Strategy Evolution Lab
Strategies that earn the right to fire — and explain why. Propose a parameter variation, validate it against real markets with bounded capital, then promote the winner to your library with a plain-English explanation of which conditions made it work.
Test parameter variations of your strategies with capital-bounded risk against live markets. When statistical evidence is conclusive, promote winners into your library with a clear suitability tag explaining the market conditions that made them work.
No closed weekly trades to summarise yet.
No active strategy experiments yet
Cards here will show each experiment with its progress toward promotion: trade count, win-rate, average return, and which market regimes it has been tested in. Populated when the AI starts proposing parameter variations to validate.
Win-rate by VIX regime
LIVE-EMPTYLoading aggregation…
Regime classifier — sanity probe
CBOE convention: ≤15 = low · 15–22 = neutral · 22–30 = elevated · >30 = crisis. Sampled at entry-bar timestamp (no look-ahead).
Ask the system about your trade history
Interpretability over autonomy. The chatbot routes structured intents to aggregation queries against the trade-context corpus. Questions will return results as trade history accumulates.
How the Strategy Evolution Lab works
1 · Propose— the reflection_agent reads the trade-context corpus (per-trade macro + sentiment + FOMC + sector + global-economy snapshots) and proposes a parameter deviation hypothesis (e.g. "RSI-2 with tighter entry threshold in elevated VIX on energy ETFs"). Semantic dedup (embedding cosine-sim ≥0.85) blocks repeated proposals; max 3 active per regime_gate prevents monoculture.
2 · Shadow trade — the deviation runs real bracket orders against the paper Alpaca account with stricter capital bounds ($10–$100/trade; 5% shadow heat cap RM-07f; max 7 per parent RM-07g; deterministic 110ms spacing for rate-limit safety). Phase 1 = paper-only, enforced by SentinelConfig __post_init__ tripwire.
3 · Validate — accumulate ≥20 closed trades; promotion gate runs Benjamini-Hochberg FDR (q=0.10) across the active cohort, requires median-trade-R ≥+0.20, max-single-loss ≤3× position_cap, ≥2 VIX regimes observed (entry-bar sampled, no look-ahead bias), backtest sanity within +5pp parent drawdown.
4 · Promote— operator clicks "Promote to live" on the deviation card. Approval modal re-runs eligibility synchronously at click moment (APP-029 friction-layer pattern) and refuses the click if any gate now fails. Audit-logged with reason ≥8 chars.
5 · Suitability tag — promoted strategy carries machine-evaluable primary_condition Boolean + human-readable explainer. Fires only when the macro confluence matches. Library cluster aggregation groups by suitability so operators see "all elevated-VIX strategies" / "all energy-sector strategies".
6 · Demote (auto)— rolling-25 Wilson LB < 0.50 → demote; 5-of-7 trailing losses → freeze; primary_condition false 30 consecutive trading days → archive. Symmetric to promotion (no dead zone). Auto-trigger gates further capital exposure; operator retains override.