Data & methodology

What Signal is built on.

Every figure in Signal resolves to an archived public source, carries a label saying how it was produced, and is served from a pipeline whose coverage you can read below — live, from the same contracts the product runs on. Nothing on this page is typed in by hand.

Coverage, measured by the pipeline

The dataset, as of right now

Days of history
loading
Trading intervals held
30-minute resolution
Archived source objects
byte-level, immutable
Provenance completeness
every interval traceable

The provenance standard

The Ridgeline Provenance Standard

Signal never asks you to take a number on trust. Every figure wears one of three badges, and the badge is binding — the product will render “not specified” before it will guess.

Behind the badges sits the archive: every source file the pipeline ingests is stored content-addressed, at the byte level, immutably — so any figure Signal has ever shown can be traced back to the exact bytes its publisher served that day. When a source publisher revises its data, Signal's derived figures change and say so; the original remains in the archive.

Licences

Licensed public data, attributed as its publishers require

Every source Signal fuses is licensed for commercial use and attributed as its publisher requires. The full register — source by source, licence by licence — is maintained at Data sources & licences. No source or government body endorses Signal.

Validation

The modelled figures are gated behind a pre-registered test

Where validation stands today. Signal's counterfactual engine is graded against a backcast protocol that was pre-registered and locked before any result existed — the bands, metrics and grading tiers were fixed first, and were not moved afterwards. The protocol has been run once, and the run-once rule means the result stands. It is published inside the product, predicted-versus-actual per metric, whatever it scored — the error is part of the product. By a deliberate ruling recorded in the product's decision log, the result is presented as a partial validation: direction-reliable, magnitude-unreliable, so the briefing shows bounded ranges, never point estimates. Forward scenarios inherit that ceiling — forward confidence is always capped below the backcast's, and no forward view is ever a single number. Accuracy figures are not quoted on this page; subscribers see them in-product, alongside the notes needed to read them honestly.