a href="https://pytorch.org/"> <img alt="PyTorch" src="https://img.shields.io/badge/PyTorch-2.4.1-blue" /> </a> <a href="https://github.com/decisionintelligence/Aurora"> <img alt="GitHub Stars" src="https://img.shields.io/github/stars/decisionintelligence/Aurora?logo=github" /> </a> <a href="https://github.com/decisionintelligence/Aurora"> <img alt="GitHub" src="https://img.shields.io/badge/GitHub-Aurora-black?logo=github" /> </a> </p> </div>

Introduction

Aurora is a highly capable multimodal time series foundation model. Based on the Modality-Guided Multi-head Self-Attention and Prototype-Guided Flow Matching, Aurora can effectively utilize the domain-specific knowledge contained in modalities and support generative probabilistic forecasting, thus covering versatile forecasting scenarios.

See Figure 1, to our best knowldege, Aurora is the first pretrained multimodal time series foundation model! Evaluated on 5 well-recognized benchmarks, including TimeMMD, TSFM-Bench, ProbTS, TFB, and EPF, Aurora is demonstrated the state-of-the-art.

Architecture

In this work, we pretrain Aurora in a cross-modality paradigm, which adopts Channel-Independence on time series data, and models corresponding multimodal interaction to inject domain knowledge. Note that the each variable of time series is first normalized through Instance Normalization to mitigate the value discrepancy. See Figure 2, Aurora mainly consists of two phases: 1) in Aurora Encoder, we tokenize and encode each modality into modal features, then fuse them to form multimodal representations; 2) in Aurora Decoder, we utilize a Condition Decoder to obtain the multimodal conditions of future tokens, leverage a Prototype Retreiver to retrieve the future prototypes based on the domain knowledge, and conduct flow matching on them to make generative probabilistic forecasts.

Quickstart

From pypi (recommended)

We have published Aurora on PyPi, you can directly install it with one line of code!

# python >= 3.10
$ pip install aurora-model==0.2.0

Then you can use the Aurora model to make zero-shot probabilistic forecasting!

Unimodal Time Series Forecasting

from aurora import load_model
import os
import torch
# os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
model = load_model()

# prepare input
batch_size, lookback_length = 1, 528 
seqs = torch.randn(batch_size, lookback_length).cuda()

# Note that Aurora can generate multiple probable predictions
forecast_length = 96 
num_samples = 100


# For inference_token_len, you can refer to LightGTS (Periodic Patching).
# We recommend to use the period length as the inference_token_len.
output = model.generate(inputs=seqs, max_output_length=forecast_length, num_samples=num_samples, inference_token_len=48)


# use raw predictions for mean/quantiles/confidence-interval estimation
print(output.shape)

Multimodal Time Series Forecasting

from aurora import load_model
from einops import rearrange
import os
import torch
# os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
model = load_model()
tokenizer = model.tokenizer

# prepare input
batch_size, n_vars, lookback_length, max_text_length = 1, 10, 528, 200
seqs = torch.randn(batch_size, lookback_length, n_vars).cuda()

text = "1983-09-12: The Federal Register provides a uniform system for making available to the public regulations and legal notices issued by federal agencies in the United States."

tokenized_text = tokenizer(text, padding='max_length', truncation=True, max_length=max_text_length, return_tensors="pt")
text_input_ids = tokenized_text['input_ids'].cuda()
text_attention_mask = tokenized_text['attention_mask'].cuda()
text_token_type_ids = tokenized_text.get('token_type_ids', torch.zeros_like(text_input_ids)).cuda()
    
batch_input_ids = text_input_ids.repeat(n_vars, 1)
batch_attention_mask = text_attention_mask.repeat(n_vars, 1)
batch_token_type_ids = text_token_type_ids.repeat(n_vars, 1)
batch_x = rearrange(seqs, "b l c -> (b c) l")

# Note that Aurora can generate multiple probable predictions
forecast_length = 96 
num_samples = 100


# For inference_token_len, you can refer to LightGTS (Periodic Patching).
# We recommend to use the period length as the inference_token_len.
output = model.generate(inputs=batch_x,text_input_ids=batch_input_ids,
                        text_attention_mask=batch_attention_mask,
                        text_token_type_ids=batch_token_type_ids,
                        max_output_length=forecast_length, 
                        num_samples=num_samples, 
                        inference_token_len=48)


# use raw predictions for mean/quantiles/confidence-interval estimation
print(output.shape)

From raw code

We release the original code of Aurora in this repo. You can also download the pretrained checkpoints in our huggingface repo and put them in the folder: aurora/.

