Research Journal of Mathematics and Statistics 2(1): 23-31, 2010

ISSN: 2040-7505
© M axwell Scientific Organization, 2010
Submitted Date: October 15, 2009 Accepted Date: November 12, 2009 Published Date: March 20, 2010

Short and Long Memory Time Series Models of Relative

Humidity of Jos Metropolis
1
M.A. C hiaw a, 2 B.K. Asare and 3 B. Audu
1
Departm ent of Mathematics and Computer Science, Benue State University, M akurdi, Nigeria
2
Departm ent of Mathematics, Usmanu D anfodiyo University, Soko to, Nigeria
3
Departm ent of Mathematics, Gombe State University, G ombe, N igeria

Abstract: The percentage monthly relative hum idity of Jos metropolis is examined in this study. Two models,
a short memory seasonal autoregressive integrated moving average model [SARIMA (1,0,1)(2,1,2)] and long
mem ory autoregressive fractional integrated moving average [ARFIMA(1,0.29,1)] are used to fit the same
hum idity data. Even though both models fit the data well, forecasts obtained from the ARFIM A(1,0.29,1)
capture the sw ing in the data and resemble the actual values better than the forecasts using
SARIM A(1,0,1)(2,1,2) mod el. This result shows that the Jos metropolitan data is better fitted by a long m emory
time series which captures the long swing in the weather data better than the short m emo ry time series m odels
whose effect quickly dies down.

Key w ords: Precipitation, autocorrelation function, autocovariance, spectrum, periodogram, fractional

integration

INTRODUCTION are especially suited for m odeling sho rt memory

processes with little or no pe rsistence shock.
The amou nt of water vapor contained in the air at a To our knowledge, ARFIM A models have not been
particular time is called its humidity. Ev en thoug h there used substantially to mod el hum idity data in Nig eria. Th is
are three different ways to measure humidity, the most is therefore one of the contributions of this research. The
frequently used is the relative humidity. Relative hum idity second contribution of the research is to be able to predict
indicates the likelihood of precipitation, dew, or fog. The the humidity pattern in Jos metropolis with some degree
relative humidity if it is high makes us feel hotter outside. of accuracy. This enables the meteorolog ist warn peop le
This is usually experienced in the sum mer, that is, in advance the next weather pattern to expect within a
between June and September. It also reduces sweating couple of m onths.
which shou ld coo l the body thu s preventing the
evaporation of perspiration from the skin. This is why Theoretical framework: Meteorological studies suggest
geograp hers usually state that ‘warmer air holds more that meteorological variables exhibit characteristics that
moisture’. In warmer air, there is more energy for more are more consistent with long memory (Haslett and
water molecules to hold themselves and overcome Raftery, 1989; M ontan ari et al., 1996; Hunt and Nason,
hydrogen bonds which seek to pull water molecules 2001; Bhardwaj and Swanson, 2004). In a long memory
together. process, the autocorre lation of a variable decay s very
In the development of space modeling of time series slow ly, in other words, the autocorrelation fun ction of a
data w ith long m emo ry dep endence , Haslett and R aftery long mem ory process typically decays at a hyp erbolic rate
(1989) assessed Ireland’s wind power resources using (Haslett and Raftery, 1989). That is to say that these
autoregressive fractional integrated moving average processes have autocovariances that are not absolutely
(AR FIM A). Since then it has become usual to model summab le (Hurst, 1951). A sequence of matrices
meteorological time series with an ARF IMA model (Hsu
et al., 1998; Hunt and N ason, 200 1). Haslett and R aftery is absolutely su mm able if each of its element
(1989) and Boutte et al. (2004 ) fit an ARFIMA m odel to forms an absolutely summable scalar sequence, that is
the Irish wind data and found seasonal component in the
series with the residuals exhibiting an autoregressive
moving average (ARM A) process. Boutte et al. (2004) for i,j = 1,2, ..., n, where is the
then went ahead to conclude that ARM A, Autoregressive
Integrated Mo ving A verage (A RIM A) and Seasonal
row i colum n j of .
Autoregressive Integrated M oving A verage (SAR IMA ),

