This program attempts to predict the next day's closing price for the S&P500 stock index using a hybrid LSTM/CNN deep learning network.
This program uses historical price data (Open, High, Low, Close), the Adjusted Closing Price (adjusted for stock splits and dividends), and trading volume to predict the next day's closing price for the S&P500 stock index ("SPY"). Stock data is downloaded using the Yahoo Finance (yfinance) library, but data can also be utilized from the SPY.txt file.
The model can be tested and visualized best using the stock_prediction.ipynb JupyterHub notebook, and architecture, parameters, and hyperparameters can be adjusted prior to incorporating into the Python script file. To train the model, the stock_prediction.py file is used so that a GPU may be utilized to speed training. This file allows for adjusting parameters and hyperparameters, institutes early stopping, and saves the model when optimally trained to the file stock_prediction.h5.
Once the model is trained and saved, datasets can be selected from the near 30 years of SPY data and tested using the model-test.ipynb Jupyter notebook file. Accuracy will be calculated, results graphed, and percent profitable trades for the test period selected shown. Follow instructions in the notebook markdown cells and comments.
This model as currently constructed does not predict profitable trades overall and will need better optimization of parameters and hyperparameters if it to successfully predict daily closing prices. Additional inputs such as technical indicators, other financial inputs such as interest rates and economic growth, and measurements of public sentiment and optimism may also be beneficial.
Future plans by the author include:
- Utilization of keras tuner to optimize network structure and hyperparameters
- Introducing additional technical analysis inputs from the TA-lib library
Modeled after Shah, A., Gor, M., Sagar, M., & Shah, M. (2022). A stock market trading framework based on deep learning architectures. Multimedia Tools and Applications, 81, 14153-14171. Retrieved from https://link.springer.com/content/pdf/10.1007/s11042-022-12328-x.pdf.