1. Name the failure before choosing a fallback
Missing means a required value is absent or unusable. Stale means an available value exceeds a declared freshness limit. Late means an expected update has not arrived by the relevant deadline. These conditions overlap but are not interchangeable: an older valid value may exist while its replacement is late. Add separate flags rather than forcing every condition into one status. A missing input is not a low reading, and a stale reading is not evidence of an unchanged economy. Neither should automatically become a named economic regime.
2. Define age relative to the research question
Observation age measures distance from the period described. Release age measures distance from publication. Receipt age measures distance from the last successful acquisition. A freshly downloaded quarterly estimate can be old in the first sense and new in the third. Choose which clock controls eligibility for each input and define calendar versus business-time counting. Use source-specific expectations where frequencies differ. Applying one short age limit to both daily and quarterly inputs can turn ordinary publication schedules into apparent failures without revealing anything about their underlying economic relevance.
3. Verify arrival independently of the schedule
FRED's release-date documentation warns that source release dates do not necessarily identify when data become available on FRED or ALFRED. A calendar entry therefore cannot stand in for a successful data check. Inspect the returned observation period, vintage and value; a successful request can still return an older observation. Record when a value passed validation and which expected event it satisfies. Distinguish an unchanged newly published estimate from an unchanged old file. The numerical value alone cannot establish that a scheduled release was received.
4. Make missing-value handling an explicit choice
Software defaults can hide data problems. The statsmodels missing-data guide demonstrates an OLS example that returns missing parameter estimates without an explicit missing-data policy, and describes options to raise an error or drop incomplete observations. That example is a software caution, not evidence about regime models. A threshold classifier also needs its own declared treatment. Rejecting the row, preserving an explicitly aged prior label or marking classification unavailable are different choices. Record the chosen rule and reason instead of assuming the numerical library supplies an economically meaningful fallback.
5. Keep imputation separate from observation
Forward-filling carries an earlier value; it does not create a new measurement. Interpolation between surrounding observations can use a later point that was unavailable at the historical cutoff. An estimated replacement introduces another model, with its own training window and assumptions. Preserve an imputation flag, original missing status and source observation date. Evaluate the classification with and without the replacement when feasible. Do not let an imputed row masquerade as a release or reset the original value's age simply because the research process touched it again.
6. Apply a fictional freshness contract
Suppose an entirely fictional exercise accepts activity inputs no more than 30 hours old by release age and requires both activity and inflation features. The activity value is present but 31 hours old, making it stale by one hour. Its expected replacement is also late. The inflation value is absent. Under a predeclared rule requiring both fresh inputs, the classification is unavailable. The table separates value presence from freshness and arrival. These invented limits are teaching assumptions, not recommended tolerances, measured delays or descriptions of any live data service.
7. Keep failures in the historical denominator
When reviewing a classifier, retain unavailable decision points and report why they were excluded from any comparison. Dropping them silently can make an unreliable process appear continuously available. Preserve the original result when a late value eventually arrives; calculate a separately labelled replay if needed. Distinguish a data-quality rejection from a boundary-based abstention, since they diagnose different problems. The useful question is whether the recorded information justified the recorded output under the declared policy. It is not whether a later-filled chart looks complete. This guide is not investment advice.
Inspect the comparison
| Fictional input | Value status | Freshness or arrival status | Policy result |
|---|---|---|---|
| Activity | Present | 31 hours old; replacement late | Fails 30-hour maximum |
| Inflation | Absent | Release age unavailable | Required input missing |
| Combined classification | Not computed | Two fresh inputs required | Unavailable, not an economic state |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.