CS 229 (Machine Learning, Spring 2019) Final Project.
This is the repository for my final project for the class. It is joint with Justin Lundgren.
Title: Algorithmic Trading using LSTM-Models for Intraday Stock Predictions
Paper:
Our final project is available at http://cs229.stanford.edu/proj2019/report/31.pdf. If you do not have access to Stanford papers, it can be found at https://www.dropbox.com/s/77ynq4ymd93o3al/CS229_final_version.pdf?dl=0
Code:
The src directory contains the pipelines for a VAR-model, a VARMAX-model, an LSTM-model and an R2N2-model which is the combination of the LSTM model and the residuals of the VARMAX-model. The src directory also contins a directory with the raw data, the output, and the evaluation methods, which are based on correlations with actual returns, a simple long/short investment strategy and mean squared error.
Setup
Make sure that you have Tensorflow 1.13 installed instead of Tensorflow 2.0. We use Keras for our code with a tensorflow backend. The current version of Keras (which I believe is 2.2.4, as of June 2019) is not compatible with Tensorflow 2.0.
Hyperparameter tuning
Here's something to play around with, namely hyperparameter tuning via a random grid search. First clone the repository, then go to src/LSTM
and run
python hyp_tuning.py.
Make sure that in the python file, the boolean rand_tuning is set to True. Now when you run hyp_tuning.py, it randomly selects the hyperparameters
- Number of LSTM units.
- Lookback period
- Learning rate
and saves the MSE value obtained on the validation set of our data into a CSV file. In the file hyp_tuning.py, on about line 140, the quantity num_trials is the number of times we want python to generate a random choice of these parameters. At the moment, num_trials is 4. We save the CSV file in src/output/LSTM_tuning/random_samples. To distinguish the different files each time we execute python hyp_tuning.py, the CSV files contain a random number in their name.
To obtain Fig. 2 in our CS 229 project linked above, go to src/output/LSTM_training/ and run the jupyter notebook Tuning_plot.ipynb. Run all the code in there, and the figure generated will be a plot of hyperparameters against MSE value.
Finally, note that the plot is made out of all the CSV files in the folder src/output/LSTM_tuning/random_samples.