lifecycle Travis build status AppVeyor build status CRAN status metacran downloads

psymonitor provides an accessible implementation of the popular real-time monitoring strategy proposed by Phillips, Shi and Yu (2015a,b;PSY), along with a new bootstrap procedure designed to mitigate the potential impact of heteroskedasticity and to effect family-wise size control in recursive testing algorithms (Phillips and Shi, forthcoming). This methodology has been shown effective for bubble and crisis detection (PSY, 2015a,b; Phillips and Shi, 2017) and is now widely used by academic researchers, central bank economists, and fiscal regulators.


You can install the stable version from CRAN


You can install the development version from GitHub

# install.packages("devtools")


For the illustration purposes we will use data on the credit risk in the European sovereign sector, that is proxied by an index constructed as a GDP weighted 10-year government bond yield of the GIIPS (Greece, Ireland, Italy, Portugal, and Spain) countries, and comes with the ‘psymonitor’ package.

Let’s walk through some basics. First load the psymonitor package and get data on GIIPS spread.


Next, define a few parameters for the test and the simulation.

y        <- spread$value
obs      <- length(y)
swindow0 <- floor(obs * (0.01 + 1.8 / sqrt(obs))) # set minimal window size
IC       <- 2  # use BIC to select the number of lags
adflag   <- 6  # set the maximum nuber of lags to 6
yr       <- 2  
Tb       <- 12*yr + swindow0 - 1  # Set the control sample size
nboot    <- 99  # set the number of replications for the bootstrap

Next, estimate the PSY test statistic using PSY() and its corresponding bootstrap-based critical values using cvPSYwmboot().

bsadf          <- PSY(y, swindow0 = swindow0, IC = IC,
                      adflag = adflag)  # estimate the PSY test statistics sequence

quantilesBsadf <- cvPSYwmboot(y, swindow0 = swindow0, IC = IC,
                              adflag = adflag, Tb = Tb, nboot = 99,
                              nCores = 2) # simulate critical values via wild bootstrap. Note that the number of cores is arbitrarily set to 2.

Next, identify crisis periods, defined as periods where the test statistic is above its corresponding critical value, using the locate() function.

dim          <- obs - swindow0 + 1 
monitorDates <- spread$date[swindow0:obs]
quantile95   <- quantilesBsadf %*% matrix(1, nrow = 1, ncol = dim)
ind95        <- (bsadf > t(quantile95[2, ])) * 1
periods      <- locate(ind95, monitorDates)  # Locate crisis periods

Finally, print a table that holds the identified crisis periods with the help of the disp() function.

crisisDates <- disp(periods, obs)  #generate table that holds crisis periods
start end
1 2008-03-01 2008-03-01
2 2008-09-01 2009-04-01
3 2010-05-01 2012-08-01

Pleas check the packages’ articles for an elaborated analysis of the spreads data, as well as a demonstration using data on the S&P 500 price-to-dividend ratio.