Predicting cryptocurrency prices has become one of the most pressing challenges in modern financial analytics. Among digital assets, Ethereum stands out due to its robust ecosystem, smart contract functionality, and significant market influence. However, its high volatility demands advanced modeling techniques capable of capturing complex, nonlinear patterns. This article explores a novel hybrid framework that integrates machine learning with signal processing to enhance Ethereum price trend forecasting accuracy.
The proposed method combines Short-Time Fourier Transform (STFT) for frequency-domain feature extraction and the Adaptive Neuro-Fuzzy Inference System (ANFIS) for intelligent decision-making. Supported by a dual-stage feature selection process using Random Forest and ReliefF, this approach identifies the most influential predictors while improving model efficiency and robustness.
Core Keywords
- Ethereum price prediction
- Machine learning in finance
- Short-Time Fourier Transform (STFT)
- ANFIS model
- Feature selection
- Cryptocurrency trend forecasting
- Time-series analysis
- Financial signal processing
Introduction: The Need for Advanced Prediction Models
Why Ethereum?
Ethereum has evolved beyond a simple digital currency into a foundational platform for decentralized applications (dApps), DeFi protocols, and NFT marketplaces. Its widespread adoption has made it a focal point for investors and analysts alike. However, this popularity comes with extreme price fluctuations, making accurate trend prediction essential for informed trading and risk management.
Traditional models like ARIMA or linear regression often fail to capture the nonlinear dynamics of crypto markets. Even standard machine learning models such as Random Forest or LSTM may overlook hidden periodicities in price movements—patterns that exist not just in time but also in frequency domains.
👉 Discover how AI-powered analytics are transforming crypto trading strategies.
Bridging the Gap with Signal Processing
This study introduces an innovative solution: combining signal processing techniques with interpretable AI models. By applying STFT, we uncover cyclical behaviors in Ethereum’s price data—insights invisible to pure time-domain models. These enriched features are then fed into ANFIS, a neuro-fuzzy system that blends the learning power of neural networks with the interpretability of fuzzy logic.
Unlike black-box deep learning models, ANFIS allows us to understand why a prediction was made—critical for financial applications where transparency matters.
Data Collection and Variable Engineering
The dataset spans from April 1, 2018, to April 9, 2023, incorporating 16 technical, behavioral, and macroeconomic indicators. Sources include historical trading data from Kaggle, Google Trends search volumes, and Economic Policy Uncertainty (EPU) indices.
Key Input Variables
Price and Volume Metrics
- Open, High, Low, Adjusted Close Prices
- Trading Volume (standardized over 7-day rolling mean and std)
- Turnover Ratio (volume / closing price)
- Daily Return (log difference between consecutive days)
Market Sentiment & Momentum Indicators
- Google Trends: Search interest for "Ethereum"
- Momentum: Price change over 9 days
- Williams %R: Measures overbought/oversold conditions
- Stochastic %K: Compares closing price to recent trading range
- Commodity Channel Index (CCI): Detects deviations from average price behavior
Liquidity and Flow Analysis
- Accumulation/Distribution Oscillator: Tracks money flow in/out of the market
- Illiquidity Measure: Absolute return divided by volume (near zero throughout sample period)
Descriptive statistics reveal strong volatility across price-related variables. Correlation analysis shows:
- Near-perfect correlation between high, low, and adjusted close prices (same-day derivation)
- Moderate positive link between trading volume and Google Trends (public attention drives activity)
- Negative relationship between turnover ratio and price levels (higher prices → lower turnover)
These insights guide both preprocessing and feature selection.
Methodology: A Two-Stage Intelligent Framework
Step 1: Feature Selection Using Random Forest & ReliefF
To avoid overfitting and improve computational efficiency, we apply a two-phase feature selection strategy:
- Random Forest Classifier determines the optimal number of neighbors (
k=1) via 5-fold cross-validation, achieving a score of 0.7323—indicating strong predictive potential in volatile environments. - ReliefF Algorithm assigns weights to all candidate features based on their ability to distinguish upward vs. downward trends (defined by a 0.001 daily return threshold).
👉 Learn how data-driven models are reshaping crypto investment decisions.
The top five most influential features identified:
- Trading Volume – Reflects market participation intensity
- Accumulation/Distribution Oscillator – Reveals capital inflows/outflows
- Williams %R – Signals short-term reversal points
- Stochastic %K – Captures momentum shifts within recent ranges
- CCI (Commodity Channel Index) – Highlights trend strength and anomalies
These features form the core input set for the ANFIS model.
Step 2: Enhanced ANFIS Model with STFT Preprocessing
Signal Transformation via STFT
Short-Time Fourier Transform converts the time series into time-frequency representations:
- FFT size: 256 points
- Overlap: 128 points (50% overlap)
- Output: Spectral components revealing hidden cycles and noise patterns
Inverse STFT reconstructs a cleaner version of the original signal—removing high-frequency noise while preserving meaningful oscillations.
