search

LedgerInvesting / bayesblend

Easily combine predictions from multiple Bayesian models using techniques including (pseudo) Bayesian model averaging, hierarchical stacking, and more!
https://ledger-investing-bayesblend.readthedocs-hosted.com/en/latest/
MIT License
13 stars 0 forks source link

Direct wrappers/class methods from other software? #28

Open cmgoold opened 3 months ago

cmgoold commented 3 months ago

BayesBlend currently works with CmdStanPy models, and ArviZ. Users can, in theory, use any modelling pipeline compatible with ArviZ (e.g. PyStan, CmdStanPy, PyMC, NumPyro etc.), and use ArviZ as the bridge. But should we provide other BayesBlend.from_* methods for other software directly?

Nathaniel-Haines commented 3 months ago

I think it makes sense for us to have the .from_* methods, even if under the hood we go through arviz? Anything that makes it easy to move from one ecosystem to bayesblend and back would, I think, be good to have.

That said, I guess the question is—is the use of arviz as a bridge going to come along with any issues re maintainability? If so, we be best to directly interface with other libraries. If not, it makes sense to me to just use arviz to add more support faster 🤔