Environmental Science and Pollution Research (2023) 30:19617–19641

https://doi.org/10.1007/s11356-023-25148-9

REVIEW ARTICLE

Autoregressive models in environmental forecasting time series:

a theoretical and application review
Jatinder Kaur1,2 · Kulwinder Singh Parmar2 · Sarbjit Singh3

Received: 13 May 2022 / Accepted: 2 January 2023 / Published online: 17 January 2023
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023

Abstract
Though globalization, industrialization, and urbanization have escalated the economic growth of nations, these activities
have played foul on the environment. Better understanding of ill effects of these activities on environment and human health
and taking appropriate control measures in advance are the need of the hour. Time series analysis can be a great tool in this
direction. ARIMA model is the most popular accepted time series model. It has numerous applications in various domains
due its high mathematical precision, flexible nature, and greater reliable results. ARIMA and environment are highly cor-
related. Though there are many research papers on application of ARIMA in various fields including environment, there is
no substantial work that reviews the building stages of ARIMA. In this regard, the present work attempts to present three
different stages through which ARIMA was evolved. More than 100 papers are reviewed in this study to discuss the applica-
tion part based on pure ARIMA and its hybrid modeling with special focus in the field of environment/health/air quality.
Forecasting in this field can be a great contributor to governments and public at large in taking all the required precautionary
steps in advance. After such a massive review of ARIMA and hybrid modeling involving ARIMA in the fields including or
excluding environment/health/atmosphere, it can be concluded that the combined models are more robust and have higher
ability to capture all the patterns of the series uniformly. Thus, combining several models or using hybrid model has emerged
as a routinized custom.

Keywords Time series analysis · Air Quality Index (AQI) · Autoregressive models · Statistical modeling · Forecasting

Introduction could be analyzed through time series. Time series forecast-

ing is an extensive quantitative technique involving collec-
Though globalization, industrialization, and urbaniza- tion and analyzation of historical observations for the devel-
tion have escalated the economic growth of nations, these opment of an appropriate model. Theoretically, analysis of
activities have played foul on the environment. Better under- time series basically contains three steps: characterization,
standing of ill effects of these activities on environment and modeling, and forecasting. While forecasting calculates
human health and taking appropriate control measures in short-term progression of the system, the modeling compo-
advance are the need of the hour. For this the relevant data nent establishes long-term behavioral features of the system.
The first step determines the fundamental properties such as
measure of randomness or degrees of freedom. This analysis
Communicated by Marcus Schulz can be employed to build predictive models with minimum
errors for forewarning. It is important to underline that the
* Kulwinder Singh Parmar
[email protected] modeling choice in any temporal, or more generally, in any
spatio-temporal prediction, is relevant and must be suitably
1
Department of Mathematics, Guru Nanak Dev University faced (De Iaco and Posa 2018; Cappello et al. 2018; De Iaco
College Verka, Amritsar, Punjab, India 143501 et al. 2013 2015).
2
Department of Mathematics, I.K. Gujral Punjab Technical Autoregressive integrated moving average (ARIMA)
University, Jalandhar, Punjab, India 144603 model is the most popular accepted time series models
3
Department of Mathematics, Guru Nanak Dev University (Shahwan and Odening 2007; Singh et al. 2020b). The
College, Narot Jaimal Singh, Pathankot Amritsar, Punjab, Box-Jenkins methodology (Box and Jenkins 1970), high
India 145026

13
Vol.:(0123456789)
19618 Environmental Science and Pollution Research (2023) 30:19617–19641

mathematical precision, and reliability are what makes moving average models. Hipel et al. (1977) have given the
ARIMA models very popular (Singh et al. 2021b). ARIMA theoretical and practical approaches for the model building
models have large number of applications. The model is stages of ARIMA.
applied to forecast various things like commodity prices ARIMA processes are a kind of stochastic techniques
(Weiss 2000); for load forecasting in the power system which are used to investigate behavioral pattern of time
(Nicolaisen et al. 2000; Hippert et al. 2001); future of energy series. ARIMA models are quite flexible in nature as they
resources, such as oil (Morana 2001) or natural gas (Bucha- can represent pure autoregressive (AR), pure moving aver-
nanan et al. 2001); daily environmental factors such as ozone age (MA), and mixed AR and MA (ARMA) series. Unfor-
levels ( Robeson and Steyn 1990; Prybutok et al. 2000); fore- tunately, ARIMA models are unable to capture nonlinear
casting various air pollutant (Kulkarni et al. 2018; Chaud- pattern of the series and thus are not suitable for approxi-
huri and Dutta 2014), noise pollution data (Garg et al. 2015), mating complex real-world problems. Time series which has
water quality time series data (Faruk 2010), for water quality either a trend or seasonal patterns do not exhibit stationary
management (Parmar and Bhardwaj 2014); etc. Thus, we behaviors. ARIMA (p, d, q) model only captures trends and
find large applications of ARIMA almost in every field. not seasonal behavior of time series. To model a seasonal
ARIMA and environment are highly correlated. Fore- pattern, we have ARIMA (p,d,q)(P,D,Q) model.
casting is needed in all the fields of environment such as
air pollution, noise pollution, fossil fuels, rainfall data, and
underground waters, as all these factors have direct relation- Mathematical formulation
ship with health. Forecasting in these fields can result in
forewarning which can be highly beneficial. Though there Moving average (MA) process
are many research papers on application of ARIMA in vari-
ous fields including environment, there is no substantial A process { zt } is said to be a moving average process of
work that reviews the building stages of ARIMA. In this order q if.
regard, the present work attempts to present three differ- zt = at − 𝜃t−1 −…. − 𝜃q at−q where 𝜃i; i = 1,2,3…,q are con-
ent stages through which ARIMA was evolved. It needs to stants and { at } is a purely random process with mean zero
be highlighted that detailed information about the process and variance 𝜎 2(Box et al. 1994).
of evolution is equally significant as knowledge about its
application. In this review paper, we have tried to review
more than 100 papers based on pure ARIMA and its hybrid Autoregressive (AR) process
modeling. The first part of this paper deals with general
introduction to ARIMA, followed by its historical overview, A process { zt } is regarded as an autoregressive process of
and the last part deals with its application review. The last order p if.
section is divided into two categories, where major emphasis zt = 𝜙1 zt−1 + 𝜙2 zt−2 + ⋯ + 𝜙p zt−p + at w h e r e 𝜙j
is laid on application of ARIMA in the field of atmosphere/ (j = 1,…,p) are constants and { at } is a purely random pro-
environment/health or factors influencing air quality such as cess with mean zero and variance 𝜎 2 (Box et al. 1994). This
air pollutants, noise pollution, and rainfall, and the last sec- model works like multiple regression model.
tion deals with application of ARIMA in other fields such as
financial data and load forecasting data. The last categories
are again divided into two further subsections: one involving Autoregressive moving average (ARMA) process
purely ARIMA models and the other development of hybrid
models with ARIMA. Both the autoregressive (AR) and the moving average (MA)
are combined to build the autoregressive moving average
(ARMA) model. The (ARMA)(p, q) model for a time series
ARIMA model analysis which contains p (AR) terms and q (MA) terms can be
expressed as.
The classes of autoregressive moving average (ARMA) ∑p ∑q
models are frequently used while modeling linear and xt = i=1 𝜙i xt−i − j=1 𝜃j at−j + at ;t = 1, … , T
stationary time series due to their outstanding results and
Here, at is known as normal white noise process. It has
effectiveness (Al-Saba and El-Amin 1999). In 1921, Yule
zero mean and variance 𝜎 2 . T is the amount of data in the
presented pure moving average process, whereas he intro-
time series (Box et al. 1994). AR parameters should satisfy
duced pure autoregressive process in 1927. Box and Jenkins
the condition for stationarity and MA parameters should sat-
(1970), Hannan (1970), and Anderson (1971) have been pio-
isfy the conditions for invertibility.
neers in building various techniques using autoregressive

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19619

Seasonal ARIMA expressly modeled by ARIMA extension known as sea-

sonal ARIMA, i.e., SARIMA. Seasonal ARIMA models are
ARIMA as such does not support seasonal data, i.e., time defined by 7 parameters p, d, q, P, Q, D, ands and mathemati-
series with repeating cycles. However, such time series are cally defined as

( )( )
1 − ϕ1 B − ϕ2 B2 − ∙ ∙ ∙ ∙ −ϕp Bp 1 − β1 Bs − β2 B2s − ∙ ∙ ∙ − βP BPs (1 − B)d (1 − Bs )D yt = c
( )( )
+ 1 − Ψ1 B − Ψ2 B2 − ∙ ∙ ∙ ∙ −Ψq Bq 1 − θ1 Bs − θ2 B2s − ∙ ∙ ∙ − θQ BQs 𝜖t

• p and P are non-seasonal and seasonal autoregressive Sometimes by using visualization tools such as ACF and
polynomial orders, respectively. PACF, it is not possible to identify the parameters p, d, q
• q and Q are non-seasonal and seasonal moving average and P, D, Q. In that case, the model with the lowest BIC
polynomial orders, respectively. (Bayesian information criterion) or AIC (Akaike informa-
• d and D are order of normal and seasonal differencing, tion criterion) is selected. Based on different information
respectively. theoretic techniques such as AIC or minimum description
• s is the period of the seasonal pattern appearing. length (MDL), various methods have been proposed by
researchers Hurvich and Tsai (1989), Ljung (1987), and
The ARMA models work only on stationary data, but in Shibata (1976) for order selection.
reality, data of the various fields is non-stationary, making Astrom and Eykhoff (1971), Van Boom and Enden (1974),
ARMA unfit for these problems. With the help of differenc- and Unbehauen and Göhring (1974) acquainted on time series
ing, the data can be made stationary, and this step leads to model identification in the engineering field. Since the begin-
the development of ARMA. ARMA is in fact generalization ning of the 1970s, various estimation-type identification meth-
of ARMA processes. ods have come into limelight. Akaike (1969) introduced final

Historical overview

Based on the principle of parsimony, statisticians Box and

Jenkins (1970) gave a practical approach to build ARMA
model. The method uses a three-step iterative approach of
model identification, parameter estimation, and diagnostic
checking to build the best parsimonious model from a gen-
eral class of ARMA models (Fig. 1). The process is repeated
until a satisfactory model is obtained which can be then used
for prediction (Singh et al. 2020a, b, 2021a, b). In this sec-
tion, historical evolution of various steps involved in ARMA
modeling is discussed in detail.