Corresponding Author: M.A. Chiawa, Department of Mathematics and Computer Science, Benue State University, Makurdi,
Nigeria
23
 Res. J. Math. Stat., 2(1): 23-31, 2010

Long memory processes have been popularized by

whe re is the standard error for , and ˆ denotes
Grange and Joyeux (1980) and Hosking (1981). A long
mem ory autoregressive fractional integrated moving estimate. The null hypothesis of unit root is accep ted if
average (ARFIM A) process is a process that must be the test statistic is greater than the critical values.
differenced d-times, where d can take on non-integer
values, to achieve a stationary process. This process also Kwiatkowski, Phillips, Schmidt, and Sh in (KP SS) test:
has the property that its spectrum is unbounded at the The integration properties of a series y t may also be
origin. The notion of fractional integration has prov en to investigated by testing the null hypothesis that the series
be quite important in modeling meteorological data since is stationary aga inst a un it root. Kw iatkow ski et al. (1992)
the study of Haslett and Raftery (1989). Whitcher (1998) derived a test for this pair of hypotheses. Assuming no
use the discrete wavelet transform to construct a test for linear trend term, the data generating process is given as
hom ogeneity of variance in time series exhibiting long
memory characteristics and applied the result to the Nile
(2)
river w ater levels.
whe re x t is a rand om w alk, x t = x t-1 +v t , v t ~iid(0, )
MATERIALS AND METHODS
and z t is a stationary process. Kwiatkowski et al., (1992)
This study was conducted at the Jos International propose the follow ing test statistic
Airport by the Meteorological Department of the Federal
Ministry of Aviation in Nigeria. Monthly percentage
relative humidity time series data was collected from (3)
January 1985 to D ecem ber 20 04. Therea fter, the da ta
generating p ro cess o f th e S ARIM A and A RFIMA m odels
are tested. T o app ly the ARFIMA and SARIM A tests the
variables are first examin ed for u nit root and stationarity.
The Augmented D ickey Fuller (ADF) test and the whe re with and an
Kw iatkow ski, Phillips, Schmidt and Shin (K PSS) test are
used for this purpose. These preliminary tests are
necessary in order to determine the order of estimator of the long run variance of
nonstationa rity of the data. If the series is integrated of
o rd er o ne th en usin g A R FIM A and SA R IM A models on
the data results in good specification of these models.
These tests are discu ssed below .

Adju sted Dickey -Fuller (ADF ) test: The ADF

regression equation due to Dickey and Fuller (1979) and
Said and Dickey (1984) is given by: The null hypothesis of the test is against the

alternative hypothesis . This test uses the

Bartlett windo w w ith a lag truncation param eter

t = p+ 1, p+2, … ,T. (1) (4)

whe re
whe re µ o is the intercept, represents the trend incase
it is present, N is the coefficient of the lagged dependent
variab le . and p lags o f with coefficients
a j are added to account for serial correlation in the
residuals. The null hypothesis H O : N =0 is that the series and
has unit root while the alternative hypothesis H 1 : N…0 is
that the series is stationary. The ADF test statistic is given
by

Reject the null hypo thesis if the test statistic is greater

than the asymptotic critical values.

24
 Res. J. Math. Stat., 2(1): 23-31, 2010

The AR FIM A m odel: ARF IMA models were first

introduced by Granger and Joyeux (1980) and Hosking
(1981). An ARFIMA (p,d’,q) process is a stationary
process that satisfies: (10)

t = 1,2,… ,T (5)
The last term on the right-hand side of (10) becomes
whe re d is the parame ter of fractional differentiation, c is
negligible when low -frequency ordinates are close to
a constant and and are autoregressive and
zero. Equation (10) the n results in the followin g simple
moving average polynomials of order p and q, regression equation
respectively. The au tocorrelations, p k fo r an A R FIM A
process for large k and d<1/2 are given by the following
approximation (Granger and Joyeux , 1980). (11)

(6) whe re the intercept , with

parame ter and the error term is

which is a monotonic, hyperbolic function. The
autocorrelation function of A RFIMA decreases slowly to
zero while its spectral density is infinite (Hosking, 1981).

