Economic data are measurements or estimates of economic activity, prices, employment and other economic conditions. A quant may use them to forecast a variable, describe a market regime or test how asset prices respond to an announcement. The useful input is more than a number: its meaning, reference period and availability history determine whether it belongs in a particular decision.
A monthly output estimate, an unemployment rate and a business survey index measure different things. Even two series with similar names may cover different populations, geographies or price bases. Before fitting an econometric model, establish exactly what each series measures.
One observation, several clocks
An observation labelled “month M” does not imply that its value was known during month M. Keep these fields distinct:
| Field | What it tells you |
|---|---|
| Observation period | The month, quarter or other interval the measurement describes |
| Publication time | When the source made this version public, with timezone |
| Receipt or usable time | When the research or trading system received it and could use it after any required processing |
| Vintage | The version of the history being used, such as the values available as of a specified release or archive date |
| Measurement definition | Series identifier, units, population, frequency and adjustment conventions |
A vintage can revise several earlier observation periods at once. It is not another name for the period being measured. The St. Louis Fed's ALFRED documentation describes historical vintages and the hierarchy used to establish release dates; its real-time-period documentation distinguishes today's information about the past from information known at an earlier date. A date-level archive alone does not establish a strategy's exact intraday receipt time.
The first release stays in the first decision
Consider a fictional, seasonally adjusted monthly economic index, measured in index points. Its previous-month level is 100. On day R, the source first publishes month M at 105 at 08:30:00 UTC. Our hypothetical system receives it at 08:30:02; assume no further processing delay.
A decision at 08:30:01 cannot use 105 merely because the source has already published it. A decision at 08:30:03 can. Using the previously available denominator of 100, its month-on-month percentage change is
On a later day S, the source revises month M to 106. Assume the previous-month level remains 100 in that later vintage. The revised growth rate is then 6%. That is a valid description of the later vintage; inserting it into the day-R decision creates look-ahead bias. Revisions to the denominator would also require recomputing the change using the selected vintage's denominator.
Original synthetic index and hypothetical UTC event times; the diagram is not to scale. The previous-month value stays 100 in both vintages, and processing after receipt is assumed instantaneous.
Preserve the original records and their version history. Adding day-S information should not change a frozen day-R input. The Point-in-Time Data lesson develops this selection rule and snapshot audits; the full lesson requires membership, with a public preview.
Read the units before the headline
The index level of 105 is neither a growth rate of 105% nor a tradable price. The move from 100 to 105 is 5 index points and, in this particular example, 5%. Those numerical values coincide only because the starting index is 100. A difference between two reported percentage rates is measured in percentage points, not index points.
A first difference subtracts levels; a percentage change divides by the earlier level. The BLS calculation guide explains the distinction for consumer price indexes. State whether a change is month-on-month, year-on-year or annualised: these are different transformations. The example's 5% is a one-month change, with no annualisation. A percentage-change calculation also needs a meaningful, nonzero denominator; applying it mechanically to an index that can cross zero can be misleading.
Check adjustment labels just as carefully. Seasonal adjustment attempts to remove recurring calendar-related movements; it does not remove every unusual event or turn the result into a price-adjusted “real” series. Nominal versus real and seasonally adjusted versus unadjusted are separate choices. Adjustment estimates can themselves change: BLS documents revisions to seasonally adjusted CPI histories when it updates seasonal factors. Use the adjustment and vintage that the research question requires.
What question is the dataset answering?
A study of announcement reactions may need the first published number, an expectation recorded before release and an accurately timed price response. The surprise is the release relative to that prior expectation; a high level or a positive growth rate is not automatically a positive surprise. Neither establishes a universal direction for market returns.
A forecasting study must separately specify its input vintage and its target: the first estimate, a later estimate or another explicitly dated outcome. A later revised estimate can be a declared evaluation target without being an eligible earlier input. Calling the latest history “the data” hides that choice.
Save the source identifier, definition, publication and receipt clocks, selected vintage, transformation and missing-data policy alongside the model input. If the required release has not arrived, the value is unavailable; a later revision cannot fill that historical hole.
Sources
- Federal Reserve Bank of St. Louis — ALFRED Help: vintage histories, release dates and growth transformations.
- Federal Reserve Bank of St. Louis — FRED API Real-Time Periods: historical information sets.
- U.S. Bureau of Labor Statistics — Calculating percent changes: index levels and percentage changes.
- U.S. Bureau of Labor Statistics — Seasonal Adjustment in the CPI: adjusted series and revisions.
- Chris Brooks, Introductory Econometrics for Finance, second edition, Cambridge University Press, 2008, §§1.2–1.3.1, pp. 2–3: data types, measurement error and revision vintages.