Information Timing, Data Vintages, and Real-Time Macro Data
Auditing what macro information was actually available at each forecast or portfolio decision
QM014 · Data & Research Design · Intermediate
Core idea. A macro observation is admissible at a forecast origin only if its value and revision state were actually available by the decision cutoff; reference month and database vintage label are not sufficient substitutes for that availability test.
Use it for. Auditing historical-vintage, pseudo-real-time, and real-time macro forecasting or allocation research.
It does not establish. Exact point-in-time availability merely because a file is labeled with the same month as the forecast, or merely because an official source release date is known.
The Question
Macroeconomic data carry several clocks at once. A GDP number may refer to one quarter, be first released weeks later, revised months or years later, entered into a database at another time, and downloaded by a researcher much later still.
A historical backtest must therefore answer:
At forecast origin \(t\), which macro values and revisions were actually admissible to the model?
This is narrower than general data leakage. QM003 — Look-Ahead Bias and Data Leakage explains the general rule that future information must not cross the forecast-origin boundary. QM014 focuses specifically on the release, vintage, and availability structure that makes macroeconomic data difficult.
Why It Matters
Using today’s revised macro history in a historical forecasting exercise can give the model information that decision-makers never observed in that form. Real-time data systems such as ALFRED are designed precisely because economic observations can be revised; the St. Louis Fed describes ALFRED as archiving FRED data together with the real-time periods in which values were originally released and later revised (Federal Reserve Bank of St. Louis, n.d.-a).
But even a historical vintage label does not solve every timing question. The FRED API documentation distinguishes real-time periods and vintage dates (Federal Reserve Bank of St. Louis, n.d.-e, n.d.-d). It also explicitly warns that source-published release dates do not necessarily equal the time at which data become available on the FRED or ALFRED websites (Federal Reserve Bank of St. Louis, n.d.-b).
That distinction is the basis of a fail-closed audit.
Intuition
Treat each macro value as carrying a small provenance record. The economic period tells you what the value describes; the release and availability fields tell you when it could enter a decision; the vintage and revision fields tell you which version was observed. A historical backtest is credible only when those fields can be aligned to the forecast-origin cutoff.
The Method
Six Clocks to Keep Separate
For each macro feature, distinguish at least the following fields.
| Clock | Meaning | Example question |
|---|---|---|
| Reference period | Economic period the observation describes | Is this CPI for May or June? |
| Release date/time | When the source officially published the release | Was the release before the portfolio cutoff? |
| Database vintage / snapshot label | Version date or file label used to reconstruct history | Which historical snapshot was downloaded? |
| Actual availability | When the exact value/file/API response was demonstrably accessible to the decision process | Could this value truly have been ingested by the cutoff? |
| Revision / backfill state | Initial, revised, benchmark-revised, or later backfilled value | Is the backtest using a value that changed later? |
| Forecast-origin cutoff | Last instant at which information may enter the decision | What was known before the signal was finalized? |
These clocks can coincide, but the research design should not assume they do.
Reference Period Is Not Availability
Suppose an observation is labeled “June industrial production.” The label identifies the economic reference period. It does not say when June industrial production was published.
Using June’s reference label inside a June month-end forecast may be impossible if the release occurs in July. The admissibility question is not “Does the observation belong to June?” but “Was the exact value available by the forecast cutoff?”
Historical Vintage Is Not Automatically Exact Point-in-Time Availability
FRED defines a real-time period as the interval during which facts were true or information was known until it changed (Federal Reserve Bank of St. Louis, n.d.-e). The API also allows vintage_dates to download data as it existed on specified dates in history (Federal Reserve Bank of St. Louis, n.d.-c). These features are valuable for reconstructing historical information sets.
However, a monthly research dataset labeled “2020-06 vintage” may provide only a snapshot convention rather than a timestamp-level proof that every component series was available before a particular June decision cutoff.
Therefore:
\[ \text{historical vintage evidence} \not\Rightarrow \text{exact within-period availability certification}. \]
The stronger claim requires stronger evidence.
Same-Month Vintage Labels
A common shortcut is
\[ \text{forecast month }t \leftarrow \text{vintage labeled }t. \]
That can be defensible if the dataset’s construction and timing contract prove that all included values were available before the forecast origin. Without that proof, the same-month label alone is insufficient.
If exact availability is unresolved, describe the design as historical-vintage or another accurate weaker term. Do not upgrade it to exact point-in-time or real-time simply because the file and forecast share the same month label.
A Fail-Closed Admissibility Rule
Let \(c_t\) be the forecast-origin cutoff and let \(A_{j,v}\) be the documented availability time of feature \(j\) in revision state \(v\).
The exact point-in-time admissibility condition is
\[ A_{j,v}\le c_t. \]
If \(A_{j,v}\) cannot be established from the retained evidence, the audit should not silently replace it with a reference date or snapshot label. Under a fail-closed standard, the claim becomes:
exact absence of timing leakage cannot be certified from the available evidence.
This is not the same as proving leakage occurred. It means the stronger timing claim has not been established.
Conservative Lag or Buffer Robustness
When exact within-period availability cannot be certified, a conservative robustness specification can deliberately move the information set backward.
