Abstract
Ethereum has gained recognition as a prominent investment asset due to recent volatility in cryptocurrency markets. Accurate price trend prediction is critical for informed decision-making. This study introduces a novel hybrid methodology combining:
- Random Forest Classifier with ReliefF feature selection to identify top predictive features.
- Short-Time Fourier Transform (STFT) for time-frequency analysis of price patterns.
- Adaptive Neuro-Fuzzy Inference System (ANFIS) to model nonlinear relationships.
The model achieves 76.56% accuracy in predicting daily Ethereum price movements, outperforming traditional approaches (LSTM, XGBoost, etc.). Key innovations include STFT-based spectral feature extraction and interpretable fuzzy logic rules, addressing gaps in cryptocurrency research.
Introduction
Background
Cryptocurrencies like Ethereum exhibit extreme volatility due to decentralized markets, speculative trading, and evolving regulations. While blockchain technology enables transparency, price forecasting remains challenging.
Research Gap
Prior works rely heavily on:
- Statistical models (ARIMA, regression) – Limited nonlinear modeling.
- Machine learning (SVMs, Random Forests) – Black-box interpretations.
- Deep learning (LSTMs) – Computationally expensive.
Our Contribution
- STFT-ANFIS Fusion: Enhances pattern detection in time-frequency domains.
Dual Feature Selection: Combines Random Forest and ReliefF to identify:
- Trading Volume
- Accumulation/Distribution Oscillator
- Williams %R
- Stochastic %K
- Commodity Channel Index (CCI)
- Interpretability: Fuzzy logic provides actionable insights.
Data & Methodology
Variables Analyzed
| Category | Metrics | Source |
|------------------|----------------------------------|----------------------|
| Price Data | Open/High/Low/Close | Kaggle |
| Volume | Daily Trading Volume | Ethereum Blockchain |
| Momentum | 9-Day Price Change | Calculated |
| Sentiment | Google Trends ("Ethereum" searches) | Google Trends |
STFT-ANFIS Pipeline
Preprocessing:
- Apply STFT to isolate cyclical trends.
- Standardize features (Z-score normalization).
ANFIS Architecture:
- Input Layer: 5 selected features.
- Fuzzification: Gaussian membership functions.
- Rule Layer: 10 fuzzy IF-THEN rules.
- Output: Price direction (↑/↓).
Key Results
Performance Metrics
| Dataset | Accuracy | Precision | Recall | F1-Score |
|-----------|----------|-----------|--------|----------|
| Training | 78.54% | 78.18% | 73.40% | 76.99% |
| Test | 76.56% | 76.47% | 73.86% | 75.14% |
Comparative Analysis
👉 Why STFT-ANFIS outperforms LSTM/XGBoost
| Model | Test Accuracy | ROC-AUC |
|------------------|---------------|---------|
| ANFIS-STFT | 76.56% | 0.7601 |
| LSTM | 71.20% | 0.7023 |
| XGBoost | 74.10% | 0.7105 |
FAQs
Q: How does STFT improve price prediction?
A: STFT reveals hidden periodicities (e.g., weekly/monthly cycles) missed by time-domain models.
Q: Which features matter most?
A: Trading volume and momentum indicators (Williams %R, CCI) are top predictors per ReliefF weighting.
Q: Can this model adapt to sudden market shocks?
A: Yes. ANFIS’s fuzzy rules dynamically adjust to volatility spikes, unlike static statistical models.
Conclusion
This study demonstrates:
- Feature Selection Matters: Volume + 4 technical indicators drive 76%+ accuracy.
- Time-Frequency Analysis: STFT uncovers latent cyclical patterns.
- Practical Applications: Useful for traders, hedge funds, and risk managers.
Future work may explore:
- Real-time deployment optimizations.
- Extension to other cryptocurrencies.
👉 Explore crypto analytics tools
By Superx | Chengdu