This repository contains a series of models designed to predict Bitcoin prices using historical time-series data. The models implement a range of machine learning and deep learning approaches, including Dense Artificial Neural Networks (ANN), Convolutional Neural Networks (Conv1D), Long Short-Term Memory (LSTM) networks, and more advanced methods like N-Beats and Ensemble Models.
Below are the performance metrics for each model:
| Model | MAE | MSE | RMSE | MAPE (%) | MASE |
|---|---|---|---|---|---|
| Model 1: Dense ANN | 576.48 | 1,181,204 | 1,086.83 | 2.60 | 1.01 |
| Model 2: Conv1D | 578.55 | 1,191,960 | 1,091.77 | 2.60 | 1.02 |
| Model 3: LSTM | 581.04 | 1,218,118 | 1,103.68 | 2.62 | 1.02 |
| Model 4: N-Beats | 579.60 | 1,158,628 | 1,076.40 | 2.65 | 1.02 |
| Model 5: Ensemble | 577.34 | 1,166,218 | 1,079.92 | 2.60 | 1.01 |
| Model 6: Turkey | 17,149.12 | 615,804,400 | 24,815.41 | 121.62 | 26.54 |
- Model 1: Dense ANN (Artificial Neural Network): Fully connected feedforward neural network for predicting Bitcoin prices.
- Model 2: Conv1D (1D Convolutional Neural Network): CNN-based model for detecting patterns in time-series data.
- Model 3: LSTM (Long Short-Term Memory): A recurrent neural network designed for sequence prediction tasks, capturing long-term dependencies in data.
- Model 4: N-Beats: Deep learning-based time-series forecasting model that uses blocks of neural networks to capture trends and seasonal patterns.
- Model 5: Ensemble: Combines the predictions of multiple models (ANN, LSTM, Conv1D) to provide more accurate forecasts.
- Model 6: Turkey Model: Predicts Bitcoin prices considering extreme events, simulating catastrophic events in the market.
- Deliverable: Cleaned and prepared Bitcoin time-series data, ready for model training.
- Deliverable: Trained models for Dense ANN, Conv1D, LSTM, N-Beats, Ensemble, and Turkey models.
- Deliverable: Evaluation of model performance based on MAE, MSE, RMSE, MAPE, and MASE metrics.
- Deliverable: Visualizations showing the comparison of predicted vs. actual Bitcoin prices.
The Naive Bayes Model provides the most robust approach by combining the strengths of individual models. However, the Turkey Model highlights how extreme events can disrupt predictions, with a very high error rate observed during its evaluation. This suggests the importance of accounting for such anomalies in time-series forecasting tasks.