Metrace

Method

We work from what was actually published, on the day it was published. Every value in our files can be traced to an archived page.

Sources

We pull the full list of archived captures for each source site from the Internet Archive's Wayback Machine, then fetch the pages, PDFs and data files that matter. We also read live sites directly: government search portals and open-government catalogues fill in documents the archive missed.

Reading

Language models read each archived version. For statistics offices, two detectors propose candidate events: one watches the stream of file names and folders, the other compares the text of successive releases and calendars. For bank rate pages and rate cards, a parser reads every table and keeps the effective date the bank stated.

Checking

Two model reviewers code each candidate independently against a written manual. A third model settles disagreements. Every event we publish is confirmed by a source outside the office, then a separate agent tries to refute it.

Dates

We keep three dates where they exist: when the change took effect, when the source announced it, and when an archive first captured it. A capture date is evidence that something was public by then, not the release date.

Corrections

When a source revises a value or a page changes, we add a new row and keep the old one, so any past date can be reconstructed as it looked then.

Validation results

CheckResult
Known statistics events recovered, list fixed before the pipeline ran17 of 22
Agreement between two independent model reviewers, 100 cases87% (kappa 0.755)
Further events found and confirmed by an outside source47
First 2022 deposit rate rises dated as in news reports7 of 7 banks
State Bank of India rate changes recovered against the bank's own history file13 of 16, no false dates
Arrears reports found in a full listing of one Shenzhen agency's catalogue4 of 4

Request a sample

Tell us which countries, banks or series you cover. We will send a sample file and the field guide.

Email us