stratevo 6.0.4

AI-native quantitative finance engine with genetic algorithm strategy evolution

Stop writing trading strategies. Evolve them.

An open-source genetic algorithm engine that evolves and walk-forward tests trading strategies across 484 market factors. No manual rules, no API keys, no PhD required.

Demo · Results · 📊 Live Performance · Docker · Styles · Ensemble · Robustness · Contribute · Discord

---

30-Second Demo

pip install stratevo
stratevo analyze AAPL          # Quick analysis with evolved strategies
stratevo evolve --quick        # See evolution in ~2 minutes

That's it. No config files, no API keys, no exchange accounts, no LLM required. Pure Python + math.

_{Real evolution run: 9 crypto assets, 600 generations. Fitness goes from random noise to 33% annual return, 6.06 Sharpe, 5.6% max drawdown. (Backtest only, not live trading)}

---

Docker

Zero-setup. No Python version issues, no dependency conflicts.

# See evolution in 30 seconds (built-in demo data)
docker compose up demo

# Run a quick evolution (~2 minutes)
docker compose up quick

# Full crypto evolution: auto-downloads data + evolves
docker compose up evolve-crypto

# Analyze any symbol
docker compose run --rm analyze analyze BTC

Manual install (without Docker)

pip install stratevo
stratevo evolve --demo            # 30-second demo
stratevo evolve --quick           # ~2 minute evolution
stratevo download-crypto          # Download BTC/ETH/SOL data
stratevo evolve --market crypto   # Full evolution run

---

What's Different Here

Most quant tools make you write the strategy. You pick indicators, set thresholds, run a backtest, tweak, repeat. It works — if you already know what to look for.

StratEvo takes the opposite approach: it evolves strategies. A genetic algorithm searches across 484 factors and finds weight combinations that actually work out-of-sample.

You don't pick the indicators. The algorithm does. You decide how long to let it search.

You write the rules        →  StratEvo discovers the rules
You tune parameters        →  GA tunes parameters  
You test on one period     →  Walk-forward tests on multiple windows
You hope it generalizes    →  Monte Carlo measures if it does

Why Genetic Algorithms?

GAs are good at problems where the search space is huge and the relationship between inputs isn't obvious — which is exactly what trading strategy design looks like. With 484 factors, the number of weight combinations is absurd. Grid search won't cut it. Manual tuning takes forever. GAs get around this by keeping what works and mutating what doesn't, generation after generation.

---

Strategy Styles

When StratEvo evolves a strategy, the resulting "DNA" (a vector of 484 factor weights) falls into a recognizable style based on which factors dominate:

Style	Heavy Factors	What the Algorithm Found
Value Seeker	P/E, P/B, earnings yield	Buys cheap, holds patient. Benjamin Graham would recognize this DNA.
Momentum Rider	ROC, acceleration, trend strength	Chases what's moving, cuts what isn't. The DNA learned that trends persist.
Mean Reverter	Z-score, Keltner position, %B	Bets on bounce-backs. Statistical arbitrage without the PhD.
Flow Reader	OBV, smart money flow, Wyckoff VSA	Follows the money. Volume precedes price — the DNA figured that out.
Volatility Hunter	ATR squeeze, vol-of-vol	Profits from volatility expansion. Sits during calm, strikes during storms.
Crypto Native	Funding rate, liquidation cascade	200 factors built for a 24/7 market that doesn't follow TradFi rules.

These aren't made up — they're what the algorithm actually converged on.

---

Multi-DNA Voting

Instead of betting on a single evolved strategy, run multiple DNAs and aggregate their signals:

stratevo analyze AAPL --ensemble 5    # 5 strategies vote on one stock

  Multi-DNA Analysis: AAPL
  ═══════════════════════════════════════
  
  Strategy Votes:
    Value Seeker       → BUY   (score: 0.72, confidence: high)
    Momentum Rider     → BUY   (score: 0.85, confidence: high)  
    Mean Reverter      → HOLD  (score: 0.45, confidence: medium)
    Flow Reader        → BUY   (score: 0.68, confidence: medium)
    Volatility Hunter  → HOLD  (score: 0.51, confidence: low)

  Consensus: BUY (3/5 bullish, 0/5 bearish)
  Combined Score: 0.64
  
  This is a research tool, not financial advice.

Each strategy was independently evolved. Agreement = stronger signal. Disagreement = mixed picture, worth looking closer.

