Keep the Revision History of News and Economic Events
By DX Research Group · · Market data
A revised-release fixture preserves the earlier agent input while allowing a later corrected account of the event.
Store an event's versions, including the original text or value and the revision that superseded it. An agent can act coherently on an initial report that later proves inaccurate. A replay built from today's corrected event history would judge a different information set. We would preserve both the historical decision input and the latest interpretation, with separate queries for each.
FRED's real-time-period documentation explains that observations can be revised and that historical vintages retrieve what was known in an earlier period. The same versioning principle is useful when designing collectors for issuer releases and news corrections, whose source contracts must be checked individually.
A correction changes the sign of a feature
Assume a fictional economic release, an expected value of 1.5%, and complete capture of two published versions. At 09:00 the source reports 2.0%. At 09:20 it corrects the value to 1.0%. An agent decision at 09:05 receives the first version.
| Version | Reported value | Surprise against 1.5% expectation |
|---|---|---|
| Initial | 2.0% | +0.5 percentage points |
| Corrected | 1.0% | -0.5 percentage points |
The correction reverses the sign. Both values should remain in storage. The earlier turn references the initial version, while a later research account can say that the source corrected the event. This is a synthetic data-contract example, rather than evidence that a particular agent traded a release.
Preserve what changed and who changed it
Use one event identity with append-only version records. Each version references its source, captured payload, predecessor, and changed fields. Distinguish a publisher correction from a collector repair. Fixing a parser that read “million” as “billion” changes the local representation; it does not establish that the publisher revised the release.
For prose, retain the material excerpt under the applicable retention policy and a content digest. A digest detects a changed payload but cannot explain the change after the original text is gone. Store structured differences when they affect features such as direction, amount, or effective date.
Corrections can also retract an event entirely. An event-state model should express retracted, superseded, and unresolved versions, allowing later consumers to handle the new status. Keep historical references stable so the original decision remains inspectable.
Two queries deserve different names
A historical-input query returns the version available to the decision collector. A latest-evidence query returns the most recent supported interpretation. Defaulting every research tool to latest evidence can silently alter old features when a dataset is refreshed.
A proposed fixture checks that the 09:05 turn keeps +0.5 percentage points after the correction is ingested, while a 09:25 turn receives -0.5. Our frozen replay note describes binding the resulting payload. Our observations and rationale note separates the recorded source claim from the agent's conclusion. Correct version reconstruction supports fair behavior evaluation; it leaves the economic value of reacting to either version as a separate question.