ANFIS Architecture Overview
The model follows a five-layer structure:
- Input Layer: Receives normalized feature vectors
- Fuzzification Layer: Applies Gaussian membership functions to convert crisp inputs into fuzzy values
- Rule Layer: Computes firing strength of each fuzzy rule (product of membership degrees)
- Defuzzification Layer: Generates weighted linear outputs per rule
- Output Layer: Aggregates results through summation and applies sigmoid activation for binary classification (up/down trend)
Each Gaussian function is parameterized by mean and variance—learned during training—to adaptively model uncertainty and nonlinearity.
Experimental Setup and Training Procedure
Data was split 80%/20% into training and testing sets using train_test_split (random_state=2023). All features were standardized using StandardScaler.
The ANFIS model was configured with:
- 10 fuzzy rules (optimized balance between complexity and performance)
- Weighted BCEWithLogitsLoss to handle class imbalance (pos_weight ≈ 1.087)
- AdamW optimizer with learning rate = 0.001 and L2 regularization
- 1000 epochs with performance tracking at each step
Model outputs were evaluated using:
- Accuracy, Precision, Recall, F1 Score
- Confusion Matrix and ROC-AUC
No missing values ensured data integrity throughout.
Results and Performance Evaluation
| Metric | Training Set | Test Set |
|---|---|---|
| Accuracy | 78.54% | 76.56% |
| Precision | 78.18% | 76.47% |
| Recall | 73.40% | 73.86% |
| F1 Score | 76.99% | 75.14% |
| Loss | 0.4581 | 0.5194 |
| ROC-AUC | – | ~0.76–0.77 |
Key observations:
- Minimal performance drop between train and test sets → strong generalization
- Balanced recall and precision → reliable detection of upward trends
- AUC > 0.76 indicates solid discriminatory power in highly chaotic markets
Visualizations confirm alignment between predicted and actual trends over a six-month window, despite frequent crossovers due to inherent volatility.
Comparative Analysis with Other Models
Compared to Gradient Boosting, LSTM, Random Forest, and XGBoost:
- XGBoost overfits: high training accuracy but significant test degradation
- ANFIS-STFT maintains consistent performance across datasets
- Highest test ROC-AUC (0.7601 vs. XGBoost’s 0.7105)
- Most balanced confusion matrix → superior handling of both classes
This demonstrates the advantage of integrating domain-aware signal processing with interpretable AI.
Frequently Asked Questions (FAQ)
Q: Why use STFT instead of traditional time-series methods?
A: STFT reveals hidden periodic patterns in price data that pure time-domain models miss. It enables detection of recurring cycles—such as weekly or monthly investor behaviors—by analyzing frequency components over time.
Q: How does ANFIS differ from standard neural networks?
A: ANFIS combines neural learning with fuzzy logic rules, offering better interpretability. You can trace predictions back to specific rules (e.g., “high volume + rising CCI → likely uptrend”), which is crucial in finance.
Q: What makes this model suitable for real-world trading?
A: Despite moderate accuracy (~76%), it consistently outperforms peers in generalization and robustness. Its resistance to overfitting and clear decision logic make it ideal for risk-sensitive applications.
Q: Can this model predict exact prices?
A: No—it predicts directional trends (up/down), not absolute prices. This binary classification suits tactical trading strategies focused on entry/exit timing.
Q: Is the model computationally efficient?
A: STFT and ANFIS training are moderately resource-intensive, limiting real-time deployment without optimization. Future work may explore GPU acceleration or cloud-based inference.
Q: Can this approach be applied to other cryptocurrencies?
A: Yes—the methodology is asset-agnostic. Early tests on Bitcoin and Binance Coin show promising transferability, especially when recalibrating thresholds and features.
Conclusion and Future Directions
This study presents a breakthrough in Ethereum price trend prediction by fusing signal processing (STFT) with an interpretable AI model (ANFIS), supported by rigorous feature selection. Achieving a test accuracy of 76.56% and stable recall across datasets, the model surpasses conventional machine learning approaches in both performance and transparency.
Key contributions:
- First application of STFT in Ethereum forecasting to extract time-frequency patterns
- Dual-stage feature selection enhances relevance and reduces noise
- ANFIS provides explainable predictions—valuable for institutional adoption
Practical implications span investment strategy development, algorithmic trading systems, and financial product innovation (e.g., ETH-linked derivatives).
Future research should focus on:
- Reducing computational load via parallelization or lightweight architectures
- Extending the framework to multi-class or multi-horizon forecasting
- Integrating on-chain metrics (e.g., gas fees, active addresses) for richer context
As crypto markets mature, hybrid models like ANFIS-STFT will play a pivotal role in turning chaotic data into actionable intelligence.
👉 Explore cutting-edge tools that bring predictive analytics to your fingertips.