In this repo I am using deep learning recurrent neural networks to model bitcoin closing prices. One model will use the FNG indicators to predict the closing price while the second model will use a window of closing prices to predict the nth closing price.
- For the closing price model, I used previous closing prices to try and predict the next closing price.
- For the Fear and Greed model, I used the FNG values to try and predict the closing price.
- For each model used 70% of the data for training and 30% of the data for testing.
- Applied a MinMaxScaler to the X and Y values to scale the data for the model. Lastly, reshaped the X_train and X_test values to fit the model's requirement of samples, time steps, and features. (example: X_train = X_train.reshape((X_train.shape[0], X_train.shape[1], 1)))
In each Notebook, I created the same custom LSTM RNN architecture. In one notebook, I fitted the data using the FNG values. In the other notebook, I fitted the data using only closing prices.
Lastly, use the testing data to evaluate each model and compare the performance.
Use the above to answer the following:
- Which model has a lower loss? - LSTM RNN Closing Prices
- Which model tracks the actual values better over time? - LSTM RNN Closing Prices
- Which window size works best for the model? - window size = 5