🔥 FiRe: Financial Reinforcement Learning by Trality

Welcome to FiRe! FiRe is a Reinforcement Learning (RL) pipeline built to research Single- and Multi-Objective trading algorithms.

What is supported?

• Critic-only Deep Q-Learning RL agent with Hindsight Experience Replay link
• Single- and Multi-Objective reward learning with generalization in the sense of [Friedman, Fontaine]
• MLP-type agent’s Neural Network
• Four reward mechanisms (last logarithmic return, average logarithmic return, Sharpe Ratio, profit only when closing)

Install

We recommend python version 3.10. (Check your version with python3 --version)

We recommend using a virtual environment: to create it, run

python3 -m venv .venv
source .venv/bin/activate

in the repository's main folder.

Install the required packages and download the Bitcoin dataset by running

make

in the main folder.

Test

To test the code you can run the sin-wave example:

python3 main.py parameters/sin_wave_test_multi.json

How to run an experiment

In order to run a particular experiments run the main.py script adding as argument a json configuration file:

python3 main.py <your configuration file>.json

the json configuration file contains information about:

• The chosen dataset
• The reward or set of rewards to use
• Q-Learning general parameters
• Neural Network hyperparameters

Experiments results are stored in the results_of_experiments folder, each subfolder contains the results of each single experiment.

Reproduce experiments in paper (Bitcoin dataset only)

We provide all configuration files for the experiments on the Bitcoin dataset that are shown in the paper. To run the experiments use:

git checkout e2bc651b07eec388ee0b03228e0acdee8550ba7d
sh scripts/run_experiments_in_folder.sh parameters/paper_plots

How to run walk forward validation

In order to run a walk forward validation run the walk_forward_validation.py script with the following arguments:

python walk_forward_validation.py --config <json configuration file> --n_folds <number of folds>

you can add the argument --anchored if you want the train set of each fold to start at the same point.

You can add the argument --jobs <number of jobs> if you want to parallelize the execution.

The given configuration file specifies the configuration of the experiment of the first fold, the following folds will be based on the same configuration but with different datasets.

Currently it's not possible to run walk forward validation with synthetic datasets.

Datasets

For the simulation we discussed in our paper we used two datasets:

• Bitcoin hourly close price (BTC-USD)
• Nifty50 minute close price

You can download the Bitcoin dataset by running:

make datasets

(this is not necessary if you already ran make)

Latex for plots

If you want to use latex to create the plots you need to install the following packages:

sudo apt-get install texlive-latex-extra texlive-fonts-recommended texlive-fonts-extra dvipng

You can now set "text.usetex": True, at the top of src/plots.py.

Acknowledgement of sources

Our code utilizes parts of the following open source repositories:

1. The Deep Q-Network implementation at https://github.com/mswang12/minDQN
2. The stocks environment at https://github.com/AminHP/gym-anytrading