TradeMaster — RL platform for quantitative trading

An open-source platform for quantitative trading powered by reinforcement learning, from the AMI research group at Nanyang Technological University (NTU), Singapore. Where tensortrade is a lightweight RL library, TradeMaster bills itself as a holistic ecosystem — the broadest instance of reinforcement-learning-trading in this wiki, spanning the whole predict → decide → execute stack in one toolkit. Apache-2.0; ~2.9k★ (v1.0.0, 2023). T3: first-party academic repo with self-reported capabilities, but with published-paper backing (the named algorithms are research papers — a T1 upgrade path).

The six modules

A full pipeline rather than a single component:

1. Multi-modal market data — stocks, crypto, futures, forex at multiple timeframes (8 bundled datasets: S&P 500, DJ30, SSE50, HS30, Bitcoin, futures, forex; 2000–2023).
2. Data preprocessing — including diffusion-model missing-value imputation and automatic feature generation/selection.
3. Market simulators — data-driven simulation environments per trading task.
4. RL algorithm library — 13+ trading-specific algorithms (DeepScalper, EIIE, SARL, DeepTrader, OPD, Investor-Imitator) plus classic baselines (PPO, DQN, SAC) via Ray RLlib.
5. Evaluation toolkit — a six-axis, 17-measure assessment framework, with the PRIDE-Star multi-dimensional visualization.
6. User interfaces — a web sandbox for agent training and market-dynamics modeling.

Tasks covered

Portfolio management, algorithmic/intraday trading, order execution, high-frequency trading, and market making — i.e. it reaches past the founding trio’s strategy/execution focus into execution-level and market-microstructure tasks.

Why it matters here

Three things it adds to the spoke:

• Deepens the RL/decide layer. A whole algorithm zoo and simulators, vs tensortrade‘s component toolkit — it is the heavyweight to TensorTrade’s lightweight, a contrast TradeMaster itself draws.
• Brings systematic evaluation. The 6-axis / 17-measure toolkit speaks to the synthesis’s open “where’s the evidence?” question — not a neutral third-party benchmark, but a structured, reproducible evaluation method beyond a single self-reported number.
• One ecosystem spans the stack. Data → simulation → RL policy → evaluation in a single platform is a data point that the predict/decide/execute layers do compose — and its FinAgent (multimodal foundation agent) and Market-GAN (synthetic market data) reach toward the foundation-model corner (Kronos).

Related

reinforcement-learning-trading · tensortrade · algorithmic-trading · backtesting · strategy-optimization · financial-time-series-foundation-model · kronos-financial-foundation-model · nanyang-technological-university · synthesis