CharlieYPeng/cryptogat-crypto-1d

CryptoGAT Cryptocurrency Daily OHLCV Dataset

  

This dataset accompanies the paper [CryptoGAT: Are Time Series Models Effective for Cryptocurrency Forecasting?](https://arxiv.org/abs/2606.27670) by Yu Peng, Matloob Khushi, and Josiah Poon.

CryptoGAT studies cryptocurrency forecasting from a cross-asset perspective: instead of relying only on temporal patterns within each coin, it treats the market as a graph of interacting assets and learns relationships across cryptocurrencies. This Hugging Face dataset provides the daily OHLCV market data, processed tensors, feature metadata, and asset ordering needed to reproduce and extend the CryptoGAT experiments.

[image]

Why This Dataset

• —Reproducible benchmark for cryptocurrency forecasting. Includes the raw long-format OHLCV table and the processed tensors used by the CryptoGAT experiments.
• —Cross-asset graph modeling ready. The processed files align assets along axis 0, making them convenient for graph neural networks, attention models, and cross-sectional forecasting baselines.
• —Both simple and enhanced features. Use the base OHLCV-derived representation for clean comparisons, or the enhanced technical-indicator representation for richer feature studies.
• —Easy to load from the Hub. The default configuration loads directly with datasets, while processed tensors can be downloaded with huggingface_hub.

Dataset At A Glance

USDCUSDT and TUSDUSDT are kept in the raw OHLCV file but excluded from the processed model tensors.

Quick Start

Load the default raw OHLCV table:

from datasets import load_dataset

dataset = load_dataset("CharlieYPeng/cryptogat-crypto-1d")
df = dataset["train"].to_pandas()

print(df.head())
print(df["symbol"].nunique())
print(df[["date", "symbol", "close", "volume"]].head())

Download a processed tensor used by the paper:

import pickle
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="CharlieYPeng/cryptogat-crypto-1d",
    repo_type="dataset",
    filename="processed/CRYPTO_1D_ALL/eod_data.pkl",
)

with open(path, "rb") as f:
    eod_data = pickle.load(f)

print(eod_data.shape)  # (66, 999, 5)

Download the asset ordering for axis 0:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="CharlieYPeng/cryptogat-crypto-1d",
    repo_type="dataset",
    filename="processed/CRYPTO_1D_ALL/coin_names.txt",
)

with open(path) as f:
    coins = [line.strip() for line in f if line.strip()]

print(coins[:10])

Repository Structure

data/raw_ohlcv.csv
processed/CRYPTO_1D_ALL/eod_data.pkl
processed/CRYPTO_1D_ALL/price_data.pkl
processed/CRYPTO_1D_ALL/gt_data.pkl
processed/CRYPTO_1D_ALL/mask_data.pkl
processed/CRYPTO_1D_ALL/coin_names.txt
processed/CRYPTO_1D_ENHANCED/eod_data.pkl
processed/CRYPTO_1D_ENHANCED/price_data.pkl
processed/CRYPTO_1D_ENHANCED/gt_data.pkl
processed/CRYPTO_1D_ENHANCED/mask_data.pkl
processed/CRYPTO_1D_ENHANCED/coin_names.txt
metadata/base_feature_names.csv
metadata/enhanced_feature_names.csv
metadata/cryptogat_model_assets.csv
metadata/manifest.json
assets/cryptogat-overview-results.png

Raw OHLCV Table

data/raw_ohlcv.csv is a long-format daily market table.

Processed Tensor Layout

The processed files are Python pickle files containing NumPy arrays. Asset order is fixed by the corresponding coin_names.txt file.

CRYPTO_1D_ALL

Feature names are listed in metadata/base_feature_names.csv:

open_norm, high_norm, low_norm, close_norm, volume_norm

CRYPTO_1D_ENHANCED

Feature names are listed in metadata/enhanced_feature_names.csv. They include normalized OHLCV fields, moving-average ratios, MACD, RSI, rate-of-change, volatility, volume-flow features, candle-shape features, and return lags.

Intended Uses

This dataset is designed for research on:

• —cryptocurrency return forecasting;
• —cross-asset graph neural networks;
• —graph attention models for financial markets;
• —time-series versus cross-sectional modeling comparisons;
• —reproducible baselines for pure price-based crypto prediction;
• —feature engineering studies on daily OHLCV data.

It can also be used as a compact benchmark for teaching or prototyping financial machine learning pipelines.

Reproducing CryptoGAT

The original training code expects processed tensors under the GitHub repository's dataset/ directory. To use this Hugging Face copy for reproduction, download the processed folders and place them as:

CryptoGAT/dataset/CRYPTO_1D_ALL/
CryptoGAT/dataset/CRYPTO_1D_ENHANCED/

Then follow the training instructions in the official implementation:

https://github.com/FanBroWell/CryptoGAT

Data Source And License Note

The raw market data are cryptocurrency OHLCV records collected from Binance USDT trading pairs and released here for academic research and reproducibility of the CryptoGAT experiments. Users should independently verify that their intended use complies with the terms of the original data source.

This dataset is provided for research and benchmarking. It is not financial advice and should not be used as the sole basis for trading or investment decisions.

Citation

If you use this dataset, code, or paper, please cite:

@misc{peng2026cryptogat,
  title = {{CryptoGAT}: Are Time Series Models Effective for Cryptocurrency Forecasting?},
  author = {Peng, Yu and Khushi, Matloob and Poon, Josiah},
  year = {2026},
  eprint = {2606.27670},
  archivePrefix = {arXiv},
  primaryClass = {cs.CE},
  doi = {10.48550/arXiv.2606.27670},
  url = {https://arxiv.org/abs/2606.27670}
}

Links

• —Paper: https://arxiv.org/abs/2606.27670
• —Code: https://github.com/FanBroWell/CryptoGAT
• —Dataset: https://huggingface.co/datasets/CharlieYPeng/cryptogat-crypto-1d