The Tidymodels Extension for Time Series Boosting Models
-
Getting Started with Boostime: A walkthrough of the tidy modeling approach with the package.
-
Boostime: Combining ARIMA with Catboost Introduction to Boostime with a practical example using ARIMA and Catboost.
-
Hyperparameter Tuning with Boostime Learn how to select the best hyperparameters for the Boostime algorithms.
Not on CRAN yet (and predictably won’t be because Catboost doesn’t look like it’s going to be any time soon)
#install.packages("boostime")Development version:
# install.packages("devtools")
devtools::install_github("AlbertoAlmuinha/boostime")Boostime unlocks boosting methods to improve modeling errors (residuals) on time series analysis.
The following algorithms are available:
-
Arima + Catboost: You can use either an automatic version of Arima (auto.arima, in which orders are selected from KPSS unit root tests or the manual version) in conjunction with Catboost to model the residuals. One of the great advantages of this model over XGBoost is that it can deal with categorical variables and you can use GPU without any configuration effort.
-
Prophet + Catboost: It uses Prophet and Catboost to model the residuals. One advantage of Prophet over Arima is that it can handle multiple seasonalities.
-
Arima + LightGBM: You can use either an automatic version of Arima (auto.arima, in which orders are selected from KPSS unit root tests or the manual version) in conjunction with LightGBM to model the residuals.
-
Prophet + LightGBM: It uses Prophet and LightGBM to model the residuals. One advantage of Prophet over Arima is that it can handle multiple seasonalities.
I would especially like to thank Matt Dancho for his contribution to the world of time series, in particular for the creation of the Modeltime package. Without this package, Boostime would never have been possible. THANKS!