This code implements a novel forecasting method based on bivariate copula. The method offers competitive results when it is compared with several popular machine learning techniques.

The algorithms are programed on R. You can execute it on Windows, Linux or Mac. The provided action uses Ubuntu 18.04.

To repdroduce it you need install R and the following packages:

r-cran-tidyverse 
r-cran-bitops 
r-cran-catools 
r-cran-vinecopula 
r-cran-ks 
r-cran-data.table 

In Ubuntu 18.04 you can use the marutter repository, which contains the needed precompiled packages.

sudo add-apt-repository ppa:marutter/rrutter3.5
sudo add-apt-repository ppa:marutter/c2d4u3.5
sudo apt-get update
sudo apt-get install r-base
sudo apt-get install r-cran-tidyverse r-cran-bitops r-cran-catools r-cran-vinecopula r-cran-ks r-cran-data.table r-cran-rapportools
sudo apt-get install r-cran-rcurl r-cran-bitops r-cran-rjson r-cran-statmod libssl-dev```

Also, the H2O R library is needed. In this case, it is compiled from the sources:

sudo R -e 'install.packages("h2o")'

At top levevel there are 2 scripts:

• H2O_Code_Without_DL.R - Which uses the autoML of H2O (a compilation of some Machine Learning methods) over the data with comparison purpose.
• copula_learning_example.R - Which train and test the proposed model over a dataset. This script uses two important functions:
  • copulaLearningMethod() in copulaLearningMethod.R - The model training function.
  • copulaLearningMethodPredict() in copulaLearningMethodPredict.R - The model scoring function.

To execute theses scripts you can type:

R < code/H2O_Code_Without_DL.R --no-save
R < code/copula_learning_example.R --no-save

In both cases you can edit the top level scripts to select other dataset.

An action that execute the whole project over a sample dataset is provided. Here you can see the result of an execution over the kdd1998 dataset.