I.

Applying a VAR model to macroeconomic data

Preparation and Data Loading
● The R environment is cleared of all existing objects and any open graphics devices
are closed to ensure a fresh start for analysis. Packages are also loaded to assist the
analysis:
# clear data & close graphs
rm(list=ls())
graphics.off()

# load package
library(tsm)
library(vars)
source("plot_varcheck.R")
● GDP and unemployment data are loaded from a CSV file and converted into time
series objects, starting from the first quarter of 1948:
# load data and create time series object ----
dat_sa <- read.csv("Datasets/blanchQua.csv")

gdp <- ts(dat$GDP, start = c(1948, 2), freq = 4) # gdp

une <- ts(dat$U, start = c(1948, 2), freq = 4) # unemployment

II. Plotting and Initial Analysis

GDP and unemployment data are plotted to visualize trends and cyclical patterns:
plot(cbind(gdp, une))
Autocorrelation functions for GDP and unemployment are computed to analyze
persistence in the data:
# look at the persistence in the data
gdp.acf <- ac(gdp, main = "output")

une.acf <- ac(une, main = "unemployent")

 ● GDP data appears to have low persistence with the impact of past values
diminishing quickly, as suggested by the ACF and PACF plots.
● Unemployment data shows high persistence, where past unemployment rates
continue to influence future rates for an extended period.

The unemployment data exhibits strong persistence, indicating the potential presence of a
unit root. To investigate this, we perform a Dickey-Fuller test, which allows us to assess
whether we can reject the null hypothesis of a unit root in the data:
# check unemployment for UR
adf.une <- ur.df(une, type = "trend", selectlags = "AIC")
summary(adf.une)

###############################################
# Augmented Dickey-Fuller Test Unit Root Test #
###############################################

Test regression trend

Call:
lm(formula = z.diff ~ z.lag.1 + 1 + tt + z.diff.lag)

Residuals:
Min 1Q Median 3Q Max
-0.70448 -0.21077 -0.04731 0.19844 1.03067

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.0059997 0.0536401 0.112 0.911
z.lag.1 -0.0972035 0.0204701 -4.749 4.67e-06 ***
tt -0.0001167 0.0005838 -0.200 0.842
z.diff.lag 0.6846610 0.0595395 11.499 < 2e-16 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
 Residual standard error: 0.3311 on 153 degrees of freedom
Multiple R-squared: 0.4776, Adjusted R-squared: 0.4674
F-statistic: 46.63 on 3 and 153 DF, p-value: < 2.2e-16

Value of test-statistic is: -4.7486 7.5583 11.3199

Critical values for test statistics:

1pct 5pct 10pct
tau3 -3.99 -3.43 -3.13
phi2 6.22 4.75 4.07
phi3 8.43 6.49 5.47

Based on the ADF test results, we can conclude that the time series data does not have a
unit root and is stationary. This conclusion is supported by both the test statistic (which is
more negative than the critical values) and the very low p-value.

III. Model Selection and Estimation

To create a combined dataset for the two time series we will analyze, we can simply
concatenate the two existing data objects by column:
dat.bv <- cbind(gdp, une)
colnames(dat.bv) <- c("gdp", "une")

After that, Information criteria (AIC, HQ, SC, FPE) are computed to select the
appropriate number of lags for the VAR model:
info.bv <- VARselect(dat.bv, lag.max = 12, type = "const")
info.bv$selectio

AIC(n) HQ(n) SC(n) FPE(n)

2 2 2 2

We can see that the criteria above all suggest using 2 lags for the model. We, therefore,
set 𝑝 = 2 to estimate the model:
# group data and choose lag length
bv.est <- VAR(dat.bv, p = 2 ,type = "const", season = NULL, exog
= NULL)
summary(bv.est)

VAR Estimation Results:

=========================
Endogenous variables: gdp, une
Deterministic variables: const
Sample size: 157
Log Likelihood: -211.593
Roots of the characteristic polynomial:
0.8361 0.8361 0.1014 0.1014
Call:
VAR(y = dat.bv, p = 2, type = "const", exogen = NULL)

Estimation results for equation gdp:

