A credible forecasting archive is designed before the outcome arrives. It records the complete question, the time at which the forecast was made and the probability or other prediction being evaluated. The point is to preserve what was actually known and claimed, rather than reconstructing a successful-looking version afterward.

Retain the losers and unresolved entries

An archive that keeps only correct calls is not a performance record. Keep the full sample, including failures and entries that have not yet resolved. If an observation is excluded, state a reason that does not depend on whether the result was convenient. A consistent inclusion rule makes the record easier to audit.

Define the outcome in advance

Specify the official resolution source and how ambiguous or revised results will be handled. If the question uses a first release, save that release. If it uses a later estimate, identify which one. Resolving a forecast with whichever value supports the preferred conclusion defeats the purpose of the archive.

Record revisions without rewriting history

A forecaster can update a probability when information changes. The archive should preserve the earlier entry and add the new one with its own timestamp. This permits both an evaluation of the initial forecast and an account of how views developed. Silent overwriting makes those tasks impossible.

Assessment should then distinguish calibration, accuracy and economic value. A forecast can be useful to understanding an event without implying a profitable transaction after fees. Closely related outcomes can also limit how much independent evidence a set of forecasts provides. Our research notes use these distinctions to keep the record honest: a specific question, a contemporaneous observation and a documented outcome, even when the final result challenges the original narrative.

Sources & context

BEA revision documentation, illustrating the importance of estimate vintage ↗

Factual references checked October 8, 2026. Interpretive discussion is editorial analysis. Examples are hypothetical unless explicitly identified otherwise.