Au-dela du PnL : comment on evalue les strategies IA sur 7 criteres

A strategy with 90% win rate can still lose money.

It happens more often than you think. A strategy wins 9 small trades, then loses everything on the 10th. The win rate says 90%. The account says -50%.

That's why Strategy Arena evaluates every AI trading strategy through 7 independent frameworks — not just PnL or win rate alone.

The 7 frameworks

Inspired by the Decision Evaluation Framework pattern from Claude Agent Blueprints, each of our 59 strategies receives a grade from A+ to F on every axis:

⚖️ Risk-Reward (25% weight)

What it measures: Sharpe ratio + profit factor + raw PnL Why it matters: A strategy that makes +10% with a Sharpe of 0.3 is worse than one making +5% with a Sharpe of 2.0. Risk-adjusted returns separate luck from skill.

📊 Activity (10% weight)

What it measures: Number of trades executed Why it matters: A strategy with 0 trades scores F. A strategy with 2,000 trades might be overtrading. The sweet spot is enough trades to be statistically significant without churning.

📈 Consistency (20% weight)

What it measures: Win rate stability + equity curve smoothness Why it matters: A strategy that alternates between +10% and -10% months has poor consistency even if the average is positive.

🛡️ Robustness (20% weight)

What it measures: Profitability combined with drawdown resistance Why it matters: Can the strategy survive a bad week? A strategy profitable at +5% but with 25% drawdown is fragile. One at +3% with 3% drawdown is robust.

⚡ Efficiency (10% weight)

What it measures: Profit per trade Why it matters: Making $100 in 10 trades (efficient) is better than making $100 in 1,000 trades (inefficient + more fees + more exposure).

🔥 Stress Resistance (15% weight)

What it measures: Maximum drawdown + worst trade survival Why it matters: This is the "will it blow up?" detector. Strategies that survived major drawdowns without catastrophic loss score high.

🏥 Overall Health

What it is: Weighted composite of all 6 scores → letter grade A+ to F What it tells you: Instant, at-a-glance health of any strategy.

How to read the scores

Real examples from our arena

Visit the Strategy Health Check page to see all 59 strategies scored live. Click any strategy card to see the detailed breakdown.

The scores update in real-time as strategies trade. A strategy that was graded C yesterday might be B today after a good run — or E after a drawdown.

Connected to the intelligence layer

The Health Check feeds into the broader Strategy Arena ecosystem:

• Evolution Lab: The Darwin Engine uses fitness scores (similar to Health Check) to decide which strategies to evolve
• Meta-Harness: Optimizes how fitness is calculated — the weights of each framework are themselves optimized nightly
• Knowledge Graph: Health scores are connected to strategy nodes, visible in the interactive graph
• Arena Brain: Ask "which strategy is healthiest?" and the RAG pulls Health Check data

Why multi-dimensional evaluation matters

Single-metric rankings are misleading: - Ranked by PnL: a lucky gambler tops the list - Ranked by win rate: a strategy that takes 1 trade and wins leads - Ranked by Sharpe: a strategy with 3 trades and 0 volatility scores infinite

Our 7-framework approach catches all of these. A strategy needs to score well across multiple dimensions to earn an A.

This is how professional hedge funds evaluate strategies internally. We just made it transparent and free.

Check the Strategy Health of all 59 strategies now. See who's really healthy — not just lucky.