====================================
gdp = gdp.l1 + une.l1 + gdp.l2 + une.l2 + const

Estimate Std. Error t value Pr(>|t|)

gdp.l1 0.06905 0.10782 0.640 0.52289
une.l1-0.61530 0.30723 -2.003 0.04699 *
gdp.l2 0.04741 0.09098 0.521 0.60309
une.l2 0.90193 0.31426 2.870 0.00469 **
const -0.01570 0.07684 -0.204 0.83833
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.9613 on 152 degrees of freedom

Multiple R-Squared: 0.2687, Adjusted R-squared: 0.2495
F-statistic: 13.97 on 4 and 152 DF, p-value: 1e-09
Estimation results for equation une:
====================================
une = gdp.l1 + une.l1 + gdp.l2 + une.l2 + const

Estimate Std. Error t value Pr(>|t|)

gdp.l1 -0.119156 0.035975 -3.312 0.001157 **
une.l1 1.333482 0.102510 13.008 < 2e-16 ***
gdp.l2 -0.029883 0.030358 -0.984 0.326497
une.l2 -0.416951 0.104853 -3.977 0.000108 ***
const -0.007501 0.025639 -0.293 0.770243
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.3207 on 152 degrees of freedom

Multiple R-Squared: 0.9429, Adjusted R-squared: 0.9414
F-statistic: 628.1 on 4 and 152 DF, p-value: < 2.2e-16

Covariance matrix of residuals:

gdp une
gdp 0.9240 -0.2022
une -0.2022 0.1029

Correlation matrix of residuals:

gdp une
gdp 1.0000 -0.6558
une -0.6558 1.0000

The results indicate that the system is stable since all the roots of the characteristic
polynomial are less than 1. Moreover, the lagged values of the unemployment rate seem
to be significant predictors for the current GDP. The analysis also reveals that the past
values of the unemployment rate are significant determinants of the current employment
rate, highlighting the persistence and interdependence of labor market conditions over
time.

IV. Diagnostic Checks\

To assess the model's goodness of fit, we can conduct diagnostic tests on the residuals.
Specifically, to test for the presence of serial correlation, we can apply the Portmanteau
test, which can be implemented as follows:
# testing serial correlation with Portmanteau-Test
bv.serial <- serial.test(bv.est, lags.pt = 12, type =
"PT.asymptotic")
bv.serial

Portmanteau Test (asymptotic)

data: Residuals of VAR object bv.est

Chi-squared = 46.579, df = 40, p-value = 0.22

Since p-value is greater than 5%, we can conclude that there is an absence of serial
correlation. We, then, conduct residuals diagnostic checks for both GDP and
Unemployment rate data:
my_varcheck_plot(bv.serial, names = "gdp")
 my_varcheck_plot(bv.serial, names = "une")

After inspecting the visualizations of GDP and Unemployment rate residuals, it is quite
clear that there is no serial correlation left in the residuals, their variance are constant, and
they follow a normal distribution. This indicates that there are no clear patterns and the
residuals are a close resemblance of white noise. The model, therefore, seems to have
adequately captured the data’s structure.
To test for heteroscedasticity in the residuals we can perform a multivariate ARCH
Lagrange-Multiplier test:
# testing heteroscedasticity
bv.arch <- arch.test(bv.est, lags.multi = 12, multivariate.only
= TRUE)
bv.arch

ARCH (multivariate)

data: Residuals of VAR object bv.est

Chi-squared = 108.4, df = 108, p-value = 0.4712
The p-value we obtained is larger than 5%, which would indicate that there is no
heteroscedasticity.
To consider the distribution of the residuals, we could apply a normality test since it may
be subjective and inaccurate to just eye-ball the plots that we have drawn above:
# testing for normality
bv.norm <- normality.test(bv.est, multivariate.only = TRUE)
bv.norm

$JB

JB-Test (multivariate)

data: Residuals of VAR object bv.est

Chi-squared = 2.3194, df = 4, p-value = 0.6772

$Skewness

Skewness only (multivariate)

data: Residuals of VAR object bv.est

Chi-squared = 2.1157, df = 2, p-value = 0.3472

$Kurtosis

Kurtosis only (multivariate)

data: Residuals of VAR object bv.est

Chi-squared = 0.20376, df = 2, p-value = 0.9031

Once again, all the p-values are greater than 5%. This suggests that the residuals are
fairly normally distributed.
Lastly, to test for the structural break in the residuals we can apply a CUSUM test:
# testing for structural stability
bv.cusum <- stability(bv.est, type = "OLS-CUSUM")

plot(bv.cusum)

Because the CUSUM lines stay within bounds of the critical lines, the coefficients of the
model are stable over the period of the sample. There is no statistical evidence of
structural breaks within the given confidence level.