Model identification

Model identification is a herculean task in any ARMA mod-

eling. In this regard, autocorrelation function (ACF) and
partial autocorrelation function (PACF) are vital statistics
in determining the order of the model. While ACF explains
correlation, PACF describes partial correlation between the
series and lags of itself. Box and Jenkins (1970) introduced
the concept of degree of differencing “d” also. Generally, d
is 0, 1, or 2. After d is selected, p and q are calculated from
the overall trend of ACF and PACF of the appropriately dif-
ferenced series. Cleveland (1972) presented inverse autocor-
relation function (IACF) for this step. But the method was not
suitable for mixed models. Fig. 1  Methodology of ARIMA model

13
19620 Environmental Science and Pollution Research (2023) 30:19617–19641

prediction error (FPE) to determine the order p of the AR (1988). Koehler and Murphree (1988) preferred Schwarz
models which is defined as information criterion (SIC) over Akaike information cri-
terion (AIC) for order selection. Hurvich and Tsai (1989)
2p 2
FPE(P) = (1 + )𝜎 (1) developed a bias-corrected method AlCc for better model
n p order choices. Koreisha and Pukkila (1993) introduced
where n is the number of observations and 𝜎 ̂2(P) is an esti- iterative procedures for determining the degree of differ-
mate of white noise variance. encing required to make time series data stationary. Zhang
For the mixed model, the criteria are expressed as and Zhang (1993) developed an algorithm involving only
autocorrelations for determining order of MA processes.
( 2 ) (p + q)g(n) Liang et al. (1993) gave a new approach based on the Eigen
𝛿(p, q) = 𝑙𝑜𝑔 𝜎
̂ + (2)
n values of covariance matrix for ARMA model order determi-
Here, 𝜎̂2 represents the maximum likelihood estimate for nation. Sreenivasan and Sumathi (1997) formulated a new
the residual variance 𝜎 2 . The values ̂p and ̂q minimizing δ generalized parameters technique for both seasonal and
(p, q) are the best approximations for p and q. Taking differ- non-seasonal ARMA model identification.
ent values of g(n) leads to different criteria (Table 1).
Another procedure known as criterion of autoregressive Model estimation
transfer function (CAT) was introduced by Parzen (1974)
where the actual model is presumed to have an AR ( ∝ ) The estimation of AR parameters is very crucial in time
representation. The order selected ̂ p is interpreted as best series analysis for the adequate information about the model.
finite order AR approximation to the true AR (∝) process. Maximum likelihood methods, ordinary least squares (OLS),
Extending the work of Woodward and Gray (1981), Glasbey and method of moments are some of the extensively used
(1982) introduced generalization of partial autocorrelations techniques for parameter estimation in time series analysis.
(GPAC) as crucial techniques in ARMA model identifica- Pure autoregressive models are either estimated by OLS
tion. To estimate d, Janacek (1982) proposed a method method or by using Levinson (1947)-Durbin (1960) algo-
using log of the power spectrum. For order estimation rithm. Durbin (1960) showed that method of OLS leads to
of a finite moving average process, Bhansali (1983) gave optimum estimates of the model provided errors are nor-
autoregressive and window estimates of the inverse correla- mally distributed. Burg (1975) came with an algorithm
tion function. Monahan (1983) used Bayesian approach for similar to forward backward prediction especially for short
determining the order (p, q) of the ARMA model. Pattern records and Huzii (1981) proposed a method based on higher
identification techniques using extended Yule-Walker equa- order moment for estimating AR coefficients. Proposing an
tions for ARMA order identification was employed during alternative to the Burg’s estimates Pukkila and Krishnaiah
early 1980s. Tucker (1982) replaced R and S arrays with (1988) calculated true correlation matrix of the lagged vari-
RS array for the ARMA model identification. By integrating ables. Basu and Das (1992) consider optimality of the maxi-
order identification approach for mixed stationary and non- mum likelihood estimator under a general set-up of roots of
stationary ARMA models, Tsay and Tiao (1984) eliminated the characteristic equation of the p th order of autoregressive
the need of differencing required for producing stationar- process.
ity in the time series. Augmenting their study, Tsay and In comparison to the parameter estimation methods avail-
Tiao (1985) introduced the smallest canonical correlation able for AR model, the number of techniques to approxi-
(SCAN) method. Broersen (1985) presented weak param- mate MA parameters is fewer. Durbin (1959) is pioneer in
eter criterion (WPC) for model order selection. Poskitt the estimation of MA parameters with a simple estimation
(1987) modified Hannan and Rissanen (1982) criterion procedure. A quadratically convergent algorithm was for-
of order selection. The order of ARMA model was deter- mulated by Wilson (1969) to estimate the parameters of
mined using white noise test by Pukkila and Krishnaiah a MA process. Whereas Gauss–Newton method was used

Table 1  Model identification criterions depending upon values of g(n) along with their mathematical expressions
Value of g(n) Criterion Mathematical expression
( )
g(n) = 2 Akaike information criterion (AIC) given by Akaike (1974) AIC(p, q) = log 𝜎 2 + 2(p+q)
( 2 ) (p+q)logn
n
g(n) = logn Bayesian information criterion (BIC) given by Schwarz (1978) BIC(p, q) = log 𝜎 +
( ) n
g(n) = clogn Hannan and Quinn’s criterion (HQ) given by Hannan and Quinn 1979 HQ(p, q) = log 𝜎 2 + (p+q)loglogn
n
where c is a constant to be
specified

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19621

to estimate the parameters of a non-linear function, Fuller Model forecasting

(1976) utilized it for the MA process. Godolphin (1977,
1978) proposed an alluring computational procedure aug- Among the wide applications of time series, the most popu-
menting the previous studies. lar is forecasting. Forecasting is easily attainable with state
For ARMA cases, myriad studies are available. Wilson space framework of which ARIMA models are special case.
(1969) and Marquardt (1963) have developed an algorithm With the state space framework, observational vectors are
to estimate the ARMA parameters which was employed by brought into a system with one element at a time (Durbin
Box and Jenkins (1970) to lay the foundation of all the and Koopman 2012). State space models have great contri-
ARMA, AR and MA processes. The research done by the bution in the science of environment (Harvey et al. 2004)
latter is regarded to be a milestone in the field of mod- and also play a pivotal in the analytical handling of time
eling. McLeod (1977) gave an easier implementing modi- series models (Harvey 1989).
fied version of Box and Jenkins (1970) which approxi- For the practical computation of forecasting, the simplest
mated exact ML estimators very precisely. Tuan (1984) and most elegant method is of difference equation from
derived many recursive relations which worked as a tool which minimum mean square error forecasts can be directly
in identifying the order and parameters of ARMA at its generated (Box and Jenkins 1970). Also, the probability
preliminary stages. Using autocovariance function, Choi limits for the forecasts can be obtained by solving recur-
(1986) presented an algorithm for the parameter estima- sively. Makridakis and Wheelwright (1977) concluded that
tion of stationary ARMA process. The iterative algorithm the adaptive filtering technique can be applied to time series
proposed by Choi in 1986 was not convergent; therefore, forecasting dealing with real data. Cartwright (1985) has
he developed a convergent Newton–Raphson solution for assessed the forecasting performance of Priestley’s model
MA parameters in the subsequent year. Saikkonen (1986) and concluded that forecasting errors can be significantly
derived two-step estimators which were asymptotically reduced with the use of broader classes of time series. Ray
efficient. (1988) concluded that forecasting performance of Bilinear
Recently, different studies have focused on maximum model is more than Box-Jenkins model and threshold autore-
likelihood (ML) procedures as a tool to estimate ARMA gression model. Tiao and Tsay (1994) studied developments
models. Basu and Das (1991) analyzed the asymptotic of time series in both the linear and non-linear domain. They
properties of least-squares estimation procedures of the were of the opinion that when the parameters involved are
parameters of an ARMA (p, q) process in the stable case. estimated adaptively, linear models provide more accurate
The SCA System can be used to estimate the parameters forecasts. Thus, we can see the model building process is
of the model. For best results, a conditional likelihood quite laborious and needs great human expertise.
function is selected along with the detection and adjust- Wrapping up all the above steps, the general statistical
ment of outliers. Mikosch et al. (1995) derived the asymp- methodology scheme of ARIMA is as presented by Fig. 1.
totic properties of the estimators. The major drawback of
ML estimates is that it often lies outside the invertibility
region. However, the generalized least-squares method for Modeling and forecasting work flow
the estimation of pure MA and ARMA model presented of ARIMA
by Pukkila et al. (1990) has succeeded in overcoming it.
As observed ARIMA finds large applications in various
fields due its high mathematical precision and great reli-
Model diagnostic checking ability. The next section of the paper deals with review of
papers employing ARIMA exclusively and hybrid ARIMA
Before the remarkable advent of Box and Jenkins (1970) models published in world’s best-class journals. The ris-
procedure, hypothesis testing methods were largely used ing disastrous threats to the atmosphere or air quality are
for ARMA model identification. McLeod (1974) examined turning to be great concerns of the twenty-first century. Its
the need to check whiteness and homoscedasticity of the hazardous effects are felt by human beings as well as by
residuals in the diagnostic checking step. Godfrey (1979) the entire wildlife. Thus, forecasting in this field can be a
proposed a new approach based upon Silvey’s (1959) great contributor to governments and public at large in tak-
Lagrange multiplier method to check the adequacy of the ing all the required precautionary steps in advance. There
ARMA model. Hokstad (1983) has proposed a diagnostic are various environmental factors affecting the air quality
test with the estimated cross correlation function (CCF) such as air pollutants, water pollutants, noise pollution due
between the observed values and the residuals. The CCF to traffic congestions, rainfalls, and surface erosions. Also,
can also be used as an indicator for the required improve- the environment exhibits close connections with energy
ment in the model. resources. Thus, our coming section deals with all these

13
19622 Environmental Science and Pollution Research (2023) 30:19617–19641

environmental factors first forecasted with ARIMA alone December were developed separately. The result revealed
followed by hybrid models involving ARIMA. The data set that past continued to impact future values of VC.
undertaken, methodology adopted, and results obtained, all Kumar and Goyal (2011) thrived to build a forecasting
are discussed elaborately for the complete understanding of AQI model. Three models are assessed, i.e., ARIMA (1,1,1),
the readers. PCR, and ARIMA combined with PCR with respect to 4
seasons of the year in Delhi. The hybrid model performed
Survey of air quality/environment/health data better in predicting AQI one-day advance. The hourly and
involving application of ARIMA models monthly concentrations of CO spread over 7 years of data of
Hong Kong were analyzed by Lau et al. (2009) using ARIMA
Kulkarni et al. (2018) studied the variation of ­SO2, RSPM, modeling. Association of hourly concentration of CO with
NOx, and SPM parameters present in the atmosphere of different days of the week was examined. The hourly data of
Nanded City (Maharashtra, India). The randomness, trend, CO was like traffic data. This strong association was proven
and seasonality present in the data were also analyzed. The by SARIMA (0, 1, 1) (0, 1, 1)24 model. It was also shown
forecasted result revealed that RSPM and SPM are exceed- that data possessed long-term memory features. Kumar and
ing the permitted limits. Jaiswal et al. (2018) observed the De Ridder (2010) took the daily maximum ­O3 concentra-
statistical trend of CO, N ­ O2, ­SO2, ­PM2.5, and P­ M10 concen- tion data of four sites of London and Brussels for study.
trations for the time span January 2013 to December 2016 They studied GARCH modeling technique in association
for the city Varanasi (India) using Mann–Kendall and Sen’s with FFT-ARIMA to make forecasts of ozone episodes at
slope estimator approach. Different ARIMA models, namely, these sites. In the study of Slini et al. (2002), ARIMA (1,1,1)
ARIMA (1,0,0), (1,0,1), and (1,1,1), were fitted on the three and ARIMA (1,1,0) were used over the data of maximum
data sets of summer, monsoon, and winter season and their daily ­O3 concentration data of Athens (Greece) from 1990
results compared. ARIMA (1,1,1) was chosen as the best fit to 1998. The forecasting performance of these two mod-
model for forecasting all the pollutants. Pohoata and Lungu els were observed under the three categories of alarm limit
(2017) tried to analyze the air quality of the city Ploiesti in greater than 180,170 and 160 μg/m3. Robeson and Steyn
Romania for the pollutants ­O3, CO, ­NO2, NOx, and ­PM10. (1990) tried to develop three predictive models, i.e., D/S
While ARIMA (3,1,3) provided good results for NOx, N ­ O2 model, ARIMA (1,1,0), and TEMPER model for the tempo-
and ­O3, it failed to give satisfactory results for CO and P ­ M10. ral variability of ozone in the lower Fraser Valley of British
Kumar et al. (2004) forecasted one-day advance O ­ 3 con- Columbia by taking 8 years data for the period 1978–1985
centration in Brunei Darussalam using ARIMA modeling from two monitoring stations T9 and T11. On comparison
approach. ARIMA (1,0,1) was the best fitted method. Liu of forecasting ability for ozone concentrations, TEMPER
et al. (2018) used ARIMA with numerical forecasts (ARI- model outperformed the other two models. Siew et al. (2008)
MAX) with the vision to improve forecast of ­O3, ­PM2.5, and compared the performance of ARMA (3, 1, 3) and integrated
­NO2 for Xingtai (China). Significant reduction in RMSE, ARFIMA (0, − 0.5, 2) models for the data of O ­ 3, ­PM10, ­NO2,
viz., 47.8–49.7%,14.3–21.0%, and 41.2–46.3%, respectively, ­SO2, and CO concentration from March 1998 to Decem-
for ­O3, ­PM2.5, and N ­ O2, was seen by employing CMAQ- ber 2003 of Selangor Malaysia. Though both models could
ARIMA for the daily 1-h and 8-h forecasting values at all the not forecast all the values completely, ARFIMA performed
stations. Dynamic hourly forecast shows that ARIMAX can slightly better than ARIMA.
also be successfully applied to forecast of 7- to 72-h ­PM2.5, Ahn (2000) applied second order differencing ARIMA
4- to 72-h ­NO2, and 4- to 6-h O ­ 3. Zhang et al. (2018) fore- models to daily groundwater head time series (1985–1990)
casted ­PM2.5 concentrations using AQI and meteorological from 7 monitoring wells located in Collier County, Florida.
parametric data for Fuzhou (China) with the help of ARIMA. Variance and autocovariance equations were derived for
It was observed that ­PM2.5 concentrations were positively the second-order time series models using ARIMA (0,2,1),
correlated with P ­ M10, ­NO2, and S­ O2 concentrations and neg- (0,2,2), (1,2,0), (2,2,0), and (1,2,1) as function of parameters
atively correlated with meteorological parameters. Average of the model under study. Mirzavand et al. (2014) worked
­PM2.5 concentrations were 52% higher in cold periods in on AR, MA, ARMA, ARIMA, and SARIMA models in
comparison to warm periods. ARIMA (6,1,1) was found to analogous to forecast groundwater levels up to 60 mo in
be best model for the data. plain expanses of Kashan aquifer in Iran. The seasonality
This is the first applicability of ARIMA model on the data- and stationarity of the data were checked. Taheri Tizro et al.
set of VC. Prior to model simulation of VC over the region (2014) analyzed several water parameters of Hor Rood River
of Delhi, trends and variations in the data set were analyzed. at Kakareza station with the help of ARIMA modeling. Based
12 ARIMA models ARIMA (0,0,1), (1,0,2), (0,0,5), (1,0,0), on ­R2, AIC, RMSE, and MAPE, ARIMA (2,1,3), (2,1,3),
(0,0,1), (1,0,0), (0,0,1), (0,0,1), (1,0,0), (0,0,2), (0,0,1), (1,1,3), (1,1,3), (2,1,1), (2,1,1), (1,1,1), and (2,1,2) were
and (0,0,3), respectively, for each month from January to found to be best suitable for parameters TDS, EC, H ­ CO3-,

