Darwin evolves trading strategies.
It doesn't just prompt them.
A population of backtestable strategy specs, scored on a deterministic, lookahead-free backtester and bred across generations into a champion — evolved on live CoinMarketCap market data, then given an on-chain identity on BNB Chain.
Champion Dual-MA Trend Follower v2 - Regime Filtered × Dual-MA Trend Follower v2 - Regime Filter · 10966ca20775 · bred across 7 generations · 21 trades
The evolved champion
Not prompt-generated — bred. This strategy is the fittest spec the genetic algorithm produced after evaluating a full population on a deterministic backtester.
Dual-MA Trend Follower v2 - Regime Filtered × Dual-MA Trend Follower v2 - Regime Filter
Adds a 100-period EMA macro-regime filter and MACD confirmation to eliminate counter-trend entries during sustained bear phases, reducing drawdown and improving trade quality at the cost of fewer but higher-conviction signals.
EMA FASTcrosses aboveEMA SLOWCLOSE>EMA REGIMERSI14>45RSI14>58.309MACD>MACD SIG
EMA FASTcrosses belowEMA SLOWCLOSE<EMA REGIMERSI14>80MACDcrosses aboveMACD SIG
- Stop loss
- 8%
- Take profit
- 28%
- Max position
- 15%
- Trailing stop
- 0.7%
- Max positions
- 4
- Sizing
- 20% eq-wt
- Execution at next-bar open — no same-bar fills, no lookahead.
- PancakeSwap-realistic fees + slippage.
- Fitness penalizes over-trading, drawdown, and rule-fighting.
- 100% adherence to user risk rules.
Performance metrics
Forward-tested out-of-sample
The strategy was designed on the first half of history and forward-tested on a held-out window it never saw. On unseen data it preserved capital with a -1.0% max drawdown and 100% rule adherence — no overtrading, no blow-up. Most "AI strategy" results are 100% in-sample and can't show this at all.
- Return
- +13.0%
- Sharpe
- 2.31
- Max drawdown
- -1.1%
- Win rate
- 79%
- Trades
- 14
- Rule adherence
- 100%
- Return
- +0.1%
- Sharpe
- 0.05
- Max drawdown
- -1.0%
- Win rate
- 43%
- Trades
- 7
- Rule adherence
- 100%
The held-out window is never used to design or select the strategy — a true forward test, the same way Track 2 judges score against an unseen market window.
A population, bred across generations
Most "AI strategy" tools prompt an LLM once. Darwin evolves a population: select, crossover, mutate, repeat. Over 7 generations the GA refined a high-churn idea into a low-drawdown champion.
- 0Dual-MA Trend Follower+16.4%fitness 0.115
- 1Dual-MA Trend Follower v2 - Regime Filtered+3.1%fitness 0.177
- 2Dual-MA Trend Follower v2 - Regime Filter+3.6%fitness 0.195
- 3Dual-MA Trend Follower v2 - Regime Filtered+13.1%fitness 0.762
- 4Dual-MA Trend Follower v2 - Regime Filtered+15.1%fitness 0.773
- 5Dual-MA Trend Follower v2 - Regime Filtered+15.1%fitness 0.773
- 6Dual-MA Trend Follower v2 - Regime Filtered+13.1%fitness 0.851champion
Watch Darwin invent a strategy
Run the genetic algorithm's real LLM operator on the spot — it breeds a fresh, backtestable candidate from the champion, live.
Sentiment, not just price
The CoinMarketCap Fear & Greed Index feeds directly into the strategy rules.
Darwin treats CMC's Fear & Greed Index as a first-class signal — champions gate entries and exits on the sentiment regime, not just price.
A real on-chain identity
Darwin holds a verifiable identity on BNB Chain
Registered via the official BNB AI Agent SDK — a portable, on-chain identity the agent can carry across the BNB ecosystem. Not a mock: it's a real testnet transaction.
View registration tx on BscScan ↗Three sponsor integrations, wired end-to-end
CoinMarketCap for data and signal, the BNB AI Agent SDK for identity, and the Trust Wallet Agent Kit for execution.
CoinMarketCap
Live OHLCV, quotes, and the proprietary Fear & Greed Index — wired into the strategy DSL as a first-class signal.
BNB AI Agent SDK
Registers Darwin's verifiable ERC-8004 on-chain identity on BSC testnet — gas-sponsored, portable across the ecosystem.
Trust Wallet Agent Kit
Live HMAC-authenticated gateway client with correct request signing and live BSC asset resolution for PancakeSwap routes.