V. Granger causality, IRFs and Variance Decompositions

Now, we can test for Granger causality:
# Granger & instantaneous causality tests
bv.cause.gdp <- causality(bv.est, cause = "gdp")
bv.cause.gdp

$Granger

Granger causality H0: gdp do not Granger-cause une

 data: VAR object bv.est
F-Test = 5.5723, df1 = 2, df2 = 304, p-value = 0.0042

$Instant

H0: No instantaneous causality between: gdp and une

data: VAR object bv.est

Chi-squared = 47.214, df = 1, p-value = 6.364e-12

Based on the p-value obtained from the result above, we can reject the null hypothesis of
no Granger causality for GDP.
bv.cause.une <- causality(bv.est, cause = "une")
bv.cause.une

$Granger

Granger causality H0: une do not Granger-cause gdp

data: VAR object bv.est

F-Test = 12.821, df1 = 2, df2 = 304, p-value = 4.511e-06

$Instant

H0: No instantaneous causality between: une and gdp

data: VAR object bv.est

Chi-squared = 47.214, df = 1, p-value = 6.364e-12

We reached the same conclusion for Unemployment rate as for GDP.

To generate impulse response functions that illustrate the dynamic response of output to
an unemployment shock, we can follow these steps:
 irf.gdp <- irf(bv.est, impulse = "une", response = "gdp",n.ahead
= 40, boot = TRUE)
plot(irf.gdp, ylab = "ouput", main = "Shock from unemployment")

It is clear that while an unemployment shock initially has a significant negative effect on
GDP, the effect is temporary and GDP is expected to recover over time back to its
pre-shock levels.
To consider the response of unemployment to output shock we do as follows:
irf.une <- irf(bv.est, impulse = "gdp", response = "une",n.ahead
= 40, boot = TRUE)
plot(irf.une, ylab = "unemployment", main = "Shock from output")
The same thing can be said for Unemployment rate, a positive shock in GDP would
typically lead to a negative effect on unemployment, and this also tapers off through time.
Similarly, we can look at the effect of an unemployment shock on unemployment:
irf.une_un <- irf(bv.est, impulse = "une", response =
"une",n.ahead = 40, boot = TRUE)
plot(irf.une_un, ylab = "unemployment", main = "Shock from
unemployment")

To generate the forecast error variance decompositions we make use of the fevd
command, where we set the number of steps ahead to ten:
# generate variance decompositions
bv.vardec <- fevd(bv.est, n.ahead = 10)
plot(bv.vardec)
 ● The variance decomposition for GDP shows that the majority of the forecast error
variance in GDP can be attributed to its own shocks. This indicates that GDP is
largely explained by its own innovations over the horizon of 10 periods.
● In the case of unemployment, the plot indicates that its own shocks and GDP
shocks both account for a substantial portion of the forecast error variance. This
suggests unemployment is influenced by both shocks.

VI. Forecasting
To generate forecasts, we can utilize the predict() command, which in this case will
produce forecasts for the next 8 time periods. Additionally, we aim to obtain 95%
confidence intervals for the forecasted values, providing a measure of the uncertainty
around the point forecasts:
# Forecasting with the VAR model
predictions <- predict(bv.est, n.ahead = 8, ci = 0.95)
plot(predictions, names = "gdp")

plot(predictions, names = "une")

Furthermore, the forecasts and their associated uncertainty can be visually represented
using fan charts, which provide a graphical depiction of the 95% confidence intervals
alongside the point forecasts:
# Fanchart for forecasts ----
fanchart(predictions, names = "gdp")
fanchart(predictions, names = "une")