Estimating AR FIM A M odels: ARFIMA is a more

The number of ordinates j to be used in the estimation
general modeling of the autoregressive integrated moving
of the regression trunc ates at n = g(T ), where g(T) satisfy
average (ARIM A) model that allows non-integer d values.
the following con ditions:
A typical ARFIM A model is given by the equation

(i) and
(7)

where
(ii)

(8) The function g (T) = T : , with 0 < : < 1 is the number

of ordina tes used to estimate d and this satisfies b oth
conditions. The estimator of d is therefore consistent
(Hurvich et al., 1998). The test of hypothesis for the
parameter d can be done based on the asymptotic
In this study a simp le method of estimating d is used. distribution of derived by Geweke and Porter-Hudak
This method is due to Geweke and Porter-Hudak (1983)
and starts the estimation of d w ith the spectral density (1983).

function of

(12)
(9)

whe re y i is the regressor ln [4 Sin 2 (8 j /2)] . Since

Take logarithms and evaluate at harmonic frequencies
and it follow s that p

lim s 2 = B 2 / 6, where s2 is the sample variance of the

residuals from the regression Eq. (11).
The value of the pow er factor, : is the determinant of
add the periodogram I(8 j ) at ordinate j, to both sides of (9) ordinates included in the regression. Traditionally the
yields number of ordinates is chosen from the interval [T0 . 4 5,

25
 Res. J. Math. Stat., 2(1): 23-31, 2010

T 0 . 5 5]. However, Hurvich et al. (1998) recently showed

that the optimal lag m is O(T 0.8 ).

S AR IM A mo del: A Seasonal Autoregressive Integrated

Moving Average (SARIMA) model denoted b y SAR IM A
(p,d,q)(P,D ,Q) is given as

(13)

whe re p deno tes the n umb er of autoregre ssive terms, q the

number of moving average terms and d is an integer
which denotes the number of times the series must be
Fig. 1: Time plot of the monthly percentage relative humidity
differenced to attain stationarity. P is the number of of Jos
seasonal autoregressive com ponents, Q denotes the
number of seasonal moving average terms and D the
number of seasonal differences required to induce
stationarity. a is a con stant, i.e. e t is a

sequence of uncorrelated normally distributed random

variables with the same mean : and variance F 2 , L is the
lag operator defined by Lk X i = X t-k ,

and

Fig. 2: The ACF and the PACF of the monthly percentage

relative humidity of Jos
Residual analysis: The mod el spec ified is statistically
tested for adequacy using a barrage o f tests. Th ese tests
include: th e Portman teau test, L ag rang e multiplie r (LM )
test, Jarque-Bera test, autocorrelations and partial
autocorrelations, and the Ljung-Box test. Details of these
tests can be found in Bro wn, D urbin and E vans (1975),
Ljung and Box (1979), Jarque and Bera (1987) and
Harvey (1990, 1993). Detailed discussions for these tests
are skipped in this section to reduce the size of the paper.

Analysis of data:
Initial analysis of data: A time plot of the original series
is conducted in Fig. 1. A visual inspection of the plot
shows that the series has constant mean and variance. The
plot however does not show any evidence of stationarity.
W e proceed to examine the autocorrelation function
Fig. 3: The Periodogram of the monthly percentage relative
(Fig. 2). The autocorrelation function (A CF) is strongly
humidity of Jos
period ic with period 12 and very persistent. Clearly the

