Regime Model Specification

This document defines the preparatory specification for the regime-aware model layer. It covers ENG-5241, ENG-5242, ENG-5243, ENG-5244, ENG-5245, ENG-5251, and ENG-5252.

It is a specification and review artifact only. It does not implement model changes or change current UI behavior.

Purpose

The model layer should convert Gordon's market thesis into auditable diagnostics:

Where is spot-vol behavior changing, does the options surface reflect that
change, and should the conditional model be trusted right now?

The model should not simply label BTC as in one global regime. The sponsor discussion made clear that the effect may live in specific parts of the surface: front month, a strike band, a delta bucket, or a smile segment.

Gordon's Thesis To Encode

Recent BTC behavior may have shifted locally from:

spot up -> vol down
spot down -> vol up

to:

spot up -> vol up

The important divergence is that the risk reversal may still show the older equity-style skew pattern while realized local spot-vol behavior is starting to behave differently. That gap can be alpha, but it can also be noise. The model therefore needs evidence, confidence, and surface-location detail.

Ticket Coverage

Lookback Policy

Current pages use inconsistent historical windows, including approximately 730-day, 365-day, 130-day, and 90-day views. This inconsistency needs to become explicit model configuration.

Gordon's comment on post-IBIT history implies this policy:

• older BTC option history is not useless
• older history should not be treated as structurally identical to the current IBIT/liquid-options era
• Model Lab should allow the user to compare longer history, post-IBIT history, and recency-weighted history
• production defaults should be stable and documented
• scenario overrides should be visible and not silently promoted to production defaults

Recommended model modes:

Every regime output should show:

• lookback window
• sample size
• data start and end dates
• whether old/pre-IBIT history is included
• weighting method if recency-weighted
• scenario override state

Surface-Location Spot-Vol Regime Detection

The spot-vol regime detector should classify behavior by surface location, not only globally.

Dimensions:

Classification states:

Minimum evidence fields:

surface_regime = {
  currency,
  as_of,
  tenor_bucket,
  dte_range,
  delta_bucket,
  moneyness_bucket,
  smile_segment,
  lookback_mode,
  sample_size,
  spot_return_measure,
  iv_change_measure,
  correlation,
  slope,
  classification,
  p_value_or_confidence_proxy,
  source_quality,
  caveat
}

The exact estimator can evolve. The product need is stable: show whether the observed behavior is global or local, and whether it is robust enough to affect trade warnings.

Risk-Reversal Versus Realized Spot-Vol Divergence

The model should explicitly compare what the risk reversal implies with what realized local spot-vol behavior is doing.

The divergence state should not automatically mean "trade now." It means the model found a potentially important mismatch that needs confidence, surface location, and trust-engine gating.

Conditional-Vs-Vanilla Danger Zones

The conditional distribution page resonated with Gordon because it translated the thesis into a clear question:

Where does the conditional model assign materially more probability than
the vanilla market-implied surface, and are we short options there?

For the sponsor discussion, the notable example was the approximate 85k-90k area on a 30-day tenor. The model should generalize this as a strike-band detection problem.

Minimum danger-zone output:

danger_zone = {
  currency,
  as_of,
  tenor_bucket,
  strike_low,
  strike_high,
  vanilla_probability,
  conditional_probability,
  probability_gap,
  gap_materiality,
  trust_state,
  source_quality,
  message,
  warning_severity
}

Interpretation rules:

Warning language should remain risk-control language, not trade advice.

Smile-Deviation Matrix Audit

Gordon thought the smile-deviation matrix was useful in principle but suspicious because every cell appeared negative. That output may mean IV is broadly near the bottom of the recent range, or it may mean the baseline logic is too crude.

The audit should decompose the matrix into separate effects:

The matrix should eventually distinguish:

• "all vol is cheap versus history"
• "calls are cheap relative to ATM"
• "puts are rich relative to calls"
• "front month is cheap versus later tenors"
• "cell is unreliable because history is sparse"

Trust Engine And Mixed Thesis

The forecasting layer asks:

What does the conditional setup imply?

The trust engine asks:

Should we trust that implication right now?

Low confidence is a valid output. If elevated-vol and lower-vol evidence are balanced, or stationarity is unstable, the system should say so instead of forcing a directional read.

Trust inputs:

• stationarity KS p-value threshold
• mean-shift threshold
• gamma structure stability
• spot-vol correlation trend
• source freshness
• surface-location sample size
• signal agreement
• lookback policy
• mixed-thesis balance
• proxy/estimated/unavailable states

Trust states:

Model Lab Provenance

Model Lab should stay central because Gordon wants to inspect and tune assumptions.

Every displayed model result should expose:

• risk-free rate
• density method
• lookback window
• post-IBIT inclusion or weighting
• stationarity KS p-value threshold
• mean-shift threshold
• blend weights
• gamma-conditioning parameters
• candidate filters
• source-quality exclusions
• scenario override flags
• calculation version

Model Lab should clearly distinguish:

Scenario outputs should be exportable or reviewable, but they should not become production defaults unless explicitly approved.

Decision Points

Preparatory Acceptance

This specification is complete when:

• lookback and post-IBIT relevance are explicit model settings
• spot-vol regime is defined by surface location
• risk-reversal divergence has named states
• conditional danger zones have a strike-band output contract
• the smile matrix audit separates broad level from skew and curvature
• Model Lab provenance is documented
• mixed thesis and low-confidence states are treated as valid model outputs