A seasonal forecast needs an issue date and a target interval
Keep forecasts out of observed-data evaluation samples unless their role is explicit.
The date type is not the time meaning
Two climate records can share a date type while answering different questions. One is an observation; another is a forecast issued today for a later season. Truthound can check the declared ordering of the target interval in a prepared input table. It cannot judge forecast accuracy or predict damage at a particular site.
The WMO seasonal interval
WMO published its November 2026–January 2027 Global Seasonal Climate Update on October 21, 2026. It describes ensemble outlooks for the coming season. The October publication date and the later target period are different facts. Reading the outlook as an October observation would change the meaning of a subsequent model evaluation.
Keep the forecast vintage
Store issue version, issue date, target start, target end and observation-or-forecast state separately. Retain earlier issues when an outlook is updated. An analyst testing what was knowable at a historical decision point needs the forecast available then, not a newer edition quietly substituted into the old input.
Check the declared interval
For the forecast edition and target-period table prepared as a Polars DataFrame, Truthound Core Python SDK 3.3.4's DateOrderValidator can inspect two date columns against a declared ordering policy. It can flag an inverted target interval after the owner defines the time convention. It does not distinguish a forecast from an observation automatically, and it cannot establish that the recorded issue is the edition actually available to the earlier decision-maker.
Schema preparation has a different role
Pandera 0.35.0 provides schema preparation, including Polars support. Typed date fields can support the workflow, but a schema API alone is not the same two-date ordering or vintage-selection rule. Keep tested preparation and explicitly define any interval predicate separately. Neither library supplies the scientific validation of the climate forecast.
Test an October issue for a later season
Test a record issued in October whose target spans November through January. It must not enter an observed October sample simply because a downstream process keeps only one date. A second test can use an updated forecast with the same target interval, checking that the issue identity remains distinct. These examples expose temporal leakage before statistical interpretation.
Repair the evaluation boundary
Approve the input only after its temporal role and vintage policy are clear. Re-evaluate samples that mixed forecasts with observations, preserving the original records and explaining the correction. The scientific hypothesis becomes more testable when issue time, target time and observed outcome remain separate, rather than when a validator merely accepts a date format.
Evidence
- Event checked
- Subject checked
- Modified
