BtcDeribitIvTermStructureZScoreCalendarRV

BTC Deribit IV Term-Structure Calendar Reversion, Delta-Neutral Vega Relative-Value (DERIBIT ATM Straddle Calendar: short/long the ~7-DTE weekly leg vs the ~30-DTE monthly leg when the front-vs-back ATM-IV slope is a z-score extreme, weekly roll, 3-parameter)

A DELTA-NEUTRAL, single-underlying, TWO-EXPIRY relative-value VEGA trade on DERIBIT BTC options. It trades the SLOPE of the ATM implied-volatility term structure (front weekly ATM IV vs back monthly ATM IV), NOT the level of vol — mechanically distinct from every options death here: not absolute-vol VRP short (died no_edge/risk_reject), not single-leg directional long-convexity (died data_wall on 12 trades), not a covered call (died on execution). Each cycle it builds an ATM STRADDLE CALENDAR pairing the ~7-DTE weekly against the ~30-DTE monthly at the same ATM strike, both near-zero delta at inception — a pure bet on the front-minus-back IV spread normalizing. Fires ~weekly when the slope is a rolling z-score extreme (~50-80 trades over Deribit's ~2y BTC history). Fills the factory's largest diversity gap (options 5.8% vs >=15% target; delta-neutral serves the under-served market-neutral bucket).

Implements the hypothesis exactly: single-underlying, two-expiry, delta-neutral vega RV on the front-vs-back ATM-IV SLOPE (not the vol level), entered at rolling z-score extremes, weekly roll cadence, three tunables (entry_z, exit_z, z_lookback). Uses parameters['rolling_options'] with structure='calendar', which is the only supported multi-expiry path — discover_deribit_calendar picks ONE common ATM strike with candle data for all four legs (front C+P, back C+P) at each weekly open and stamps both expiries with the same entry date, so the front and back genuinely share an entry and a strike. Per the hypothesis I made the entry threshold a rolling z-score rather than a fixed vol-point level: prior sibling calendars used absolute IV-spread thresholds, which are not scale-invariant across vol regimes. Two deliberate deviations, both flagged: (1) the chain lookback is 540 days (~77 weekly cycles), not 730 — the option-leg count (~300 contracts) and the Deribit chain-probe cost are the binding constraint on the 600s Layer-3 options budget, and 18 months still spans several vol regimes; (2) the back leg resolves to the Friday ~28 days out (the expander picks the nearest listed Friday to open+30), which is the liquid ~1-month tenor. Everything in the hot path is O(1) (per-leg quote slots, monotone schedule cursor, deque z-score with running moments, 24-iteration IV bisection), primary and legs both on 1-DAY bars since the legs only mark daily. Honest caveat for the analyst: this mechanism's siblings died 'structurally unmeasurable' on short option windows, so the trade count from this 18-month chain (~25-45 calendars) may still be below the hypothesis's 60-trade success criterion.

data_wall: unmeasurable / insufficient statistical power: the strategy produced only 12 calendar trades over the FULL available Deribit BTC option history (~18 tradeable months of chain), with metrics_reliable=false, only 3 days moving the equity curve (effective_daily_obs=3), sharpe_ci_low=0.0 and PSR 0.64 — non-diagnostic. This is NOT a fee/expectancy failure (avg_trade_return_pct +12.7% premium-relative, PF 11.36, positive expectancy) and the mechanism is genuinely new and correctly implemented (delta/vega-neutral term-structure SLOPE RV, distinct from prior VRP/directional/covered-call options deaths, entries faithful: 6 long/6 short, ~zero net delta). The problem is structural: a weekly, z-extreme-gated cadence over a chain that only starts ~2025-02 cannot produce a statistically meaningful sample, and 3 parameters cannot enlarge it. Two hours of 3-phase optimization would only curve-fit ~4 trades per walk-forward window and a handful in holdout. QA's trade-count/statistical-power concern is borne out on full history (12 < 30 < the hypothesis's own 60-trade criterion). Same wall as the EthDeribitTrendLongConvexityCheapIvRoll data_wall abandon and L27. abandon_class = unmeasurable.

Delta-neutral, vega-neutral relative-value trade on the SHAPE of the Deribit BTC ATM implied-vol term structure. Every daily bar it Black-Scholes-inverts the front (~7-DTE) and back (~28-30-DTE) ATM straddles from their own option candles against the perp spot, takes the slope IV(front)-IV(back), and standardizes it into a rolling z-score. At |z| >= entry_z it builds a same-strike ATM STRADDLE CALENDAR: z high (front rich / curve inverted) -> SELL the front straddle and BUY the back; z low (front cheap / steep contango) -> BUY the front and SELL the back. The back leg is sized vega-neutral (contracts_back = contracts_front * vega_front / vega_back), both legs are ATM straddles at ONE shared strike so net delta is ~0 (and is measured and capped before entry). It exits when the z-score converges back inside the exit band, or is force-rolled roll_buffer_hours before the front expiry. Real Deribit option legs are resolved by the rolling_options 'calendar' chain expander; the BTC perp is the spot/roll clock only and is never traded.

No change required; optionally reword the hypothesis text to say 'weekly roll cadence' to avoid tripping the automated consistency checker.

Layer-1.5 flagged 'hypothesis states weekly bars but config.bar_type is 1-DAY'. This is a false positive: the hypothesis uses 'weekly' to denote the front option tenor (~7-DTE weekly Deribit contract) and the roll cadence, not the signal-bar timeframe. The code correctly samples the IV-slope z-score on 1-DAY bars (option legs only mark daily) and rolls weekly, which is a faithful implementation. No genuine timeframe mismatch.

