Time series foundation models (TSFMs) are increasingly designed to accept covariate inputs, according to a new paper on arXiv. Among the possible covariates, calendar information stands out as the most common categorical type in practice, since it is available for every point in time.
The paper, titled "On the Encoding of Calendar Features in Covariate-Informed Time Series Foundation Models," appears to investigate how these calendar features are represented within such models. However, the abstract provided ends mid-sentence, so no specific results or conclusions are available from the source.
As a result, this digest can only report the paper's stated motivation and scope. Readers interested in the encoding mechanisms or experimental outcomes will need to consult the full manuscript.