Skywatch · Analytica DSS

Track record

Every call the monitor has made across the walk-forward back-test, graded against data it never saw — hits and misses both, because a record you can prune is not a record. The calls made on the live site, open and graded, are under “Live calls” below: the open ones are what we're on the hook for right now. Everything on this page is read live from the same committed artifacts the atlas runs on.

The headline, honestly denominated

The named ledger — every top-5 call, graded

A HIT means the place was still running anomalously hot in the three months after the call — months the model had never seen. The next-window ratio is what its volume actually did against its own expectation. Misses stay on the page; they are most of it, which is what a single-digit precision against a ~3% base rate looks like when you write down every row.

as ofplacezgradednext-window ratio

Forecast bands vs what happened

The final 12 months were held out; the model forecast them; here is the 80% band against the count that actually arrived.

Live calls — made on real data, graded in public

Everything above is a back-test: the rule re-run on history. These are different. Each is a call the monitor published on the live site, recorded to an append-only ledger the moment it was made, and graded by the back-test's exact rule once the following three months of FAA data land. Nothing here can be tuned after the fact, which is what makes it the record that counts.

What we withdrew

Every finding we published and later retracted is listed here; each links to the methodology section that explains what re-examination showed. Four of the corrections made the product look worse.