Hi timecopilot/impermanent team! It was nice seeing your work at ISF 2026.
So far Impermanent only includes time series derived from github user activity data. While I understand there are plans to add more data sets (I think electricity data?) I think there is value in considering non-traditional sources of time series data as well. For example, environmental and climate scientists publish a lot of time series data that does not seem to be well-reflected in time series benchmark datasets.
As an explicit example, UC San Diego publishes atmospheric carbon dioxide concentrations on a weekly basis in various frequencies, see this link. Since the data is just a static URL to a csv file, it should be pretty trivial to chron job a scrape.
If there is consideration for less-frequently published data, the Canadian government publishes annual time series of air pollution measurements from 616 different stations. Unfortunately, they only publish once per year and tend to lag a couple years (public funding issues), but this could be used to assess a model's ability to produce very long-horizon forecasts.
Happy to see this benchmark, as I feel leading models might be overfitting to Gift-Eval. Hope you can add more types of data soon.
Hi timecopilot/impermanent team! It was nice seeing your work at ISF 2026.
So far Impermanent only includes time series derived from github user activity data. While I understand there are plans to add more data sets (I think electricity data?) I think there is value in considering non-traditional sources of time series data as well. For example, environmental and climate scientists publish a lot of time series data that does not seem to be well-reflected in time series benchmark datasets.
As an explicit example, UC San Diego publishes atmospheric carbon dioxide concentrations on a weekly basis in various frequencies, see this link. Since the data is just a static URL to a csv file, it should be pretty trivial to chron job a scrape.
If there is consideration for less-frequently published data, the Canadian government publishes annual time series of air pollution measurements from 616 different stations. Unfortunately, they only publish once per year and tend to lag a couple years (public funding issues), but this could be used to assess a model's ability to produce very long-horizon forecasts.
Happy to see this benchmark, as I feel leading models might be overfitting to Gift-Eval. Hope you can add more types of data soon.