How accurately can a complex option trade be signed?
The production package signer, graded per leg against the exchange's own record
To build a dealer book from the tape you must decide, print by print, who bought. For a simple order the quote rule does that — a print at the ask was a buyer — and we have graded that rule against exchange truth before. But on the day measured here 40.2% of all SPX option prints were legs of multi-leg packages, and on those legs the quote rule is not noisy — it is invalid by construction. An exchange fills a spread at a net price and then allocates that net across the legs by convention; a leg can print anywhere inside, or outside, its own market regardless of who initiated the package.
The terminal's answer is to sign the package: legs that print on the same millisecond are reassembled into their parent order, the package's net price is compared to a net bid/ask derived from the leg quotes, and the quote rule is applied at that level — with the one free choice, the sign pattern of the structure, calibrated once from 1.9 million packages (a vertical trades on the +− orientation, an iron condor on +−−+, and so on). The idea is not unique to us; what has never existed, for ours or anyone's, is an accuracy number. Academic work that reconstructs complex orders validates the reconstruction, never the sign.
The truth set
Cboe publishes a free trade-by-trade sample: 3% of executions, and for each one the side, the capacity of the participant (customer, market maker, firm), the trade type (who removed liquidity, who was resting, who initiated an auction), and an identifier tying together the legs of one complex execution. That is per-leg ground truth. Join our tape to it on (second, expiry, strike, right, price, size), keeping only keys unique on both sides, and 12,422 complex legs carry a verdict.
The number
| rule, graded on customer legs | per print | volume-weighted |
|---|---|---|
| package signer, as shipped | 80.4% | 79.1% |
| excluding legged-in contras (unrecoverable in principle) | 87.9% | 84.7% |
| single-leg quote rule on the same legs | 75.1% | 75.1% |
| coin flip | 50% | 50% |
Two comparisons put the 80–88% in place. First, the single-leg rule on the very same legs — the thing a feed that ignores package structure implicitly does — reads 75.1%. Second, our own simple-leg signing, graded against the same kind of truth, reads 74.5% volume-weighted: the package legs, supposedly the unsignable part of the tape, are now signed more accurately than the simple prints. By structure: iron condors 88.2%, strangles and risk reversals 91.0%, butterflies 82.1%, verticals 78.1%. Even the legged-in category — packages executed leg by leg against the simple book, which Cboe itself declines to attribute — comes out at 69.1% rather than the coin flip we had budgeted.
The mirror check
A signing rule that claims to find the customer should be reliably wrong about the market maker on the other side. It is: on legs where the sampled participant is a market maker, the rule agrees with that side only 10.9% of the time — it books the opposite of the MM in about nine cases of ten, which is exactly what a customer-versus- dealer book wants.
What the rule actually measures
One finding rearranges how the rule should be described. Graded against time-priority truth the package rule fails: against legs whose trade type marks the true remover of liquidity it scores 16.9% — inverted — while against complex-auction initiators it scores 90.4%. Both make sense at once only one way: the package quote rule identifies the side that paid the spread against the leg-derived net mid, not the side that arrived second. A resting customer spread still rests at a price that gives the market maker edge, so the edge-payer is the customer whether they added or removed liquidity. For a book whose object is customer-versus-dealer positioning, that is not a bug in the rule; it is the reason the rule works.
Honesty section
One truth day (the sample is free; more days are not), so the binomial standard error on the headline is about 0.7 points, and day-to-day variation is unmeasured. The signer abstains rather than guesses: 54.7% of customer complex prints (49.3% of their volume) get a sign; the rest — packages printing at their net mid, or structures whose orientation the calibration cannot identify — are dropped, not mis-signed. Trades between two customers have no single customer side, and matched keys sampled from both sides of one execution are excluded by construction. All of this grades the signer we ship, on the day the truth exists for; it is a validation, not a theorem.
Reproduce it
The truth file is Cboe's C1 options trade-by-trade 3% sample, a free download from Cboe DataShop. The other half is the full options tape for the same day — per-print prices, sizes and the prevailing quote — which is licensed data; in our stack it comes from ThetaData. Reassemble packages by exact print timestamp, derive the package net bid/ask additively from the leg quotes, sign the net price against it, propagate the package sign to the legs, and grade exactly the matched legs on capacity-customer rows. Every threshold used here is stated in the text. The book this signing produces is the measured book on every archived session — the terminal's MEASURED view is exactly these signs, accumulated.
Part of gex.live research. Measured on the free session archive; every session is free to replay.