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19623

­SO42−, ­Ca+, ­Na+, ph, and SAR generation, respectively. The for the span 1995–2015. Employing histogram judg-
increasing trend of majority of parameters showed a picture ment method, it was shown that data contains non-linear
of deteriorating water quality conditions in the region. sequencing. This led to the development of GM (1,1)
Garg et al. (2015) in their paper simulated daily mean model on the data. Unfortunately, this method was also
LDay (06–22 h) and LNight (22–06 h) in A- and C-weight- precluded because of 21 data points in the data set. The
ings in conjunction with single-noise metrics, day-night optimal requirement of data points for application of gray
average sound level (DNL) for a period of 6 months for the model is 5–10. In the next step, gray metabolic forecast
station East Arjun Nagar in Delhi using ARIMA methodol- model was tried. After the successful application of gray
ogy. ARIMA (0,0,14), (0,1,1), (7,0,0), (1,0,0), and (0,1,14) metabolic model on the basis of relative error, a series of
are chosen to be fit models for LDay dBA, LNight dBA, residuals was obtained. Residual corrections are done with
LDay DB(C), LNight dBC, and DNL dBA, respectively. the help of ARIMA model. This process enhanced the pre-
Augmenting their work, the authors in 2016 compared diction accuracy significantly, and this was the major inno-
ARIMA and ANN on the same problem and found ANN vation of this study. Aamir and Shabri (2016) employed
model gave better results than ARIMA model. Guarnaccia ARIMA, GARCH, and ARIMA-Kalman to predict crude oil
et al. (2017) presented two methods for the acoustic data set rates in Pakistan by undertaking average monthly prices of
of Nice (France) international airport for the year 2000. The crude oil for the time February 1986 to March 2015. The
two methods utilized were DD-TSA (deterministic decompo- ARIMA Kalman filter technique proved to be best approach
sition model) and SARIMA (0,1,1) (0,1,1)24. While the for- as MAE and RMSE were minimum in this case as com-
mer method captured long-term behavior of the data set, the pared to ARIMA and GARCH.
latter captures short-term behavior. To quantify the forecast- It is very much evident from the above discussion that
ing errors, residual analysis was carried out. The DD-TSA ARIMA is highly significant in the field of environment and
gave slightly better results in terms of low standard devia- health. ARIMA being simple and reliable has tremendous
tion. Williams and Hoel (2003) applied SARIMA (1,0,1) potential to forecast in these areas. Thus, it was important
(0,1,1)672,(1,0,2)(0,1,1)672, (2,0,1)(0,1,1)672 for M25 station to review contribution of ARIMA in these domains.
and (1, 0, 1)(0, 1, 1)672 , (3, 0, 0)(0, 1, 1)672, (1, 0, 2)(0, 1, 1)672
for I-75 stations on the vehicular traffic data. The authors
concluded that one-step seasonal ARIMA predictions con- Hybrid modeling
stantly outperformed ARIMA and random walk forecast
results. The real-world problems are usually complex rather than
Ab Razak et al. (2018) using Mann–Kendall trend analy- being simple. Thus, linear predictive methods do not per-
sis found ARIMA (0, 1, 2) is the best suitable model for the form as desired when used to process data from a non-
daily and monthly rainfall and stream flow data of stations linear system (Patil 1990). To find out whether the series
of Malaysia for the period 2000–2010. Zakaria et al. (2012) is linear or non-linear requires enormous efforts of the
developed four ARIMA models (3,0,2)(2,1,1)30, (1,0,1) forecasters. The researchers try different combination
(1,1,3)30, (1,1,2)(3,0,1)30, and (1,1,1)(0,0,1)30 for the weekly of models based on different theoretical and practical
rainfall data for the stations Sinjar, Mosul, Rabeaa, and Tala- approaches and various other factors such as sampling
far in north west Iraq for the period 1990–2011 stations. variation, model structure and uncertainty to develop a
Benvenuto et al. (2020) found ARIMA (1, 0, 4) and model which yields more accurate results and enhanced
ARIMA (1, 0, 3) to be the best model for determining the forecasting (Jenkins 1982; Makridakis 1989). Bates and
prevalence and incidence of COVID-19, respectively, from Granger (1969) are considered to be pioneer in introduc-
January 20, 2020, to February 10, 2020. Logarithmic trans- ing combining forecasts as an alternative to use one single
formation was done out to check the seasonality influence forecast. The literature indicates that performance of time
on the prediction. series forecasting increases through combining forecasts
Suresh and Priya (2011) took 57 years data from (Makridakis et al. 1982).
(1950–195 l) to (2007–2008) of sugarcane area, produc- The next part of our study deals with review of
tion, and productivity for Tamil Nadu for analyzing. Vari- papers based on ARIMA hybrid modeling in the field of
ous ARIMA models with p and q varying from 0, 1, and environment/atmosphere/health.
2 were fitted. While ARIMA (1, 1, 1) model was found
suitable for sugarcane area and productivity, ARIMA (2,1, Analyzing ARIMA hybrid‑based studies
2) was appropriate for sugarcane production.
Li and Li (2017) applied GM (1,1), ARIMA (1,2,0), Mani and Volety (2021) forecasted three air pollutants, namely,
metabolism GM (1,1), and GM-ARIMA to forecast future CO, ­NH3, and ­O3 of Vijayawada station by employing both
energy needs of Shandong province using energy data ARIMA and the LSTM models. Kalman filters are also used

13
19624 Environmental Science and Pollution Research (2023) 30:19617–19641

to enhance the performance and forecasting abilities. The lat- spatial features of the residual of ARIMA model. And finally
ter model proved to have higher accuracy using RMSE and long-term dependencies of these features were captured by
MAE as performance indices. Wang et al. (2017) came up with LSTM. It was evident from the experimental analysis that the
hybrid-GARCH model to overcome conditional heteroscedas- hybrid ARIMA-CNN-LSTM outperformed the individual mod-
ticity almost present in every hybrid model. The authors used els. The dataset of carbon future prices from 7 April 2008 to
ARIMA and SVM models to explain the linear and non-linear 6 May 2019 was taken. Koutroumanidis et al. (2009) showed
components, respectively, of the AQI data comprising P ­ M2.5, that the hybrid ARIMA-ANN model has a better adaptability
­SO2, ­NO2, CO, and ­O3 concentrations for six stations from the and can make better predictions of future selling prices of the
Shenzhen air quality monitoring network (China) for the time fuelwood produced by Greek state forest farm as compared to
period 01 Sep 2013 to 10 Sep 2013. For estimating the coef- both the ARIMA model and the simple ANN model.
ficients of individual models, GARCH model is introduced. Faruk (2010) observed that the best fit models for water
The accuracy level of hybrid-GARCH model in terms of met- temperature ­( 0 C), boron (­mgl −1), and DO (­mgl −1) are
rices MAE and RMSE were higher than the individual models. SARIMA (1,1,1) (0,0,1)12, (1,1,1) (0,0,1)12, and (1,1,1)
Samia et al. (2012) attempted to foresee one day in advance the (0,0,1)12, respectively. They further found hybrid modeling
max 24-h ma PM concentrations in the region of Sfax South- approach of combining SARIMA with NNBP can give more
ern Suburbs using MLP, ARIMAX, and the hybrid model. The reliable predictions for these parameters of a river than the
study revealed that hybrid ARIMAX-ANN outperformed the Neural Network and traditional baseline ARIMA modeling
individual models since it can explore both linear and non- approach individually.
linear patterns. Zhu et al. (2017) stated that the most chal- Chattopadhyay and Chattopadhyay (2010) and Somvanshi
lenging problem while forecasting AQI is of data being highly et al. (2006) examined the rainfall data of India and found
complex and non-stationary. Thus, they presented two hybrid that hybrid models outperformed individual models in terms
models EMD-SVR-Hybrid and EMD-IMFs-Hybrid for AQI of forecasting efficiencies. Wei et al. (2016) found that the
data of Xingtai, China, collected from June 2014 to August most appropriate model for the morbidity data for hepatitis
2015. To obtain smooth IMF, EMD technique is employed. from the Heng County CDC from January 2005 to Decem-
Then, SVR is used to predict total sum of IMF’s and finally ber 2012 is ARIMA (0,1,2) (1,1,1)12. (Table 2).
S-ARIMA is applied for analyzing residual sequences obtained From Table 3, it can be clearly seen that errors are
from the two proposed models. In this paper where ARIMA is reduced when hybrid models are implemented on the various
used for modeling AQI data, S-ARIMA is kept for forecasting data sets. Thus, both theoretical and empirical findings are in
IMF5 as well as for analysis of the residues. The predicted strong favor of combining different methods to achieve effec-
outcomes of AQI are sum of EMD-IMFs and S-ARIMA. The tive and efficient forecasts (Newbold and Granger 1974).
IMF4, IMF5, IMF6 and IMF7 chosen from IMFs are fore- Consequently, hybridizing different models has become the
casted by Holt-Winters (0.9,0.2,0.3), S-ARIMA (1,1,1) (0,1,1), latest research objectives in the field of modeling and fore-
Holt Winters (0.1,0.2,0.5), and GM (1,1) respectively. Chelani casting. Table 2 provides concise results of sections “Survey
and Devotta (2006) examined the effectivity of their proposed of air quality/environment/health data involving application
hybrid model based on chaos theory with ARIMA and non- of ARIMA models,” “Analyzing ARIMA hybrid-based stud-
linear models for the time series data of NO2 concentration ies,” “Reviewing application of ARIMA alone,” and “Inves-
present in the air from 1999 to 2003 at a site in Delhi. The tigating ARIMA with hybrid methodology” and some of the
prediction performance based on MAPE, RMSE, and RE con- other papers involving ARIMA in tabular form which would
firmed that the hybrid modeling is more effective than indi- be easier to comprehend.
vidual models in forecasting the air pollutant concentrations.
Díaz-Robles et al. (2008) studied hourly and daily time series
of ­PM10 and meteorological data during 2000–2006 at the Las Examining role of ARIMA in studies other
Encinas monitoring station in Temuco. Hybrid model outper- than air quality/environment/health
formed ARIMAX and ANN in terms of RMSE, MAE, and concerns
BIC. Prybutok et al. (2000) studied a neural network model
for forecasting daily maximum ozone levels and showed that ARIMA and its hybrid models have got such wide utilization
the neural network model is superior to the two conventional that they cannot be restricted to merely one or two arenas
statistical models, regression and Box-Jenkins ARIMA models. of study. Therefore, in this section, few research studies in
The maximum ozone data for the year 1994 from Houston was the field of financial data or load forecasting or some other
the research object. fields are briefly reviewed to provide a glimpse of its varied
Ji et al. (2019) forecasted the future carbon prices using the applications. Following the previous chronology, we review
hybrid ARIMA-CNN-LSTM model. The linear features of the papers with ARIMA alone and then followed by review of
data set were modeled by ARIMA. CNN model extracted the papers involving ARIMA and its hybrid methodology.

