|
|
--- |
|
|
license: apache-2.0 |
|
|
viewer: false |
|
|
task_categories: |
|
|
- time-series-forecasting |
|
|
tags: |
|
|
- finance |
|
|
pretty_name: Apogée |
|
|
size_categories: |
|
|
- 10B<n<100B |
|
|
--- |
|
|
<div align="center"> |
|
|
<a href="https://www.duonlabs.com" target="_blank"> |
|
|
<img src="https://www.duonlabs.com/static/front/logo/duon_white.png" width="30%" alt="Duon Labs Logo" /> |
|
|
</a> |
|
|
</div> |
|
|
<h1 align="center" style="font-size: 3rem;">Apogée: Crypto Market Candlestick Dataset</h1> |
|
|
<hr> |
|
|
|
|
|
## Overview |
|
|
|
|
|
Most traders believe crypto is random, but deep learning scaling laws suggest otherwise. Apogée is an open-source research initiative exploring the **scaling laws of crypto market forecasting**. While financial markets are often assumed to be unpredictable, modern deep learning suggests that increasing data and compute could uncover measurable predictability. |
|
|
Our goal is to **quantify how many bits of future price movement can be inferred** from historical candlestick data. |
|
|
[More informations on Apogée](https://www.duonlabs.com/apogee) |
|
|
|
|
|
This dataset serves as the foundation for training large-scale deep learning models on historical crypto market data. |
|
|
|
|
|
## Dataset Description |
|
|
|
|
|
The dataset contains **1-minute interval candlestick data** for the top 100 crypto pairs, sourced from Binance. It is stored in an optimized format that allows for high-performance training and multi-scale aggregation. |
|
|
|
|
|
### Assets Included: |
|
|
- **BTC/USDT** (Bitcoin) |
|
|
- **ETH/USDT** (Ethereum) |
|
|
- **SOL/USDT** (XRP) |
|
|
- **XRP/USDT** (Binance Coin) |
|
|
- **DOGE/USDT** (Solana) |
|
|
- **LTC/USDT** (Dogecoin) |
|
|
- **BNB/USDT** (Cardano) |
|
|
- **ADA/USDT** (Tron) |
|
|
- **AVAX/USDT** (Chainlink) |
|
|
- **LINK/USDT** (Avalanche) |
|
|
|
|
|
and [many more](https://huggingface.co/datasets/duonlabs/apogee/raw/main/metadata.csv) |
|
|
|
|
|
### Data Properties: |
|
|
Total dataset size: |
|
|
- **~2.188 billion candles** |
|
|
- **~43.76 billion tokens** (after uint8 tokenization) |
|
|
|
|
|
## Storage Format |
|
|
|
|
|
The dataset is stored as **NumPy memory-mapped buffers** (`.npy`) to allow for **efficient streaming and real-time aggregation**. This approach enables high-speed data access without requiring the full dataset to be loaded into RAM. |
|
|
This efficiency allows **real-time lazy aggregation** to generate different timeframes on demand |
|
|
|
|
|
### Baseline implementation |
|
|
The official dataloader used in project apogee is available at: [https://github.com/duonlabs/apogee/blob/master/apogee/data/loading.py](https://github.com/duonlabs/apogee/blob/master/apogee/data/loading.py) |
|
|
|
|
|
We tested performances under: |
|
|
- **batch_size**: 32 |
|
|
- **context_size**: 960 (tokens) |
|
|
- **Aggregation**: 1m, 5m, 30m, 2h, 8h, 1d |
|
|
|
|
|
On a consumer-grade laptop with an SSD: |
|
|
- **254.79 batches/sec** |
|
|
- **8,153.39 samples/sec** |
|
|
- **3,921,782.52 tokens/sec** |
|
|
|
|
|
## Applications |
|
|
|
|
|
This dataset is designed for training **deep learning models** on crypto price prediction, particularly in the context of **scaling law research**. Potential applications include: |
|
|
|
|
|
- **Autoregressive price forecasting** using models like **Transformers** or **State-Space Models** (SSMs). |
|
|
- **Analyzing predictability limits** in crypto markets. |
|
|
- **Developing trading algorithms** based on learned patterns in candlestick sequences. |
|
|
- **Exploring market efficiency** by testing if deep learning models can systematically extract information from past price movements. |
|
|
- **Scaling law analysis** to determine how predictive power improves with increased dataset size and model capacity. |
|
|
|
|
|
## Binance Data Disclaimer |
|
|
|
|
|
Please note that all our data services strictly follow the [Binance Terms of Use](https://www.binance.com/en/support/faq/how-to-download-historical-market-data-on-binance-5810ae42176b4770b880ce1f14932262) |
|
|
> Without written consent from Binance, the following commercial uses of Binance data are prohibited: |
|
|
> |
|
|
> - Trading services that make use of Binance quotes or market bulletin board information. |
|
|
> - Data feeding or streaming services that make use of any market data of Binance. |
|
|
> - Any other websites/apps/services that charge for or otherwise profit from (including through advertising or referral fees) market data obtained from Binance. |
|
|
> |
|
|
> You hereby understand and agree that Binance will not be liable for any losses or damages arising out of or relating to: |
|
|
> - (a) Any inaccuracy, defect, or omission of digital asset price data. |
|
|
> - (b) Any error or delay in the transmission of such data. |
|
|
> - (c) Interruption in any such data. |
|
|
> - (d) Regular or unscheduled maintenance carried out by Binance and service interruption and change resulting from such maintenance. |
|
|
> - (e) Any damages incurred by other users’ actions, omissions, or violation of these terms. |
|
|
> - (f) Any damage caused by illegal actions of third parties or actions without authorization by Binance. |
|
|
> - (g) Other exemptions mentioned in disclaimers and platform rules issued by Binance. |
|
|
|
|
|
## Citation & References |
|
|
|
|
|
If you use this dataset in your research, please cite the Apogée project: |
|
|
``` |
|
|
@misc{apogee2025, |
|
|
title={Apogée: Scaling Laws for Crypto Market Forecasting}, |
|
|
author={Duon Labs}, |
|
|
year={2025}, |
|
|
url={https://github.com/duonlabs/apogee} |
|
|
} |
|
|
``` |
|
|
|
|
|
|
|
|
For more details, refer to: |
|
|
- **Apogée GitHub Repo**: [https://github.com/duonlabs/apogee](https://github.com/duonlabs/apogee) |
|
|
- **Apogée Community**: [https://t.me/DuonLabs](https://t.me/DuonLabs) |
