How Darwin Lab Works

1. Genetic Evolution (the engine room)

Most trading bots run one hand-coded strategy. We don't. We run a population of 2800+ strategies that live, mutate, mate, and die in an endless loop — exactly like Darwin's natural selection, applied to trading.

• Population: 35 strategies per mode × 4 evolution modes = generalists, regime specialists, liquid, pair specialists
• Genes: timeframe, direction, leverage, indicators (BBands, RSI, ADX, MACD, FVG, divergence, breakout, Supertrend, StochRSI, Choppiness), SL/TP multipliers, filters
• Fitness function: Sharpe ratio 0.30 + Profit Factor 0.20 + expectancy 0.25 + win rate 0.15 + drawdown 0.10
• Anti-overfit: k-fold walk-forward (5 folds, must profit 3/5) + stability penalty + complexity penalty + 30% niche forcing
• Cycle: new generation every 4 hours, 60 to 90 days of historical data

2. Paper Arena (forward-only proving ground)

Backtests lie. Forward-tests don't. Every champion that survives evolution enters a paper arena that runs every 2 minutes on rotating pairs. Only forward performance decides who promotes to the live roster — never backtest numbers.

• Minimum 15 arena trades before consideration
• Win rate ≥ 25%, PnL > 0, Profit Factor ≥ 1.2
• Auto-demotion: WR < 20% on last 50 trades, or negative PnL on last 30
• Stale champions removed (tiered by timeframe: 1m=2d, 15m=5d, 4h=14d)
• Max roster: 10 — with quotas for snipers (WR ≥ 50%) and home-run hitters (PF ≥ 2.0)

3. Confidence Engine (XGBoost meta-model)

Even the best strategy can be wrong in the wrong context. Our confidence engine is an XGBoost classifier trained on 48 features that grades every signal 0-100 before it ever reaches your Telegram.

• ~92% backtest accuracy predicting whether a signal will profit — with the honest caveat that backtest overstates: live, the score's correlation to outcomes has been close to flat, so we lean on the regime and fee gates, not the confidence number alone
• Top features: has_trend, atr_pct, session win-rate interactions, volatility state
• Gate: signals below 45% confidence are killed, never broadcast
• Retrained weekly (Sunday 05:00 UTC) with automatic rollback if calibration degrades

4. Regime Awareness

Markets don't behave the same in bull, bear, and crash. Our classifier runs daily at 01:00 UTC and sorts the market into 7 regimes — execution adapts accordingly:

5. Risk Management

• Hardware SL: every position has a real Binance stop-market order (not a software stop)
• Kelly ÷ 4 sizing: fractional Kelly with drawdown guard
• 3-layer loss shield: max loss cap, auto pair blacklist, time-decay SL
• Position limits: 10 total, 3 per pair (5 for BTC/ETH/SOL), 5-min cooldown
• Correlation-adjusted Kelly: 0.85x / 0.70x / 0.55x penalty per same-category same-direction position
• Kill switch: daily drawdown triggers full halt
• Funding killer: no entries ±4 min around 00/08/16 UTC settlements

6. Signal Delivery

You receive signals on Telegram in < 1 second after the engine decides. Every signal contains:

• Direction (long/short), pair, entry, SL, TP1, TP2
• Grade 0-10: 💎 PRECISION (≥8), ✅ SOLID (≥6), ⚡ SPECULATIVE (≥4), ⚠️ EXPERIMENTAL
• 🔥 DEGEN PLAY tag when SL width exceeds 15% raw — reduce your size
• Percentages shown leveraged and raw — no ambiguity
• Strategy ID for traceability and cause tags (adx_strong, div_bull, fvg_bull, htf_confirmed, etc.)

FAQ