FinRL-Meta icon indicating copy to clipboard operation
FinRL-Meta copied to clipboard

FinRL­-Meta: Dynamic datasets and market environments for FinRL.

FinRL-Meta: A Universe of Market Environments and Benchmarks for Data-Driven Financial Reinforcement Learning

Downloads Downloads Python 3.6 PyPI

FinRL-Meta (website) builds a universe of market environments for data-driven financial reinforcement learning. We aim to server the users in our community to help them to build environments.

  1. FinRL-Meta provides hundreds of market environments.
  2. FinRL-Meta reproduces existing papers as benchmarks.
  3. FinRL-Meta provides tens of demos/tutorials, organized in a curriculum.

Previously called Neo_FinRL: Near real-market Environments for data-driven Financial Reinforcement Learning.


  • News and Tutorials
  • Our Goals
  • Design Principles
  • Overview
  • Plug-and-Play
  • Training-Testing-Trading
  • Our Vision

News and Tutorials

Our Goals

  • To provide benchmarks and facilitate fair comparisons, we allow researchers to evaluate different strategies on the same dataset. Also, it would help researchers to better understand the “black-box” nature (deep neural network-based) of DRL algorithms.
  • To reduce the simulation-reality gap: existing works use backtesting on historical data, while the actual performance may be quite different.
  • To reduce the data pre-processing burden, so that quants can focus on developing and optimizing strategies.

Design Principles

  • Plug-and-Play (PnP): Modularity; Handle different markets (say T0 vs. T+1)
  • Completeness and universal: Multiple markets; Various data sources (APIs, Excel, etc); User-friendly variables.
  • Layer structure and extensibility: Three layers including: data layer, environment layer, and agent layer. Layers interact through end-to-end interfaces, achieving high extensibility.
  • “Training-Testing-Trading” pipeline: simulation for training and connecting real-time APIs for testing/trading, closing the sim-real gap.
  • Efficient data sampling: accelerate the data sampling process is the key to DRL training! From the ElegantRL project. we know that multi-processing is powerful to reduce the training time (scheduling between CPU + GPU).
  • Transparency: a virtual env that is invisible to the upper layer
  • Flexibility and extensibility: Inheritance might be helpful here


Overview image of FinRL-Meta We utilize a layered structure in FinRL-Meta, as shown in the figure above, that consists of three layers: data layer, environment layer, and agent layer. Each layer executes its functions and is independent. Meanwhile, layers interact through end-to-end interfaces to implement the complete workflow of algorithm trading. Moreover, the layer structure allows easy extension of user-defined functions.


DataOps applies the ideas of lean development and DevOps to the data analytics field. DataOps practices have been developed in companies and organizations to improve the quality and efficiency of data analytics. These implementations consolidate various data sources, unify and automate the pipeline of data analytics, including data accessing, cleaning, analysis, and visualization.

However, the DataOps methodology has not been applied to financial reinforcement learning researches. Most researchers access data, clean data, and extract technical indicators (features) in a case-by-case manner, which involves heavy manual work and may not guarantee the data quality.

To deal with financial big data (unstructured), we follow the DataOps paradigm and implement an automatic pipeline in the following figure: task planning, data processing, training-testing-trading, and monitoring agents’ performance. Through this pipeline, we continuously produce DRL benchmarks on dynamic market datasets.

Supported Data Sources:

Data Source Type Range and Frequency Request Limits Raw Data Preprocessed Data
Alpaca US Stocks, ETFs 2015-now, 1min Account-specific OHLCV Prices&Indicators
Baostock CN Securities 1990-12-19-now, 5min Account-specific OHLCV Prices&Indicators
Binance Cryptocurrency API-specific, 1s, 1min API-specific Tick-level daily aggegrated trades, OHLCV Prices&Indicators
CCXT Cryptocurrency API-specific, 1min API-specific OHLCV Prices&Indicators
IEXCloud NMS US securities 1970-now, 1 day 100 per second per IP OHLCV Prices&Indicators
JoinQuant CN Securities 2005-now, 1min 3 requests each time OHLCV Prices&Indicators
QuantConnect US Securities 1998-now, 1s NA OHLCV Prices&Indicators
RiceQuant CN Securities 2005-now, 1ms Account-specific OHLCV Prices&Indicators
Tushare CN Securities, A share -now, 1 min Account-specific OHLCV Prices&Indicators
WRDS US Securities 2003-now, 1ms 5 requests each time Intraday Trades Prices&Indicators
YahooFinance US Securities Frequency-specific, 1min 2,000/hour OHLCV Prices&Indicators

OHLCV: open, high, low, and close prices; volume

adjusted_close: adjusted close price

Technical indicators users can add: 'macd', 'boll_ub', 'boll_lb', 'rsi_30', 'dx_30', 'close_30_sma', 'close_60_sma'. Users also can add their features.

Plug-and-Play (PnP)

In the development pipeline, we separate market environments from the data layer and the agent layer. A DRL agent can be directly plugged into our environments. Different agents/algorithms can be compared by running on the same benchmark environment for fair evaluations.

The following DRL libraries are supported:

  • ElegantRL: Lightweight, efficient and stable DRL implementation using PyTorch.
  • Stable-Baselines3: Improved DRL algorithms based on OpenAI Baselines.
  • RLlib: An open-source DRL library that offers high scalability and unified APIs.

A demonstration notebook for plug-and-play with ElegantRL, Stable Baselines3 and RLlib: Plug and Play with DRL Agents

"Training-Testing-Trading" Pipeline

We employ a training-testing-trading pipeline. First, a DRL agent is trained in a training dataset and fine-tuned (adjusting hyperparameters) in a testing dataset. Then, backtest the agent (on historical dataset), or depoy in a paper/live trading market.

This pipeline address the information leakage problem by separating the training/testing and trading periods.

Such a unified pipeline also allows fair comparisons among different algorithms.

Our Vision

For future work, we plan to build a multi-agent-based market simulator that consists of over ten thousands of agents, namely, a FinRL-Metaverse. First, FinRL-Metaverse aims to build a universe of market environments, like the XLand environment (source) and planet-scale climate forecast (source) by DeepMind. To improve the performance for large-scale markets, we will employ GPU-based massive parallel simulation just as Isaac Gym (source). Moreover, it will be interesting to explore the deep evolutionary RL framework (source) to simulate the markets. Our final goal is to provide insights into complex market phenomena and offer guidance for financial regulations through FinRL-Meta.

Citing FinRL-Meta

    author = {Liu, Xiao-Yang and Rui, Jingyang and Gao, Jiechao and Yang, Liuqing and Yang, Hongyang and Wang, Zhaoran and Wang, Christina Dan and Guo Jian},
    title   = {{FinRL-Meta}: Data-Driven Deep ReinforcementLearning in Quantitative Finance},
    journal = {Data-Centric AI Workshop, NeurIPS},
    year    = {2021}



Disclaimer: Nothing herein is financial advice, and NOT a recommendation to trade real money. Please use common sense and always first consult a professional before trading or investing.