26
 Res. J. Math. Stat., 2(1): 23-31, 2010

data shows seasonal behaviour. The spectral analysis is Tab le 1: Re sults of the AD F test do wn proce dure
Actual no. of lags 5% Test statistics Decision
then used to test the existence of periodicity in the time
4 -2.86 -5.14 Stationa ry
series.
Figure 3 shows the periodogram of the series. It can Tab le 2: R es ults of S A RIM A mo del id entific ation for th e hu mid ity of
be seen that there is a frequency of 1 cycle every 12 jos
months. W e therefore co nclud e that the data generating Mod el Loglikelihood AIC
SARIM A(2,0,2)(2,1,2) -723.44 1466.88
process is seaso nal and the length is 12 months, which
SARIM A(2,0,2)(1,1,1) -737.16 1490.31
coincides with the result of the ACF. SARIM A(1,0,1)(2,1,2) -724.65 1465.30

Un it root test for the humidity data: The ADF test is Table 3: Re sults of SARIMA model estimation for h um idity time
series
conducted at levels with seasonal dummies because the
Va riable Coefficient Std. Error T-Stat P-Value
time series is strongly periodic. The result of the test Constant 0.2702 0.120 02.246 0.02500
given in Table 1 sh ows that the series at lag 4 is stationary Phi-1 0.9300 0.070 13.033 0.00001
since at 5% the test statistic is -5.14 with the critical Phi-1 - 1.0140 0.080 - 13.121 0.00001
value of -2.86 . Phi-2 - 0.0150 0.076 - 00.200 0.84400
The ta-1 - 0.8600 0.100 - 08.960 0.00001
The ta-1 0.0700 0.060 01.183 0.24000
Formulation of the S AR IM A m odel for hu mid ity data: The ta-2 - 0.8700 0.050 - 16.800 0.00001
The order of seasonal integration as shown in Fig. 3 of the
series is one as there is a single spike in the periodogram. Table 4: Sum mary of Diagn ostic Tests of SARIMA (1,0,1) (2,1,2)
Since the series is stationary (see results of the A DF test) Mod el
Test p-value (P 2 )
d = 0. The result of the spectral analysis shows that s = 12
Ljung-Box 0.3580
and the order o f seaso nal integ ration of the series is 1. The Jarque -Bera 0.4137
optimal lag lengths for both the AR and M A are selected ARCH-LM 0.0621
using the information criteria. T hese are also used to
Table 5: Re sults of S AR IM A ( 1,0 ,1) (2 ,1,2) forecast values for t wo
determine the lag lengths of seasonal integration for the
years
AR and MA . Three candidates models are obtained using In sample forecast Out of sam ple forecast
p = 1, 2; q = 1, 2; P = 1, 2; Q = 1, 2. O ut of these models --------------------------------------- ------------------------------------------
a parsimonious model is obtained w hich has the lowest Hu mid ity Prediction Hu mid ity Prediction
AIC and log likelihood function. The value of the log 22 16.54 20 17.69
18 17.12 27 21.69
likelihood function and the AIC for three candidates 23 21.45 19 30.77
SAR IMA mod els are comp uted and rep orted in Table 2. 58 48.06 45 50.46
Before the choice of our m odel, standard diag nostic 70 65.17 59 70.97
test are conducted using the residual ACF and PACF of 72 70.16 72 83.56
80 81.72 76 86.10
all the candidate models. As seen in Fig. 4, 6 and 8 all the 80 80.66 80 91.31
mod els pass the stand ard diagnostic testing criteria as 66 69.61 76 81.69
their residual values lie within the 95% confidence 59 53.75 40 69.00
interval band. The residuals are also plotted against time 26 28.86 22 50.44
16 19.98 24 42.41
to ascertain whether or not the models have passed the 14 17.92 17 36.21
standard test criteria of being w hite noise. These are 13 17.52 15 29.80
reported in Fig. 5, 7 and 9. The results show that the 20 19.60 20 36.57
residual plots for all the models are white noise. Since all 20 37.59 42 27.97
66 62.19 61 82.68
the residuals of the candidates models behave well, we 76 65.35 66 104.14
proceed to select SARIMA (1,0,1)(2,1,2) which has the 72 79.10 74 113.98
lowest AIC an d log likelihoo d as the best mod el. 82 77.32 82 102.29
62 70.90 66 91.94
30 42.79 50 65.41
Parameter estimates of the model using hum idity data: 20 23.03 24 34.35
After the best model has been chosen, the parameters of 18 15.08 24 27.24
the model are next estimated. The result of the parameter
estimates of the optimal model are shown in Table 3. The results of the tests show that the model is
After fitting the model, we check the model for adequacy. adeq uate at 5% as all the p-values are greater than 0.05.
This is what we are going to se e in the next section. The model is then used to forecast future values of the
series.
Model checking for SA RIM A(1,0,1)(2,1,2): The model
is next tested for adequacy using three diagnostic tests, SARIM A(1,0,1)(2,1,2) Forecast Evaluation: After a
Ljung-Box test, Jarque-Bera test and AR CH -LM test. The good SARIM A m odel has been fitted, its ability to
results of these tests are given in Table 4. forecast the time series data is tested. This further testifies

