ResearchRL/open-rl-trading-binance-dataset

Reinforcement Learning Dataset for Volatility-Driven Futures Trading

This dataset collection contains four curated subsets of minute-level Binance Futures market data designed specifically for training, validating, testing, and backtesting reinforcement learning agents in high-volatility environments. The data is used in a production-grade project implementing a Dueling Double Deep Q-Network (D3QN) with Prioritized Experience Replay (PER), tailored for financial markets.

---

💡 Project Context

The dataset supports a full RL pipeline aimed at developing an intelligent trading system capable of:

• Making profitable trading decisions in highly volatile conditions;
• Learning from localized market impulses rather than continuous streams;
• Operating in realistic conditions, including slippage and transaction fees.

🔗 Full codebase: GitHub Repository
📄 Research article (English): RL Agent for Algorithmic Trading on Binance Futures — Architecture, Backtest, and Results
📄 Research article (Russian): RL-агент для алгоритмической торговли на Binance Futures: архитектура, бэктест, результаты
🤖 Real-time RL predictions: Telegram

---

📁 Dataset Structure

Each data sample is a 150-minute window centered around a strong volatility impulse.

• Shape: (150, 7) — 150 minutes × 7 features
• Channels: open, high, low, close, volume, volume_weighted_average, num_trades
• Format: np.ndarray wrapped in .npz
• Metadata: unique keys (TICKER, datetime) per session

Subset	Samples	Period	Purpose
Train	24,104	2020-01-14 → 2024-08-31	Training
Validation	1,377	2024-09-01 → 2024-12-01	Hyperparameter tuning
Test	3,400	2024-12-01 → 2025-03-01	Final evaluation
Backtest	3,186	2025-03-01 → 2025-06-01	Realistic simulation

Each session contains:

• 90 minutes of pre-impulse history (for state construction)
• 60 minutes of post-impulse trading session (for reward computation)

---

🧠 Dataset Motivation

This dataset departs from standard continuous sampling. Instead, it focuses only on high-volatility episodes that represent actual market decision points. Specifically:

• Price moves >5% within a 10-minute window
• Preceded by 90 minutes of relative stability
• Selected using a contrast ratio filter to remove noisy signals

These sessions serve as atomic training units for reinforcement learning agents operating in short-term trading strategies.

---

🧰 Data Pipeline Tools

All preprocessing logic is encapsulated in reusable utilities (as part of the open project):

• load_npz_dataset(path) — loads session list and metadata
• select_and_arrange_channels(data, channels) — filters and arranges input
• calculate_normalization_stats(data) — computes per-channel stats
• apply_normalization(data, stats) — standardizes for agent consumption

📂 Utilities

All preprocessing tools are included in data_utils.py.

Data undergoes:

• Channel filtering
• Relative scaling
• Log transforms
• NaN/outlier protection

---

📊 Visualizations

Each sample comes with an optional visualization (example graphs available in the project):

• Line plots with volatility impulse marker at minute 90
• Session metadata in title (ticker + UTC timestamp)
• Used to audit signal quality and alignment

---

🔐 License

License: MIT License
This dataset is released under the MIT license — you are free to use, modify, and distribute it for any purpose, including commercial use, provided that the original copyright and permission notice are included.

---

📦 Download and Usage

You can load the dataset using HuggingFace’s datasets library:

from datasets import load_dataset

# Example: load training split
train_dataset = load_dataset("ResearchRL/open-rl-trading-binance-dataset", split="train_data")