If you want to pretrain an Aurora on your own time series corpus, you need to install the following important packages:

$ pip install torch==2.4.0
$ pip install torchvision==0.19.0
$ pip install transformers[torch]

from huggingface_hub import snapshot_download
import os
import torch
# os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'

# --- Configuration ---
repo_id = "DecisionIntelligence/Aurora" 

# Target directory for the download. "." represents the current working directory.
local_dir = "./work_dir" 

# Optional: Set repo_type to "dataset" or "space" if you are not downloading a model.
repository_type = "model" 
# ---------------------

# Ensure the local directory exists before starting the download
if not os.path.exists(local_dir):
    os.makedirs(local_dir)
    print(f"Created directory: {local_dir}")

print(f"Starting download from '{repo_id}' to '{local_dir}'...")

try:
    # snapshot_download handles the download of all files in the repository
    download_path = snapshot_download(
        repo_id=repo_id,
        local_dir=local_dir,
        # Set to False to download actual files instead of symbolic links
        local_dir_use_symlinks=False, 
        repo_type=repository_type,
        # Use your HF access token for private/gated repositories
        token=None 
    )
    print(f"\nSuccess! All files downloaded to: {download_path}")

except Exception as e:
    print(f"\nAn error occurred during download: {e}")

# Then you can easily make zero-shot forecasts using Aurora

from modeling_aurora import AuroraForPrediction

model = AuroraForPrediction.from_pretrained("./",trust_remote_code=True)

# prepare input
batch_size, lookback_length = 1, 528 
seqs = torch.randn(batch_size, lookback_length).cuda()

# Note that Aurora can generate multiple probable predictions
forecast_length = 96 
num_samples = 100


# For inference_token_len, you can refer to LightGTS (Periodic Patching).
# We recommend to use the period length as the inference_token_len.
output = model.generate(inputs=seqs, max_output_length=forecast_length, num_samples=num_samples, inference_token_len=48)


# use raw predictions for mean/quantiles/confidence-interval estimation
print(output.shape)

Experiments

You should refer to our github repo for the complete experimental pipelines. For benchmarking (TSFM-Bench, ProbTS, TimeMMD, TFB, and EPF), you can install additional packages based on the requirement files under folders, and the datasets can be fetched from this link. All experimental results can be reproduced by running the scripts in the benchmark folder：

# TimeMMD
TimeMMD/scripts/run_aurora_timemmd_zero_shot.sh

# EPF
EPF/scripts/run_aurora_short_term_zero_shot.sh

# ProbTS
ProbTS/scripts/run_aurora_probts.sh

# TSFM-Bench
TFB/scripts/run_aurora_tfb.sh

# TFB univariate
TFB/scripts/run_aurora_uni.sh

Performance

Aurora ahieves consistent state-of-the-art performance on these 5 benchmarks:

Citation

If you find this repo useful, please cite our paper.

@inproceedings{wu2026aurora,
  title     = {Aurora: Towards Universal Generative Multimodal Time Series Forecasting},
  author    = {Wu, Xingjian and Jin, Jianxin and Qiu, Wanghui and Chen, Peng and Shu, Yang and Yang, Bin and Guo, Chenjuan},
  booktitle = {ICLR},
  year      = {2026}
}

Contact

If you have any questions or suggestions, feel free to contact:

Or describe it in Issues.

Task: Time Series Forecasting
Type: aurora
License: mit
Languages: en
Created: Jan 27, 2026
Updated: Apr 29, 2026