27
 Res. J. Math. Stat., 2(1): 23-31, 2010

Fig. 4: ACF and PACF of SARIMA(2,0,2)(2,1,2)

Fig. 6: Residual ACF and PACF of SARIMA (2,0,2) (1,1,1)

Fig. 5: Residual plot of SARIMA(2,0,2)(2,1,2)

the validity of this model. Table 5 contains the two years

in-sam ple and o ut of sample hum idity forecasts using two Fig. 7: Residual plot against time of SA RIM A (2,0,2)
benchm arks. (1,1,1)
The prediction is consistent with the observed
hum idity and therefore the model is adequate, valid and foreca sts are better than th e out-o f-sample forecasts. This
good. Forecast erro rs are the n com puted in order to assess shows one of the inadequacies of the model even though
the perform ance of the forecasts, in particular, the smaller it has passed all the tests. One of those things that may
the Root Mean Square E rror (RMSE), M ean A bsolute cause such behavior is choosing d=0 when it is a fraction
Error (MAE), Mean Absolute Percentage Error (MAPE) (Haslett and Raftery, 198 9).
the better the forecasts. These values are repo rted in An alternative way of modeling the meteorological
Table 6 for both in-sample and out-of-sample forecast data is by fitting a n A R FIM A model (Haslett and Raftery,
performance of the SAR IMA(1,0,1)(2,1 ,2) mod el to 1989). In the subsequent section this task is accomplished
determine its forecast ability and to mak e inference on and the resu lt compared with the SA RIM A m odel.
which mode of forecast is better for the m odel.
Since th e RM SE, MAE and the MA PE of the in- A RF IM A modelling of the relative humid ity data:
sample forecast are smaller than those of the out-of- Here, an ARFIM A model is fitted to the meteorological
sample forecast, the result suggests that the in-sam ple
data. Before estimating the d the ACF of the m onthly

28
 Res. J. Math. Stat., 2(1): 23-31, 2010

Fig. 11: The smoothed Periodogram

Fig. 12: The ACF of the fractionally differenced humidity Time

Series

Table 6: Re sult of forecast comparison of SARIMA (1,0,1) (2,1 ,2)

Mod el
Fig. 8: Residual ACF and PACF of SARIMA (1,0,1) (2,1,2) Mo de of forecast RMSE MAE MAPE
In-sa mp le 6.56 4.82 0.16
Ou t-of-sa mp le 19 .1 16.71 0.46

