Known Future Features
A class of inputs in time-series models referring to variables whose future values are known at forecast time (such as calendars and planned events), which can be safely used for full-horizon attention; mistakenly labeling observed values as this type causes information leakage.
Timeline (last 90 days)
The Tinestral paper states that known future features participate in attention computation across the full horizon.
Known future covariates are variables whose future values are genuinely known at forecast time, such as calendar features (holidays, day of week) and planned promotional activities, and the model can safely use these features for attention computation across the entire forecast horizon.
If actual observed values (such as actual temperatures) are mistakenly labeled as known future covariates, the model will "see" future information, causing backtest performance to be inflated.
The checklist for practitioners includes: verifying the temporal availability of features, distinguishing "known future" from "unknown future," using archived forecast data, checking the split details of public results, and keeping backtest logs.
All Facts (4)
The Tinestral paper states that known future features participate in attention computation across the full horizon.
50%UnverifiedKnown future covariates are variables whose future values are genuinely known at forecast time, such as calendar features (holidays, day of week) and planned promotional activities, and the model can safely use these features for attention computation across the entire forecast horizon.
50%UnverifiedIf actual observed values (such as actual temperatures) are mistakenly labeled as known future covariates, the model will "see" future information, causing backtest performance to be inflated.
50%UnverifiedThe checklist for practitioners includes: verifying the temporal availability of features, distinguishing "known future" from "unknown future," using archived forecast data, checking the split details of public results, and keeping backtest logs.
50%