Open candalfigomoro opened 1 year ago
Having similar issue.
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.
Sure will do Thanks @kdgutier !!!
@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?
Using the same data (unfortunately I can't share it), standard MinTrace performs the reconciliation in about 1 minute, while Non-Negative MinTrace (
MinTrace(method="ols", nonnegative=True)
) keeps running after hours (basically it hangs forever).hierarchicalforecast 0.3.0
Do you have any hints on how I could debug the issue? Thanks