Table 7: Results of Estimating the d parameter

Co efficie nt Estim ate Asy mpto tic standard Stand ard error
deviation deviation
D 0.2911554 0.2196708 0.1770737

relative humidity of the long memory process is

conducted as shown in Fig. 10. The autocorrelation
function of the humid ity decreases slowly at a hyperbolic
rate in conformity with a fractional integrated series
(Haslett and R aftery, 1989; G il-Alana, 2008).
A spectral density of the series is then conducted to
determine the homogeneity of variance the smoothed
periodogram. Figure 11 is the resu lt of the smoothed
Fig. 9: Residual plot against time of SARIMA (1,0,1) (2,1,2) periodogram.
The smoothed periodogram of the se ries sho w that it
has constant variance. This is a third property of a long
mem ory series. (H aslett and Raftrey, 1989). The degree of
autocorrelation in the frac tionally differenced h umid ity
series is examined using the autocorrelation function as
show n in Fig . 12.
Even though the series has been fractionally
differenced. and is stationary with a zero infinite long
mem ory characteristics, it still exhibits the behavior of an
AR MA process (Ha slett and Raftery, 198 9).

Estimating d using Geweke Porter-Hudak method:

The long memory parameter d is estimated using the
Geweke and Porter-Hudak (1983) method. The estimated
value of the parame ter, its asymptotic deviation value and
regression standard deviation values are rep orted in
Table 7.
Fig. 10: The autocorrelation function of the humidity of Jos
After estimating the long mem ory param eter d and
Metropolis
using it to fractionally difference the humidity time series,

29
 Res. J. Math. Stat., 2(1): 23-31, 2010

we mov e ahead to m odel the series as an A R FIM A Tab le 8: T he re sult o f A IC a nd lo g like lihoo d tests
Mod el AIC value SE Log likelihood
process. First we start by identifying the ARFIM A model
ARFIM A (2, 0.29, 1) 727 1.09 -359.59
that best fits the series. ARFIM A (2, 0.29, 2) 727 1.09 -358.90
ARFIM A (3, 0.29, 3) 723 1.07 -354.62
The AR FIM A m odel identification: The particular ARFIM A (1, 0.29, 1) 725 1.09 -359.88
optimization routine of the log-likelihood function shows
Tab le 9: R esu lts of d iagn ostic te sts
that among the ARFIMA models, the optimal model for No rma lity ARCH-LM Por tm anteau
the fractionally difference series is ARFIM A (1, 0.29, 1) Mod el (p-value) (p-value) (p-value)
since this model exhibit the smallest values of AIC and ARFIM A(2, 0.29, 1) 0.7621 0.2010 0.3523
log likelihood (Ta ble 8). ARFIM A(2, 0.29, 2) 0.7446 0.2044 0.4221
ARFIM A(3, 0.29, 3) 0.6645 0.2068 0.3693
The results in the Ta ble 8 show that ARFIM A (1, ARFIM A(1, 0.29, 1) 0.7850 0.2341 0.4252
0.29,1) model is the best candidate model with the least
value of AIC and the log likelihood. Before we draw any Tab le 10: R esults of th e estimated AR FIM A (1 , 0.29, 1) p arame ters
