Volatility regimes

Definition

Volatility regimes are persistent states of market turbulence: quiet periods cluster, turbulent periods cluster, and transitions between them reshape which strategies work, how positions should be sized, and what correlations hold. Volatility clustering — “large changes tend to be followed by large changes, of either sign” (Mandelbrot 1963) — is among the most robust facts in financial data, and the regime framing is the platform’s primary state variable.

How it works / structure

• The documented facts: clustering (autocorrelated magnitude — the ARCH insight, Engle 1982), asymmetry (volatility rises faster on declines than rallies — the leverage effect), mean reversion (extreme vol decays toward the long-run level), and regime persistence (states last weeks to months, not days).
• State definition (engine-executable): realized vol or ATR percentile vs trailing distribution (atr_14_pct bands), IV state (iv_rank), VIX curve shape (contango/backwardation — instrument-vix-futures), transition rules with hysteresis (separate entry/exit thresholds to stop flip-flopping). The platform pins a small state set (e.g. low / normal / elevated / crisis) rather than a continuum.
• What switches with the state: strategy gates (reversion fades degrade in vol expansions; trend systems earn their keep in sustained directional turbulence — strategy-mean-reversion vs strategy-futures-trend-following), sizing (risk-volatility-targeting mechanizes it), correlation assumptions (risk-correlation-exposure — correlations rise with vol), and options economics (premium selling’s compensation and its tail both scale with the state).

When it applies

Everywhere — this is the KB’s master regime entry: strategy entries cite it for their regime filters, sizing entries for their estimators, and options entries for the IV state. Regime classification is descriptive and current-state; regime PREDICTION is a much weaker claim the platform treats separately per thesis.

Risk profile & failure modes

• Transition lag: every state estimator is backward- looking; the first crisis day arrives inside the “normal” state at normal size — regime systems manage the second week (risk-volatility-targeting shares this limit).
• Threshold flip-flop: boundary-adjacent states without hysteresis whipsaw every regime-gated strategy simultaneously.
• Regime overfit: many states + fitted thresholds = storytelling; the platform keeps the state set small and the thresholds replay-audited.
• “This time is different” in both directions: treating a new vol floor as permanent (2017) or a spike as permanent (2020) — states end; persistence is not permanence.

Evidence & limits

Clustering and asymmetry are documented across a century of data and every liquid market (Mandelbrot 1963; Engle 1982 and the vast GARCH literature); regime-switching models improve allocation decisions in-sample (Ang-Bekaert 2002) with the usual out-of-sample humility. Current-state classification is reliable; turning-point prediction is not — no cited method calls transitions dependably, and claims otherwise are folklore.

Falsifiable-thesis examples

Illustrations only, not signals:

• “The index’s 20-day realized vol, in its top decile today, will be below its top decile within 60 sessions (mean reversion of vol)” — falsified by the vol series.
• “Gating strategy S off during ‘elevated’ states improves its replay Sharpe this decade” — falsified by the paired replay.

Cross-references

• Measurement: indicator-atr, indicator-realized-vs-implied-vol, opt-iv-rank-percentile
• Curve-based state: instrument-vix-futures
• Mechanized responses: risk-volatility-targeting, risk-correlation-exposure
• Strategy gates: strategy-mean-reversion, strategy-futures-trend-following, strategy-iron-condor