13
 Table 2  Concise results of application of ARIMA and its hybrid methodology in the field of environment/health/air quality
Author Methodology Data set Findings

Kulkarni et al. (2018) ARIMA SO2, RSPM, NOx, and SPM concentration data set 1. The randomness, trend, and seasonality present
of Nanded City (Maharashtra, India) from 2011 to in the data were analyzed using 95% confidence
2016 with the frequency of 2 days in a week interval
2. The forecasted result revealed that RSPM and SPM
are exceeding the permitted limits and thus requires
immediate attention of the govt. officials
Jaiswal et al. (2018) ARIMA (1,0,0), (1,0,1), (1,1,1) The AQI data in terms of CO, ­NO2, ­SO2, ­PM2.5, and 1. Mann–Kendall and Sen’s slope estimator approach
­PM10 for the time span January 2013 to December were employed
2016 for the city of Varanasi (India) is taken 2. Different ARIMA models, namely, ARIMA (1,0,0),
(1,0,1), (1,1,1) were fitted on the three data sets of
summer, monsoon, and winter seasons and their
results compared
3. ARIMA (1,1,1) was chosen as the best fit model for
forecasting
Pohoata and Lungu (2017) ARIMA Ambient concentrations of SO, ­NO2, NOx, CO, ­O3, 1. Principal component analysis was carried out
­PM10, and ­PM2.5 from 6 monitoring stations of 2. While ARIMA (3,1,3) provided good results for
Ploiesti (Romania) for the year 2013 NOx, ­NO2, and ­O3, it failed to give satisfactory
results for CO and ­PM10. This showed limitations of
Environmental Science and Pollution Research (2023) 30:19617–19641

using ARIMA
Chaudhuri and Dutta (2014) ARIMA The daily data is collected for the parameters ­SO2, 1. To identify the trends in data, nonparametric test
­NO2, CO, ­PM10, CO, ­O3, surface temperature Mann–Kendall (MK) is employed. The test depicted
(°C), and surface relative humidity (%) from 2002 seasonality and monthly variability in the data
to 2012 over Kolkata, India, from WBPCB 2. Box plot for each of the pollutants ­SO2, ­NO2, CO,
­PM10, CO, and ­O3 from 2002 to 2014 is plotted.
ARIMA (1,1,1), ARIMA (0,2,2) and ARIMA (2,1,2)
are employed, and their white noise (WN) variances
are calculated
3. ARIMA (0, 2, 2) is the best fit model for the daily
forecasting of these pollutants with AIC, BIC,
MAPE, FPE, and AICc as evaluation criterions
Guarnaccia et al. (2017) Seasonal ARIMA Hourly CO concentration (ppm) data values for the SARIMA (0,1,1) x (0,1,1)24 and SARIMA (1,0,1)
year 2012 of San Nicolas de Garza x (2,0,1)24 models were applied to the hourly CO
concentrations data for statistical and time series
analysis. The models showed good predictive
results for CO concentrations
Kumar et al. (2004) ARIMA Hourly concentration data of ­O3 at the airport in ARIMA (1,0,1) was the best suitable method. NMSE,
Brunei Darussalam from July 8, 1998, to March Fractional Bias and MAPE were used as deciding
15, 1999 performance criterion. χ2 goodness fit test indi-
cated the model is most appropriately fit

13
19625
 Table 2  (continued)
19626

Author Methodology Data set Findings

13
Liu et al. (2018) ARIMAX, CMAQ-ARIMA, hybrid model O3, ­PM2.5, and ­NO2 concentration data values from 1. Employing CMAQ-ARIMA led to significant
October 2014 to September 2015 for three stations reduction in RMSE for the daily 1-h and 8-h
of Hong Kong forecasting values at all the stations 47.8–49.7%,
14.3–21.0%, and 41.2–46.3%, respectively, for
­O3, ­PM2.5, and ­NO2. For the 72 hourly advance
forecasts, RMSE values are slashed respectively by
18.2%, 32.1%, and 36.7% for P ­ M2.5, ­NO2, and ­O3 by
employing ARIMAX using CMAQ
2. A great reduction in RMSE values is observed for
rural ­NO2 and nonrural ­PM2.5 when ARIMAX is
employed with three numerical models
3. Dynamic hourly forecast shows that ARIMAX can
also be successfully applied to forecast of 7- to 72-h
­PM2.5, 4- to 72-h ­NO2, and 4- to 6-h ­O3
4. Augmenting the advantages of ARIMA and numeri-
cal modeling, a more efficient and consistent hybrid
ARIMAX model is developed for real-time air
quality forecasting
Zhang et al. (2018) ARIMA AQI data and meteorological parametric data for 1. ­PM2.5 concentrations were positively correlated
Fuzhou (China) from August 2014 to July 2016 with ­PM10, ­NO2, and ­SO2 concentrations and nega-
tively correlated with meteorological parameters.
Also, high seasonal fluctuations were observed dur-
ing cold periods as compared to warm periods
2. ARIMA (6,1,1) best fitted the data
Slini et al. (2002) ARIMA The data of maximum daily ­O3 concentration of The forecasting performance of ARIMA (1,1,1) and
Athens (Greece) from 1990 to 1998 ARIMA (1,1,0) were compared and the limitations
of model over the false alarms also explained
Kumar and Goyal (2011) ARIMA Daily AQI data of RSPM, ­SO2, ­NO2, and SPM for Three models are assessed, i.e., ARIMA (1,1,1), PCR
a period of 7 years ranging from 2000 to 2006 at and ARIMA combined with PCR with respect to 4
ITO station, Delhi seasons of the year in Delhi. The hybrid model per-
formed better in predicting AQI one-day advance
Saha et al. 2019 ARIMA Monthly average and seasonal variation of ventila- This is the first applicability of ARIMA model on the
tion coefficient values over Delhi from Jan 2006 dataset of VC. 12 ARIMA models for each month
to Feb 2014 of the year were developed using Box-Jenkins
methodology and used for forecasting. The average
long-term value of VC was calculated as 1249 ± 236
­m2/s
Lau et al. (2009) ARIMA, SARIMA CO concentration data of district of Hong Kong, 1. The hourly data of CO was found to behave like
between January 1999 and December 2005 traffic data, and this strong association was proven
by SARIMA (0, 1, 1) (0, 1, 1)24 model
2. Long-term memory were features found in data
Environmental Science and Pollution Research (2023) 30:19617–19641
 Table 2  (continued)
Author Methodology Data set Findings

Díaz-Robles et al. (2008) ARIMA, ANN, hybrid model Hourly data of ­PM10 for the time span 2000–2006 Hybrid model outperformed ARIMAX, ANN, and
of the Las Encinas monitoring station situated in MLR on the basis of experimental results
Temuco
Kumar and Jain (2010) ARIMA, ARMA Daily mean concentration ­O3, CO, NO, and ­NO2 at ARIMA is a successful tool in modeling and fore-
an urban traffic site (ITO) of Delhi, India, from casting ­O3, CO, NO, and N ­ O2 and thus can be effec-
August 2006 to July 2007 tively utilized for the forewarning purposes
Kumar and Ridder (2010) FFT-ARIMA and FFT-ARIMA-GARCH Daily maximum ­O3 concentration at Brussels B1 (1 Modeling heteroskedastic effects using GARCH in ­O3
Jan 1998 to 31 July 2007), Brussels (B2) (1 Jan time series helps in making
2002 to 30 Jun 2007), London (L1) (1 May 1996 a) More accurate forecast
to 30 Sep 2007) and London (L2) (1 Jan 2000 to b) Improving confidence intervals
30 Sep 2008) c) Reduction in false alarms at the given sites
Siew et al. (2008) ARIMA, ARFIMA Air pollution index from March 1998 to December The MAPE, RMSE, and MAE values of the ARMA
2003 in Shah Alam, Selangor (3,1,3) and integrated ARFIMA (0, − 0.5, 2) models
were compared. ARFIMA gave slightly better
performance than ARIMA
Abhilash et al. (2018) ARIMA Pollution data of ­NO2, ­PM10, and ­SO2 from January RSPM is the paramount contributor to air pollution,
2013 to March 2016 of Bengaluru (India) while ­SO2 is the most submissive one in this study.
Separate analysis is done for each pollutant and
Environmental Science and Pollution Research (2023) 30:19617–19641

results analyzed. ARIMA model is apt for short-term

predictions
Gocheva-Ilieva et al. (2014) SARIMA, ARIMA Data for concentrations of ­PM10, SO, ­NO2, NO, Factor analysis and SARIMA approach are suitable
NOx, and ­O3 in Blagoevgrad from 1 September techniques in examining air pollution levels for
2011 to 31 August 2012 small urban areas
Chelani and Devotta (2006) ARIMA, nonlinear model, hybrid model on the basis NO2 concentration data set of Delhi from 1999 to Hybrid models are better substitute to individual
of chaos theory 2003 models in forecasting air pollutants
Kim (2010) AR, ARX, T-TARX, P-TARX Daily ozone concentration data (in ppb) for the time The proposed model could capture non-linearity as
period January 2001 to December 2003 in San well as non-stationarity more efficiently and also
Bernardino County forecast with one-day advance more accurately
Ediger and Akar (2007) ARIMA, SARIMA (of seasonal order 1, 2, 3, 4, and Total primary energy data of Turkey (hard coal, ARIMA and SARIMA are an efficient tool in forecast-
5) asphaltite lignite, Petro coke, wood, animal and ing energy demand of Turkey. The accurateness of
plant remains, oil, natural gas, geothermal, hydro- the fitted model is checked with MSE
power, heat, electricity, and solar) from 2005 to
2020 was used
Ediger et al. (2006) ARIMA, SARIMA, regression model Fossil fuel reserves of Turkey from 1950 to 2003 1. Various forecasting models were developed for
different fossil fuel types
2. Fossil fuel production of Turkey will end in 2038