conclusion, it is nece ssary to examine the resid uals of all Estimates Coefficient SE T-ra tio Approx. prob.
the models to see if they pass th e diagnostic tests. Table AR1 -0.455 0.39 -1.15 0.25
MA1 -0.546 0.37 -1.54 0.12
8 shows the results of the diagnostic tests. Constant 0.055 0.08 0.72 0.47
The diagn ostic test results in Tab le 9, sho ws that all
the mod els have passed the diagnostic tests of normality, Tab le 11 : M isspe cifica tion te sts
serial correlation and residual autocorrelation, since the Type o f test Test stat p-value De cision rule
Serial C orr LM -type 0.8661 0.2341 Reject
p-values are greater than 0.05 . W e therefore pro ceed to
Resid Co rr Ljung-Box 7.7556 0.4252` Reject
estima te the parameters of the ARFIMA (1, 0.29,1) model (LB)-type
selected. ARCH ARCH-LM 8.0547 0.5286` Reject
No rma lity Skew ness 0.1555 0.7850 Accept
No rma lity Ku rtosis 3.0440 0.7850 Accept
Estimating the parameters of ARFIM A (1,0.29,1)
No rma lity JB test 0.9872 0.7850 Accept
mo del: The results of the estimated param eters of A R M A
(1,0.29,1) model are shown in Table 10. Having fitted a Table 12: Results of forecast evaluation of ARFIM A (1,0.29,1) model
model to the fractionally differenced humidity time series, In Samp le forecast Out of sam ple forecast
we check the model for adequacy. ------------------------------------ ------------------------------------------
Hu mid ity Prediction Hu mid ity Prediction
22 21.7020 20 19.83
ARF IMA (1,0.29,1) model checking: The residual plots 18 17.4852 27 26.99
of the ARFIM A(1,0.29,1 ) mod el is first perfo rmed to 23 23.4589 19 18.88
examine the adequacy of the model. The residual ACF 58 57.5854 45 45.02
70 70.2214 59 58.91
and PACF of the fitted ARFIM A(1,0.29,1) model of the
72 71.9539 72 72.04
fractionally differenced h umid ity time series are given in 80 79.8470 76 75.92
Fig. 13. The plots show that there is no serial correlation 80 80.2920 80 80.05
observed in the residuals of the series therefore the model 66 65.6393 76 75.93
59 59.3617 40 40.06
is adeq uate and good. 26 25.7297 22 21.93
The results of the diagnostic tests are show n in 16 16.1518 24 24.06
Table 11. The existence of serial correlations in the 14 14.0085 17 16.94
residuals is rejected based on the resu lts of the Lagrange 13 12.8771 15 15.06
20 20.2443 20 19.94
multiplier tests. The Ljung-Box test also rejects the 20 20.2621 42 42.06
presence of residuals correlation. The results of the 66 66.2726 61 60.94
Jarque-B era test, the skewness and kurtosis which are 76 75.8219 66 66.06
tests for norm ality of the residuals, show that the residuals 72 72.1016 74 73.94
82 82.0332 82 81.95
are norm ally distributed as the skew ness is close to zero 62 61.8988 66 65.95
and kurtosis close to 3. 30 30.1976 50 50.05
20 19.8100 24 23.95
Forecast evaluation of ARFIM A(1,0.29,1) mo del: After 18 18.2024 24 24.05

