For now our codebase is inspired by the paper Universal Trading for Order Execution with Oracle Policy Distillation. If you are interested to see the main concepts I wrote a summary on Medium. We made a few changes to make it more generic & robust:
- The setup can we used on any kind of asset timeseries.
- The actions are taken only at one time step based on some lagged observations.
- You can use it at any timescale: 1 year, 1 month, 1 day
- We have improved the preprocessing method so the data is stationary.
- We have added a supervised head to help learning.
- We have added as input reference indexes, like S&P500 & Nasdaq100 to help the agent to find the trend.
- We have improved the model to help the agent to pick features from different variables.
A detailed explanation of those improvements can be found at the following paper.
Here are some examples of how the agent is performing:
This is how the agent is taking the actions:
