search

LinkedEarth / Pyleoclim_util

Python Package for the Analysis of Paleoclimate Data. Documentation at
https://pyleoclim-util.readthedocs.io
GNU General Public License v3.0
88 stars 33 forks source link

Better timeseries padding with `darts` #150

Open CommonClimate opened 3 years ago

CommonClimate commented 3 years ago

Is your feature request related to a problem? Please describe. Timeseries padding is important to minimize edge effects, particularly with filtering. Currently, pyleoclim uses tspad, which fits an ARIMA model with statsmodels with fixed parameters.

Describe the solution you'd like Use Darts: https://github.com/unit8co/darts which has the ts.fit(), ts.predict() syntax and allows more objective ways to choose a model.

Describe alternatives you've considered leave as is

CommonClimate commented 2 years ago

Note that CWT padding is currently limited to zero-padding, so this could help there too

CommonClimate commented 2 years ago

ts_pad should be moved from filter.py to tsbase.py

alexkjames commented 2 years ago

Meta time series forecasting package: https://facebook.github.io/prophet/

alexkjames commented 1 year ago

also consider just using LSTMs, could be more lightweight in sktime or pytorch

CommonClimate commented 1 year ago

This is another option for padding: https://github.com/uber/orbit. Seems to work well in the example shown in the readme, and this is the sort of behavior we'd want for padding. It might be way overkill, however, if it takes 3h to run a 10-point extension of a 100-pt series.