The question
Amazon publishes a 90-day history of EC2 spot prices per availability zone (AZ). Within one region, several AZs price the same instance, and their prices move together. The sharp question: when one AZ’s price reverts toward the others, is that genuine cross-AZ coupling — the zone adjusting because it diverged from its neighbours — or is it just each zone independently tracking a shared regional factor (region-wide GPU demand, a common pricing rule) that happens to move them in step? The first is a structural finding about price discovery; the second is the null. Telling them apart is the entire task.
What we did
We verified the source on real data (a 14-day probe of g4dn.xlarge / Linux across five us-east-1 AZs), pre-registered the coupling test blind, and built an offline self-test harness before any analysis-window data was examined — the discipline the whole program runs on. The primary control was a geography placebo: compare each zone’s reversion toward its own region’s consensus against its reversion toward a different region’s consensus, on the reasoning that only genuine same-region coupling should survive the difference.
What the harness found — before any data run
The harness is built to try to break its own positive result on synthetic controls, and here it did, three times over. A naïve circular-shift null false-detected under a shared drift; a single foreign-region placebo false-detected under differential regional drift and collapsed when the foreign region was sparse. Each looked patchable. The third finding was not.
At a realistic per-block sample size, an independent no-coupling panel — zones that only track a shared factor, with no cross-AZ coupling built in — false-detects as coupled against a realistic foreign region, every time. The diagnosis is structural, not a bug: for a handful of zones sharing a regional factor, “reverts toward the others” and “reverts toward the shared factor” are the same observable (their correlation runs about 0.87). A foreign region is a weaker proxy of a zone’s short-term reversion target than its own region’s consensus, so the difference doesn’t cancel — it manufactures apparent coupling. Move the other way and use the exact shared factor as the reference, and it over-corrects, erasing coupling that is genuinely there. There is no reference in between that a blind design can pick without already knowing the answer.
This is a textbook reflection problem: an endogenous “because of the others” effect and a correlated “shared influence” effect are not separable from contemporaneous cross-sectional data alone, absent exogenous variation. The geography placebo does not resolve it.
What this is — and what it is not
We closed the domain here, methodology-grade. The finding is an identification failure: the design cannot answer the question, and we established that before collecting the analysis window, so no result — true or false — was ever shipped.
To be exact about the claim, because the temptation runs both ways: this is not a finding that cross-AZ coupling is absent, and it is not a claim that these zones are efficiently or independently priced. The data rules coupling neither in nor out; it cannot separate the two mechanisms. AWS sets spot prices per zone algorithmically, and nothing here speaks to coordination of any kind — this is a statement about what a measurement can and cannot distinguish, and nothing more.
The reason to publish it is the reason the program exists. An instrument that will only report a signal it can actually identify — and that refuses, pre-data, to convert a shared drift into a false discovery on a question it cannot answer — is the product. This is that instrument declining a question it was honest enough to recognise it could not answer.