13
19627
 Table 2  (continued)
19628

Author Methodology Data set Findings

13
Li and Li (2017) GM (1,1), ARIMA (1,2,0), GM-ARIMA Energy demand of Shandong province from 1995 1. GM- ARIMA model has higher accuracy in terms
to 2015 of average relative error in comparison to GM (1,1).
ARIMA (1,2,0) and metabolism GM (1,1)
2. Series of residuals are obtained with gray meta-
bolic model. Residual corrections are done with the
help of ARIMA model. This process enhanced the
prediction accuracy significantly, and this was the
major innovation of this study
Aamir and Shabri (2016) ARIMA, GARCH andARIMA-Kalman The average monthly prices of Pakistan crude oil The ARIMAKalman filter technique proved to best
from February 1986 to March 2015 approach as MAE and RMSE were minimum as
compared to ARIMAand GARCH
Zhu and Wei (2013) ARIMA, LSVM, ANN, ARIMALSSVM Two main carbon future prices one maturing on ARIMALSSVM is a better forecasting model than
December 2010 (April 22, 2005, to December the other models. To calculate the optimal param-
20, 2010) and other on December 2012 (April 22, eters, PSO technique is used with an aim of better
2005, to June 27, 2011) forecasting accuracy
Koutroumanidis et al. (2009) ARIMA, ANN, ARIMA -ANN hybrid model Mean annual prices data for coniferous fuelwood for Result revealed better adaptability of ARIMA–ANN
the period 1964–2006 in Greece as compared to ARIMA model
Narayanan et al. (2013) ARIMA Monthly rainfall data for March, April, and May for 1. To investigate the rainfall trend, pre-whitened
six stations Abu, Ahmedabad, Ajmer, Amritsar, Mann Kendall test was employed
Bikaner, and Jodhpur for a period of 60 years were 2. Forecast results for 2000 to 2009 were compared
analyzed from 1949 to 1999 using RMSE
3. Univariate ARIMA modeling showed there is major
increase in the pre-monsoon rainfall over the north-
west part of the country
Ab Razak et al. (2018) ARIMA Daily and monthly rainfall and stream flow data for ARIMA (0, 1, 2) is the best fit model. Mann–Kendall
Bandar Segamat, Kemelah, and Ladang Sungai trend analysis was used to observe the trend in the
Labis stations (Malaysia) for the period 2000 to series
2010
Geetha and Nasira (2016) ARIMA Five years rainfall data of Trivandrum from 2009 For the given time series, ARIMA is the best suited
to 2014 model with forecast accuracy above 80%
Zakaria et al. (2012) Seasonal ARIMA Rainfall data for Sinjar, Mosul, Rabeaa, and Talafar Box-Jenkins (ARIMA) model methodology could be
(3,0,2)x(2,1,1)30, (1,0,1)x(1,1,3)30, (1,1,2) for the time period 1990–2011 used as a suitable tool to forecast the weekly rainfall
x(3,0,1)30 and (1,1,1)x(0,0,1)30 in semi-arid region like North West of Iraq for the
up-coming 5 years (2012–2016)
Chattopadhyay and Chat- ARIMA, ARNN Average rainfall data over India for the time period 1. ARIMA (0,1,1) is the most suitable method
topadhyay (2010) from 1871 to 1999 2. ARNN is superior to ARIMA (0,1,1) based on
Willmott’s index and Pearson correlation
Somvanshi et al. (2006) ARIMA, ANN Average annual rainfall data of Hyderabad for the ANN is an apt technique for modeling the rainfall
period of 1901 to 2003 data of Hyderabad based on RMSE, MAE, AIC,
BIC
Environmental Science and Pollution Research (2023) 30:19617–19641
 Table 2  (continued)
Author Methodology Data set Findings

Wang et al. (2018) ARIMA model and grey model GM (1,1) Monthly incidence of hepatitis B in China from 1. Incidence of hepatitis B in China has seasonal vari-
March 2010 to October 2017 ation. AIC is lowest in ARIMA (3,1,1) (0,1,2)12
2. RMSE, MAE, and MAPE of ARIMA (3,1,1)
(0,1,2)12 model was lower than GM (1,1) model so
ARIMA model showed better hepatitis B fitting and
forecasting performance than GM (1,1) model
Duan and Zhang (2020) ARIMA Daily new confirmed cases of the COVID-19 infec- ARIMA (6,1,7) and ARIMA (2,1,3) was found to be
tions in Japan and South Korea from January 20, best fit model for Japan and South Korea, respec-
2020, to April 26, 2020 tively
Wei et al. (2016) ARIMA, GRNN The monthly morbidity data for hepatitis from the 1. The most appropriate model for the data under
Heng County CDC from January 2005 to Decem- consideration is ARIMA (0,1,2) (1,1,1)12
ber 2012 2. The hybrid ARIMA-GRNN model showed better
forecasting performance than the individual model
in terms of lower MAPE, MAE, MSE, and RMSE
Singh et al. (2020a, b) ARIMA, Hybrid Wavelet – ARIMA The dataset consisting of death cases by COVID-19 The performance of hybrid model is nearly 80% bet-
of Italy, Spain, France, the UK, and the USA from ter than ARIMA model for the countries Italy, Spain,
21 January 2020 to 11 April 2020 and the UK. However, it is approximately 50%
improved for the countries France and the USA
Environmental Science and Pollution Research (2023) 30:19617–19641

Ispriyanti (2018) ARIMA, GARCH Cayenne production monthly data from January For the given data ARIMA ([1,3],0,0)-GARCH (1,0)
2003 to November 2015 in Central Java is the best model
Singh et al. (2013) ARIMA Annual paddy production and its area under cultiva- ARIMA (2, 1, 2) was found the best fit model for
tion from 1974–1975 to 2010–2011of Chhattis- Paddy Area while ARIMA (2, 1, 0) was the most
garh (India) appropriate model for Paddy production in Chhat-
tisgarh (India) with AIC as selection criterion
Assis et al. (2010) ARIMA (3,1,2), GARCH (1,1), mixed ARIMA Monthly average data from January 1992—Decem- The GARCH model gave better results than the expo-
(3,1,2)/GARCH (1,1) exponential smoothing, ber 2006 of Bagan Datoh cocoa bean prices nential smoothing, ARIMA and the mixed ARIMA/
GARCH in forecasting given data using RMSE,
MAE, and MAPE
Ali (2013) EGARCH, GJR-GARCH, TGARCH, AVGARCH, FIB time series from 2006 to 2008 Keeping Akaike, Bayes, Log Likelihood information,
NGARCH, IGARCH, and APARCH and Hannan-Quinn as criterions TGARCH proved
to be slightly better than the other models in captur-
ing response of the pathogen indicator variable.
Nyblom Statistic is also used for each model
Williams and Hoel (2003) ARIMA, SARIMA (1,0,1)(0,1,1)672, (1,0,2) Vehicular traffic Data of M25 motorway of London Since MAD and MAPE values were lowest in case
(0,1,1)672, (2,0,1)(0,1,1)672 for M25 station and and I-75 freeway of Atlanta of SARIMA as compared to ARIMA and Random
SARIMA (1,0,1)(0,1,1)672, (3,0,0)(0,1,1)672, Walk therefore the experimental results supported
(1,0,2)(0,1,1)672 for I-75 stations, random walk the theory of SARIMA as the best model for mod-
eling and forecasting traffic data
Nie et al. (2012) ARIMA, SVM, ARIMA-SVM The electric load data of power company in Hei- On the basis of MAPE and RMSE hybrid, ARIMA-
longjiang of China from March 1 to May 31, 1999 SVM model is better than the two separate models
alone

13
19629
 Table 2  (continued)
19630

Author Methodology Data set Findings

13
Zhang (2003) ARIMA, ANN, Hybrid The Wolf’s sunspot data, the Canadian lynx dat,a Combined model has greater forecasting accuracy on
and the British pound and US dollar exchange rate the basis of MAD, MSE
data from 1980 to 1993
Chen and Wang (2007) SVM, SARIMA, SARIMASVM (integration of Production data of the Taiwanese machinery indus- When forecasted errors were calculated, the hybrid
SARIMA and SVM) try from January 1991 to December 1996 model SARIMASVM2 gave better results
Khashei and Bijari(2011) ARIMA, ANN, Zhang’s hybrid models, novel hybrid Wolf’s sunspot data (1700–1987), Canadian lynx The new model developed outperformed ARIMA,
model data (1821–1934), and the British pound/US dol- ANN, and Zhang’s hybrid models across varied
lar exchange rate data (1980–1993) time horizons
Tseng et al. (2002) SARIMA and SARIMABP Total production revenues time series data of SARIMABP outperformed SARIMA model
Taiwan machinery trade for the time span 1991
to 1996
Wang (2011) ARIMA and fuzzy time series by heuristic models Taiwan exports data for the time period January The heuristic fuzzy time series model is an appro-
1990 and 30 March 2002 priate tool when information is lacking and vital
decisions are needed
Conejo et al. (2005) Wavelet ARIMA Electricity market data of mainland Spain for the Wavelet-ARIMA outperforms ARIMA alone in terms
year 2002 of weekly error (%)
Pai and Lin (2005) ARIMA, SVM, hybrid model Closing price of 10 stocks from Oct. 21, 2002, to By calculating MAE, MSE, MAPE, RMSE for each
Dec. 31, 2002 of the ten companies under consideration and each
model, it was observed that prediction performance
of single ARIMA and SVM was less than the
hybrid model. Hybridizing two different models
with optimal parameters resulted in reduction of the
forecasting errors
Wang et al. (2012) ESM, ARIMA, BPNN, the equal weight hybrid Monthly SZII closing index and DJIAI opening Improved results were obtained with the help of
model (EWH), and the random walk model index from China and the USA, respectively, from hybrid models with RMSE, MAPE, and ME as
(RWM) January 1993 to December 2010 measuring errors in both the cases
Ho et al. (2002) ARIMA, RNN, MFNN The failure time data for a repairable compressor RNN at the optimal weighting factor gives satisfac-
system at a Norwegian process plant was investi- tory performances compared to the ARIMA model.
gated in the time period between 1968 and 1989 MAD and MSE were calculated
Aslanargun et al. (2007) ARIMA, Neural Network, hybrid model Number of monthly tourist arrivals to Turkey (12:1:1) MLP& (5:1:1:1) MLP and (12:1:1) MLP&
between January 1984 and December 2003 (6:9:1) RBFN hybrid gave better results
Faruk (2010) ARIMA, ANN, Hybrid model The dataset of Aydın measurement, comprising Hybrid modeling approach can give more reliable
three water quality parameters monitored over predictions of water temperature, boron, and dis-
9 years (1996–2004), was analyzed solved oxygen time series of a river than the ANN
and ARIMA modeling approach individually
Environmental Science and Pollution Research (2023) 30:19617–19641
Environmental Science and Pollution Research (2023) 30:19617–19641 19631

Table 3  Application of performance evaluation measures on various ARIMA and its hybrid models
Authors Research object area/selection crite- Model used Performance measures
ria/research Data