---

How It Works

  Random DNA population (484 factor weights + risk parameters)
       │
       ▼
  ┌──────────────────────┐
  │  Walk-Forward Test   │  Multi-window out-of-sample validation
  │  each DNA candidate  │  Real fees, slippage, position caps
  └──────────┬───────────┘
             │
             ▼
  Keep the survivors (Sharpe × Return / MaxDD)
             │
             ▼
  Mutate + Crossover → next generation
             │
             ▼
  Repeat for N generations

Each DNA is a weight vector across 484 factors plus risk/position parameters — all evolvable:

Parameter	Range	What it controls
Factor weights (×484)	0.0–1.0	Which factors matter and how much
hold_days	2–60	Day trades through swing trades
trailing_stop	%	Lock in profits, trail below peak
market_regime	sensitivity	Auto-reduce exposure in bear markets
kelly_fraction	0–1	Position sizing based on recent win rate

The GA tries combinations you wouldn't. That's the point.

stratevo evolve --market crypto --generations 50
stratevo evolve --pareto                          # Multi-objective (NSGA-III)
stratevo evolve --logic-guided                    # Semantically guided mutations

---

Anti-Overfitting

We learned the hard way. An early version showed 25,000% returns — it was a bug (look-ahead bias). Here's what prevents that now:

Defense	What it does
Walk-Forward	Multi-window OOS validation. Must profit on data it never trained on.
Monte Carlo	1,000 shuffled iterations. p-value < 0.05 or it's luck.
Arena Mode	Multiple strategies compete simultaneously. Crowded signals get penalized.
Correlation Guard	Rejects picks correlated above 0.85. No fake diversification.
Factor Diversity	HHI-based penalty for over-relying on one factor group.
Turnover Penalty	Excessive trading gets punished in fitness. Real transaction costs modeled.
Bias Detection	Look-ahead, snooping, survivorship — flagged automatically.

stratevo check-backtest    # Verify any result's integrity

Transparency note: Our early versions had inflated returns from look-ahead bias. We fixed it, disclosed it publicly, and added check-backtest so you can verify results yourself. Honest 45% beats fabricated 25,000%.

---

Real Results

Actual numbers from our running evolution engines. Not cherry-picked. Updated live.

🇺🇸 US Stocks (50 S&P 500 stocks, V3 — Gen 31)

Metric	Best DNA
Annual Return	45.0%
Sharpe Ratio	1.67
Max Drawdown	16.2%
Win Rate	59.2%
Profit Factor	1.63
Total Trades	232

Walk-Forward Validation — All 4 Windows Profitable ✅

Window	Period	OOS Annual Return	AAPL Buy&Hold	OOS Sharpe	Trades	Win Rate
W0	Dec 2022 – Jul 2023	+80.1%	+88.9%	2.36	55	60.0%
W1	Jul 2023 – Feb 2024	+24.1%	-8.9%	1.62	45	60.0%
W2	Feb 2024 – Sep 2024	+23.2%	+39.6%	0.88	69	56.5%
W3	Sep 2024 – Apr 2025	+60.5%	-18.0%	1.83	63	60.3%

_{All 4 windows OOS-profitable. W0 trailed AAPL buy-and-hold during the 2023 rally; W1 and W3 significantly outperformed in flat/down markets — where active strategies should add value.}

---

₿ Crypto (20 assets, V12 — Gen 108)

Metric	Best DNA
Annual Return	105.6%
Sharpe Ratio	2.89
Max Drawdown	14.8%
Win Rate	44.0%
Profit Factor	1.89
Total Trades	150

Walk-Forward Validation — 4 of 5 Windows Profitable

Window	Period	OOS Annual Return	BTC Buy&Hold	OOS Sharpe	Trades	Win Rate
W0	Jan 2025 – Mar 2025	-42.8%	-62.2%	-1.87	30	16.7%
W1	Mar 2025 – May 2025	+10.4%	+195.7%	0.84	30	43.3%
W2	May 2025 – Jul 2025	+98.2%	+111.0%	2.83	30	43.3%
W3	Jul 2025 – Sep 2025	+524.0%	+2.4%	6.50	30	53.3%
W4	Sep 2025 – Oct 2025	+370.3%	+20.5%	6.13	30	63.3%

_{W3/W4 massively outperformed BTC buy-and-hold. W1 missed the Mar–May 2025 rally (BTC +195% annualized vs strategy +10%). W0 lost money but less than BTC itself (-42.8% vs -62.2%). We're working on cross-regime robustness (GT-Score fitness function, arXiv:2602.00080).}

---

Caveats: These are backtests, not live trades. Real results will differ. We hide nothing.

