A collection of machine learning projects for financial market analysis, covering time-series forecasting, clustering, reinforcement learning, and dynamic pricing strategies.
Hybrid deep learning model combining LSTM with Residual Network architecture for Bitcoin price prediction.
Architecture:
- Two LSTM layers with dropout regularization
- TimeDistributed Dense layer applied to each time step
- ResNet-style residual connections via custom ResidualWrapper
Results:
| Metric | Value |
|---|---|
| MAE | 0.015 |
| MSE | 0.0005 |
| RMSE | 0.023 |
Key Features:
- Predicts price changes rather than absolute prices for robustness
- 6-day lookback window predicting 6-day forward changes
- Includes swing trading simulation with SMA-30 indicator
- Compared against CNN-Transformers, vanilla LSTM, regression, and ConvNet variants
Unsupervised clustering of cryptocurrencies using K-Means based on market data attributes.
Pipeline:
- Data retrieval via CoinGecko API
- Preprocessing and feature engineering
- K-Means clustering with PCA for visualization
Results:
- Silhouette Score: 0.914
Transformer-based architecture for financial time-series forecasting.
Branches:
focal-loss-gpu-opts- Focal loss implementation with GPU optimizationsgtn-transformer-variant- Gated Transformer Network variant
Reinforcement learning agent for automated trading strategies.
Work in progress
Random Forest Regression model for predicting optimal pricing based on market conditions.
Features:
- Number of riders/drivers
- Vehicle type
- Expected ride duration
Pipeline:
- Categorical encoding and outlier detection
- EDA with Plotly visualizations
- Random Forest training and evaluation
- Deep Learning: PyTorch, TensorFlow/Keras
- ML: scikit-learn, XGBoost
- Data: Pandas, NumPy
- Visualization: Matplotlib, Seaborn, Plotly
- APIs: CoinGecko
For detailed findings and proprietary implementations, contact: masihmoafi12@gmail.com