Provides eigenvector-based (geometric) forecast combination methods; also includes simple approaches (simple average, median, trimmed and winsorized mean, inverse rank method) and regression-based combination. Tools for data pre-processing are available in order to deal with common problems in forecast combination (missingness, collinearity).
(written by Chris E. Weiss and Gernot Roetzer)
The R package GeomComb presents functions to pool individual model forecasts using geometric (eigenvector-based) forecast combination methods. The package also provides functions for simple forecast combination methods (inverse rank approach, simple average, trimmed mean, and winsorized mean - including the option of a criterion-based optimisation of the trimming factor) and regression-based forecast combination methods.
The forecast combination methods allow for 3 different input types:
Only training set
Training set + future forecasts
Full training + test set
Accuracy measures are provided accordingly, summary and plot functions have been created for the S3 classes. The function auto.combine() is an automated selection of the best combination method based on criterion optimisation in the training set.
The package is still in the development stage -- updates on CRAN release will be shared here in the future.
If you are interested in using the provided functions for your research in the meantime, you are welcome to email us: email@example.com
You can also install the development version from Github
This package is free and open source software, licensed under GPL (>= 2).