Regime-Aware Trade Copilot Strategy

This document captures the strategy coming out of the 2026-05-11 sponsor discussion with Gordon. It should be read as the planning bridge between the current Vol Dashboard and the eventual trade copilot.

The important framing is that the existing dashboard is already useful as a consolidated display and measurement tool. Gordon validated that he periodically clicks through it and finds value in having the market, model, measurement, and reference material in one place. The next phase does not replace that base. It builds on it by turning the daily data backbone, model diagnostics, conditional distributions, and portfolio feed into an explicit decision-support layer.

The eventual product question is:

Given the current volatility regime, surface shape, conditional distribution,
portfolio exposure, liquidity, and data quality, does a proposed trade or
restructuring action make sense?

Executive Direction

The project has moved from "show useful market data" toward "measure whether trades match the regime." That expansion is deliberate and appropriate for the problem, provided the model remains transparent, auditable, and conservative when inputs are incomplete.

The near-term objective is not to claim the model is fully decision-grade. The near-term objective is to make the system reliable enough to:

• prove daily data capture is complete or explain why it is not
• map trades to the correct derivative-equivalent identity or proxy
• detect where spot-vol behavior is changing by surface location
• compare vanilla market-implied probabilities with conditional probabilities
• identify dangerous trades under the current regime before optimizing for alpha
• let Gordon inspect and tune model assumptions rather than treating the model as a black box

Current Dashboard Utility

Gordon explicitly validated the current dashboard as useful before the advanced model layer is production-ready. That matters because the project already has practical value as an aggregate display and audit tool.

The current dashboard should keep doing three things well:

1. Show live and persisted market setup clearly.
2. Show whether imports, snapshots, and derived metrics are healthy.
3. Explain calculations, assumptions, and data gaps through reference docs and hover/help text.

The reference material at the top of pages and the hover explanations should remain part of the product strategy. The aim is that every artifact, calculation, and derived label can be audited by Gordon or by the system itself when answering questions.

Core Market Thesis To Model

Gordon's central thesis is that BTC implied volatility may be transitioning from the recent classic regime:

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

to a local regime where:

spot up -> vol up

The interesting part is not only that spot and vol may now be rising together. The important anomaly is that the risk reversal still appears to reflect the older regime, while realized local behavior may be showing the new one.

That divergence is where the potential alpha lives. If the smile says upside vol should be cheaper than ATM, but observed behavior says front-month upside vol can reprice higher with spot, then selling that upside vol may be dangerous even when the old risk-reversal intuition says it looks attractive.

The model therefore cannot stop at a global "regime changed" label. It needs to locate the behavior across:

• maturity and tenor
• moneyness
• delta bucket
• smile segment
• front-month versus later-dated expiries
• low-delta wings versus options migrating toward ATM

The first practical application is accident avoidance: warn the trader before they add short vol in the part of the surface where the regime says vol may rise with spot.

Updated Strategic Pillars

1. Reliable Measurement Backbone

   Keep prioritizing daily Amberdata capture, import health, coverage diagnostics, and source quality. This is the foundation: Greeks, IV surface, spot, OI, BTC/ETH snapshots, Alpha trades, and eventually SMA trades.

2. Derivative Identity And Proxy Mapping

   Build an explicit mapping layer for ICOI, IMST, SMA, OTC-like trades, custom strikes, custom expiries, adjusted contracts, and proxy tickers. Without this, portfolio analytics remain structurally incomplete.

3. Regime Detection By Surface Location

   The core model should detect spot-vol regime change not just globally, but by tenor, moneyness, delta bucket, and smile segment. The key question is where spot up / vol up is live and where the old risk-reversal logic still dominates.

4. Conditional Distribution Decision Layer

   The conditional-vs-vanilla CDF view should become a decision tool. If conditional probability is materially higher than vanilla around 85k-90k, the UI should say that shorting options in that band is dangerous, subject to trust/confidence.

5. Trade Restructuring Evaluator

   The eventual copilot should compare choices like buy back May, sell July, use June, trade a vertical, sell a 1x2, or do nothing. It should rank choices by regime fit, surface edge, term structure, liquidity, confidence, and portfolio exposure.

6. Model Transparency And Scenario Tuning

   Model Lab should stay central. Gordon should be able to inspect risk-free rate, density method, lookback window, stationarity thresholds, KS p-value settings, blend weights, and candidate filters.

Product Strategy

The work should progress in layers. Each layer creates value on its own and reduces the risk of building a convincing but unsupported model.

Accident Avoidance First

Gordon's most concrete use case was not "find the perfect trade." It was "do not let me do the wrong trade for the regime."

Example:

• A trader is short May upside calls.
• Spot rallies and the trader wants to reduce exposure.
• A mechanically tempting action is to buy back the current call and sell a larger amount of a higher strike, such as a 1x2 call spread.
• If the local regime is now spot up / vol up, that action can leave the book shorter implied vol exactly where vol may reprice higher.

