Menu
How to read market regime classifier output (HMM + crisis overlay) and what each regime implies for trading.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
How to pick single-leg options for directional setups. Long premium only for now — SELL-to-open is hard-blocked at the order tools until multi-leg combos get atomic sizing (R5b/R5c will govern short legs when the block lifts).
Framework for trading around earnings — detecting the earnings lock (R6), picking IV-aware structures, and recording post-earnings lessons.
Hard sizing rules (R1-R7) for every proposed paper order. The order tools re-validate these numerically inside the moomoo MCP server — if you skip them, the order is blocked.
Invalidation discipline, exit rules, and portfolio heat — how to manage risk on open paper positions.
End-to-end order flow — thesis → sizing → paper order. Any step failing aborts the whole flow.
Every paper order requires a thesis recorded in the last 10 minutes. Use journal-mcp before calling any moomoo order tool — the order guard blocks otherwise.
| name | regime-detection |
| description | How to read market regime classifier output (HMM + crisis overlay) and what each regime implies for trading. |
The autonomous system classifies the current market into one of 5 regimes. This shapes every downstream decision.
| Label | Trigger conditions (rough) | Trading posture |
|---|---|---|
BULL_TREND | SPY 50d > 200d, 5d return > 0, vol < 75th pctile, breadth positive | Full long bias; full size multiplier (1.0) |
RANGE_LOW_VOL | SPY trendless, 20d realized vol < 60th pctile, breadth flat | Long with smaller size (0.7) |
VOLATILE_TRANSITION | Vol elevated, breadth ambiguous, possible flip | Long allowed but heavily downsized (0.4); LLM risk council mandatory |
BEAR_TREND | SPY 50d < 200d, 5d return < 0, vol elevated, breadth negative | Long puts only; size 0.5 |
CRISIS | VIX >85th pctile + breadth collapse + correlated drawdown | NO new entries; only reduce-risk |
A RegimeState object on the LangGraph state has:
label: one of the 5 aboveconfidence: 0.0–1.0 (combined from HMM posterior + crisis overlay)pending_label: a state the system is transitioning to (after 3 consecutive higher-confidence reads)crisis_flags: list (vix_spike_85th, breadth_collapse, correlated_drawdown_5pct, vix_term_inversion)gate.size_multiplier: 0.0–1.0 multiplier applied to deterministic sizinggate.allow_new_entries: bool (false in CRISIS)gate.require_llm_risk_review: bool (true in TRANSITION/BEAR)Confidence rises slowly: a regime change to a more risk-on tier requires 3 consecutive days of higher posterior. Confidence drops fast: any single CRISIS-overlay trigger immediately moves to CRISIS regardless of HMM.
Treat the regime layer as drawdown control, not alpha. The walk-forward record (2017 → 2026-06-12, 5 seeds): overlay Sharpe 0.908 ± 0.029 vs SPY 0.837. The +0.071 difference is ~0.2 standard errors — not distinguishable from zero — and the overlay trails SPY buy-and-hold by ~142pp cumulatively (avg 0.69× exposure). What IS demonstrated: 2022-bear protection (+9–10pp) and max drawdown ~−14–15% vs SPY's ~−25%.
Two caveats when reasoning about regime output:
src/trading_agent/regime/holdout.py) — never run a
model-selection evaluation into it without --break-glass.scripts/promote_hmm.py). To measure
the deployed model itself, use scripts/validate_production_hmm.py.Practical implication: lean on the regime label to SIZE DOWN and to block entries (CRISIS/BEAR gates), not as a directional return forecast.