Mani and Volety (2021) Air pollutant ARIMA RMSE = 0.13

CO MAE = 0.12
LSTM RMSE = 0.13
MAE = 0.09
NH3 ARIMA RMSE = 0.44
MAE = 0.39
LSTM RMSE = 0.14
MAE = 0.13
O3 ARIMA RMSE = 13.6
MAE = 13.13
LSTM RMSE = 2.06
MAE = 1.70
Wang et al. (2017) Station 1.ARIMA RMSE = 8.1242
1. Fuyong MAE = 6.0771
2. SVM RMSE = 8.2779
MAE = 5.8432
3.Hybrid GARCH RMSE = 7.8458
MAE = 5.5375
2. Longhua 1.ARIMA RMSE = 5.2084
MAE = 4.4061
2. SVM RMSE = 4.7392
MAE = 3.8072
3.Hybrid GARCH RMSE = 4.6373
MAE = 3.7256
3. Henggang 1. ARIMA RMSE = 7.1915
MAE = 6.1382
2. SVM RMSE = 7.1910
MAE = 5.8290
3. Hybrid GARCH RMSE = 6.4584
MAE = 5.2359
4. Pingshan 1. ARIMA RMSE = 7.1943
MAE = 5.5786
2. SVM RMSE = 7.2744
MAE = 5.4968
3.Hybrid GARCH RMSE = 7.0182
MAE = 5.1594
Samia et al. (2012) PM10 ARIMAX RMSE = 51.094
R2=0.675
MSE = 2610
MAE = 25.677
ANN RMSE = 29.939
R2=0.888
MSE = 896.2
MAE = 20.573
ARIMAX-ANN RMSE = 11.656
R2=0.983
MSE = 135.828
MAE = 8.268

13
19632 Environmental Science and Pollution Research (2023) 30:19617–19641

Table 3  (continued)
Authors Research object area/selection crite- Model used Performance measures
ria/research Data

Zhu et al. (2017) AQI ARIMA MAPE = 186.400

RMSE = 208.6240
MAE = 198.545
SVR MAPE = 30.200
RMSE = 42.7060
MAE = 35.812
Wavelet-SVR MAPE = 30.890
RMSE = 42.7388
MAE = 36.2368
GRNN MAPE = 25.91
RMSE = 36.0930
MAE = 26.7732
Wavelet-GRNN MAPE = 24.24
RMSE = 33.0731
MAE = 25.9814
EMD-GRNN MAPE = 21.86
RMSE = 27.6400
MAE = 22.3637
EMD-IMFs-Hybrid MAPE = 15.600
RMSE = 25.767
MAE = 17.240
EMD-SVR-Hybrid MAPE = 15.600
RMSE = 24.462
MAE = 18.101
Kumar and Goyal (2011) Summer ARIMA RMSE = 35.30
NMSE = 0.0106
PCR RMSE = 35.70
NMSE = 0.0113
MODEL3 (ARIMA + PCR) RMSE = 31.99
NMSE = 0.0086
Monsoon ARIMA RMSE = 62.55
NMSE = 0.0537
PCR RMSE = 68.93
NMSE = 0.0694
MODEL3 (ARIMA + PCR) RMSE = 61.94
NMSE = 0.0531
Post-monsoon ARIMA RMSE = 57.70
NMSE = 0.0354
PCR RMSE = 62.93
NMSE = 0.0416
MODEL3 (ARIMA + PCR) RMSE = 56.55
NMSE = 0.0341
Winter ARIMA RMSE = 26.74
NMSE = 0.0054
PCR RMSE = 58.39
NMSE = 0.0301
MODEL3 (ARIMA + PCR) RMSE = 65.76
NMSE = 0.0390

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19633

Table 3  (continued)
Authors Research object area/selection crite- Model used Performance measures
ria/research Data

Ji et al. (2019) Carbon prices CNN MAPE = 0.0623

RMSE = 0.8616
LSTM MAPE = 0.0456
RMSE = 0.7251
ARIMA MAPE = 0.0423
RMSE = 0.7015
ARIMA-CNN-LSTM MAPE = 0.0421
RMSE = 0.6940
Chelani and Devotta (2006) NO2 data ARIMA MAPE = 17.3
RMSE = 58.78
RE = 0.24
Nonlinear prediction MAPE = 11.6
RMSE = 55.37
RE = 0.23
Hybrid model MAPE = 5.37
RMSE = 13.93
RE = 0.19

13
19634 Environmental Science and Pollution Research (2023) 30:19617–19641

Table 3  (continued)
Authors Research object area/selection crite- Model used Performance measures
ria/research Data

Kim (2010) AR at location AR RMSE = 15.32

1. Crestline R2=0.74
2. Fontana AR RMSE = 19.45
R2=0.66
3. Redlands AR RMSE = 17.51
R2=0.70
4.SB-4th St AR RMSE = 17.22
R2=0.68
5. Upland AR RMSE = 18.05
R2=0.64
ARX at location 1. Crestline ARX RMSE = 13.37
R2=0.80
2. Fontana ARX RMSE = 14.89
R2=0.80
3. Redlands ARX RMSE = 13.96
R2=0.81
4. SB-4th St ARX RMSE = 13.96
R2=0.81
5. Upland ARX RMSE = 13.11
R2=0.81
P-TARX at location P-TARX RMSE = 12.92
1. Crestline R2=0.82
2. Fontana P-TARX RMSE = 13.91
R2=0.83
3. Redlands P-TARX RMSE = 13.41
R2=0.83
4. SB-4th St P-TARX RMSE = 12.28
R2=0.84
5. Upland P-TARX RMSE = 12.37
R2=0.83
T-TARX at location T-TARX RMSE = 12.56
1. Crestline R2=0.83
2. Fontana T-TARX RMSE = 12.25
R2=0.85
3. Redlands T-TARX RMSE = 13.10
R2=0.84
4. SB-4th St T-TARX RMSE = 11.60
R2=0.86
5. Upland T-TARX RMSE = 11.89
R2=0.85
Somvanshi et al. (2006) Rainfall data ANN RMSE = 5.5186
(4,2,1) R2=0.9841
MAE = 0.9841
ARIMA RMSE = 35.882
(4,0,0) R2=0.953
MAE = 24.388

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19635

Table 3  (continued)
Authors Research object area/selection crite- Model used Performance measures
ria/research Data

Wei et al. (2016) Hepatitis data ARIMA MAPE = 0.1115

RMSE = 1.5561
MSE = 2.4215
GRNN MAPE = 0.0.0150
RMSE = 0.4726
MSE = 0.2233
ARIMA-GRNN MAPE = 0.0878
RMSE = 0.9391
MSE = 0.8820
Zhu and Wei (2013) 1.Carbon prices Dec 10 ARIMA RMSE = 0.2474
LSSVM RMSE = 0.2473
ANN RMSE = 0.2568
ARIMALSSVM2 RMSE = 0.0311
ARIMASVM RMSE = 0.0636
ARIMAANN RMSE = 0.0967
2.Carbon prices Dec12 ARIMA RMSE = 0.2678
LSSVM RMSE = 0.2676
ANN RMSE = 0.2712
ARIMASSVM RMSE = 0.0628
ARIMAANN RMSE = 0.0963
ARIMALSSVM2 RMSE = 0.0309
Koutroumanidis et al. (2009) 1. Coniferous species ARIMA MAPE = 16.9133
RMSE = 0.051308
ANN MAPE = 14.583646
RMSE = 0.05233171
ARIMA-ANN hybrid model MAPE = 14.0036172
RMSE = 0.050151798
2. Broadleaved species data ARIMA MAPE = 16.6856
RMSE = 0.147595
ANN MAPE = 13.016263
RMSE = 0.0966550
Ho et al (2002) 1. Short term RNN(α = 0.3) MAE = 58.7
RNN(α = 0.6) MAE = 58.5
RNN(α = 0.8) MAE = 161.2
MFNN MAE = 297.6
ARIMA MAE = 54.1
2. 1. Long term RNN(α = 0.3) MAE = 89.5
RNN(α = 0.6) MAE = 88.7
RNN(α = 0.8) MAE = 177.6
MFNN MAE = 187.2
ARIMA MAE = 86.7
Díaz-Robles et al. (2008) PM10 data ARIMAX RMSE = 28.46
MAE = 19.87
ANN RMSE = 28.57
MAE = 20.65
Hybrid RMSE = 8.80
MAE = 6.74

13
19636 Environmental Science and Pollution Research (2023) 30:19617–19641

Table 3  (continued)
Authors Research object area/selection crite- Model used Performance measures
ria/research Data

Faruk (2010) Water temperature (°C) ARIMA MAPE = 42.076

RMSE = 0.102
ANN MAPE = 21.114
RMSE = 0.048
Hybrid MAPE = 18.282
RMSE = 0.039
DO (mg l−1 ) ARIMA MAPE = 47.10
RMSE = 0.113
ANN MAPE = 29.612
RMSE = 0.061
Hybrid MAPE = 26.985
RMSE = 0.051
Boron (mg l−1) ARIMA MAPE = 57.512
RMSE = 0.165
ANN MAPE = 36.813
RMSE = 0.074
Hybrid MAPE = 33.611
RMSE = 0.063
Singh et al. (2020a, b) Italy ARIMA MSE = 2.565 ×106
MAE = 1.243 ×103
Wavelet-ARIMA hybrid model MSE = 0.398 ×106
MAE = 0.464 ×103
Spain ARIMA MSE = 0.782 ×106
MSE = 0.782 ×106
Wavelet-ARIMA hybrid model MSE = 0.028 ×106
MAE = 0.136 ×103
France ARIMA MSE = 11.006 ×106
MAE = 24.640 ×103
Wavelet-ARIMA hybrid model MSE = 5.245 ×106
MAE = 1.627 ×103
USA ARIMA MSE = 16.540 ×106
MAE = 2.822 ×103
Wavelet-ARIMA hybrid model MSE = 3.900 ×106
MAE = 1.341 ×103
UK ARIMA MSE = 29.742 ×106
MAE = 1.316 ×103
Wavelet-ARIMA hybrid model MSE = 0.064 ×106
MAE = 0.193 ×103

Reviewing application of ARIMA alone for mainland Spain and California with the conclusion that
GARCH model is more accurate than ARIMA. Yaziz et al.
CONTRERAS et al. (2003) proposed two ARIMA models to (2016) surveyed the modeling and forecasting performances
predict hourly prices in the electricity markets of Spain and of gold prices using ARIMA-TGARCH with Gaussian, Stu-
California. The Spanish data showed volatility. Jadevicius dent’s t skewed, Student’s t, GED, and skewed GED innova-
and Huston (2015) generated around twenty ARIMA models tions. Hybrid ARIMA (0,1,0)-TGARCH (1,1) with t-inno-
ranging from ARIMA (1, 0, 0) to ARIMA (4, 0, 4). ARIMA vation was chosen as the best model.
(3, 0, 3) model outperformed all the models to model Lithua- Ariyo et al. (2014) and Wadi et al. (2018) applied vari-
nian house price index. Garcia et al. (2005) applied ARIMA ous ARIMA models on the stock exchange data sets to find
and GARCH model to forecast one-day ahead electricity ARIMA (2,1,0), ARIMA (1,0,1), and ARIMA (2,1,1) best

