/DDPG_for_finance

Primary LanguageJupyter Notebook

DDPG_for_finance

Installation steps

conda create -n ddpg_trading python=3.8
conda activate ddpg_trading
brew install swig
pip install -r requirements.txt

Install pytorch based on system from here:

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Make sure you have an active weights and biases account login:

  • run 'wandb login' in terminal

Downloading DJI data for testing

Run the following command with required start_date and end_date:

python download_data.py --start-date 2009-01-01 --end-date 2022-11-30

This command will download DOW30 data, process it, and add technical indicators defined in "config.py". The output csv file will be saved in './data' directory and filename will be printed as output.

For Generating results

  • Change "model_type" variable in config_results.py based on whether you want results for daily or monthly trading

  • Change "processed_csv" variable in config_results.py (line 7) to select the data file required

  • Download trained models from google drive and place them in '../trained_models/'

Run:

python get_results_artefacts.py

Note: make sure your wandb log in is done

For hyperparameter tuning before training [Optional]

  • Make changes in "config_tuning.py" if you want to change num_trials etc.
  • Check train test split in hyperparemeter_tuning.py script and Run:
python hyperparemeter_tuning.py

Hyperparameters returned can then be updated in config.py before running training in next step.

For training on stock market data [without retraining]

  • Change config.py for hyperparameters, artefact storage directories, wandb project name etc.

Run:

python train_ddpg_agent.py

For training on stock market data [with retraining]

  • Change config.py for hyperparameters, artefact storage directories, wandb project name etc.

Run:

python train_ddpg_agent_with_retraining.py

Note: if the parameter 'do_hyp_tuning' is set to True in train_ddpg_agent_with_retraining.py, you don't need to mention hyperparams in config as they won't be used. For every training, optimal hyperparameters will be found and used.