[ENH] Hierarchical reconcilers for point forecasts

sktime / sktime

A unified framework for machine learning with time series

https://www.sktime.net

BSD 3-Clause "New" or "Revised" License

7.88k stars 1.36k forks source link

[ENH] Hierarchical reconcilers for point forecasts #2157

Open ciaran-g opened 2 years ago

ciaran-g commented 2 years ago

Some initial thoughts, first two will be a good place to start

[x] Bottom up (aggregate levels ignored)
[ ] Top down (historical proportions)
[x] Top down (forecast proportions)
[x] Unweighted OLS combination
[x] Weighted LS combination
[x] Trace minimisation (MinT) (various options for covariance matrix , including things like regularisation)
[ ] Middle out (?)

Think it might be best for now to focus on spatial hierarchies? i.e. get these up and running and worry about grouped or temporal hierarchies afterwards?

More info here

@fkiraly update: there are now two estimators, Reconciler (transformer) and ReconcilerForecaster (forecaster) where the two remaining reconciling algorithms could/should be added. Since this prescribes a good templated structure, I think it's a good first issue (for individuals familiar with, or interested in, the reconciliation algorithms)

fkiraly commented 2 years ago

Thanks for the list!

What would be extremely useful is a quick analysis which of these reconciliation strategies can be mapped on transformers (e.g., apply after forecasting, use only forecasts to fit and reconcile), and which need to be wrappers/compositors (e.g., needs the original or historical data, needs to call methods of the forecaster as part of the algorithm).

ciaran-g commented 2 years ago

Good shout Franz - started working on the bottom-up and OLS if you want to have a look

Transformers:

bottom up
OLS
Weighted LS (structural)

Wrappers/compositors:

Weighted LS (variance) - uses residuals
MinT (sample) - uses residuals
MinT (shrinkage) - uses residuals
Top down - historic proportions

fkiraly commented 2 years ago

linking this here, some open work items related to the above: https://github.com/alan-turing-institute/sktime/issues/2416

fkiraly commented 1 year ago

@ciaran-g, just to check - these are all implemented now, right? Part of ReconcilerForecaster.

ciaran-g commented 1 year ago

@fkiraly the middle_out and top_down_historic proportions methods have not been implemented. I'm not currently working on these at the minute though if you have a plan in mind to get them implemented

fkiraly commented 1 year ago

ok, we'll leave this open then - I think it might be a good first issue since it could be added into an existing estimator.