CS230 Winter 2019: Prediction of Apple shares relative relative performance to SPY
Contributors: Sebastian Hurubaru, Ruben Ignacio Contesti, Jun Yan
We have generated in total a number of eight models:
- GRU/LSTM: a single layer of GRU/LSTM with 128 units
- Wavelet GRU/Wavelet LSTM: the previous model with performing two wavelet transformation on the features
- SAE Wavelet GRU/SAE Wavelet LSTM: the model above with encoding the twenty features in only ten deep features
- Tweets SAE Wavelet GRU/Tweets SAE Wavelet LSTM: the model above with adding extracted sentiment from the daily tweets as feature
The project is structured in the followings:
- /data: folder containing the datasets in original and then pre-processed to fit the models
- /output: contains automatically created folders with the some models and their associated scalers
- prediction_model.py: file containing the implementation of the models enumerated above
- prediction_model_generator.py: generates the model file for the hyperparameters specified in the globals.py file
- globals.py: file containing the hyperparameters and the rest of the global values
- keras_extensions.py: contains variouse custom metric functions: RMSE, Theil U, MAPE(customized with an epsilon value of 1)
- utils.py: file containing utility functions used throughout the project
- sae_model.py: file containing the implementation of the Stacked Auto Encoders models
- Predicting_Movie_Reviews_with_BERT_on_TF_Hub-Modified.py: sample file from Google that trains BERT on a movie review dataset
- bert_sentiment_predictor.py: file containing the sentiment extraction from a daily list of tweets using BERT