An R Package for Forecasting Models with Real-Time Data.
The nowcasting package contains useful tools for using dynamic factor models. In this version of the package we present three methods, based on the articles of Giannone et al. 2008 and Bańbura et al. 2011. Furthermore, the package offers auxiliary functions to treat variables, constuct vintages, visualize results, etc.
The package is in development. Reviews, comments and pull requests are welcome.
github
devtools::install_github('nmecsys/nowcasting')
CRAN
install.packages('nowcasting')
Two examples of how the nowcasting package can be used are discussed below. In the first example we use the dataset from Giannone et al. (2008), which uses a time series panel for the US economy. In the second example we show how to make forecasts based on pseudo real time vintages using Brazilian data and how to use information criteria for determining the number of factors and shocks to use in the model.
The dataset used in Giannone et al. (2008) can be added to the environment with the following command.
library(nowcasting)
data(USGDP)
USDGP is a list with two data frames:
USGDP$base
: this is an unbalanced panel with the model´s variables;USGDP$legend
: this contains the legend with the relevant information for each variable in the USGDP$base
dataframe.In order to use the nowcast
function we require stationary variables. This can be done by using Bpanel
. This function creates a balanced panel using an unbalanced panel as input. The default option is to substitute missing observations and outliers using the outlier correction methodology from Giannone et al. (2008). The function includes most usual transformations to obtain stationary variables. For this particular example, the object USGDP$legend
contains all the transformations used in Giannone et al. (2008). All the explanatory variables are monthly while the GDP observation is quarterly as captured by the frequency vector. Note however, that the data object is a monthly mts. It is usual to include quarterly variables as monthly time series where the two first months of each quarter appear as NA. In this example the quarterly value is repeated.
data <- Bpanel(base = USGDP$base,
trans = USGDP$legend$Transformation,
aggregate = FALSE)
frequency <- c(rep(12, ncol(data) -1), 4)
Once these variables have been treated, the nowcast
function can be used to estimate the model´s parameters according to the selected estimation method, the number r of dynamic factors, the lag order of the factors p and the number q of shocks to the factors. For this example we use the Two-Stage - With aggregation method that will aggregate monthly factors as in Mariano and Murasawa (2003). The arguments r, p and q were defined according to Giannone et al. (2008).
nowcastUSGDP <- nowcast(formula = RGDPGR ~ ., data = data, r = 2, p = 2, q = 2,
method = '2s_agg', frequency = frequency)
The in sample evaluation from Giannone et al. (2008) could be reproduced by looking at the ACF of the residuals of the model specified above.
res <- ts(nowcastUSGDP$reg$residuals, start = start(data), frequency = 4)
acf(window(res, start = c(1985,1), end = c(2004,4)))
The results can be accessed from the object nowcastUSGDP
.
# y forecasts
tail(nowcastUSGDP$yfcst,8)
# the regression between y and its factors can be accessed using `$reg`.
summary(nowcastUSGDP$reg)
# the results related to the estimation of factors
tail(nowcastUSGDP$factors$dynamic_factors) # factors
head(nowcastUSGDP$factors$Lambda) # Lambda matrix
nowcastUSGDP$factors$A # A matrix
nowcastUSGDP$factors$BB # BB': u's variance covariance matrix (factor equation)
diag(nowcastUSGDP$factors$Psi) # Psi: epsilon's variance covariance matrix (x equation)
# the forecasts of the explanatory variables are in `$xfcst`.
tail(nowcastUSGDP$xfcst[,1:5]) # x forecasts (first 5 variables)
The graphs available with the nowcast.plot
function allow the to visualize some results of interest.
# y fcst
nowcast.plot(nowcastUSGDP, type = "fcst")
# factors
nowcast.plot(nowcastUSGDP, type = "factors")
# how much of the variability in the dataset is explained by each factor
nowcast.plot(nowcastUSGDP, type = "eigenvalues")
# importance of each variable in the first factor
nowcast.plot(nowcastUSGDP, type = "eigenvectors")
In this example we work with the data the Federal Reserve of New York uses in its weekly nowcasting report. The explanatory variables are mixed frequencies including both monthly and quarterly series.
library(nowcasting)
data(NYFED)
Similarly to the previous working example, the object NYFED contains all the necessary information to run the nowcast
function. The block structure, the transformations to make the variables stationary and the frequencies of the variables can be loaded as illustrated below.
base <- NYFED$base
blocks <- NYFED$blocks$blocks
trans <- NYFED$legend$Transformation
frequency <- NYFED$legend$Frequency
The data-set x can be prepared by using the function Bpanel
. For the EM algorithm, we do not want to replace the missing values that are not part of the jagged edges as was the case with the Two-Stage method. This can be done by telling the function not to replace those particular missing values, i.e. NA.replace = F
. We also do not want to discard series with many missing values and therefore use na.prop = 1
.
x <- Bpanel(base = base, trans = trans, NA.replace = F, na.prop = 1)
The same setting as the NY FED is used. We therefore limit the number of factors, r, per block to one and define the factor process as a VAR(1). The algorithm displays the convergence of the loglikelihood function every 5 iterations. As opposed, to the Two-Stage method, the x represents the entire data-set and y represents the name of the variable that is being forecast.
nowEM <- nowcast(formula = GDPC1 ~ ., data = data, r = 1, p = 1,
method = "EM", blocks = blocks, frequency = frequency)
The forecasts can be visualized using the function nowcast.plot
as illustrated below.
nowcast.plot(nowEM)
A vintage is a dataset observed on a specific date. The latter is useful to evaluate the out of sample performance of our model. The nowcasting package contains a dataset of Brazilian economic time series.
library(nowcasting)
data(BRGDP)
The PRTDB
function is intended to construct pseudo real time vintages of any dataset. The function excludes observations from time series based on the lag information provided by the user and simulates what would be observed on the reference date. In this case we have construct a 10 year database ending in 2015-06-01.
vintage <- PRTDB(mts = BRGDP$base, delay = BRGDP$delay, vintage = "2015-06-01")
base <- window(vintage, start = c(2005,06), frequency = 12)
x <- Bpanel(base = base, trans = BRGDP$trans)
The variable to be forecast is then made stationary. We also use the month2qtr
function to cast GDP as a quaterly variable.
GDP <- base[,which(colnames(base) == "PIB")]
y <- diff(diff(GDP,3),12)
Information criteria can be used in order to help determine the number of factors r and shocks to the factors q that the model should have.
ICR1 <- ICfactors(x = x, type = 1)
ICR2 <- ICfactors(x = x, type = 2)
ICQ1 <- ICshocks(x = x, r = 2, p = 2)
The user is now ready to forecast the variable of interest. The summary of the regression can be accessed as illustrated below.
data <- cbind(y,x)
colnames(data) <- c("y",colnames(x))
frequency <- c(4,rep(12,ncol(x)))
now <- nowcast(formula = y~., data = data, r = 2, q = 2 , p = 2, frequency = frequency)
summary(now$reg)
Finally the in- and out of sample forecasts for this particular vintage can be visualized using the nowcast.plot
function.
nowcast.plot(now, type = "fcst")