Every price in this system is a traded price, and that changes which tests are even meaningful

The constraint

This warehouse is not entitled to option NBBO. The REST quotes endpoint and the quotes_v1 flat file both return 403, and so do all index levels (I:SPX, I:VIX). There is no bid, no ask, no mid and no displayed depth anywhere in the data. The only price available for a contract is its last trade, falling back to the day close.

Entitlements are probed rather than assumed: a job walks every endpoint the system relies on and exits non-zero on drift in either direction, because a capability silently gained is as much a change to reason about as one lost.

What I expected

I expected this to be a precision problem. Without a midpoint, the inverted volatility carries an unknown offset bounded by roughly half the spread, so I expected to characterise that bound once, carry it as an error bar, and otherwise proceed as normal.

What happened

The offset is real and is bounded roughly as expected. What I did not anticipate is that the missing quote breaks a validation step rather than only degrading an estimate.

The daily canary checks three identities on at-the-money names, where at-the-money means |K/F − 1| ≤ 5%:

• Reprice. The stored engine price must reproduce the input price. Independent for American-style SPY rows, which are priced on a binomial tree while the check runs against the stored value.
• Pair symmetry. For European pairs, call and put gamma must agree, and call and put vega must agree, at a shared volatility.
• Put–call parity. The call minus the put must equal the discounted forward less the discounted strike.

The third one cannot be run when either leg carries a last-trade price. Put–call parity is a statement about a call and a put observed at the same instant. Two last trades are two different instants, sometimes hours apart on an illiquid strike, and a parity violation computed across them measures the gap between the prints rather than any error in the pricing. So the check is skipped for those rows rather than being run and quietly failing.

Diagnosis

There is a general lesson here that I had to be shown rather than reason my way to. A missing input does not degrade every consumer equally. It degrades estimates smoothly and it breaks tests discontinuously, because a test is usually an equality that holds under an assumption the missing input was quietly supplying. Here that assumption was simultaneity.

Two related decisions follow from the same reasoning.

Vendor implied volatility and Greeks are copied into the warehouse but are never inputs. They are unit-aligned for comparison and written into the daily report as diagnostics. A null vendor volatility is not a failure and does not fail the job. The reason is that a vendor number computed from a quote midpoint I cannot see is not a check on my arithmetic; it is a second, differently-wrong number, and treating it as ground truth would hide exactly the errors the canary exists to find. The live test is my own math on new prices.

Inversion raises rather than returning a NaN. A price outside the discounted no-arbitrage bounds has no implied volatility, and Newton with a Brent fallback will not find one. Returning a NaN would let that row flow downstream and be dropped silently somewhere later, at which point the dataset has a hole nobody chose. Raising forces the decision to be made where the information is.

The forward is recovered from put–call parity rather than from a spot price, since no index level is entitled either. Pricing then happens in the forward measure, which reduces Black–Scholes–Merton to Black-76 — a spot and a dividend yield are precisely the two things this tier does not provide, so the parameterisation that avoids needing them is the correct one rather than a workaround.

What this forbids downstream

The site previously claimed the cleaning layer dropped stale and crossed quotes and filtered by spread width, and that the cost model charged half-spread crossing scaled by quoted width and displayed depth. Crossed and locked are definitionally bid-and-ask states. Quoted width and displayed depth are NBBO fields. None of it was implementable on this tier, and the claim has been removed rather than softened.

The cost consequence is the one that matters, and it is worse than an approximation. In options the spread is frequently wider than the entire modelled edge, so a cost model that cannot measure the spread cannot be calibrated at all — it can only be bounded. A backtest published under those conditions would be reporting the size of its own assumption. That is why no backtest result will be published under this protocol until the input situation is resolved and stated on the result itself.

What changed as a result

The parity check now skips rather than fails when either leg is a last trade, and the skip is recorded rather than silent. The canary's pass/fail rests on median reprice error and the fraction of at-the-money names failing identities, not on far-out-of-the-money ticks, which are noise rather than signal. The public pages now distinguish the hygiene that runs on entitled inputs from the hygiene that is blocked behind an entitlement, and the build-status page carries that distinction as a separate field so that a reader can tell unwritten code from an unavailable input.

Every diagnostic this project publishes will state the kind of input it was computed from. Traded-price analytics and quote-based analytics are not comparable, and a report that omits which one it is invites the reader to assume the better of the two.