Output E — Regime State¶

The system's self-diagnosis of the current market regime. Output E is the meta-confidence input for every other output: when regime is unstable or out-of-distribution, downstream signals are suppressed or attenuated rather than fired with stale assumptions.

Cadence¶

Hourly, computed alongside Output A.

Schema¶

The pydantic model is memenet.schemas.regime.RegimeState.

Example payload:

{
  "timestamp": "2026-05-02T14:00:00Z",
  "regime_class": "stable_rotating",
  "indicators": {
    "mean_pairwise_correlation": 0.42,
    "correlation_stability_24h": 0.81,
    "cluster_persistence_7d": 0.78,
    "universe_churn_pct_7d": 8.2,
    "feature_drift_score": 0.12,
    "out_of_distribution_score": 0.18
  },
  "circuit_breakers": {
    "high_correlation_breaker": false,
    "low_correlation_breaker": false,
    "high_churn_breaker": false,
    "high_drift_breaker": false,
    "ood_breaker": false
  },
  "confidence_multiplier": 0.94,
  "signal_status": "active",
  "last_model_retrain": "2026-04-15T00:00:00Z",
  "model_version": "laggard_v3"
}

Regime classes¶

Class	Definition	Signal behaviour
`stable_rotating`	normal regime, mean correlation 0.3–0.6, persistent clusters	full confidence
`risk_on_blowoff`	mean correlation > 0.7, everything pumping together	rotation suppressed; laggards attenuated
`risk_off_crash`	strong negative co-movement, capital fleeing	all signals suppressed
`decorrelated`	mean correlation < 0.2, no co-movement structure	all signals suppressed
`structural_break`	drift or OOD beyond historical bounds	all signals suppressed pending retrain
`degraded`	one or more indicators in caution range but not breaker territory	signals fire with confidence multiplier < 0.5

Circuit breakers¶

Hard rules that override the classifier. If any breaker fires, signal_status is forced to "suppressed" regardless of what the model predicts. Default thresholds:

Breaker	Default threshold
`high_correlation_breaker`	mean pairwise correlation > 0.85
`low_correlation_breaker`	mean pairwise correlation < 0.15
`high_churn_breaker`	7-day universe churn > 40%
`high_drift_breaker`	feature drift score > 0.5
`ood_breaker`	OOD score beyond the 99th percentile of training

Confidence multiplier¶

A scalar in [0, 1] that multiplies the base statistical confidence on every downstream signal. Trained as a calibrated logistic regression on the validation split; calibration is evaluated with reliability diagrams and Brier score, not just AUC.

How to consume¶

Snapshots are written to partitioned Parquet under outputs/regime/. Consumers that depend on the toolkit's signal output should always check signal_status before acting on a fired alert.