Macro Signals Need the Number That Was Published, Not the Revised One

Many macroeconomic series are estimated, published, and then revised as more complete source data arrive. Output, employment and income figures can move between the first release and later vintages, and benchmark revisions can redefine a series. The current database shows the latest estimate for every past period: the right input for describing the economy, and the wrong one for simulating a decision made at the time.

Consider an illustrative rule that reduces equity exposure when quarterly output growth turns negative. Run on the latest vintage, the rule reacts to quarters that are negative today. Some of those quarters may have been first reported as positive and revised later, and the reverse can also happen. The two versions are different strategies, and only one of them could have been followed.

Vintage archives exist for this purpose. The St. Louis Fed's ALFRED lets users retrieve each data release that was available on a specific date in history, and the Philadelphia Fed's Real-Time Data Set for Macroeconomists provides vintages, or snapshots, of major macroeconomic series. With such sources, a backtest can look up, for each decision date, the last vintage published before it.

Publication timing matters as much as the value. A monthly figure for March becomes known on its release date, not on the last day of March. The release calendar belongs in the data model; a series aligned to its reference period instead of its release date carries a look-ahead of weeks.

Real-time and latest-vintage results can both be useful when labelled. The real-time run estimates what could have been traded; the latest-vintage run shows how much of the result depended on information that arrived later. A large difference indicates that the signal leans on revisions and supports a narrower claim.

Where no archive covers a series, the limitation should be stated rather than hidden. A cautious design might lag the input by more than its usual revision window, test sensitivity to that lag, and avoid presenting the outcome as fully point-in-time. The same discipline applies to central-bank statistics as to national accounts.