Spreading the 0DTE fly across SPX and QQQ

The design, in plain terms

Adding a second liquid 0DTE surface (QQQ) running the identical S045 fly (ATM shorts, 0.4% wing, 09:35 parity-centered, 25% PT, hold-to-settlement) diversifies the SPX fly's idiosyncratic session noise, lifting the risk-adjusted return of a combined SPX+QQQ book (the classic √(2/(1+ρ)) Sharpe benefit). Motivated by the QQQ-enrichment probe (result record: QQQ is the only non-SPX surface that clears the crossing cost) and the Rajamony/Hoffstein podcast's reminder that diversification is the only true free lunch — but never trust a low correlation; check the tail (the correlation floor).

How it earned a forward slot

QQQ standalone (SR +0.45 gross, worst session only −$115/contract) looked suspiciously clean. Checked, honestly: - Not a data error — it's scale. Worst QQQ session −$115 ≈ SPX worst (~−$1,500) ÷ 13 (QQQ ≈ 1/13 the size). The bounded tail is real: a 0DTE fly is defined-risk with no overnight gap (09:35→16:00), so the loss is capped by the wing by construction. The 2024/2025 vol spikes (Aug-24 unwind, Apr-25 tariffs) are in the sample and hit it, but stayed bounded. - But the SR level is regime. By year: 2022 +1.05 (n=57), 2023 +0.33, 2024 +0.22 (weakest), 2025 +0.37, 2026 +0.65. Everyday SR is a modest 0.2–0.4; the +0.45 headline is lifted by late-2022 and 2026-YTD. - No genuine correlated crash in the 2022-26 window — it was an exceptional low-vol tech melt-up. The low ρ (0.03 OOS) is partly regime; in a real co-crash both sleeves lose together (both bounded, but simultaneously). The diversification smooths the everyday path more than it protects the true tail. - Unmodeled QQQ frictions: QQQ options are American-style / physically settled → assignment & pin risk on ITM finishes (SPX is cash-settled — none); the flat 4% cost hides spread blow-outs on stress days (2022 was 8–20%); mid-fills are optimistic. Net of these the real edge is thinner than +0.45 (already +0.34 post-commission). - Takeaway: a real signal, but use it as a diversifier (the correlation benefit that smooths the everyday path), not a "QQQ beats SPX" claim — and size to that benefit, not the flattered SR (hence 5, deliberately below the in-sample optimum).

Gate frozen by the owner on 2026-07-14 (window = 63 forward sessions; moderate strictness; QQQ silent capture-only). The design below is now locked; the forward window has not started — it begins on the first captured session once the collector is deployed. Do not edit these values retroactively (a change = a new identifier, per CLAUDE.md §2).

Pre-registration

Why forward, not a historical holdout. All 819 historical common sessions (incl. the 2025-26 "OOS" split) were examined during exploration — no untouched block remains. Per CLAUDE.md §6 the only 0DTE data is intraday history and it is now seen. The genuinely clean, unseen block is therefore the forward period: the confirmation is a pre-registered forward-paper test, run alongside the S-045 SPX fly. (This is the same

Amendment (2026-07-14, pre-window — nothing collected yet, so no new identifier per §2A): the QQQ sleeve closes at 15:55 ET, it does not hold to settlement. QQQ options are American-style / physically settled, so holding into the cash print risks assignment / exercise / pin (you end up long-or-short QQQ shares with an overnight gap). SPX (European, cash-settled) keeps hold-to-settlement. This is the last clean structural change: once the window opens, a further change to the exit is a new identifier.

