I can't seem to find much on smoothing across a large number of correlated sparse time series. I have some working ideas, but is there a standard approach?
This sounds like something that a hierarchical bayesian version of something like @seanjtaylor's prophet would work for?
-
-
possibly! i do have a hierarchy (of sorts) of the series. my baseline is just to desparsify by aggregating using the hierarchy. in anycase i certainly need to drill into prophet more :)
-
To be clear my thinking is that there would be hyperparameters k' and δ' (with priors as in the paper, ie δ' ~ laplace), and then specific k and δ for each time series where, eg, δ ~ normal(δ',σ).
-
That might work well with one experiment I've been trying for the sparseness; modeling the process as two parts; a poisson for the run of zeros & a gaussian for the non zero values.
-
ok yeah, so model the parameters of the individual poissons and gaussians as generated by some distributions that are themselves globally parameterized...
-
note: I think I'm over my head in terms of model correctness, but I'm the master of empirical testing so I should be ok :)
End of conversation
New conversation -
-
-
We don’t handle multiple related time series yet but there’s a neat hack to do that. You first fit the aggregate or clusters or SVD components. Then you use the forecasts generated by those models as regressors in the individual level models.
-
awesome. that makes sense & is easy to do! cheers!
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.