Non-Negative MinTrace hangs forever about hierarchicalforecast HOT 4 OPEN

Hey @iamyihwa and @candalfigomoro,

There is work to be merged that greatly improves the computational efficiency of the MinTrace method.

In between if you are interested in non-negative coherent predictions I suggest you to take a look to the HINT method we recently added to the NeuralForecast library. You can combine non-negative distributions (Poisson, NegativeBinomial, Tweedie, PoissonMixture) and BottomUp or MinTrace (with efficiency improvements) type reconciliations.

• HINT Usage Example
• Distribution Losses Catalog

from hierarchicalforecast.

Having similar issue.

from hierarchicalforecast.

Sure will do Thanks @kdgutier !!!

from hierarchicalforecast.

@kdgutier Looks cool!

Unfortunately with my current forecasting problem, reconciliation with MinTrace or BottomUp doesn't seem to work well and the only reconciliation method that seems to work well in my specific case seems to be MiddleOut (lower levels are very intermittent so a top-down method from an intermediate level works better, while a bottom-up method works better for higher levels which are otherwise underestimated by NHITS).

Are there any plans to integrate other reconciliation methods into HINT?

from hierarchicalforecast.