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19637

fits Nokia stock index, Zenith bank stock index, and Amman and MFNN to the failure time data for a repairable compres-
Stock Exchange, respectively. Siregar et al. (2017) based sor system at a Norwegian process plant. RNN at the optimal
on ACF and PACF concluded ARIMA (3.0, 2) is the most weighting factor gives satisfactory performances compared to
appropriate model for predicting the sales of the factory in the ARIMA model. The simple and wide use of ARIMA mod-
the city of Bandung. Wabomba et al. (2016) found ARIMA els led Mondal et al. (2014) to study 56 Indian stock markets
(2,2,2) as the best model for Kenyan GDP data. ARIMA spread over different sectors. The high lightening feature of this
(1,1,1) was found the best suitable method on the monthly research was analysis of sector-based ARIMA models, thus
gold price data by Guha and Bandyopadhyay (2016). In covering larger portion of Indian stocks. Different ARIMA
his research work, Lin (2018) gave an extensive review on models were generated and their AIC compared.
GARCH models as well as detailed analysis of stock markets All the above narration can be summarized that in recent
in general and particularly in China. past hybrid modeling techniques have substituted single mod-
eling processes. Also, these emerging nonlinear soft computing
techniques are robust, parsimonious in their data requirements
Investigating ARIMA with hybrid methodology and provide good long-term forecasting. Though with the
advent of new soft computing methods many difficulties while
To capture the volatility present in the data, Babu and Reddy implementing ARIMA have been overcome, still ARIMA
(2014) decomposed the series into two sets using MA filter. remains the benchmark in the field of modeling and forecast-
The proposed hybrid model resulted in improved one-step ing due its high level of simplicity and great level of reliability.
and multi-step ahead prediction accuracy. Nie et al. (2012) Comprehension of ARIMA models along with its
compared hybrid of ARIMA model and SVMs for short- hybrid requires knowledge of performance evaluation cri-
term load forecasting with individual models via simulation teria also. Thus, our coming section deals with various
for the electric load data of power company in Heilongjiang performance evaluation metrices.
of China from March 1 to May 31, 1999. Hybrid ARIMA-
SVM model performed better than the two separate models
alone. Chen and Wang (2007) observed that the values of Performance evaluation of hybrid models
NMSE, R (correlation coefficient), and MAPE were lowest
for hybrid model (SARIMASVM2) for the production data Different statistical metrices such as MAD (mean absolute
of the Taiwanese machinery industry from January 1991 to deviation), SSE (sum squared error), RMSE (root mean
December 1996. Zhang (2003) compared ARIMA and ANN squared error), MSE (mean squared error), MAPE (mean
for the Wolf’s sunspot data from 1700 to 1987, the Canadian absolute percentage error) are employed while evaluating
lynx data, and the British pound spanning from 1821 to 1934 the performance of the proposed model. MAD, RMSE and
and the US dollar exchange rate data extending from 1980 to MAPE are defined by
1993 and concluded combined model has greater forecast- ∑n
ing accuracy. Khashei and Bijari analyzed the same data as MAD = 1n t=1 ��y(t) − ̂ y(t)�� ⎫
�
used by Zhang (2003) in the year 2011. A new hybrid model ∑n 2⎪
RMSE = 1n t=1 ��y(t) − ̂ y(t)�� ⎬
better than Zhang’s model, ARIMA and ANN alone, was ∑ � y(t) � ⎪
MAPE = 1n nt=1 � y(t)−̂ � × 100 ⎭
developed. Tseng et al. (2002) concluded that SARIMABP � y(t) �
model outperformed SARIMA models for the total production
where ̂ y(t) denotes the predicted value of y(t) and n is the
revenues of Taiwan machinery industry. Wang (2011) com-
number of points of the training and testing data sets. It is
pared ARIMA and fuzzy time series by heuristic models with
very much evident from Table 3 that errors are reduced when
Taiwan exports data from January 1990 and 30 March 2002.
hybrid models are implemented on the various data sets.
For longer analyzing time period, the MSE of the time series
Table 3 represents the various models employed and the
ARIMA model are lower, while for shorter analyzing period,
criteria chosen for performance evaluation of these various
the MSE are more. The heuristic fuzzy time series model is
models by various authors.
an appropriate tool when information is lacking and an urgent
decision is needed. Wang and Leu (1996) used a hybrid model
to conclude that ANN provides better results with differenced
data than raw data in case of Taiwan stock exchange.
Conclusion
Wang et al. (2012) observed data for monthly closing index
Though ARIMA finds its application extensively in the field
of SZII and opening index of DJIAI from China and the USA,
of forecasting, it too has shortcomings. ARMA models are
respectively, from January 1993 to December 2010. It was seen
linear models, but the time series involving environmental/
that hybrid model could more effectively capture various rela-
atmosphere/air quality/financial data, etc. are rarely pure
tionships in the data. Ho et al. (2002) applied ARIMA, RNN,

13
19638 Environmental Science and Pollution Research (2023) 30:19617–19641

linear combinations. A great competence, expertise, and Abhilash MSK, Thakur A, Gupta D, Sreevidya B (2018) Time series
experience of a researcher is required while implement- analysis of air pollution in Bengaluru using ARIMA model.
Ambient Commun Comp Sys. Springer, Singapore, pp 413–426
ing results of the Box-Jenkins procedure. Moreover, results Ahn H (2000) Modeling of groundwater heads based on second-order
are strongly affected by the path chosen and hence are path difference time series models. J Hydrol 234(1–2):82–94
dependent (Weigend and Gershenfeld 1994). But almost all Akaike H (1969) Fitting autoregressive models for prediction. Ann Inst
the models though different from each other have similar Stat Math 21(1):243–247
Akaike H (1974) A new look at the statistical model identification.
estimated correlation patterns, resulting in the arbitrary IEEE Trans Autom Control 19(6):716–723
choice of the model (Box et al. 1994). Another drawback of Ali G (2013) EGARCH, GJR-GARCH, TGARCH, AVGARCH,
Box-Jenkins procedure is that it is sometimes highly time NGARCH, IGARCH and APARCH models for pathogens at
consuming in the model identification step. ARMA models marine recreational sites. J Stat Econ Methods 2(3):57–73
Al-Saba T, El-Amin I (1999) Artificial neural networks as applied to
are highly unsuitable for running policy simulations. long-term demand forecasting. Artific Intell Eng 13:189–197
After such a massive review of ARIMA and hybrid mod- Anderson TW (1971) The stationarity of an estimated autoregressive
eling involving ARIMA in the fields including or excluding process. STANFORD UNIV CA DEPT OF STATISTICS
environment/health/atmosphere, it can be concluded that the Ariyo AA, Adewumi AO, Ayo CK (2014) Stock price prediction using
the ARIMA model. In: UKSim-AMSS 16th Int Conf Comp Mod-
combined models are more robust and have higher ability to elling Simu: 106–112
capture all the patterns of the series uniformly. Thus, com- Aslanargun A, Mammadov M, Yazici B, Yolacan S (2007) Com-
bining several models or using hybrid model has emerged parison of ARIMA, neural networks and hybrid models in
as a routinized custom, though ARIMA still remains the time series: tourist arrival forecasting. J Stat Comput Simul
77(1):29–53
benchmark of many baseline models. Assis K, Amran A, Remali Y (2010) Forecasting cocoa bean prices
using univariate time series models. Res World 1(1):71
Acknowledgements The authors are thankful to the I K Gujral Punjab Astrom KJ, Eykhoff P (1971) System identification—a survey. Auto-
Technical University, Government of Punjab, for the research facilities. matica 7(2):123–162
The corresponding author is also thankful to SERB-DST, Government Babu CN, Reddy BE (2014) A moving-average filter-based hybrid
of India, for the financial support with the research project MATRICS ARIMA–ANN model for forecasting time series data. Appl Soft
MTR/2020/000479. Comput 23:27–38
Basu AK, Das JK (1991) On estimation and asymptotic properties
Author contribution Ms. Jatinder Kaur wrote the paper and studied all of the parameters of ARMA (p, q) process in the stable case.
the previous published research works. Dr. Kulwinder Singh Parmar Calcutta Statist Assoc Bull 41(1–4):45–64
developed the idea and did the computational work. Mr. Sarbjit Singh Basu AK, Das JK (1992) Optimality of the maximum likelihood esti-
also did the computational work and helped in writing the manuscript. mator in AR (p) model under a general set-up of the roots. Cal-
cutta Statist Assoc Bull 42(1–2):1–18
Funding The corresponding author received funding from Bates JM, Granger CW (1969) The combination of forecasts. Journal
SERB-DST, Government of India, under the MATRICS project of the Operational Research Society 20(4):451–468
(MTR/2020/000479). Benvenuto D, Giovanetti M, Vassallo L, Angeletti S, Ciccozzi M
(2020) Application of the ARIMA model on the COVID-2019
Data availability There is no use of raw or published data for this cur- epidemic dataset. Data Brief 29:105340
rent review article. Bhansali RJ (1983) A simulation study of autoregressive and window
estimators of the inverse correlation function. J Roy Stat Soc: Ser
Declarations C (appl Stat) 32(2):141–149
Box GEP, Jenkins G (1970) Time series analysis, forecasting and con-
Research ethics/ethics approval and consent to participate We further trol. Holden-Day, San Francisco, CA
confirm that there is no aspect of the work covered in this manuscript Box GEP, Jenkins GM, Reinsel GC (1994) Time series analysis: fore-
that has involved human patients. casting and control, 3rd edn. Prentice-Hall, Englewood Cliffs, NJ
Broersen P (1985) Selecting the order of autoregressive models from
Consent for publication The authors give consent for the publication small samples. IEEE Trans Acoust Speech Signal Process
of this manuscript. 33(4):874–879
Buchananan WK, Hodges P, Theis J (2001) Which way the natural gas
Competing interests The authors declare no competing interests. price: an attempt to predict the direction of natural gas spot price
movements using trader positions. Energ Eco 23(3):279–293
Burg JP (1975) Maximum entropy spectral analysis. Stanford
University
References Cappello C, De Iaco S, Posa D (2018) Testing the type of non-sep-
arability and some classes of space-time covariance function
Aamir M, Shabri A (2016) Modelling and forecasting monthly crude models. Stoch Env Res Risk Assess 32(1):17–35
oil price of Pakistan: a comparative study of ARIMA, GARCH Cartwright PA (1985) Forecasting time series: a comparative analysis
and ARIMA Kalman model. In AIP Conf Proc 1750(1):060015 of alternative classes of time series models. J Time Ser Anal
Ab Razak NH, Aris AZ, Ramli MF, Looi LJ, Juahir H (2018) Tem- 6(4):203–211
poral flood incidence forecasting for Segamat River (Malaysia) Chattopadhyay S, Chattopadhyay G (2010) Univariate modelling of
using autoregressive integrated moving average modelling. J summer-monsoon rainfall time series: comparison between
Flood Risk Managt 11:794–804 ARIMA and ARNN. CR Geosci 342(2):100–107

