In DataScienceForPublicPolicy/hypML: 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)

#Set splits
  N <- 500 
  splits <- c(0.5,0.3, 0.2)
  train_index <- 1:round(N*splits[1])
  val_index <- (round(N*splits[1])+1):round(N * sum(splits[1:2]))
  test_index <- round(N * sum(splits[1:2]) + 1):N

Define

#Set DGP parameters if Single scenario
  dgp_params <- list(list(dgp = "dgp_enders3", sd = 1, gamma = 1),
                     list(dgp = "dgp_enders2", sd = 1, alpha0 = 1,  gamma = 1),
                     list(dgp = "dgp_enders1",  sd = 1, alpha0 = 1, alpha2 =  .005, gamma = 1))

Generate

#Simulate train and test time series
  ts_data <- gen_bank(iter = N,
                       sample_prob = .50,
                       t = c(5,50),
                       freq = 12,
                       nur_ur = c(0.90000,.99999),
                       run_par = TRUE,
                       dgp_params = dgp_params) 
  save(ts_data, file = "0_ts_series.Rda")

Create input features

#Construct feature set for each set
  train_feat <- gen_features(ts_data[train_index])
  val_feat <- gen_features(ts_data[val_index])
  test_feat <- gen_features(ts_data[test_index])

#Save out (change the directory!)
  save(train_feat, test_feat, val_feat, file = "1_ts_features.Rda")

Train models for test

#Set algorithm parameters
  model_params <- list(list(method = "ranger",
                            cv_folds = 5,
                            tune_grid = expand.grid(mtry = c(3, 9, 27),
                                                    splitrule = "gini",
                                                    min.node.size = c(2, 4, 16))),
                       list(method = "xgbTree",
                            cv_folds = 5,
                            tune_grid = expand.grid(nrounds = 300,
                                                    eta = c(0.01, 0.03, 0.1, 0.3, 0.5),
                                                    gamma = 0,
                                                    colsample_bytree = c(0.8, 1),
                                                    min_child_weight = 1,
                                                    subsample = c(0.8, 1),
                                                    max_depth = c(4, 6))))

#Train model
  custom_set <- gen_test(train_set = train_feat, 
                         val_set = val_feat,
                         model_params = model_params)

  save(custom_set, file = "custom_set.Rda")

Apply custom ML Test to new time series object

#List of time series 
  x <- list(norm1 = ts(rnorm(120, 10,10), freq=12),
            with_trend = ts(rnorm(120, 10,10) + 1:120, freq=12),
            uroot = ts(cumsum(rnorm(120,10,10)), freq = 12))

#ML test
  out <- ml_test(bank = x, 
                 pvalue = 0.05,
                 original = FALSE,
                 custom_model = custom_set)

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

DataScienceForPublicPolicy/hypML documentation built on Dec. 17, 2021, 4:06 p.m.