About This Project
Developed and compared a suite of neural network architectures for Bitcoin price forecasting using historical time series data. Implemented naive baseline, dense, LSTM, CNN, N-BEATS, and ensemble models to evaluate performance across varying horizon and window sizes. Applied time series preprocessing including sliding window construction, normalization, and train-test splitting with no data leakage. Evaluated models using MAE, RMSE, and MAPE metrics to identify the best-performing architecture for financial time series prediction.
Key Features
- Compared 6+ architectures: Naive, Dense, LSTM, CNN, N-BEATS, Ensemble
- Sliding window preprocessing with configurable horizon & window sizes
- N-BEATS implementation for interpretable neural forecasting
- Ensemble model combining predictions across architectures
- Evaluated with MAE, RMSE, and MAPE for realistic financial benchmarking
- No-leakage train/test split for robust evaluation
