In ha0ye/rEDM: Applications of Empirical Dynamic Modeling from Time Series

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/"
)

This version of rEDM is no longer being maintained (as of October 2019) - please see https://github.com/SugiharaLab/rEDM/ for the latest version of the package

rEDM

Binder Demo

Try out the package without installation (after loading, try clicking on README.Rmd in the Files tab).

Overview

The rEDM package is a collection of methods for Empirical Dynamic Modeling (EDM). EDM is based on the mathematical theory of recontructing attractor manifolds from time series data, with applications to forecasting, causal inference, and more. It is based on research software previously developed for the Sugihara Lab (University of California San Diego, Scripps Institution of Oceanography).

Installation

You can install rEDM from CRAN with:

install.packages("rEDM")

OR from github with:

# install.packages("remotes")
remotes::install_github("ha0ye/rEDM")

If you are on Windows, you may need to install Rtools first, so that you have access to a C++ compiler.

Example

knitr::opts_chunk$set(fig.width = 6, fig.height = 4)
par(mar = c(4, 4, 1, 1), mgp = c(2.5, 1, 0), oma = c(0, 0, 0, 0))

We begin by looking at annual time series of sunspots:

dat <- data.frame(yr = as.numeric(time(sunspot.year)), 
                  sunspot_count = as.numeric(sunspot.year))

plot(dat$yr, dat$sunspot_count, type = "l", 
     xlab = "year", ylab = "sunspots")

First, we use simplex to determine the optimal embedding dimension, E:

library(rEDM)                             # load the package

n <- NROW(dat)
lib <- c(1, floor(2/3 * n))               # indices for the first 2/3 of the time series
pred <- c(floor(2/3 * n) + 1, n)          # indices for the final 1/3 of the time series

output <- simplex(dat,                    # input data (for data.frames, uses 2nd column)
                  lib = lib, pred = lib,  # which portions of the data to train and predict
                  E = 1:10)               # embedding dimensions to try

summary(output[, 1:9])

It looks like E = 3 or 4 is optimal. Since we generally want a simpler model, if possible, let's go with E = 3 to forecast the remaining 1/3 of the data.

output <- simplex(dat,
                  lib = lib, pred = pred, # predict on last 1/3
                  E = 3, 
                  stats_only = FALSE)     # return predictions, too

predictions <- na.omit(output$model_output[[1]])

plot(dat$yr, dat$sunspot_count, type = "l", 
          xlab = "year", ylab = "sunspots")
lines(predictions$time, predictions$pred, col = "blue", lty = 2)
polygon(c(predictions$time, rev(predictions$time)), 
        c(predictions$pred - sqrt(predictions$pred_var), 
        rev(predictions$pred + sqrt(predictions$pred_var))), 
        col = rgb(0, 0, 1, 0.5), border = NA)