In DataScienceForPublicPolicy/unitrootML: unitrootML - Machine Learning-Based Hypothesis Testing for Time Series

In this vignette, we illustrate how to apply the pre-trained unitrootML unit root test to a set of time series.

Set up

Set up requires installation from the Github page

#Load package
  library(unitrootML)
  library(forecast)
  library(ggplot2)

For this example, we will simulate five time series: two with a unit root (gamma = 1) and three without a unit root (gamma < 1). These time series are stored in a list object. Let's take a look at a few of these time series using autoplot from the forecast package and arranged into a grid using grid.arrange.

 #Example time series
    set.seed(123)
    out <- list(series1 = ts(dgp_enders1(gamma = 1), freq = 12, start = c(2010,1)),
                series2 = ts(dgp_enders1(gamma = 1), freq = 12, start = c(2010,1)),
                series3 = ts(dgp_enders1(gamma = 0.98), freq = 12, start = c(2010,1)),
                series4 = ts(dgp_enders1(gamma = 0.8), freq = 12, start = c(2010,1)),
                series5 = ts(dgp_enders1(gamma = 0.5), freq = 12, start = c(2010,1)))

  #Plot
    gridExtra::grid.arrange(autoplot(out[[1]]), 
                            autoplot(out[[2]]), 
                            autoplot(out[[3]]), 
                            autoplot(out[[4]]), 
                            autoplot(out[[5]]), 
                            ncol = 2)

To screen these time series, apply the ml_test_all function, which expects a list object containing time series objects (bank =) and apvalue = 0.05 to return results evaluated at a 5% level.

  res <- ml_test(bank = out, 
                 pvalue = 0.05)

The res object captures the outputs of a battery of tests, including pre-trained ML-based unit root tests. The first element summarizes the verdict that each unit root test reached when evaluating each time series ("yes" = "unit root"). While each of the traditional tests may disagree, it is worth noting that the tests built on Random Forest (verdict_ranger) and Gradient Boosting (verdict_xgbTree) offer markedly improved diagnostic accuracy.

  verdicts <- res$verdicts
  print(verdicts[,1:5])

The test statistics and their critical thresholds are recorded in the results element.

  results <- res$results
  print(results[,1:5])

Lastly, all the input features for the ML-based tests that were calculated from the time series list are available in the features element.

  feats <- res$features
  feats[,1:4]

For more in-depth background on the construction and properties of the ML-based unit root test, read Cornwall, Chen, and Sauley (2021).

DataScienceForPublicPolicy/unitrootML documentation built on Dec. 17, 2021, 4:07 p.m.