A Deep Reinforcement Learning framework for trading & order execution in financial markets. The goal of the project is to speed up the development & research of financial agents by building a modular and scalable codebase. The framework supports the following main features:

• Loading & preprocessing data directly from different APIs
• Training & evaluating deep reinforcement learning agents
• Use specific financial metrics or quickly implement your own
• Visualizing the performance of the agent with some intuitive graphs

The nice part is that everything is configurable within a config file.
The code is using popular packages like:

• pytorch
• pandas
• stable-baselines3
• gym
• wandb
• mplfinance

The architecture is split into 4 main categories:

• Data
• Environment
• Reinforcement Learning Agents
• Specific Task Layer

The Specific Task Layer is a glue code module that is used for training & backtesting. It can be further be extended into the applicaton layer.

Visual representations of the actions taken by the agent & the current status of the agent:

• Code tested under Python 3.8, pytorch 1.13.0, and cuda 11.6 on Ubuntu 20.04

1. Create a conda environment and activate it:

conda create --name yacht python=3.8
conda env activate yacht

2. Install torch and cuda separately with conda:

conda install pytorch=1.13.0 torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia

3. Ultimately, install other requirements with pip:

pip install -r requirements.txt

• The configuration system is built with google protobuf. If you want to recompile / change the protobuf files, you should install the protoc compiler on your system:

sudo apt  install protobuf-compiler libprotobuf-dev

• Run the compilation command from the root folder:

  protoc -I=. --python_out=. yacht/config/proto/*.proto

• Create a file called .env at the root directory level. If you want to fully use the market APIs and experiment trackers you should add the secret keys.
• Look at .env.default for the supported env vars.
• Not all env vars are mandatory. For example the free version of Yahoo Finance does not require any credentials.

• Currently, we have support for:
  • Binance
  • Yahoo Finance.
• You should set the api keys in the .env file for full support.

• S&P 500
• Dow 30
• Nasdaq 100
• Russell 2000

You can set tickers: ['NASDAQ100'] in the configuration file and all the tickers from the index will be expanded. You can also set something like ['NASDAQ100', 'S&P500', 'AAPL'] or any combination you like.

The data is stored in h5 files.

python main.py download --config-file-name download_4years.config.txt --storage-dir ./storage/download_4_years --market-storage-dir ./storage

• The --market-storage-dir CLI argument is optional. If you add it the market will be placed in a different location than your storage-dir. This is helpful because it can be accessed by multiple experiments in parallel during training (the h5 file will be set in a read only mode). Otherwise, while training, only one experiment can access a specific file.
• --market-storage-dir should be used also during training & backtesting
• You can use the market_mixins: [...] from the config file to preprocess the data before it is stored.

All the supported configs can be found at ./yacht/config/configs. You should only add the config path relative to the root directory.

python main.py train --config-file order_execution/all/single_asset_all_universal_silent.config.txt --storage-dir ./storage/yacht --market-storage-dir ./storage

python main.py backtest --config-file order_execution/all/single_asset_all_universal_silent.config.txt --storage-dir ./storage/yacht --market-storage-dir ./storage

You can download the pretrained weights from here.

cd /root/directory
mkdir storage
--> place the downloaded weights in ./storage 

Suppose you downloaded and placed the pretrained weights and data correctly as showed above. In that case, you can run the following command to resume the agent:

python main.py train --config-file order_execution/all/single_asset_all_universal_silent.config.txt --storage-dir ./storage/yacht --resume-from best-train --market-storage-dir ./storage

For the parameter --resume-from we support the following combinations:

• Absolute path to the checkpoint.
• latest-train = resume the agent from the latest checkpoint saved during training
• best-train = resume the agent from the best checkpoint saved during training

NOTE: For the best-train parameter, you can choose a metric on which the agent was monitored. You can do that with the meta.metrics_to_load_best_on parameter from the configuration file. For example, metrics_to_load_best_on: ['PA', 'GLR'] will load two agents: The agent monitored on the metric PA & the one who performed the best on GLR.

• Get started

• For now, we support wandb for experiment tracking and logging.
• Just add the api key in the .env file. Also, in the configuration file you should add:

meta: {
  experiment_tracker: 'wandb'
}

If you want to add a specific project_entity add:

meta: {
  project_entity: 'your_project_entity'
} 

NOTE: Be aware that this name is unique between all the users that use wandb. For example, I use project_entity=yacht. If you try to use it will through an unauthorized error because you do not have access to my entity.
Here is an example of how it looks:

• If you don't want to log a specific experiment on the experiment tracker just remove the config field or replace it with the empty string ''.

• We support yyperparameter optimization with weights & biases sweeps.
• Weights & biases should work as a simple experiment tracker before using this.
• You can use any other config from tools/tuning/configs or generate your own.

wandb sweep tools/tuning/configs/single_asset_order_execution_crypto.yaml
wandb agent id-given-by-generated-sweep

For further reading go to:

• Available Components
• Config Explanation
• Trading
• Order Execution
• Metrics
• Teacher-Student