As noted here, this is done in a few places already, but conceptually this is not that clean.
Observations with allowing optional quantile values:
it will make scoring and ensembling not as straight-forward, as it requires subsetting to just models that have a shared set of quantiles. Complicates downstream analyses.
several hubs have optional "0" and "1" quantiles (indicating a "min" and "max") of the distribution. Not all distributions will have these (e.g. many distributions are unbounded), so it's not a concept that is general for all probabilistic representations. If we allow 0/1 quantiles (some might argue we should not), then they definitely should be optional, as not all distributions will have finite min/max.
There is a "sorting" issue with pmf and cdf category names that is not a problem with quantiles (because they can always be alphanumerically sorted), so having a list broken up into required and optional sections would not hamper sorting.
Per https://github.com/orgs/hubverse-org/discussions/24, the thinking here is that it does not make sense for a hub to allow submissions where some quantile output_type_id fields are optional.
As noted here, this is done in a few places already, but conceptually this is not that clean.
Observations with allowing optional quantile values: