Seasonal Unit Root Test

Nicolás Ronderos

June 25, 2015

This document explains how to perform a seasonal unit root test using an Eviews add-in. The add-in test works
with biannual, quarterly and monthly data. The add-in contains the option to obtain the p-value via Monte Carlo
methods.

1 HEGY Test

The seasonal unit root test of the add-in is based on [4], which introduces the test but concentrates in quarterly
time series. The test for biannual and monthly data can be found in [2, 3].
The goal of the test is to determine what class of seasonal processes is the generator of the seasonality, these
could be: a deterministic seasonal processes, a stationary seasonal processes or an integrated seasonal processes. A
series st is an integrated seasonal processes if it has a seasonal unit root and a peak at any seasonal frequency in its
spectrum (excluding the zero frequency), more generally it is integrated of order d at frequency θ i.e. st ∼ Iθ (d).
The method of performing the test is explained below for biannual, quarterly and monthly time series.

1.1 Biannual

For these temporal aggregation frequency the test consist of estimating the parameters in equation (1) by OLS.

Φ(L)s3,t = µt + π1 s1,t−1 + π2 s2,t−1 + et (1)

Where µt are deterministic regressors, in the add-in four options of these regressors are included: intercept,
intercept and trend, intercept and seasonal dummies and intercept, trend and seasonal dummies, whenever the
seasonal dummies are included these will be centered dummy variables, and where:

s3,t = (1 − L2 )st (2)

s1,t = (1 + L)st (3)

s2,t = −(1 − L)st (4)

And Φ(L) is a polynomial in the lag operator, in the add-in the order of Φ(L) is selected using information
criteria. If π1 = 0 then the presence of the non-seasonal unit root cannot be rejected. If π2 = 0 then the presence
of a seasonal unit root cannot be rejected. This hypothesis can be tested using (one-sided) t-test, nevertheless the
distribution of these tests are nonstandard.

1.2 Quarterly

The test consist of estimating the parameters in equation (5) by OLS.

Φ(L)s4,t = µt + π1 s1,t−1 + π2 s2,t−1 + π3 s3,t−1 + π4 s3,t−2 + et (5)

Where:

s4,t = (1 − L4 )st (6)

s1,t = (1 + L + L2 + L3 )st (7)

1
 s2,t = −(1 − L)(1 + L2 )st (8)

s3,t = −(1 − L2 )st (9)

For this temporal aggregation there could be two seasonal unit roots. The hypothesis on the coecients π1 and
π2 are tested using t-tests as before. If π3 = π4 = 0 then the presence of a seasonal unit root cannot be rejected,
these hypothesis can be tested using a joint F-test, which has a nonstandard distribution.

1.3 Monthly

For this the test consist of estimating equation (10) by OLS.

Φ(L)s8,t = µt + π1 s1,t−1 + π2 s2,t−1 + π3 s3,t−1 + π4 s3,t−2 + π5 s4,t−1 + π6 s4,t−2 + π7 s5,t−1 + (10)

π8 s5,t−2 + π9 s6,t−1 + π10 s6,t−2 + π11 s7,t−1 + π12 s7,t−2 + et

Where.

s8,t = (1 − L12 )st (11)

s1,t = (1 + L)(1 + L2 )(1 + L4 + L8 )st (12)

s2,t = −(1 − L)(1 + L2 )(1 + L4 + L8 )st (13)

s3,t = −(1 − L2 )(1 + L4 + L8 )st (14)

√
s4,t = −(1 − L4 )(1 − 3L + L2 )(1 + L2 + L4 )st (15)

√
s5,t = −(1 − L4 )(1 + 3L + L2 )(1 + L2 + L4 )st (16)

s6,t = −(1 − L4 )(1 − L2 + L4 )(1 − L + L2 )st (17)

s7,t = −(1 − L4 )(1 − L2 + L4 )(1 + L + L2 )st (18)

The hypothesis on the coecients π1 ,π2 ,π3 and π4 are tested as before. If πi = πi+1 = 0 then the presence of a
seasonal unit root cannot be rejected for i = 5, 7, 9, 11, this hypothesis can be tested using a joint F-test, which has
a nonstandard distribution.

2 Critical values and Monte Carlo

The add-in contains two options to obtain a rule to decide if the null must be rejected or not. The rst consist to
compare the value of the statistics with the critical values tabulated in [3, 4], which are showed in the add-in. And
the second consist on simulate the distributions of the statistics using the Monte Carlo method, the steps for the
simulations performed in the add-in are showed below.
1. Simulate the process st = st−s + et where et ∼ nid(0, 1), s = 2 for the biannual test, s = 4 for the quarterly
test and s = 12 for the monthly test. The simulated process contain the same number of observations than
the tested time series, this gives more precision than compare the statistics with the tabulated critical values,
given that they are available only for some values of the number of observations.
2. Obtain the value of the statistics in accordance with the temporal aggregation of the data.
3. Repeat steps 1 and 2 a one preferably large number of times.
4. Obtain the associated probability of the statistical considering their empirical distribution.

2
3 Using the add-in

The add-in can be used via point and clicking or via command line. Some examples will be showed using data of the
money aggregate M1 of Colombia and of the critical radio frequencies (CRF) in Washington D.C. both present a
strong seasonal pattern, the time series are plotted in gure 1. A workle with the data comes with the installation.

Figure 1: CRF and M1

After installing the add-in the user has to click in the time series object proc->Add-ins->HEGY to perform
the test. Then the window showed in gure 2 will appear. As mentioned the add-in contains several options of
deterministic regressors.

Figure 2: Add-in window

3
 To access the add-in via command line the user just has to follow the instructions given in table 1. The defaults
of the command line are the same as the ones showed in gure 2.

series_name.hegy(options)
Options
Intercept,trend and seasonal dummies itd
Intercept and seasonal dummies id
Intercept and trend it
Intercept i
Akaike criteria to select optimum lag length aic
Schwarz criteria to select optimum lag length schwarz
Hannan-Quinn criteria to select optimum lag length hq
Maximum lags to test maxlag=8
Simulate via Monte Carlo mc
Number of simulations simulations=1000

Table 1: Command line

Figure 3: HEGY test on M1

Figure 4: HEGY test on CRF

For example, to perform the unit root test on M1 via command line the user has to use the following command
m1.hegy(hq,maxlag=6,mc,simulations=2000). The results of the test are showed in gure 3, it can be seen that

4
there is evidence of seasonal unit roots at almost all seasonal frequencies, therefore to model the nonstationary
seasonality of this series a lter of the form (1 − L12 )M 1t is suggested.
The non-seasonal unit root statistic corresponds to the t-statistic of π1 , the seasonal unit root statistic with
2 month per cycle corresponds to π2 , the statistics reported by the add-in corresponds respectably to the tested
coecient or pair of coecients, this also holds for the other temporal aggregations.
The HEGY test on CRF is showed in gure 4, this was carried out with the following command crf.hegy(id,aic,maxlag=4).
It can be seen that there is evidence of one unit root at zero frequency, therefore to model the time series the usual
rst dierence lter is suggested.

References

[1] J Joseph Beaulieu and Jerey A Miron. Seasonal unit roots in aggregate us data. Journal of econometrics,
55(1):305328, 1993.

[2] Philip Hans Franses. Seasonality, non-stationarity and the forecasting of monthly time series. International
Journal of Forecasting, 7(2):199208, 1991.

[3] Philip Hans Franses and Bart Hobijn. Critical values for unit root tests in seasonal time series. Journal of
Applied Statistics, 24(1):2548, 1997.

[4] Svend Hylleberg, Robert F Engle, Clive WJ Granger, and Byung Sam Yoo. Seasonal integration and cointegra-
tion. Journal of econometrics, 44(1):215238, 1990.