This project aims to forecast Bitcoin prices using various Recurrent Neural Network (RNN) architectures. It compares the performance of different models, including a standard RNN, an Alpha-RNN, GRU, and LSTM.
The project uses historical Bitcoin price data to train and evaluate different RNN models. It includes data preprocessing, model building, training, and evaluation steps.
- Data preprocessing and normalization
- Implementation of multiple RNN architectures:
- Standard RNN
- Alpha-RNN (custom implementation)
- GRU (Gated Recurrent Unit)
- LSTM (Long Short-Term Memory)
- Hyperparameter tuning using GridSearchCV
- Model training with early stopping
- Performance comparison of different models
- Visualization of predictions and errors
- Python
- TensorFlow
- Keras
- Pandas
- NumPy
- Matplotlib
- Scikit-learn
- Statsmodels
- Ensure all required libraries are installed.
- Load the Bitcoin price data (CSV file) into the project.
- Run the Jupyter notebook or Python script to execute the entire pipeline.
- Data loading and preprocessing
- Feature engineering (creating lagged features)
- Model definition and compilation
- Cross-validation and hyperparameter tuning
- Model training and evaluation
- Results visualization
The project compares four different RNN architectures:
- Simple RNN
- Alpha-RNN (a custom RNN implementation)
- GRU
- LSTM
The models are compared based on their Mean Squared Error (MSE) on both training and test datasets. Visualizations are provided to compare the predicted values against the observed values, as well as the prediction errors for each model.
- Incorporate additional features (e.g., trading volume, market sentiment)
- Experiment with more advanced architectures (e.g., attention mechanisms)
- Implement ensemble methods to combine predictions from multiple models