AI Strategy Optimizer

Automated trading strategy generation system using LLM-powered analysis and backtest-kit framework.

Purpose

Demonstrates AI-driven strategy optimization capabilities for:

  • Multi-timeframe market data analysis (1h, 30m, 15m, 1m candles)
  • LLM-based pattern recognition and strategy generation
  • Automated strategy comparison and validation
  • Data-driven trading decision recommendations

Key Features

  • Multi-Timeframe Analysis: 4 data sources with different granularity levels

    • Long-term (1h candles): 48h lookback with Fibonacci levels
    • Swing-term (30m candles): 96 candles with volume/volatility analysis
    • Short-term (15m candles): Fast indicators and ROC metrics
    • Micro-term (1m candles): 60 candles with squeeze momentum and pressure index

  • LLM Integration: Ollama-powered strategy generation (gpt-oss:20b)

  • Training/Testing Split: 7-day training period → 1-day validation

  • Walker Framework: Automatic strategy comparison and ranking

  • Progress Monitoring: Real-time optimization tracking

  • Rich Indicators: 50+ technical indicators across all timeframes

    • RSI, Stochastic RSI, MACD, ADX, ATR, CCI
    • Bollinger Bands, Moving Averages (SMA, EMA, DEMA, WMA)
    • Support/Resistance levels, Momentum, ROC
    • Volume analysis, Volatility metrics

Technology Stack

  • Runtime: Node.js (ES Modules)
  • Framework: backtest-kit 1.4.3
  • AI Provider: Ollama (gpt-oss:20b model)
  • Utilities: functools-kit 1.0.93
  • Data Source: node-ccxt-dumper API

Project Structure

demo/optimization/
├── src/
│   └── index.mjs        # Main optimizer configuration
├── generated/           # AI-generated strategy code (output)
├── package.json         # Dependencies and scripts
└── README.md           # This file

Installation and Setup

# Navigate to project directory
cd demo/optimization

# Install dependencies
npm install

# Set environment variables
export OLLAMA_API_KEY=your_ollama_api_key

# Run optimizer
npm start

Configuration

Environment Variables

Create a .env file or set environment variables:

OLLAMA_API_KEY=your_ollama_api_key
CCXT_DUMPER_URL=node-ccxt-dumper-instance

Training Configuration

The optimizer is pre-configured for BTCUSDT with:

  • Training Period: November 24-30, 2025 (7 days)
  • Testing Period: December 1, 2025 (1 day)
  • Symbol: BTCUSDT
  • Data Source: node-ccxt-dumper

Data Sources

  1. long-term-range - 1h candles with Fibonacci analysis
  2. swing-term-range - 30m candles with volume/volatility
  3. short-term-range - 15m candles with ROC metrics
  4. micro-term-range - 1m candles with pressure index

Usage Examples

Basic Usage

Run the optimizer to generate strategies:

npm start

Output:

Progress: 14.285714285714286%
Progress: 28.571428571428573%
Progress: 42.857142857142854%
Progress: 57.14285714285714%
Progress: 71.42857142857143%
Progress: 85.71428571428571%
Progress: 100%

Generated file: ./generated/btc-optimizer_BTCUSDT.mjs

Running Generated Strategies

After generation, execute the strategy comparison:

node ./generated/btc-optimizer_BTCUSDT.mjs

This will:

  1. Initialize 7 strategies (one per training day)
  2. Run Walker comparison on test data
  3. Rank strategies by performance (Sharpe Ratio)
  4. Output best strategy statistics

Customizing Symbols

Modify src/index.mjs to analyze different cryptocurrencies:

await Optimizer.dump(
  "ETHUSDT",  // Change symbol
  {
    optimizerName: "btc-optimizer",
  },
  "./generated"
);

Adjusting Time Ranges

Edit training/testing periods in src/index.mjs:

const TRAIN_RANGE = [
  {
    note: "Custom period 1",
    startDate: new Date("2025-01-01T00:00:00Z"),
    endDate: new Date("2025-01-01T23:59:59Z"),
  },
  // Add more training periods...
];

const TEST_RANGE = {
  note: "Validation period",
  startDate: new Date("2025-01-08T00:00:00Z"),
  endDate: new Date("2025-01-08T23:59:59Z"),
};