Prefer resolving the premium unit from the venue/instrument path (native-options vs linear-USD) rather than a magnitude threshold, if the leg universe is ever broadened.

_premium_usd() uses a magnitude heuristic (px < 1.0 => coin premium * spot, else USD premium). For a near-dated ATM BTC option coin premium is a small BTC fraction so this classifies correctly, but a deep/long-dated option whose coin premium exceeds 1.0 BTC would be misread as already-USD. Not reachable for this ATM ~7/30-DTE straddle universe, but brittle if the leg universe widens.

For the full backtest, run against the maximum available Deribit BTC option history rather than the 540-day sandbox window so the sample is as large as the data allows before the analyst judges edge.

lookback_days=540 (18 months) is shorter than the hypothesis's implied ~2-year Deribit BTC option history, chosen to keep ~300 legs inside the Layer-3 budget. This directly constrains the achievable sample; combined with weekly cadence and the z-extreme gate it is the root of the trade-count concern above.

Trade-count / statistical-power risk. The developer honestly estimates ~25-45 calendars from the 540-day (18-month) chain lookback, below the hypothesis's own 60-trade success criterion; the sandbox produced only 12 trades over 178 days with metrics_reliable=false (only 3 days moved the equity curve, sharpe_ci [0.0, 3.63]). This mechanism's option siblings previously died 'structurally unmeasurable' on short Deribit windows. The mechanism is a genuinely NEW and correctly-implemented one (term-structure SLOPE RV, distinct from the prior VRP/directional/covered-call options deaths), so it deserves a full-history backtest, but the analyst should check the realized trade count first: a sub-30-trade result cannot distinguish edge from noise regardless of PF/Sharpe. Also note the sandbox win metrics are premium-relative and healthy (avg_trade_return_pct +12.7%, positive per-notional), so this is a sample-size concern, NOT a fee-viability failure.

Genuinely NEW and correctly-implemented mechanism: IV term-structure SLOPE relative value (front-week vs month ATM straddle calendar), delta- and vega-neutral by construction, distinct from every prior options death (VRP short, directional convexity, covered call). Fills a real diversity gap.

Per-notional edge is healthy, so this is NOT a fee-viability failure: avg_trade_return_pct +12.7% (premium-relative), profit_factor 11.36, positive expectancy +$212/trade, low 10.1% max drawdown.

Entries look faithful to the hypothesis: 6 long / 6 short calendars, balanced, ~near-zero net entry delta (-0.25), avg moneyness ~1.02, avg DTE ~15.8 (front-vs-back straddles).

DECISIVE: only 12 trades over the ENTIRE available Deribit BTC option chain. The option legs span ~2025-02 to ~2026-09 (the perp is 2018+, but the tradeable window is ~18 months), and the weekly, z-extreme-gated cadence extrapolates to only ~12-45 calendars over ALL of history.

metrics_reliable=false; only 3 days moved the equity curve (effective_daily_obs=3); sharpe_ci_low=0.0 (straddles zero at the bottom), probabilistic_sharpe 0.64. The Sharpe 2.34 / PF 11.36 / 67% win rate are non-diagnostic on 12 observations.

Sample cannot be enlarged by tuning: 3 parameters cannot raise the trade count, and a 3-phase walk-forward would leave ~4 trades per window and a handful in holdout — curve-fitting noise. The hypothesis's own 60-trade success criterion is not met.

BtcDeribitIvTermStructureZScoreCalen-9fd024ca24

This strategy introduced a genuinely new and correctly-implemented mechanism for the factory's under-served options bucket: a delta/vega-neutral bet on the front-vs-back ATM IV slope normalizing, distinct from the prior VRP-short, directional-convexity, and covered-call options deaths. Its per-trade economics were healthy (avg_trade_return_pct +12.7%, PF 11.36, positive expectancy), so it was not a fee-viability failure and entries were faithful to the thesis. The fatal problem was structural: only 12 calendar trades over the entire ~18-month tradeable Deribit chain, with metrics_reliable=false and only 3 days moving equity, so Sharpe 2.34 and the win rate carried no statistical weight. The analyst abandoned it at backtest-review as a data_wall — the same wall that killed the EthDeribit convexity roll — after a single iteration, before any optimization ran.

A mechanically sound, novel edge is still unshippable if the tradeable data window structurally caps the sample below statistical significance; check achievable trade count against the data history before building, not after.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate, classifying it as a data_wall: the weekly, z-extreme-gated cadence over a Deribit option chain that only begins ~2025-02 cannot generate a statistically meaningful sample, and 3 parameters cannot enlarge it. Optimization, analyst post-opt review, and risk review were never reached.

A delta- and vega-neutral relative-value trade on the SLOPE of Deribit BTC's ATM implied-vol term structure, pairing a ~7-DTE weekly straddle against a ~30-DTE monthly straddle as a calendar whenever the front-vs-back IV z-score hit an extreme.

Over the full available option history it produced only 12 trades (6 long / 6 short calendars) with avg_trade_return_pct +12.7% (premium-relative), profit factor 11.36, 67% win rate, Sharpe 2.34 and 10.1% max drawdown. But metrics_reliable was false: only 3 days moved the equity curve (effective_daily_obs=3), sharpe_ci_low=0.0 and probabilistic Sharpe 0.64, making those headline figures non-diagnostic.