What is the training cutoff date for the real data?

[enussl]

Hi, could you please specify until which date real data was used in training? For example, up to which month or year datasets like FRED-MD were included. This would help define an appropriate out-of-sample evaluation period.

Thanks,
Emanuel

[shchur]

For Chronos-2, the FRED-MD datasets were not used at all during training. The list of datasets used for training is available in Table 6 of the publication: https://arxiv.org/pdf/2510.15821v1

[enussl]

FRED-MD is part of the data used by Ansari et al. (2024), which is part of Table 6 of the publication, no? And irrespective of the concrete data source, it would be very useful to publish a cutoff date in case people want to evaluate on the series used in the pre-training.

[shchur]

Table 6 provides an exhaustive list of real-world datasets used for model training

FRED-MD is not one of these datasets. It is possible to check the corresponding sources (all datasets are publicly available under https://huggingface.co/datasets/Salesforce/GiftEvalPretrain and https://huggingface.co/datasets/autogluon/chronos_datasets) and determine the cutoff based on that

[enussl]

Hi, thank you, that is very useful, I did not see that!

However, the list of sources includes monash_fred_md, but now I can see until which date, ty.

Related to that, is there a way to see to which time series a given id corresponds? For example, there might be an inflation series in m4_monthly, but I do not see a way of judging this. Connected to that, the max timestamp in m4_monthly is 2124-04-30T12:00:00.000Z, making it hard to judge the maximum date of the real data that was used.

[shchur]

To clarify, even though the linked resources include more datasets than listed in the table, only the datasets listed in Table 6 were used for training of Chronos-2. You can disregard other datasets. M4 competition datasets include dummy timestamps. The competition took place in 2017, so no data beyond this year was used.