1. Can monthly and quarterly data define one regime?
Yes, provided the classification explicitly describes an asynchronous information set rather than pretending that all inputs measure one instant. A quarterly activity estimate can coexist with a newer monthly price release at a decision deadline. Preserve each reference period, publication timestamp and transformation. The resulting label means the rule evaluated those eligible inputs together, not that quarterly activity was newly observed every month. This distinction is essential when counting updates, transitions and periods of apparent stability. A mixed-frequency dashboard may legitimately change between quarterly releases because another input changes; the archive should identify that input instead of implying that every component was refreshed simultaneously.
2. Declare the update clock before joining inputs
Choose whether decisions occur on scheduled calendar dates, after specified releases or whenever any required source changes. Each policy creates different opportunities to produce labels. A calendar policy needs rules for ordinary days with no new release. An event policy needs rules for simultaneous publications, late ingestion and revisions to older periods. Keep the most recent eligible source identity attached to every output. FRED's real-time documentation distinguishes information vintages, while its observation interface documents frequency aggregation options. Neither a timestamp join nor an API aggregation option decides the economic meaning of a mixed-frequency label; that responsibility remains in the research contract.
3. Stocks, flows and rates aggregate differently
An end-of-month stock, a monthly flow and an index level should not automatically share an averaging rule. A constructed quarterly flow of thirty, forty and fifty units totals one hundred twenty units; its average monthly flow is forty. A quarterly end-stock observation would instead select the declared endpoint. Averaging growth rates generally differs from first aggregating compatible levels and then calculating growth. Rates also need their quotation conventions preserved, particularly annualized quarterly changes. Write the desired quarterly object in words and units before choosing sum, average or end-of-period aggregation. Otherwise a mechanically valid transformation can answer a different question than the classifier's stated purpose.
4. Inspect an event-driven synthetic update
In the fictional table, a quarterly activity feature of positive 0.4 is already eligible at the first decision. A monthly price-change feature of negative 0.2 is also eligible. At the next event only the price feature updates to positive 0.1, so the joint sign label changes even though activity information is unchanged. A later activity revision to negative 0.1 changes the label again. There are three evaluation records, two subsequent input events, and no invented monthly activity measurements. All feature values are abstract scores in declared comparable units, not actual growth or inflation observations, and every timestamp is a constructed teaching assumption.
5. A ragged edge is not an invitation to fill with hindsight
At a common cutoff, the newest observation period can differ across series. This incomplete boundary is often called a ragged edge. Forward-filling may be a declared carry policy, but interpolation using a later release introduces future information. A nowcast supplies a model estimate rather than an observation and needs a distinct identity, estimation window and uncertainty statement. Statsmodels' mixed-frequency dynamic-factor documentation discusses monthly and quarterly inputs and missing-data patterns. That supports recognizing the separate modeling problem, not claiming this publication has fitted such a model. Preserve missingness and classify a result unavailable when the chosen rule cannot legitimately bridge the gap.
6. Separate event coverage from repeated display rows
Repeating the same information snapshot on successive dates can be useful for display or alignment, but those rows are not separate source updates. Report evaluation counts and distinct input-version counts together. Otherwise a quarterly feature repeated through many sessions can create impressive-looking persistence statistics with little new measurement. Revisions should identify the old reference period they amend, not masquerade as a current-period release. A comparison across classifiers should hold the eligible information identities fixed when isolating method differences. If one method evaluates on every session and another only after releases, their transition counts cannot be compared without acknowledging the different opportunities to switch.
7. Review questions and common failures
Which event can change the label, and which source can remain carried? Does every transformation preserve stock-versus-flow meaning and rate units? Is an estimated current value explicitly marked as a nowcast? What happens when one source arrives after the deadline? Check whether a revision restarts freshness improperly, whether daily rows inflate release counts, and whether an end-of-quarter date has been substituted for an actual availability timestamp. A useful update record names the triggering event and lists unchanged components alongside changed ones. This supports an inspectable information history without claiming a forecasting advantage, institutional feed or continuously updated production service.
Inspect the comparison
| Fictional UTC decision | Activity identity/value | Price identity/value | Changed input | Joint signs |
|---|---|---|---|---|
| 2032-05-01 16:00 | Q1 initial: +0.4 | March initial: -0.2 | Initial snapshot | + / - |
| 2032-05-15 16:00 | Q1 initial: +0.4 | April initial: +0.1 | Price only | + / + |
| 2032-05-30 16:00 | Q1 revised: -0.1 | April initial: +0.1 | Activity revision only | - / + |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.