---

Paper Trading

Test strategies with real market data, simulated execution:

stratevo paper start --exchange okx --balance 100000
stratevo paper status
stratevo paper report

Track P&L, win rates, and drawdowns before risking real money.

---

484 Factors

Click to expand full factor list

All normalized to [0, 1]:

Category	Count	Examples
Crypto-Native	200	Funding rate, session effects, whale detection, liquidation cascade
Momentum	14	ROC, acceleration, trend strength
Volume & Flow	13	OBV, smart money, Wyckoff VSA
Volatility	13	ATR, Bollinger squeeze, vol-of-vol
Mean Reversion	12	Z-score, Keltner channel position
Trend Following	14	ADX, EMA golden cross, MA fan
Qlib Alpha158	11	KMID, KSFT, CNTD (Microsoft Qlib compatible)
Risk Warning	11	Consecutive losses, death cross, gap-down
Quality Filter	11	Earnings momentum, relative strength
Price Structure	10	Candlestick patterns, support/resistance
Sentiment	2	EN/ZH keyword sentiment
DRL Signal	2	Q-learning buy probability

All factor weights are discovered by evolution — no manual tuning required.

---

Data Sources

Market	Source	API Key?
Crypto	ccxt (100+ exchanges)	No
US Stocks	Yahoo Finance	No
A-Shares	AKShare + BaoStock	No

---

MCP Server

StratEvo exposes its tools to any MCP-compatible AI client (Claude Desktop, Cursor, etc.). The MCP server itself doesn't need an LLM — your AI client provides the intelligence, StratEvo provides the data and computation:

{
  "mcpServers": {
    "stratevo": {
      "command": "stratevo",
      "args": ["mcp", "serve"]
    }
  }
}

10 tools: get_quote, run_backtest, analyze_portfolio, screen_stocks, and more.

---

AI Copilot (Optional)

StratEvo includes an optional AI copilot for natural-language strategy exploration. This is the only feature that requires an LLM API key — everything else (evolve, analyze, backtest, paper trading) works without one.

# Set any ONE of these — StratEvo auto-detects the provider
export OPENAI_API_KEY="sk-..."           # OpenAI (GPT-4o-mini)
export GEMINI_API_KEY="AIza..."          # Google Gemini
export ANTHROPIC_API_KEY="sk-ant-..."    # Anthropic Claude
export DEEPSEEK_API_KEY="sk-..."         # DeepSeek
export GROQ_API_KEY="gsk_..."            # Groq (Llama 3)
# Or run local models — no API key needed:
# export OLLAMA_BASE_URL="http://localhost:11434/v1"

stratevo copilot

Supported: OpenAI, Anthropic, Gemini, DeepSeek, Groq, Mistral, Moonshot, Azure OpenAI, Ollama (local).

---

For Contributors

StratEvo's plugin architecture makes contributing easy. Adding a new factor is often one file, one function, one PR:

stratevo scaffold factor --name my_signal --category momentum
# Edit the generated file, implement compute(), submit PR

Component	Difficulty
Factors	Easy
Fitness functions	Medium
Data sources	Medium
Strategy DNA configs	Advanced

git clone https://github.com/NeuZhou/stratevo.git
cd stratevo && pip install -e ".[dev]" && pytest

CONTRIBUTING.md · Discord · Issues

---

Roadmap

• 484-factor evolution engine
• Walk-forward + Monte Carlo + Arena mode
• Multi-DNA ensemble voting
• Strategy style classification
• Paper trading
• MCP server for AI tools
• Plugin architecture
• Streamlit dashboard (live evolution tracking)
• DEX execution (Uniswap V3)
• Multi-timeframe evolution
• Community strategy marketplace

---

License

Dual-licensed under AGPL-3.0 and a Commercial License.

• 🆓 Open Source (AGPL-3.0): Free for personal use, research, education, and open source projects
• 💼 Commercial License: Required for paid signal services, proprietary SaaS, or closed-source commercial use

👉 See COMMERCIAL.md for details or contact [e-mail] for licensing.

---

If this is useful, a star helps others find it.

English · 中文 · 한국어 · 日本語