For monthly macro data, examples include:
- use a prior-month database vintage rather than the same-month vintage;
- cap the latest admissible macro reference period one additional month earlier; or
- impose a fixed release buffer that is known to exceed the uncertain availability interval.
The exact rule should be pre-specified and applied uniformly rather than tailored series by series after seeing performance.
The robustness result answers a narrower but valuable question:
Does the research conclusion survive when the macro information clock is made deliberately conservative?
It does not retroactively certify the primary specification as exact real-time.
Market-Price Clock versus Macro-Information Clock
Market prices and macro data should not be forced onto the same clock.
A month-end asset price may be observable at the market close on date \(t\). A monthly macro series carrying reference month \(t\) may not be released until later. The admissible information set can therefore contain current market-price history while using an older macro reference period.
A clean decision timeline is
\[ \text{market data available by }c_t + \text{macro data proven available by }c_t \rightarrow \text{model / signal} \rightarrow \text{portfolio decision} \rightarrow \text{future holding-period return}. \]
This separation prevents an overly conservative macro lag from being mechanically imposed on market prices that were actually observable.
Revisions and Backfills
Vintage-aware research must also record which version of a value was used.
The FRED series/vintagedates endpoint identifies dates when a series value was revised or new values were released (Federal Reserve Bank of St. Louis, n.d.-d). A later snapshot can differ from an earlier one because of:
- routine revisions;
- benchmark or seasonal-adjustment revisions;
- historical backfills;
- source methodology changes; or
- changes in the set of available series.
A reproducible pipeline should retain enough metadata to distinguish these possibilities rather than merely store the final numeric matrix.
FRED-MD and Monthly Vintages
FRED-MD provides a large monthly macroeconomic database for empirical research (McCracken and Ng 2016). Historical-vintage use materially improves realism relative to a single current revised panel. But a project still needs its own decision-time contract: which vintage file was selected, which reference periods were allowed, and whether that selection is sufficient to certify actual availability before each forecast origin.
The database is an input. The information-timing proof belongs to the research design.
How to Interpret Timing Labels
A useful vocabulary is deliberately tiered:
Current revised / latest snapshot. Values reflect information available today; historical decisions are not reconstructed.
Historical-vintage. A historical database state is used, but exact availability before every decision cutoff may not be fully certified.
Exact point-in-time / real-time certified. The project retains sufficient release/availability evidence to establish the admissible information set at each forecast origin.
If evidence is incomplete, choose the weaker accurate label. This is preferable to making a stronger claim that the data contract cannot support.
Financial / Economic Example
Assume a portfolio signal is finalized at the close of June 30.
- June 30 market prices are available at that close and can be part of the price-information clock if execution occurs afterward.
- A macro dataset is labeled “June vintage.”
- One underlying June-reference series is not officially released until July.
- The vintage label alone does not provide an exact intraday availability timestamp for every included value.
The safe interpretation is not “all June macro data were real-time available on June 30.” A conservative robustness design might instead use the May vintage and cap the latest macro reference period at April, while preserving price data through June 30 if their market timestamp is valid.
The dates are schematic. Actual admissibility must be established from the source-specific release and availability evidence retained by the project.
Implementation
A data contract should represent clocks explicitly rather than infer them from filenames.
from dataclasses import dataclass
from datetime import datetime
@dataclass
class MacroObservation:
reference_period: str
release_time: datetime | None
vintage_label: str
actual_available_time: datetime | None
revision_state: str
def admissible(obs, cutoff):
return (
obs.actual_available_time is not None
and obs.actual_available_time <= cutoff
)When actual_available_time is missing, the exact real-time test returns unresolved rather than guessing from reference_period or vintage_label.
Common Mistakes
Using today’s revised values in a historical forecast. This contaminates the historical information set when revisions matter.
Equating reference month with release month. Economic period and publication time are different fields.
Equating source release date with database availability. FRED explicitly notes these need not be identical (Federal Reserve Bank of St. Louis, n.d.-b).
Treating a same-month vintage label as proof of intramonth availability. A snapshot label can be weaker evidence than a timestamped availability record.
Lagging market prices just because macro data need a lag. Market and macro clocks can have different admissible cutoffs.
Calling uncertainty “proof of leakage.” Fail-closed means the absence of leakage cannot be certified; it does not claim leakage definitely occurred.
When Not to Use It
QM014 is unnecessary for a strictly price-only strategy with no macro or revised non-market data. That strategy still needs market-data timing and execution auditing under QM003 and QM007, but FRED-style vintage governance is not the relevant risk.
It is also not a substitute for checking transformations, target alignment, or train/test leakage. Those remain general research-design problems.
Used in SlackQuant Research
- The Decision Value of Price and Macro Information: distinguishes historical-vintage macro evidence from exact within-month availability and uses a conservative lagged specification when stronger timing certification is unavailable.
Reproducibility
The Python example encodes the fail-closed rule directly: exact admissibility requires an actual-availability timestamp no later than the forecast cutoff. A missing timestamp returns an unresolved state rather than substituting the reference period or vintage label.