search

DynamicsAndNeuralSystems / pycatch22

python implementation of catch22
https://time-series-features.gitbook.io/catch22/python
GNU General Public License v3.0
75 stars 15 forks source link

Forecasting/Regression Method #4

Closed gdevos010 closed 2 years ago

gdevos010 commented 2 years ago

Is there a recommended way to run catch22 for time series forecasting?

benfulcher commented 2 years ago

Hi there, pycatch22 just implements time-series feature extraction. It's up to the user to adapt that for their specific application. FYI: Some infrastructure for running feature extraction across a time-series dataset in R is theft. Some other related packages are here Some summaries of using time-series features for applications is in this paper. Ben

gdevos010 commented 2 years ago

@benfulcher, if I have a univariate pandas dataframe, is there some recommended method for calculating these catch22 features on a lookback window?

For example, say I have 1000 samples, and I have a lookback window of 10 samples. I want to use the catch22 features as covariates in forecasting the future. I believe the pandas rolling method works, but it is very slow. Is there a faster approach for calculating these features? Or do I miss understand how the catch22 library works?

benfulcher commented 2 years ago

I have no recommendations on doing this—maybe others will have experience with this. In general, computing 22 features from just 10 samples may be suboptimal, relative to traditional forecasting using, say, the 10 values themselves.