Bitcoin Price Forecasting: Comparative Analysis of Artificial Neural Network Architectures and Training Algorithms
This study provides a comprehensive evaluation of artificial neural networks for Bitcoin price prediction, comparing architectural designs and optimization algorithms. Through rigorous testing on historical data and blockchain metrics, we identify optimal configurations for cryptocurrency market forecasting.
Methodological Framework for Cryptocurrency Forecasting
Neural Network Architectures
The experimental design incorporated two fundamental ANN types:
- Feedforward Neural Network (FNN): Three-layer structure with log-sigmoid activation functions
- Nonlinear Autoregressive (NAR) Network: Recurrent connections with delayed feedback loops
NAR configurations demonstrated superior temporal modeling through hidden state equations:
\( h_t = \sigma(W_{hh}h_{t-1} + W_{xh}x_t) \)
\( \hat{y}_t = W_{hy}h_t \)
Optimization Algorithms
| Algorithm | Memory Complexity | Convergence Rate | Hessian Handling |
|---|---|---|---|
| Levenberg-Marquardt | O(n²) | Quadratic | Exact Approximation |
| Scaled Conjugate Gradient | O(n) | Superlinear | Numerical Approximation |
Data Infrastructure and Preprocessing
The dataset incorporated critical blockchain metrics:
- Daily closing prices (2012-2018)
- Network hash rate (EH/s)
- BTC transaction volume
Stationarity transformation through differencing:
\( \Delta P_t = P_t – P_{t-1} \)
Performance Evaluation
| Metric | NAR-LM | FNN-SCG | Improvement |
|---|---|---|---|
| MAE | 65.88 | 89.42 | 26.3% |
| RMSE | 202.34 | 254.71 | 20.5% |
| Training Time | 47min | 32min | -31.9% |
Critical Insights
- LM optimization achieved 23% faster convergence than SCG
- NAR networks showed 32% better MASE during high volatility
- Memory requirements favored SCG for large datasets
Future Research Directions
- Transformer networks for long-range dependency modeling
- Hybrid CNN-LSTM architectures
- Explainable AI integration for regulatory compliance
Key Takeaways
- NAR-LM configuration achieved best accuracy (RMSE 202.34)
- Recurrent networks outperform feedforward models in temporal modeling
- Algorithm choice depends on hardware constraints and volatility regimes
- Blockchain metrics improve prediction accuracy by 18-22%