Solana Price Prediction - TypeScript Version

A TypeScript implementation of the Solana price prediction model with enhanced stacking, walk-forward validation, and hybrid windows.

Features

• ✅ Enhanced Stacking Method: Combines multiple base models with a neural network meta-learner
• ✅ Walk-Forward Validation: Tests models on multiple time periods without data leakage
• ✅ Automatic Feature Engineering: Creates technical indicators, lagged features, and financial metrics
• ✅ Hybrid Windows: Rolling window (short-term) + Expanding window (long-term patterns)
• ✅ TEMPO-Based Improvements (NEW):
  • STL Decomposition: Separates time series into Trend, Seasonal, and Residual components
  • Component-Specific Models: Separate models for each component with additive combination
  • Soft Prompts: Learnable prompts for component-specific guidance
  • Time Series Patching: Efficient patch-based processing
• ✅ Modular Architecture: Clean separation of concerns

Project Structure

meth/
├── src/
│   ├── data/
│   │   └── DataLoader.ts          # Data loading and preprocessing
│   ├── features/
│   │   └── FeatureEngineer.ts     # Automatic feature engineering
│   ├── models/
│   │   ├── BaseModel.ts           # Base model interface and implementations
│   │   └── EnsembleStacker.ts    # Enhanced stacking ensemble
│   ├── validation/
│   │   └── WalkForwardValidator.ts # Walk-forward validation
│   ├── training/
│   │   └── Trainer.ts             # Main training module
│   ├── utils/
│   │   └── RollingWindow.ts       # Rolling window utilities
│   └── scripts/
│       ├── train.ts               # Training script
│       └── predict.ts             # Prediction script
├── package.json
├── tsconfig.json
├── merged_dataset.csv             # Required data file
└── README.md

Installation

npm install
npm run build

Quick Start

1. Install dependencies:

   npm install
   npm run build

2. Train the model:

   # Standard training
   npm run train
   
   # Enhanced training (recommended)
   npm run train -- --enhanced
   
   # Return-based (best for volatile crypto)
   npm run train -- --enhanced --returns

3. View results:

   • Check console output for metrics
   • Open validation_results.html for visualizations

📖 See USER_GUIDE.md for complete step-by-step guide!

Usage

Training Options

Standard Training (Original Method):

npm run train

• Uses ensemble of Neural Network + Linear Regression
• Walk-forward validation
• Good baseline for comparison

Enhanced Training (TEMPO-based):

npm run train -- --enhanced

• STL decomposition (Trend, Seasonal, Residual)
• Historical decomposition patterns as features
• More sophisticated feature engineering

Return-Based Training (Recommended for Crypto):

npm run train -- --enhanced --returns

• Predicts returns (percentage change) instead of prices
• More stable for volatile assets
• Often more accurate: predicted_price = current_price * (1 + predicted_return)

See TEMPO_IMPROVEMENTS.md for technical details.

Making Predictions

npm run predict
# or
node dist/scripts/predict.js

Architecture

Enhanced Stacking Method

The stacking ensemble uses:

• Base Models: Multiple models trained on different windows (rolling + expanding)
• Meta-Learner: Neural network that learns to combine base model predictions
• Out-of-Fold Predictions: Prevents overfitting by using validation set predictions

Walk-Forward Validation

• Tests models on multiple time periods
• No data leakage (future data never used for training)
• Provides realistic performance estimates

Hybrid Windows

• Rolling Window: Last N days (captures short-term trends)
• Expanding Window: All historical data (captures long-term patterns)
• Both windows are used to train separate models, then combined via stacking

Automatic Feature Engineering

Creates:

• Technical indicators (RSI, MACD, Bollinger Bands)
• Volatility and Sharpe ratio
• Moving averages
• Lagged features
• Momentum indicators
• Cross-asset ratios

Model Types

• Neural Network: Deep learning model with dropout and regularization
• Linear Regression: Simple but effective baseline model

Configuration

Edit src/scripts/train.ts to adjust training parameters:

• rollingWindow: Size of rolling window (default: 365 days)
• stepSize: Step size for walk-forward validation (default: 30 days)
• minTrainSize: Minimum training set size (default: 180 days)
• testSize: Percentage of data for testing (default: 0.2)

Notes

This is a TypeScript implementation that demonstrates the architecture. For production use, you may want to:

• Add model serialization/deserialization
• Implement more sophisticated base models
• Add hyperparameter tuning
• Implement real-time prediction API
• Add comprehensive error handling

License

MIT