FinRL-Meta is a comprehensive framework designed to provide a universe of near-real market environments for data-driven deep reinforcement learning (DRL) in quantitative finance. The framework addresses key challenges in DRL-based market simulation, such as noisy historical data and the need for diverse agents, by separating data processing from the DRL pipeline and offering open-source data engineering tools. It follows the DataOps paradigm to automate data access, cleaning, and feature extraction from multiple financial data sources, including Yahoo Finance, CCXT, Alpaca, and others. FinRL-Meta provides hundreds of market environments for various trading tasks, such as stock trading, cryptocurrency trading, and portfolio allocation, and supports multiprocessing training on thousands of GPU cores to accelerate agent training.

The framework employs a plug-and-play design, allowing any DRL agent to be easily integrated and evaluated on benchmark environments. It uses a training-testing-trading pipeline to prevent information leakage and ensure fair comparisons. Experimental results demonstrate the effectiveness of FinRL-Meta: in stock trading, agents trained with PPO (from ElegantRL and Stable-baselines3) outperformed the DJIA index in both backtesting and paper trading, achieving higher annual returns and Sharpe ratios. In cryptocurrency trading, the ElegantRL agent significantly outperformed a buy-and-hold strategy on Bitcoin, with an annual return of 360.823% and Sharpe ratio of 2.992 in backtesting. These results validate the framework's ability to generate realistic market environments and facilitate the development of high-performing DRL trading strategies. Future work aims to build a multi-agent market simulator with tens of thousands of agents, termed FinRL-Metaverse, to further explore complex market dynamics and regulatory insights.