search

Nixtla / hierarchicalforecast

Probabilistic Hierarchical forecasting 👑 with statistical and econometric methods.
https://nixtlaverse.nixtla.io/hierarchicalforecast
Apache License 2.0
588 stars 76 forks source link

[<Library component: Models|Core|etc...>] Enhancement request: Include Elasso Algorithm #275

Open lanzalibre opened 4 months ago

lanzalibre commented 4 months ago

Description

Enclude Elasso algorithm as per https://www.monash.edu/__data/assets/pdf_file/0003/3611901/Optimal-forecast-reconciliation-with-time-series-selection.pdf

Use case

No response

elephaint commented 4 months ago

Thanks - good suggestion, I'll have a look at it.

elephaint commented 3 months ago

@lanzalibre It's computationally a very intensive method requiring an additional dependency on a MIP solver, for what seems to be relatively minor gains. For very small datasets, this could be interesting, but for anything larger than a few timeseries I'm hesitant to include something that is so intensive - the relatively small performance gain would then be completely offset by the time required to get an answer (the time saved after one already gets a normal MinT answer can then be usefully spent on improving the base forecast estimators, which in general is far more fruitful than squeezing performance out of a reconciliation method).

I'll leave this open, if this is something that more users like to see I'll have another look at it, and open to suggestions to improve the computational complexity of this method!