The first decision-support version should therefore detect and explain this type of risk:

This adjustment reduces near strike exposure, but increases short upside-vol
exposure in a surface region where conditional probability and spot-vol behavior
suggest upside options are underpriced. Treat as dangerous unless overridden.

The warning should be conditioned on confidence. If the trust engine says stationarity is unstable, source coverage is incomplete, or signals are mixed, the system should say that clearly. The model should not force conviction.

Conditional Distribution Interpretation

The conditional-vs-vanilla distribution view resonated strongly with Gordon because it translated the thesis into a visible decision rule.

The useful read is:

Where does the conditional model assign materially more probability than the
vanilla options surface, and are we short options in that band?

For the current discussion, Gordon highlighted the 85k-90k region as meaningful. If the conditional CDF implies higher probability of reaching or finishing near those higher spot levels than the vanilla surface implies, then selling options in that band is unattractive unless the trust engine says the conditional model is weak.

This page should evolve from a model visualization into an explicit decision layer:

• show the vanilla market-implied baseline
• show the conditional overlay
• highlight material probability gaps by strike/price band
• map those gaps to "shorting here is dangerous" or "market appears to compensate the risk"
• attach stationarity, source quality, and confidence states

Trust Engine Role

The forecasting layer asks:

What does the conditional setup imply?

The trust engine asks:

Should we trust that implication right now?

This distinction should remain explicit. The model can produce a conditional probability shift while the trust engine says the shift should be discounted.

Trust should be controlled by:

• gamma structure and regime stability
• spot-vol correlation and whether it is becoming less negative or turning positive
• stationarity tests, including KS p-value threshold and mean-shift checks
• source freshness and snapshot completeness
• signal agreement versus mixed thesis
• lookback-window consistency
• whether the relevant surface segment has enough observations

When elevated-vol and lower-vol signals are balanced, the correct output is a mixed thesis and lower confidence. Gordon agreed this workflow makes sense. The system should not always force a directional answer.

Decision Points For Gordon / Team

These decisions should be tracked explicitly. Some work can proceed before answers are final, but the model should not present unsupported outputs as sourced facts.

Derivative-equivalent identity and proxy policy are maintained in Derivative Identity And Proxy Mapping. Source-authority details for marks, P&L, estimates, proxies, and unavailable states are maintained in the Source Authority Register.

The preparatory implementation specifications are maintained separately so Gordon can review the plan before code changes:

• Data Ingestion Readiness Plan for reliable measurement, durable jobs, coverage reasons, and provider-validation questions.
• Regime Model Specification for lookbacks, post-IBIT history, surface-location spot-vol, smile audit, divergence, danger zones, Model Lab provenance, and mixed-thesis trust states.
• Trade Copilot Evaluation Specification for accident warnings, trade-action taxonomy, restructuring comparison, and candidate filter diagnostics.
• SMA Ingestion Requirements for future SMA source fields and derivative-equivalent mapping requirements.

Work That Can Proceed Before Business Answers

The following work is safe because it improves reliability, transparency, or scaffolding without forcing a business assumption:

• document the strategy and decision matrix
• add source-authority placeholders and quality-state enums
• create the derivative identity and proxy mapping schema
• build coverage probes that explain provider gaps without valuing trades
• standardize configurable lookback windows and expose them in Model Lab
• audit the smile-deviation baseline and explain why all cells may be negative
• improve mixed-thesis and low-confidence messaging
• separate pre-trade regime warnings from post-trade P&L reporting
• add an accident-avoidance warning taxonomy without making live recommendations

The following work should wait for answers or ship only as explicitly caveated scaffolding:

• SMA trade ingestion into production scope
• decision-grade execution quality
• decision-grade forward, terminal, or realized P&L
• use of proxy tickers in valuation or recommendation outputs
• strong accident-avoidance warnings that could influence live trading
• ranked trade restructuring recommendations

Initial Ticket Map

Product Boundaries

This is a decision-support system, not an auto-trading system.

The dashboard should not imply more precision than the source data supports. In particular:

• missing economics are not zero
• proxy-mapped data must be labelled as proxy
• estimated marks must be labelled as estimates and excluded from sourced aggregates unless approved
• unsupported contracts should remain unsupported rather than forced through a weak mapping
• mixed evidence should produce low confidence rather than a forced trade direction
• conditional probability should be anchored to vanilla market-implied probability and gated by trust state

Target End State

The target end state is a copilot that can review a proposed trade or portfolio adjustment before execution and answer:

Does this trade fit the current regime, surface, conditional distribution,
liquidity, and portfolio exposure?

For Gordon's Friday example, the system should eventually compare alternatives such as:

• buy back May and do nothing else
• buy back May and sell July
• use June instead of July
• trade a vertical spread
• sell a 1x2
• choose another restructuring action

The output should rank those alternatives by:

• regime fit
• surface edge
• conditional-vs-vanilla probability gap
• term-structure behavior
• liquidity and open interest
• portfolio exposure and gamma sensitivity
• confidence and trust state
• source-data caveats

That is the destination: a transparent, auditable copilot that helps the trader see whether the proposed action matches the regime actually in front of them.