Questions Regarding the Training Data of Chronos-2

[xingxing445]

Hello!

First of all, thank you for your great work!

1. Regarding the real univariate datasets used for pretraining on univariate tasks in Chronos-2, which are sourced from the Chronos and GIFT-Eval pretraining corpora, does it imply that the multivariate data within these corpora is also split into individual univariate time series for training? What are the key considerations behind this data processing approach?
2. For datasets with a massive number of time series (e.g., Buildings 900K), only 100,000 time series were utilized in the training process. Could you elaborate on the factors or trade-offs that led to this selection of data volume?
3. When incorporating real univariate datasets from existing pretraining corpora, were there any specific criteria for filtering or selecting the time series (e.g., data quality, temporal frequency, domain relevance)? How did these criteria align with Chronos-2’s pretraining objectives?
4. Given that Chronos-2 aims to handle multivariate and covariate-informed forecasting tasks, why was the pretraining primarily based on univariate data (including split multivariate data) rather than using more native multivariate datasets? What advantages does this approach offer for the model’s generalization capabilities?

[abdulfatir]

@xingxing445 sorry for the delayed response.

• Yes, all datasets were treated as univariate because based on our preliminary evals we found limited multivariate characteristics in this data. Most datasets have long context which makes multivariate less useful. This may not be the ideal decision, but it was the design choice we made.
• We did not want to use too many time series from the same domain/dataset. The main goal for all data exercises was to introduce diversity.
• We do not plan to share specifics here. The goal was to enhance diversity of patterns, frequencies, interactions, etc.
• The multivariate capabilities come strictly from the synthetic data. Unlike real data, where a claimed multivariate dataset may have little multivariate behavior, interactions in synthetic data are easier to control.