Searching for an edge in same-day SPY options
Is there a rule, known before the trade, that picks the days on which buying or selling same-day SPY options pays after real bid/ask costs?
Still running. The main test asked whether a morning label (a calm, quiet open versus a trending one) picks good days for a same-day iron condor or a same-day debit spread; in the first, discovery part of the data neither passed, and the label picked worse days than chance for both. Two pre-registered searches of option-order flow (stock opening put/call buying, and index option flow in the first half hour) found no usable information in discovery about where SPY or the stocks went next; their validation was not read. Related tests also did not pass: an every-day condor (one counted look), an intraday breakout spread, and an overnight condor and put spread (the breakout and the overnight pair were run in discovery only). No paper or live trading was started. Trials on other underlyings are registered and waiting for data. The flow families' holdout is forward-only: any flow trial that passed validation could only be confirmed on sessions after 2026-11-20, unless the owner rules otherwise.
Backtested by QS Research
Pre-registration
Pre-registered and approved on 2026-11-19, before any option price for the rule was fetched. Each morning at 10:00 ET a session is labelled from facts known then (event days, a hidden-Markov volatility state from earlier daily bars, the opening gap, and how much of the day's priced move the first half hour used, with the gap and that share each ranked against its own previous sixty sessions): a quiet-open label trades a same-day SPY iron condor, a trending label trades a same-day debit spread in the open's direction, and the rest are skipped. Both trades are priced at the real bid/ask with commissions, exit at 15:45 ET, must pay on the labelled days, and must beat the same structure and rule on every other tradable day (days not skipped for an event, a half day, a crash state or a missing quote). Sessions are split in time into discovery, validation and holdout, with a month-clustered bootstrap and a deflated Sharpe check; every trial counts whatever its result. Later amendments added two flow-filter families judged on the next move of the stock or index, fixed before any of their numbers were computed.
Registered 2026-11-19Amendments
- 2026-11-19 A model of dealer gamma rebuilt from open interest, added as a reported breakdown only and never part of the label; not run for a failed label. approved: Decided 2026-11-19 on the owner's delegation, before any option price was fetched
- 2026-11-19 Five option-flow filters fixed as one counted family, judged on SPY's move after the open, before any print was joined to a return. approved: Recorded 2026-11-19, before any flow print was joined to a return
- 2026-11-19 Two premium bars in the flow filters set to the steps the data source actually offers. approved: Recorded 2026-11-19, before any return
- 2026-11-19 One session scored by hand before the run is disclosed and excluded from every judged statistic. approved: Recorded 2026-11-19, before the run; a disclosed prior look
- 2026-11-19 The flow family replaced, before it was run, by the five layouts actually watched on screen; the earlier five filters withdrawn; a flow holdout window sealed. approved: Decided 2026-11-19 on the owner's delegation, before the flow family was run
- 2026-11-20 Two new flow families fixed before any of their numbers were computed: stock opening put/call buying (long-short), and index option flow in the first half hour against SPY's move to the close. approved: Approved 2026-11-20, before any of its numbers was computed
- 2026-11-20 The half-year consistency rule changed to two of the three discovery half-years (discovery spans only three); ties and single-leg prints defined. approved: Recorded 2026-11-20, before any of its numbers was computed
- 2026-11-20 Bookkeeping: where the flow search code lives; the tests are unchanged. approved: Recorded 2026-11-20; bookkeeping, no rule changed
- 2026-11-20 How the main test is run, fixed before any label or return was computed: a single volatility-state fit on data before the window, the warm-up period, strike rounding and quote validity. approved: Recorded 2026-11-20, before any label or return was computed
- 2026-11-20 Bookkeeping: the new data inputs the runs read are published as pinned, checksummed snapshots. approved: Recorded after the runs; bookkeeping, no rule changed
- 2026-11-20 A separate study read the flow holdout window, so under the pre-registered rule that holdout is forward-only from 2026-11-20, unless the owner rules otherwise. approved: Recorded 2026-11-20 as the pre-registered holdout rule requires
A holdout sample is sealed: its results stay unread until the study is final.
Limitations
- One underlying is judged (SPY); SPX and QQQ move with it, so they could only be checks, not independent evidence.
- The label's low and high thirds are ranks against the previous sixty sessions, so in a long quiet or wild stretch the thresholds drift with the stretch.
- Prices are one-minute quotes at the stamped minute, not fills; a four-leg order at the natural price (sell at the bid, buy at the ask) can be pessimistic for a patient order and optimistic in a fast market.
- The option sample covers late 2022 to late 2026, a single volatility cycle, and part of that window had already been seen in an earlier look at selling the at-the-money straddle.
- The flow data is a vendor's selection of notable prints, not the full tape, and its selection rules may have changed over time.
- The flow families' holdout (2026-06-01 to 2026-11-18) was read on 2026-11-20 by a separate study that joined the same flow prints to the same SPY returns. Under the pre-registered rule, that holdout is now forward-only: a flow trial that passed validation could only be confirmed on sessions after 2026-11-20, unless the owner rules otherwise.
- The main test's later sessions (validation and holdout) were not read by it, because it failed in discovery; a separate one-look test of an every-day condor later used those sessions (with the 2022-11 to 2023-02 warm-up), so the main test's holdout is no longer blind.
- One session (2026-11-19) was scored by hand before any run; it is excluded from every judged statistic.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 8bc4f791d3f80f45, published 2026-11-21