regime-detection

name	regime-detection
description	How to read market regime classifier output (HMM + crisis overlay) and what each regime implies for trading.

Regime Detection

The autonomous system classifies the current market into one of 5 regimes. This shapes every downstream decision.

States

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

Inputs available

A RegimeState object on the LangGraph state has:

• label: one of the 5 above
• confidence: 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 sizing
• gate.allow_new_entries: bool (false in CRISIS)
• gate.require_llm_risk_review: bool (true in TRANSITION/BEAR)

Slow-on/fast-off rule

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.

How LLMs use this

• LLMs CANNOT re-classify the regime upward. They can only DOWNGRADE confidence or DEFER.
• Every analyst report prefaces its read with the current regime so the chain stays consistent.

What the evidence supports (and what it doesn't)

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:

• That OOS window was reused across the v3–v5 design iterations, so it's burned for model selection. Data from 2026-06-13 forward is a frozen holdout for v6+ (src/trading_agent/regime/holdout.py) — never run a model-selection evaluation into it without --break-glass.
• The walk-forward validated an auto-calibrated architecture; production runs a human-calibrated artifact (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.