How It Works

Phase 1: Data Collection

For each training period (7 days):

  1. Fetches data from 4 sources (long/swing/short/micro-term)
  2. Formats data with technical indicator descriptions
  3. Builds conversation context for LLM

Phase 2: Strategy Generation

For each training period:

  1. LLM analyzes historical patterns
  2. Generates trading strategy with entry/exit rules
  3. Creates fundamental analysis with market recommendations
  4. Stores strategy prompt and configuration

Phase 3: Code Generation

  1. Generates complete executable .mjs file
  2. Includes:
    • Exchange configuration (Binance via CCXT)
    • Frame definitions (training + testing periods)
    • 7 strategy implementations
    • Walker setup for comparison
    • Event listeners for progress tracking

Phase 4: Validation

Generated code runs Walker to:

  1. Execute all strategies on test data
  2. Calculate performance metrics (Sharpe Ratio, Win Rate, PNL)
  3. Rank strategies by selected metric
  4. Output best-performing strategy

LLM Prompt Engineering

The system uses strategic prompting for better strategy generation:

// System prompt
"В ответ напиши торговую стратегию где нет ничего лишнего,
только отчёт готовый для копипасты целиком

**ВАЖНО**: Не здоровайся, не говори что делаешь - только отчёт!"

// User prompt
`На каких условиях мне купить ${symbol}?
Дай анализ рынка на основе поддержки/сопротивления, точек входа в LONG/SHORT позиции.
Какой RR ставить для позиций?
Предпочтительны LONG или SHORT позиции?

Сделай не сухой технический, а фундаментальный анализ,
содержащий стратигическую рекомендацию, например, покупать на низу боковика`

This encourages:

  • Concise, actionable strategies
  • Support/resistance analysis
  • Risk/reward ratios
  • Long/short position preferences
  • Fundamental (not just technical) analysis

Performance Metrics

Generated strategies are evaluated by:

  • Sharpe Ratio: Risk-adjusted returns
  • Win Rate: Percentage of profitable trades
  • Average PNL: Mean profit/loss per trade
  • Total PNL: Cumulative profit/loss
  • Certainty Ratio: avgWin / |avgLoss|
  • Max Drawdown: Largest peak-to-trough decline
  • Expected Yearly Returns: Annualized profit estimate

Economic Benefits

  • Automated Strategy Development: LLM generates strategies from raw data
  • Data-Driven Decisions: 50+ indicators across 4 timeframes
  • Backtesting Validation: Historical performance verification
  • Strategy Diversity: 7 different approaches per optimization run
  • Time Savings: Minutes vs. days of manual strategy development
  • Reproducibility: Deterministic code generation from data

Advanced Customization

Adding New Data Sources

SOURCE_LIST.push({
  name: "custom-indicators",
  fetch: async ({ symbol, startDate, endDate, limit, offset }) => {
    // Fetch custom data
    return data.rows;
  },
  user: (symbol, data) => str.newline(
    "=== CUSTOM INDICATORS ===",
    JSON.stringify(data)
  ),
  assistant: () => "Custom data received"
});

Changing LLM Model

const response = await ollama.chat({
  model: "llama3:70b",  // Use different model
  messages: [...]
});

Custom Optimizer Callbacks

addOptimizer({
  optimizerName: "btc-optimizer",
  // ... existing config
  callbacks: {
    onSourceData: (symbol, sourceName, data) => {
      console.log(`✓ Fetched ${data.length} rows from ${sourceName}`);
    },
    onData: (symbol, strategies) => {
      console.log(`✓ Generated ${strategies.length} strategies`);
    },
    onCode: (symbol, code) => {
      console.log(`✓ Code generated: ${code.length} bytes`);
    }
  }
});

Related Projects

License

MIT © tripolskypetr

Settings

Member Visibility

On This Page

AI Strategy Optimizer