A transformer model that predicts the next values of multiple, related, time series

To run the model

1. Install the required packages listed in requirements.txt
2. Run the train.py script to train the model
   • You may want to adjust the hyperparameters in the script, in particular the number of epochs, depending on your hardware
   • I have not tested the model extensively with different epochs, the predictions are statistically significant with 30,000 epochs and default parameters but this takes a long time to train
   • Stock data is downloaded from Yahoo Finance and cached in the data_filename file.
   • The model is saved in the model_filename file.
   • The script will also save the scaler used to normalize the data in the scaler_filename file.
3. Run the test_model.py script to test the model
   • This will generate a plot of the model's predictions, run a T-test to determine if the predictions are statistically significant, and print the mean squared error of the model's predictions
4. Run the virtual_trading.py script to simulate trading based on the model's predictions
   • This will simulate trading in February based on the data from January with default parameters
   • Note: Yahoo Finance data seems to update at EOM for some stocks, so setting the end data to any day in the current month will not work
   • Note: You will have to run python split_data.py the first time you run this script to generate the data

Additional information

• The docker file will build an image that runs the training script only.
• The model will be deleted after training, if you want to keep it you should move/rename it.