Regime engine — model card

v1.0 · Tier 1Live Cross-asset regime classification with AR(1) persistence half-life.

Purpose

Classifies the current market state into one of three regimes — risk-on, risk-off, transition — feeding the LS6 layer-conditional weighting (engine v76) so factor weights shift with the regime. Drives the risk-color on every scoreboard surface and the playbook's "when NOT to act" disqualifier (transition → wait).

Inputs

DXY (dollar index) trend + level
10Y-2Y yield curve spread
Equity breadth (% above 200d SMA across SPX components)
VIX level + slope
Credit spreads (HY OAS via FRED)
Cross-asset correlation matrix (SPY × TLT × GLD × crypto)

Outputs

{
  state:        'risk-on',           // 'risk-on'|'risk-off'|'transition'|'recovery'
  daysInState:  18,
  P_persist_5d: 0.83,                // probability state persists 5 more days
  halfLifeDays: 42,                  // state-specific
  drivers: [
    { factor: 'DXY trend',         contribution: -0.4, sigma: -0.4 },
    { factor: 'Yield curve',       contribution: -0.18, sigma: -0.18 },
    { factor: 'VIX level',         contribution: -0.3, sigma: -0.3 }
  ],
  confidence: 0.74
}

Half-life by state (Bible §6.4)

State	Half-life	Typical drivers
risk-on	42 days	Tight credit, low VIX, positive breadth
risk-off	28 days	Wide credit, VIX spike, breadth collapse
transition	10 days	Crossing thresholds; whipsaw zone
recovery	21 days	Post-shock, breadth rebuilding

Math

Persistence: P(persist N | seen t) = exp(-N/τ_state) × (1 + seasoning(t)). Seasoning multiplier penalizes stale regimes (over-staying threshold = forced re-evaluation). Closed-form, no Monte Carlo.

Limitations

What this is NOT:

Not a precise transition predictor — first 24-48h of a flip is whipsaw, low confidence by design
Not international — primarily US macro signals. Foreign markets read through ADRs and currency proxies
Not crypto-native — crypto regime detection sits in v77 specialty module, not here

Known failure modes

Fed-day whipsaw: regime label can flip momentarily as VIX/DXY spike then revert
Earnings-week breadth distortion: single mega-cap names can dominate breadth signal
Fast tightening cycles: historical half-life calibration may underestimate persistence in extended regimes

Versioning

Version	Date	Change
1.0.0	2026-04-22	Initial 4-state classification + half-life persistence

Source

engine/regime.js, engine/regimeShift.js, engine/ensembleRegime.js, engine/hmmRegime.js (alternative HMM path)

Reviewed: 2026-04-27 · Next: 2026-07-27 · Per Bible §15.61