checking the mod el for adequacy, w e finally study its

Table 13: Results of forecast for ARFIM A(1, 0.29, 1) model
forecast values. Both the in-sample and out-of-samp le Mo de of forecast RMSE MAE MAPE
foreca sts are computed and shown in Table 12. The In-sa mp le 0.280 0.230 0.008
observed and forecasted values are ve ry close which that Ou t-of-sa mp le 0.069 0.063 0.002
ARFIM A(1,0.29,1) model is adequate, valid and good.
Table 14: Results of the benchmark evaluation for the two types of
W e next compare the in-sample forecast with the out- mo dels
of-sam ple forecast of the A RFIMA(1,0.29,1 ) to evaluate Mo del forecast RMSE MAE MAPE
their performance. RMSE, MAE and MA PE are used for SARIM A(1,0,1)(2,1,2) 6.56 80.36 0.1621
this purpose (T able 13). ARFIM A(1,0.29,1) 0.28 00.226 0.0080

30
 Res. J. Math. Stat., 2(1): 23-31, 2010

Dickey, D.A. and W.A. Fuller, 1979. Distributions of the

estimators for autoregressive time series w ith a unit
root. J. Am. Stat. Assoc., 74: 427-431.
Gil-Alana, L.A., 2008. Fractional integration and
structural breaks at unknown periods of time. J. Time
Series Anal., 29(1): 163-185.
Grang er, C.W.J. and R. Joyeux, 1980. An introdu ction to
long memory time series models and fractional
Fig. 13: Residuals ACF and PACF of the ARFIMA (1,0.29,1) differencing. J. Time Series Anal., 1: 15-29.
Model Geweke, J. and S. Porter-Hudak, 1983. The estimation
and application of long memory time series models.
W e notice that for the ARFIMA (1, 0.29, 1) model the J. Time Series Anal., 4(4): 221-238.
out-of-sample forecast is better than the in-sample Harvey, A.C ., 1990 . The Econometric Analysis of Time
counterpa rt since it has smaller values for the RMSE, Series. 2nd Edn., Philip Allan, Hemel Hempstead.
MAE an d M APE. This shows that ARFIMA (1, 0.29, 1) Harvey, A.C., 1993. Time Series Mo dels. 2nd Edn., M IT
mode l has good prediction pow er. Press.
W e next com pare the SA RIM A(1,0,1)(2,1,2) short Haslett, J. and A.E. Raftery, 1989. Space-time modelling
mem ory mode l and the A RFIM A(1,0.29 ,1) long memory with long memory dependence: Assessing Ireland
mod el. The three benchmarks used are the RMSE, MAE wind power resources. Appl. Stat., 38(1): 1-50.
and MA PE. The results of these tests are given in Hosking, J., 1981. Fractional differencing. Biometrica,
Table 14. The results show that AR FIM A(1,0.29,1) has 68(1): 165-176.
smaller values of three benchmarks hence it is a better Hsu, N.J., B.K. Ray and F.J. Breidt, 1998. Bayesian
Estimation of Common Long Range D ependent
mode l for humidity data than SA RIM A(1,0,1)(2,1,2).
Mo dels. Prob. Theory M ath. Stat. B. Grigelionis et
al. (Eds.). VSP/TEV.
CONCLUSION
Hunt, K. an d G.P. Nason , 2001 . W ind speed m odelling
and short-term prediction using wavelets. W ind Eng.,
A short memory model SARIM A(1,0,1)(2,1,2) and a 25: 55-61.
long mem ory model AR FIM A(1,0.29 ,1) are used to fit the Hurst, H.E., 1951. Long-term storage capacity of
same humidity data. Even though both models fit the data reservoirs. Trans. Am . Soc. C ivil Eng., 116: 770-
well, forecasts obtained using the long memory model 779.
tend to captu re the swing in the data and resemble the Hurvich, C.M., R.S. Deo and J. Brodsky, 1998. The mean
actual values better than the forecasts using the short squared error of geweke and Porter-Hudak’s
memory model. The conclusion we draw from these estimator of the memory parameter of a long-
results is that the Jos humidity data is better fitted by a mem ory time series. J. Time Ser. A nal., 19: 19-46.
long memory time series which captures the long swing Jarque, C .M . and A.K. Bera, 1987. A test of normality of
in the weather data better than th e short mem ory time observations and regression residuals. Int. Stat. Rev.,
series models whose effect quickly dies down. 55: 163-172.
Kw iatkow si, D., P.C.B. Phillips, P. Schmidt and Y. Shin,
REFERENCES 1992. Testing the nu ll hypo thesis of stationarity
against the alternative of a unit root. J. Econometrics,
54: 159-178.
Bhardwaj, G. and N.R. Swanson, 2004. An empirical
Ljung, G.M. and G.E.P. Box, 1979. On measure of lack
investigation of the usefulness of arfima models for
of fit in time series models. Biometrika, 65: 297-303.
predicting macroeconom ic and financial time series.
Montan ari, A ., R . R osso and M . Taqqu, 1996. Some long-
J. Econometrics, 131(2): 539-578. run properties of rainfall records in Italy. J. Geophys.
Bouette, J.C ., J.F. Ch assag neux, D. Sibai, R. Terron and Res. Atmosphere, 101(23): 431-438.
A. Charpen tier, 2004. W ind in Ireland: Long mem ory Said, S.E. and D. Dickey, 1984. Testing for unit roots in
or seasonal effect? Stoch Environ. Res. Risk Assess., autoregressive mov ing average models with
1: 12. unknown order. Biometrika, 71: 599-607.
Brown, R.L., J. Durbin and J.M. Evans, 1975. Techniques W hitcher, B., 1998. Assessing nonstationary time series
for testing the constancy of regression relationships using wavelets. Unpublished Ph.D. T hesis ,
over time. J. Royal Stat. Soc., Series B37: 149-192. University of Washington.

31