Abstract
As a highly volatile cryptocurrency characterized by extreme nonlinearity and non-stationarity, Bitcoin price prediction remains a critical focus for investors and researchers. This study constructs an LSTM model to predict Bitcoin prices using six key indicators, identifying significant lag effects. To address this, a CNN model is integrated to extract deep-layer features, leveraging its superior dynamic capture capabilities. However, vertical errors persist in CNN predictions. The final CNN-LSTM hybrid model demonstrates enhanced accuracy, outperforming standalone models in overall trend alignment, local detail prediction, and lag reduction.
Keywords
- Bitcoin
- LSTM (Long Short-Term Memory)
- CNN (Convolutional Neural Network)
- Price Prediction
- Hybrid Model
- Cryptocurrency
- Deep Learning
Introduction
With evolving investment landscapes, Bitcoin's nonlinear and non-stationary nature demands advanced predictive tools. Deep learning approaches, particularly LSTM and CNN, have shown promise in financial time-series analysis. This paper evaluates individual LSTM and CNN models before proposing a combined CNN-LSTM framework to mitigate their respective limitations.
Methodology
1. LSTM Model
Structure:
- Input Layer: 6 indicators (RSI14, DIFF, DEA, MACD, Up20, Down20).
- Hidden Layers: 3 LSTM layers + 3 Dense layers + 5 Dropout layers.
- Evaluation: MAPE = 10.14.
Limitations:
- Observed lag in price tracking (Figure 3).
- Residual volatility due to Bitcoin's non-smoothness (Figure 4).
2. CNN Model
Structure:
- Input Matrix: 14-day sliding window + feature rates (e.g., ΔRSI).
- Layers: 6 2D-Conv + 6 Avg-Pooling + 3 Dense + 2 Dropout.
- Evaluation: MAPE = 9.29.
Advantages:
- Better dynamic feature capture.
Drawbacks: - Vertical prediction errors (Figure 6).
3. CNN-LSTM Hybrid Model
Integration:
- Weighted outputs (α = 0.1 for CNN, β = 0.9 for LSTM).
- Combined loss minimization via RMSE.
Performance:
- MAPE = 4.74 (vs. 8.20 for LSTM, 7.09 for CNN).
- Improved trend alignment and local detail (Figure 14).
Results
- Hybrid Model: Reduces lag and vertical errors (Figures 14–15).
- Key Insight: CNN complements LSTM by extracting hierarchical features, while LSTM refines temporal trends.
FAQs
Q1: Why combine CNN and LSTM for Bitcoin prediction?
A: CNNs excel at feature extraction from spatial data, while LSTMs model temporal sequences. The hybrid leverages both to address nonlinearity and lag.
Q2: What metrics evaluate model performance?
A: Mean Absolute Percentage Error (MAPE) is primary. Lower values indicate higher accuracy (e.g., Hybrid: 4.74 vs. LSTM: 8.20).
Q3: How is data preprocessed?
A: Time-series data is stacked with a 14-day sliding window, and outlier-prone initial samples are removed.
Q4: What are the limitations of this study?
A: Extreme Bitcoin volatility still causes residuals (e.g., Figure 15). Future work may incorporate sentiment analysis.
👉 Explore advanced crypto analytics
Conclusion
The CNN-LSTM hybrid effectively merges CNN’s feature-depth with LSTM’s sequential modeling, achieving a 42% lower MAPE than standalone models. This approach sets a benchmark for volatile asset prediction.