What it does Personae is a Python toolkit that implements seven academic papers—four reinforcement learning algorithms (DDPG, Double DQN, Dueling DQN, Policy Gradient) and three supervised models (DA-RNN, TreNet, naive LSTM)—then drops them into a bare-bones simulated market for stocks and futures. It also includes data spiders, MongoDB storage, and Docker images so you can run the whole pipeline without manually installing TensorFlow 1.4-era dependencies.

The interesting bit The author was brutally honest: reconstruction was supposed to end by September 2018, or whenever they landed a job. That deadline passed, and the repo still warns that its input features are “naive” and day-frequency data is “clearly not enough.” Refreshing candor in a field that usually promises alpha.

Key highlights

• Ships with a gym-like market environment (Market, Trader, Positions) that works without OpenAI Gym installed
• Docker image inherits from CUDA 8.0 / cuDNN 6 runtime for GPU training
• Includes spiders to crawl Chinese stock and futures data via tushare into MongoDB
• All TensorFlow models support checkpoint persistence and TensorBoard summaries
• Experiments include profit curves against baseline and price-prediction plots on four bank stocks

Caveats

• Explicitly frozen in a 2018 dependency stack: Python 3.5, TensorFlow 1.4, CUDA 8.0
• Author warns features are naive and daily data is insufficient for serious trading
• Short selling is noted as “still implementing”
• README contains broken links (e.g., DDPG experiment points to algorithm/SL/DualAttnRNN.py)

Verdict Useful if you need a quick, opinionated scaffold to swap your own features into established RL/SL algorithms. Skip it if you want production-grade backtesting or modern PyTorch/TensorFlow 2.x code; the value here is the wiring, not the models themselves.