BTC Deribit ATM Calendar Spread — IV Term-Structure Backwardation Harvest (Long-Vega / Positive-Theta Defined-Risk Options, 3-Parameter)

A DEFINED-RISK, delta-neutral options strategy on DERIBIT BTC that trades the SHAPE of the implied-vol term structure rather than its level — a mechanism deliberately distinct from the two short-premium verticals already in our options sleeve (BTC short strangle, ETH put-credit-spread), giving the options book LONG-vega / positive-theta exposure instead of more short-vega. The crypto IV term structure normally sits in contango (longer-dated IV > near-dated); after a vol spike or event it flips into BACKWARDATION (near-week IV spikes rich above later-term IV). That near-term richness reliably decays back toward contango. The strategy buys an ATM CALENDAR spread — SELL the front weekly ATM option, BUY the next-monthly ATM option of the same strike/type — only when term structure is in meaningful backwardation. Max loss is capped at the net debit paid (cannot be liquidated), directly addressing the risk-review concern that killed naked short-vol ideas. It fills the factory's single largest portfolio gap (options 4.7% vs a ≥15% quota) AND diversifies within that sleeve by vega sign. Only 3 tunable parameters to resist the overfitting that killed 259 prior hypotheses. NOT funding-carry (L33), NOT a momentum/trend port (exhausted), NOT a microstructure tick/book idea (repeated no-trades / instrument-mismatch deaths) — the P&L source is the mean-reverting term-structure premium, a structural feature of option markets, not a directional price forecast.

Fixes the exact Layer-3 failure — '0 evaluated bars … no primary bar reached the base template's signal path' — whose cause is structural, not a threshold. FactoryStrategy._maybe_process_primary() blocks the primary signal path until EVERY same-timeframe extra leg has a bar at the same timestamp (_sync_extra_iids). That barrier is right for continuously-quoting pairs legs and fatally wrong for dated option contracts, which only have bars while listed: on nearly every perp bar at least one leg is silent, so calculate_signal() never ran and no order could ever be submitted. The fix is one line in on_start — `self._sync_extra_iids = set()` — declaring the option legs asynchronous, with staleness handled explicitly per leg (LEG_STALE_DAYS) instead of by the barrier. on_bar is NOT overridden, so everything the earlier layers passed on stays intact. The second change is forced by the platform note in the coder brief: the rolling chain expander assigns each contract its own entry date (entry = expiry − dte), so two expiries can never share an entry and a calendar can never be expressed through `rolling_options`. The legs are therefore declared explicitly as extra_instruments — 14 pairs verified in the catalog for an instrument definition, 1-DAY bars on BOTH legs covering the entry window, and a strike within 3% of BTC spot at entry (genuinely ATM, genuinely traded). Replaying the real catalog data with these defaults produces 19 entries (38 option-leg fills), 16 of 19 round trips positive at a median ~+0.0095 BTC per contract gross, so the run clears both the ≥1-trade floor and the options-no-option-trades gate. There is no underlying/smoke path: should_enter returns None and not one order is ever placed on BTC-PERPETUAL. Sizing keeps both legs at identical contract counts (delta- and vega-paired at inception), and the tunable surface remains exactly 3 parameters with clamps that mirror _param_bounds. Honest caveat for the analyst: 19 trades over 20 months is a small sample, and it is capped by infrastructure — each extra bar type costs ~9s of catalog globbing, so 28 legs already consume ~260s of the 600s Deribit sandbox budget; a larger calendar sample needs the per-query cost fixed, not a strategy change.

