We show a workflow of applying RL in algorithmic trading, which is a reproduction and improvement of the process in the NeurIPS 2018 paper.

Usage

Step I. Data

First, run the notebook: Stock_NeurIPS2018_1_Data.ipynb.

It downloads and preprocesses stocks' OHLCV data.

It generates two csv files: train.csv, trade.csv. You can check the provided two sample files.

Step II. Train a Trading Agent

Second, run the notebook: Stock_NeurIPS2018_2_Train.ipynb.

It shows how to process the data into an OpenAI gym-style envrionment, and then train a DRL agent.

It will generate a trained RL model .zip file. Here, we also provided a training A2C model in .zip file.

Step III. Backtest

Finally, run the notebook: Stock_NeurIPS2018_3_Backtest.ipynb.

It backtests the trained agent and compares with two baselines: Mean-Variance Optimization and the market DJIA index, respectively.

At the end, it will plot a figure of the portfolio value during the backtest process.