Keep a history of what was known
A production slowdown reported in July may look less severe in the database downloaded in December. Both numbers can describe precisely the same quarter. In the US accounts, the advance, second and third estimates draw on progressively more information. Later updates can alter the history again. Keep a table with one row per reference quarter and one column per publication date. Preserve an empty cell when no observation was available; filling it retrospectively is itself a forecast.
Reconstruct the signal
Suppose a team investigates a rule triggered by real growth below 1.0% at an annual rate. An illustrative advance reading of 0.7% triggers the rule; a third estimate of 1.2% does not. A backtest built from the latter records a decision the team could never have made. Rebuild the signal from the data column available at the decision time. If using growth calculated from levels, take both adjacent quarters from that same column.
Distinguish forecasting from explanation
For a forecast review, score the prediction against the first release and then against the later estimate, with each target identified. The first comparison evaluates a genuinely observable event; the second addresses the subsequent view of activity. Do not compare an annualised US quarterly change directly with an unannualised European one. A large revision can move a turning point without invalidating the original reasoning or proving that a revised series is definitive.
Practical check
Record the release timestamp, growth convention and vintage for every signal, not only for quarters that were subsequently revised.