Darwin Lab went live on Binance Futures mainnet on 2026-04-10. This is what 30 days actually looks like — not the highlight reel, the full picture.
The Numbers
From /api/stats.json, as of 2026-04-30:
- 369 real Binance Futures trades
- 59% win rate
- ~95% annualized ROI on deployed capital
- Profit factor: 1.4× (gross wins / gross losses across all 369 trades)
- 5 open positions at time of writing
The profit factor of 1.4× is not spectacular. It is consistent. For context on what that means at scale: a system sustaining 1.4× profit factor over hundreds of trades is statistically significant — the breakeven threshold is 1.0×, and most retail algo systems degrade below it within weeks of going live.
Live capital allocation and current equity are accessible at /api/stats.json.
What the Numbers Hide
The ~95% annualized ROI figure required surviving several stretches that looked nothing like a smooth equity curve.
The most instructive was a 72-hour window in mid-April. The system had been running with entry parameters loosened (minimum confidence score 40%, minimum signal score 2, cooldown 3 minutes). Over those 72 hours, 241 fills were recorded and 10 out of 12 pairs ended negative. The drawdown over that window was approximately 26% of account equity at the time.
The fix was a direct rollback of the permissive entry settings: confidence back to 45%, signal score back to 3, cooldown back to 5 minutes. No new algorithm. No architectural change. Just tighter filters on what was already there.
This matters because it illustrates how a system can degrade silently. Win rate looked fine in aggregate. PnL was eroding in the tail. The bleed was only visible when you looked at per-pair performance over a 72-hour window, not the headline number.
How Signals Actually Get Generated
The short version: 3,119 strategies are competing in a continuous paper arena that runs every 2 minutes. Each strategy has roughly 30 genetic parameters — timeframe, direction, leverage, which indicators fire entry/exit signals, stop-loss multipliers.
The arena runs forward-only simulations on rotating pairs. When enough strategies agree on a signal, and their combined "authority" (a score weighted by forward win rate, profit factor, and trade count) crosses a threshold, the signal goes to the executor. The executor then runs it through a confidence gate (an XGBoost model trained on 48 features), a regime filter, a correlation guard, and a position-sizing layer before any order hits Binance.
The current live roster has 10 champion strategies. They were promoted from the arena based on forward performance, not backtest numbers. This distinction turns out to matter a lot — more on that in a separate post about one strategy that had 82% backtest win rate and 49% live win rate.
The Architecture in One Paragraph
Every 4 hours, a genetic algorithm runs a new generation of strategies. Each generation selects the fittest from the previous one, applies crossover and mutation, and produces a new population. Fitness is scored as: Sharpe ratio (0.30 weight) + profit factor (0.20) + expectancy (0.25) + win rate bonus (0.15) + drawdown penalty (0.10). To pass, a strategy must also survive 5-fold walk-forward validation (profiting in at least 3 of 5 folds). The ones that survive enter the arena. The arena decides who goes live.
There are currently 3,119 alive strategies and 471 killed ones. The killed ones are not deleted — their DNA is archived with the regime they traded in, in case a future market regime matches their original environment.
What the System Does Not Do
It does not predict price direction. It detects recurring market structures (Bollinger Band squeezes, RSI divergences, ATR-based breakouts, Fair Value Gaps) and sizes positions according to how often those structures have historically resolved in a particular direction, on a particular pair, in the current market regime.
It also does not run 24/7 without intervention. Since live deployment, there have been roughly 91 development sessions logged — each one either fixing something that broke, tuning a parameter that was creating noise, or adding a new guard rail. A few of those sessions were because of bugs. Most were because the market changed and something that worked in March did not work in April.
What the Next 30 Days Look Like
The system is currently in a weak_bear regime. The regime-aware sizing applies a 0.80x multiplier to position sizes on long entries. The kill switch threshold is calibrated to the real balance, not the config default. The loss shield will force-close any position losing more than 8% of account equity.
The full trade history is available at /track-record. The raw stats endpoint is /api/stats.json.
If you want to see signals when they fire — before this blog gets updated — the free Telegram channel posts every live entry: t.me/DarwinLabSignals.
Risk disclaimer: Trading futures involves substantial risk of loss. The returns described above represent actual past performance and do not guarantee future results. Full disclaimer.