Skip to main content

Strategy Evolution

Understanding how Gopher discovers and evolves trading strategies.

Overview​

Gopher uses an AI-driven evolutionary approach to discover profitable trading strategies. Instead of manually designing rules, the AI:

  1. Generates trading strategies in natural language
  2. Backtests them against historical market data
  3. Analyzes performance and identifies weaknesses
  4. Evolves strategies based on feedback
  5. Repeats until finding high-performing strategies

The Evolution Loop​

┌─────────────────────────────────────────────────────────────┐
│                   Strategy Evolution Loop                    │
│                                                              │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐           │
│   │ Generate │────▶│ Backtest │────▶│ Analyze  │           │
│   │ Strategy │     │ Strategy │     │ Results  │           │
│   └──────────┘     └──────────┘     └──────────┘           │
│        ▲                                 │                  │
│        │           ┌──────────┐          │                  │
│        └───────────│  Evolve  │◀─────────┘                  │
│                    │ Strategy │                             │
│                    └──────────┘                             │
└─────────────────────────────────────────────────────────────┘

Phase 1: Strategy Generation​

The AI generates trading strategies in natural language. For example:

"Follow momentum breakouts. Enter long when price breaks above
the 20-period high with RSI above 50 and increasing volume.
Use a 2% stop loss from entry and target 3:1 reward-to-risk."

What the AI Considers​

  • Market conditions: Trending, ranging, volatile
  • Technical indicators: RSI, MACD, Bollinger Bands, moving averages
  • Volume patterns: Confirmation signals, divergences
  • Support/resistance: Key price levels
  • Recent price action: Candlestick patterns, momentum

Configure which indicators are active per run in the UI or via config:

Phase 2: Backtesting​

Each strategy is tested against historical market data:

  1. Load data: Fetch candles for selected timeframes
  2. Warm-up: Calculate indicators (requires ~20 candles)
  3. Decision loop: At each evaluation interval:
    • Build context with market data
    • AI decides: long, short, close_long, close_short, or wait
    • Execute trade in simulated wallet
  4. Track performance: Record PnL, win rate, drawdown, Sharpe

Example Backtest Sequence​

Evaluation 1 (2025-01-01 09:00):
  Market: BTC @ $95,000, RSI=45, trending up
  Decision: WAIT (waiting for momentum confirmation)

Evaluation 2 (2025-01-01 10:00):
  Market: BTC @ $95,500, RSI=55, breakout forming
  Decision: LONG @ $95,500, SL=$93,590, TP=$101,230

Evaluation 3 (2025-01-01 11:00):
  Market: BTC @ $96,200, in profit
  Decision: WAIT (holding position)

... continues until end date or TP/SL hit

Phase 3: Analysis​

After backtesting, the AI analyzes the results:

Metrics Evaluated​

MetricDescription
PnL %Total return percentage
Win RatePercentage of winning trades
Sharpe RatioRisk-adjusted return
Max DrawdownLargest peak-to-trough decline
Trade CountNumber of completed trades

Performance Feedback​

The AI receives structured feedback:

Backtest Results:
- PnL: +15.3%
- Win Rate: 58%
- Trades: 24
- Sharpe: 1.8
- Max Drawdown: -8.2%

Analysis: Strategy performs well in trending markets but
struggles during consolidation periods. Consider adding a
filter to avoid ranging markets.

Phase 4: Evolution​

Based on analysis, the AI evolves the strategy using different approaches:

Refinement​

Adjust existing parameters:

  • Tighten or widen stop losses
  • Modify indicator periods
  • Change entry/exit thresholds

Extension​

Add new conditions:

  • Volume confirmation
  • Additional timeframe alignment
  • Market regime filters

Pivot​

Try a different approach:

  • Switch from trend-following to mean reversion
  • Change from breakout entries to pullback entries
  • Explore different indicator combinations

Example Evolution Sequence​

Iteration 1:

"Enter long on RSI oversold bounce"
→ PnL: +5.2%, many false signals

Iteration 2 (Refined):

"Enter long on RSI oversold bounce WITH volume spike"
→ PnL: +12.1%, fewer but better signals

Iteration 3 (Extended):

"Enter long on RSI oversold bounce with volume spike,
confirmed by bullish divergence on 4H timeframe"
→ PnL: +18.7%, high-quality setups

Multi-Timeframe Analysis​

Gopher analyzes multiple timeframes simultaneously for better signals:

┌─────────────────────────────────────────────────┐
│              Multi-Timeframe Stack               │
├─────────────────────────────────────────────────┤
│ 4H:  Trend direction (primary filter)           │
│ 1H:  Entry setup identification                 │
│ 15m: Entry timing (precision)                   │
└─────────────────────────────────────────────────┘

This allows strategies like:

  • "Only take long trades when 4H trend is bullish"
  • "Enter on 15m signals that align with 1H structure"
  • "Use higher timeframe for trend, lower for entry"

When Evolution Stops​

Evolution continues until one of:

  • Max iterations reached: Your configured limit
  • User intervention: You click Stop
  • Convergence: AI determines diminishing improvements

Best Strategy Selection​

The best strategy is determined by a composite score:

Score = (PnL × 0.4) + (Sharpe × 0.3) + (WinRate × 0.2) + (1/Drawdown × 0.1)

This balances:

  • Raw profitability (PnL)
  • Risk-adjusted returns (Sharpe)
  • Consistency (Win Rate)
  • Capital preservation (low Drawdown)

Monte Carlo Validation​

Top strategies undergo statistical validation:

  1. Run N simulations (typically 100)
  2. Introduce variation: Slight timing differences, slippage
  3. Calculate statistics:
    • Mean PnL across all runs
    • Standard deviation
    • 95% confidence interval
    • Probability of profit

Validation Criteria​

A strategy is considered robust if:

  • Mean PnL > 0
  • 95% confidence interval lower bound > -5%
  • Probability of profit > 60%

Tips for Better Evolution​

1. Data Quality​

  • Use assets with high coverage (>90%)
  • Include periods with different market conditions
  • Longer date ranges = more reliable results

2. Configuration​

  • Start with multiple timeframes (15m + 1h + 4h)
  • Use reasonable risk settings (2% per trade)
  • Allow enough iterations (50+) for meaningful exploration

3. Strategy Prompts​

If providing a starting prompt:

  • Be specific about conditions
  • Include both entry AND exit rules
  • Define risk management parameters

4. Patience​

  • Evolution takes time
  • Early iterations often underperform
  • Best results typically appear mid-to-late in evolution
  • Don't stop too early!

Understanding Results​

Good Signs​

  • Steadily improving PnL across iterations
  • Sharpe ratio > 1.5
  • Win rate 50-65% with good R:R
  • Reasonable trade frequency (not too many, not too few)

Warning Signs​

  • Perfect results (likely overfitting)
  • Very high trade count (overtrading)
  • High win rate but negative PnL (poor risk management)
  • Inconsistent results across Monte Carlo runs

Next Steps​