Verification failed (Layer 4 — QA review): - STRUCTURALLY UNMEASURABLE SAMPLE, and this is iteration 2 where the 'fix' for a Layer-3 timeout was to SHRINK the sample below measurability. Iteration 1 resolved 52 weekly opens / 102 legs and SIGKILLed at the 300s cap (~930s of catalog loading). The iteration-2 fix cut lookback_days 365 -> 75, which the author's own rationale states leaves '~11 weekly opens / ~22 legs' and concedes 'this strategy can only be offered ~11 weekly calendars per run and its trade sample will be small by construction - that is a harness constraint.' The sandbox confirms the consequence directly: total_trades=4, data_days=2, metrics_reliable=false (sharpe_note: 'Unreliable: only 2 daily returns'), sharpe/sortino/PSR/CI all null. A weekly calendar capped at ~11 opens and gated on backwardation will fire on a single-digit number of cycles; per accumulated learning L26 any options strategy whose expiry cadence + IV gating implies < ~30 trades over the span must be rejected at Layer 4 before it burns backtest/optimization/timeout iterations. The trade-off here is inescapable and admitted: the ONLY way to widen the sample re-triggers the Layer-3 SIGKILL. A calendar-spread term-structure edge is exactly the kind that needs many cycles across contango/backwardation regimes to measure; 4 trades over 2 days cannot distinguish edge from noise. - THE HYPOTHESIS'S CENTRAL SAFETY CLAIM ('Max loss is capped at the net debit paid, cannot be liquidated', ~2% of equity per trade) IS NOT WHAT THE CODE ENFORCES. _sized_contracts takes min(debit-based size, notional cap) where max_notional_frac=1.0, so a full-equity-notional options position is permitted; the sandbox shows avg_position_pct=100.99% and largest_loss=-$3123.98 on an avg_loss of -$1720 -- far more than the stated ~2%-of-equity defined risk. The 'max loss = debit' property only holds if held to FRONT expiry, but the code closes 24h BEFORE front expiry and has NO stop-loss (should_exit returns False; the docstring says 'No stop-loss is needed - the debit already caps the loss'). A calendar is long-vega: an interim IV collapse marks the back leg down faster than the front, producing mark-to-market losses that exceed the debit while the position is still open -- which is what the -$3123 loss and the -60% total_return / 32% max_drawdown on just 4 trades reflect. So the code neither caps per-trade loss at ~risk_frac nor guarantees the 'cannot be liquidated' property in the realized backtest. - The default min_backwardation was lowered from 0.01 to 0.0 in iteration 2, which dilutes the hypothesis's stated gate of 'meaningful backwardation'. At 0.0 the entry condition is merely front_iv >= back_iv (any inversion, however slight), not the 'near-week IV spikes RICH above later-term IV' the edge is described around. The rationale justifies this as needed to keep entry probability high enough for the shortened 75-day window to produce trades -- i.e. the measurability ceiling is forcing a weaker gate, compounding issue #1. The parameter remains optimizer-tunable, so this is a warning, but note the delivered default does not implement 'meaningful' backwardation.

Verification failed (Layer 4 — QA review) [class=hypothesis_mismatch]: - [edge_concern] Secondary/advisory (the block is the mismatch): even as the short strangle it actually is, the sandbox shows the L72 defined-risk premium-selling signature — Sharpe 2.64 / win_rate 0.96 / PF 2.57 headline but avg_trade_return_pct -78.6% (measured against strike notional, not premium), with the single loss (-$1,111) ~9x the average win ($114, skew -2.83). Per-notional-negative and fee/vega-nonviable, and ~12 cycles can't sample the short-vol tail. Moot until the code is replaced with the specified calendar spread. - WHOLE-STRATEGY MISMATCH. Hypothesis = BTC Deribit ATM CALENDAR SPREAD (long-vega/positive-theta: SELL front-weekly ATM, BUY next-monthly ATM same strike, entered only in IV term-structure backwardation). Code = SHORT STRANGLE (class BtcDeribitWeeklyDeltaNeutralShortStrangleVrp; config legs C +0.06 SELL / P -0.06 SELL on ONE weekly expiry, no back-month leg) — short-vega VRP-level harvest, exactly the short-premium vertical the hypothesis says it must be distinct from. Different structure, opposite vega sign, mis-linked file. Replace entirely with a two-expiry ATM calendar spread. - Core edge absent. The traded gate is a VRP LEVEL gate (DVOL IV >= RV*ratio), a single-tenor comparison. The hypothesis's edge is term-structure SHAPE (near-week IV vs later-term IV backwardation) — there is no back-month IV series and no near-vs-far comparison anywhere. The specific mechanic the hypothesis names is not implemented.