Forecastable Component Analysis

Implementation of Forecastable Component Analysis ('ForeCA'), including main algorithms and auxiliary function (summary, plotting, etc.) to apply 'ForeCA' to multivariate time series data. 'ForeCA' is a novel dimension reduction (DR) technique for temporally dependent signals. Contrary to other popular DR methods, such as 'PCA' or 'ICA', 'ForeCA' takes time dependency explicitly into account and searches for the most ''forecastable'' signal. The measure of forecastability is based on the Shannon entropy of the spectral density of the transformed signal.

ForeCA R package

ForeCA implements Forecastable component analysis in R. For details on algorithm & methodology see Forecastable Component Analysis, JMLR, Goerg (2013).

In a nutshell: ForeCA finds linear combinations of multivariate time series that are most forecastable, where forecastability is measured by the spectral entropy of the resulting signal (linear combination of input).

Installation

You can install the stable version from CRAN:

install.packages('ForeCA')

Alternatively, you can also install the latest version of ForeCA package directly from github as

library(devtools)
devtools::install_github("gmgeorg/ForeCA")

Usage

The workhorse function is ForeCA::foreca() which works just like the built-in princomp function for PCA.

library(ForeCA)
citation("ForeCA")

For a tutorial on how to use foreca() and the entire ForeCA suite of functions see the introductory vignette on CRAN.

References

• ForeCA references & applications in the literature (non-exhaustive; see here for full list of ForeCA citations)

  • Very interesting application of ForeCA to historical time series data of temperature/climate to extract predictable climate signals. Fischer, Matt. (2016). Predictable components in global speleothem δ18O. Quaternary Science Reviews. 131. 380-392. 10.1016/j.quascirev.2015.03.024.
  • ForeCA's forecastability measure, spectral entropy of a time series, can be useful as a feature to characterize/visualize/predict performance of different algorithms applied to a set of time series. Kang, Yanfei & Hyndman, Rob & Smith-Miles, Kate. (2017). Visualising forecasting algorithm performance using time series instance spaces. International Journal of Forecasting. 33. 345-358. 10.1016/j.ijforecast.2016.09.004.

• Cross-validated & SO posts (non-exhaustive)

  • How to determine forecastability of time series

• Blog posts (by others)

  • Stock Forecasting with Machine Learning - Are Stock Prices Predictable? (2016/04/20)
  • Are stocks predictable? (2014/02/20)