Frozen design (fixed before any forward session is scored): - Structure: the identical S-045 fly — ATM shorts, 0.4% fixed wing, 09:35 parity-centered, 25% profit-take. Exit: SPX sleeve hold-to-settlement; QQQ sleeve exit 15:55 ET (physical-settlement avoidance). No active price stop on either. Threshold-free — there is no tunable entry parameter to overfit. - Surfaces: SPX + QQQ, every regular full session (holiday-aware; skip half-days per the trading-calendar guard). - Sizing: 1 SPX fly + 5 QQQ flies, frozen (the exploration's choice, deliberately below the in-sample SR optimum; ≈36% QQQ risk-share). - Cost model (fixed, identical both surfaces): the measured ~4% crossing cost per surface + flat commission ~$1.15/contract, open-only ($0 to close). No flattering fills. Entry (09:35) and QQQ exit (15:55) are charged the realized net spread from the captured bid/ask (not just the 4% proxy) — see below. - Data capture (forward): SPX via the existing spx_0dte_chain_capture; QQQ via qqq_0dte_chain_capture, same 5-min cadence extended to 15:55 ET so the exit mark and its spread are captured. Every snapshot stores per-leg bid/ask/mid → the forward window is the spread study. Combined-book P&L computed offline.

Statistics computed (fixed): per-session net P&L of each sleeve and the combined 5:1 book; combined per-trade Sharpe vs SPX-alone; max drawdown ($); per-session P&L correlation + the correlation floor (ρ on the worst-decile SPX days); both-sleeves-independently-positive check. Plus — for production sizing of the fill margin: the realized net-spread distribution (p50/p75/p90) at 09:35 entry and 15:55 exit, so the production limit-walk margin is set from measured spreads, not a guess. The go/no-go charges a conservative (p75) realized crossing at both entry and exit (no assumed price improvement on the forced exit).

Gate — FROZEN 2026-07-14 (moderate). A pass requires all of — 1. Forward window: the gate reads after 63 forward sessions (≈ one quarter). Window starts on the first captured session (TBD on deploy). 2. Diversification helps (moderate margin): combined net per-trade Sharpe > SPX-alone by ≥ 10% (relative) over the window. (Operationalizes the owner's "moderate positive margin"; adjustable only until the window opens, then locked.) 3. QQQ carries its weight: both sleeves independently net-positive over the window. 4. Correlation floor holds: worst-decile-SPX-day tail correlation < +0.35. 5. The book works: combined book net-positive over the window. - Pass → promote to a live-tracked dual-surface cell (still paper, still not sold). Fail → logged honestly in the open; QQQ diversification not promoted. - Signals: QQQ is silent capture-only for the whole window — no Telegram signal (S-045 SPX keeps its signal). A QQQ signal is considered only on a pass.

Forward record so far

Forward paper-trading against the frozen gate is under way; the record here is machine-generated from the live track and updates on its own. It is interim — no result is claimed until the gate's window completes.

The live forward track has not yet accumulated a reportable window — the frozen promotion gate below states what it must show.

Conditions & caveats

• The dollar tail and drawdown are NOT reduced. CVaR₅ is ≈ flat (−$1,491 at 5:1 vs −$1,431 SPX-alone; −$1,840 at 13:1) and the dollar maxDD is slightly deeper — adding QQQ deploys more total risk, so the win is a better risk-adjusted return + the correlation floor, not a lighter worst case. (The earlier σ-unit "lighter tail" read was an artifact of risk-normalisation; the earlier "−12% drawdown" read was an artifact of one QQQ session — see Correction 2 in the banner.)
• The +88% OOS lift is inflated by a weak SPX denominator (SPX-alone SR 0.07 OOS). The real read is that the 5:1 book held ~0.13–0.23 across both regimes while SPX-alone collapsed.
• Non-stationarity / regime risk (Rajamony's point): the low OOS correlation and QQQ's recent strength are a 2025-26 regime; if tech/QQQ reverses, the decoupling may not persist. Don't assume ρ stays ~0. Sizing 5 (not 13) deliberately limits this exposure.
• Per-trade Sharpes are annualisation-free and in-sample-flattered — the RELATIVE lift + the OOS survival are the read, not the levels.

The promotion gate

See the specialist appendix below.