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19639

Chaudhuri S, Dutta D (2014) Mann-Kendall trend of pollutants, tem- Godolphin EJ (1977) A direct representation for the maximum likeli-
perature and humidity over an urban station of India with forecast hood estimator of a Gaussian moving average process. Biom-
verification using different ARIMA models. Environtl Monitng etrika 64(2):375–384
Assesst 186(8):4719–4742 Godolphin EJ (1978) Modified maximum likelihood estimation of
Chelani AB, Devotta S (2006) Air quality forecasting using a Gaussian moving averages using a pseudoquadratic convergence
hybrid autoregressive and nonlinear model. Atmos Environ criterion. Biometrika 65:203–206
40(10):1774–1780 Guarnaccia C, Quartieri J (1836) Tepedino C (2017) Deterministic
Chen KY, Wang CH (2007) A hybrid SARIMA and support vector decomposition and seasonal ARIMA time series models applied
machines in forecasting the production values of the machinery to airport noise forecasting. In: AIP Conf Proc 1:020079
industry in Taiwan. Expert Sys App 32(1):254–264 Guha B, Bandyopadhyay G (2016) Gold price forecasting using
Choi BS (1986) An algorithm for solving the extended Yule-Walker ARIMA model. J Adv Manage Sci 4(2)
equations of an autoregressive moving-average time series (Cor- Hannan EJ (1970) Multiple time series Wiley. New York
resp.). IEEE Transac Inform Theor 32(3):417–419 Hannan EJ, Rissanen J (1982) Recursive estimation of mixed autore-
Cleveland WS (1972) The inverse autocorrelations of a time series and gressive-moving average order. Biometrika 69(1):81–94
their applications. Technometrics 14(2):277–293 Harvey A (1989) Forecasting, structural time series models and the
Conejo AJ, Plazas MA, Espinola R, Molina AB (2005) Day-ahead elec- Kalman filter. Cambridge University Press
tricity price forecasting using the wavelet transform and ARIMA Harvey A, Koopman SJ, Shephard N (2004) State space and unob-
models. IEEE Transac Power Sys 20(2):1035–1042 served component models. Cambridge University Press
Contreras J, Espinola R, Nogales FJ, Conejo AJ (2003) ARIMA mod- Hipel KW, McLeod AI, Lennox WC (1977) Advances in Box-Jenkins
els to predict next-day electricity prices. IEEE Trans Power Sys modeling: 1. Model Construction. Water Resour Res 13(3):567–575
18(3):1014–1020 Hippert HS, Pedreira CE, Souza RC (2001) Neural networks for short-
De Iaco S, Myers DE, Palma M, Posa D (2013) Using simultaneous term load forecasting: a review and evaluation. IEEE Trans
diagonalization to identify a space–time linear coregionalization Power Syst 16:44–55
model. Math Geosci 45(1):69–86 Ho SL, Xie M, Goh TN (2002) A comparative study of neural network
De Iaco S, Palma M, Posa D (2015) Spatio-temporal geostatistical and Box-Jenkins ARIMA modeling in time series prediction.
modeling for French fertility predictions. Spatial Stat 14:546–562 Comp Industl Engg 42(2–4):371–375
De Iaco S, Posa D (2018) Strict positive definiteness in geostatistics. Hokstad P (1983) A method for diagnostic checking of time series
Stoch Env Res Risk Assess 32(3):577–590 models. J Time Series Anal 4(3):177–183
Díaz-Robles LA, Ortega JC, Fu JS, Reed GD, Chow JC, Watson JG, Hurvich CM, Tsai CL (1989) Regression and time series model selec-
Moncada-Herrera JA (2008) A hybrid ARIMA and artificial neu- tion in small samples. Biometrica 76(2):297–307
ral networks model to forecast particulate matter in urban areas: Huzii M (1981) Estimation of coefficients of an autoregressive process
The case of Temuco. Chile Atmos Environ 42(35):8331–8340 by using a higher order moment. J Time Ser Anal 2(2):87–93
Duan X, Zhang X (2020) ARIMA modelling and forecasting of irregu- Ispriyanti D (2018) Modelling of cayenne production in Central
larly patterned COVID-19 outbreaks using Japanese and South Java using ARIMA-GARCH. Int J Phy: Conference Series
Korean data. Data Brief 31:105779 1025(1):012120
Durbin J (1959) Efficient estimation of parameters in moving-average Durbin J, Koopman SJ (2012) Time series analysis by state space meth-
models. Biometrika 46(3/4):306–316 ods, 2nd edn. Oxford University Press
Durbin J (1960) Estimation of parameters in time-series regression Jadevicius A, Huston S (2015) ARIMA modelling of Lithuanian house
models. J Roy Stat Soc: Ser B (methodol) 22(1):139–153 price index. Int J Housing Markets Anal.
Ediger VŞ, Akar S (2007) ARIMA forecasting of primary energy Jaiswal A, Samuel C, Kadabgaon VM (2018) Statistical trend analysis
demand by fuel in Turkey. Energy Policy 35(3):1701–1708 and forecast modeling of air pollutants. Global J Environtl Sci
Ediger VŞ, Akar S, Uğurlu B (2006) Forecasting production of fos- Managt 4(4):427–438
sil fuel sources in Turkey using a comparative regression and Janacek GJ (1982) Determining the degree of differencing for time
ARIMA model. Ener Policy 34(18):3836–3846 series via the log spectrum. J Time Ser Anal 3(3):177–183
Faruk DO (2010) A hybrid neural network and ARIMA model for water Jenkins GM (1982) Some practical aspects of forecasting in organiza-
quality time series prediction. Engg Appln Artif Intell 23:586–594 tions. J Forecast 1:3–21
Fuller WA (1976) Introduction to statistical time series, new york: Ji L, Zou Y, He K, Zhu B (2019) Carbon futures price forecasting based
Johnwiley. FullerIntroduction to Statistical Time Series1976 with ARIMA-CNN-LSTM model. Procedia Comp Sci 162:33–38
Garcia RC, Contreras J, Van Akkeren M, Garcia JBC (2005) A GARCH Khashei M, Bijari M (2011) A novel hybridization of artificial neural
forecasting model to predict day-ahead electricity prices. IEEE networks and ARIMA models for time series forecasting. Appl
Trans Power Sys 20(2):867–874 Soft Comput 11(2):2664–2675
Garg N, Soni K, Saxena TK, Maji S (2015) Applications of Autore- Kim SE (2010) Tree-based threshold modeling for short-term forecast
gressive integrated moving average (ARIMA) approach in time- of daily maximum ozone level. Stoch Environ Res Risk Assess
series prediction of traffic noise pollution. Noise Control Engg 24(1):19–28
J 63(2):182–194 Koehler AB, Murphree ES (1988) A comparison of the Akaike and
Geetha A, Nasira GM (2016) Time-series modelling and forecasting: Schwarz criteria for selecting model order. J Roy Stat Soc: Ser
modelling of rainfall prediction using ARIMA model. Int J Soc C (Appl Stat) 37(2):187–195
Sys Sci 8(4):361–372 Koreisha SG, Pukkila TM (1993) New approaches for determining
Glasbey CA (1982) A generalization of partial autocorrelations use- the degree of differencing necessary to induce stationarity in
ful in identifying ARMA models. Technometrics 24(3):223–228 ARIMA models. J Stat Plan Infer 36(2–3):399–412
Gocheva-Ilieva SG, Ivanov AV, Voynikova DS, Boyadzhiev DT (2014) Koutroumanidis T, Ioannou K, Arabatzis G (2009) Predicting fuelwood
Time series analysis and forecasting for air pollution in small prices in Greece with the use of ARIMA models, artificial neu-
urban area: a SARIMA and factor analysis approach. Stoch Env ral networks and a hybrid ARIMA–ANN model. Energ Policy
Res Risk Assess 28(4):1045–1060 37(9):3627–3634
Godfrey LG (1979) Testing the adequacy of a time series model. Biom- Kulkarni GE, Muley AA, Deshmukh NK, Bhalchandra PU (2018)
etrika 66(1):67–72 Autoregressive integrated moving average time series model

13
19640 Environmental Science and Pollution Research (2023) 30:19617–19641

for forecasting air pollution in Nanded city, Maharashtra, India. Newbold P, Granger CWJ (1974) Experience with forecasting univari-
Model Earth Syst Environ 4(4):1435–1444 ate time series and the combination of forecasts. J R Stat Soc Ser
Kumar A, Goyal P (2011) Forecasting of daily air quality index in A 137:131–164
Delhi. Sci Total Environ 409(24):5517–5523 Nicolaisen JD, Richter Jr, CW, Sheblé GB (2000) Price signal analy-
Kumar K, Yadav AK, Singh MP, Hassan H, Jain VK (2004) Fore- sis for competitive electric generation companies. In Proc Conf
casting daily maximum surface ozone concentrations in Brunei Elect Utility Deregulation and Restructuring and Power Technol,
Darussalam-an ARIMA modeling approach. J Air Waste Managt London, U.K. 66–71
Assoc 54(7):809–814 Nie H, Liu G, Liu X, Wang Y (2012) Hybrid of ARIMA and SVMs for
Kumar U, De Ridder K (2010) GARCH modelling in association short-term load forecasting. Ener Procedia 16:1455–1460
with FFT–ARIMA to forecast ozone episodes. Atmos Environ Pai PF, Lin CS (2005) A hybrid ARIMA and support vector machines
44(34):4252–4265 model in stock price forecasting. Omega 33(6):497–505
Kumar U, Jain VK (2010) ARIMA forecasting of ambient air pollut- Parmar KS, Bhardwaj R (2014) Water quality management using sta-
ants (O 3, NO, NO 2 and CO). Stoch Environ Res Risk Assess tistical analysis and time-series prediction model. Appl Water
24(5):751–760 Sci 4:425–434
Lau JC, Hung WT, Yuen DD, Cheung CS (2009) Long-memory Parzen E (1974) Some recent advances in time series modeling. IEEE
characteristics of urban roadside air quality. Transp Res Part d: Trans Autom Control 19(6):723–730
Transp Environ 14(5):353–359 Patil RB (1990) Neural networks as forecasting experts: test of dynamic
Li S, Li R (2017) Comparison of forecasting energy consumption in modeling over time series data. Thesis, Oklahoma State Univer-
Shandong, China Using the ARIMA model, GM model, and sity, M.Sc
ARIMA-GM model. Sustainability 9(7):1181 Pohoata A, Lungu E (2017) A complex analysis employing ARIMA
Liang G, Wilkes DM, Cadzow JA (1993) ARMA model order estima- model and statistical methods on air pollutants recorded in Ploi-
tion based on the eigenvalues of the covariance matrix. IEEE esti. Romania Rev Chim 68(4):818–823
Trans Signal Process 41(10):3003–3009 Poskitt DS (1987) A modified Hannan—Rissanen strategy for mixed
Lin Z (2018) Modelling and forecasting the stock market volatility of autoregressive-moving average order determination. Biometrika
SSE Composite Index using GARCH models. Future Gen Comp 74(4):781–790
Sys 79:960–972 Prybutok VR, Yi J, Mitchell D (2000) Comparison of neural network
Liu T, Lau AK, Sandbrink K, Fung JC (2018) Time series forecasting models with ARIMA and regression models for prediction of
of air quality based on regional numerical modeling in Hong Houston’s daily maximum ozone concentrations. European J
Kong. J Geophysical Res Atmos 123(8):4175–4196 Oper Res 122:31–40
Ljung L (1987) System Identification Theory for the User. Prentice- Pukkila T, Koreisha S, Kallinen A (1990) The identification of ARMA
Hall, Englewood Cliffs, NJ models. Biometrika 77(3):537–548
Makridakis S (1989) Why combining works? Int J Forecast 5:601–603 Pukkila TM, Krishnaiah PR (1988) On the use of autoregressive order
Makridakis S, Andersen A, Carbone R, Fildes R, Hibon M, Lewan- determination criteria in univariate white noise tests. IEEE Trans
dowski R, Newton J, Winkler R (1982) The accuracy of extrapo- Acoust Speech Signal Process 36(5):764–774
lation (time series) methods: results of a forecasting competition. Ray D (1988) Comparison of forecasts: an empirical investigation
J Forecast 1:111–153 Sankhyˉa 50B:258-277
Makridakis S, Wheelwright SC (1977) Forecasting: issues & chal- Robeson SM, Steyn DG (1990) Evaluation and comparison of statisti-
lenges for marketing management. J Mark 41(4):24–38 cal forecast models for daily maximum ozone concentrations.
Mani G, Volety R (2021) A comparative analysis of LSTM and Atmos Environ 24(2):303–312
ARIMA for enhanced real-time air pollutant levels forecast- Saikkonen P (1986) Asymptotic properties of some preliminary esti-
ing using sensor fusion with ground station data. Cogent Engg mators for autoregressive moving average time series models. J
8(1):1936886 Time Ser Anal 7(2):133–155
Marquardt DW (1963) An algorithm for least-squares estimation of Samia A, Kaouther N, Abdelwahed T (2012) A hybrid ARIMA and
nonlinear parameters. J Soc Ind Appl Math 11(2):431–441 artificial neural networks model to forecast air quality in urban
McLeod I (1974) Contributions to applied time series, master’s thesis. areas: case of Tunisia. Adv Mater Res 518:2969–2979
Univ. of Waterloo, Ont Schwarz G (1978) Estimating the dimension of a model. The annals
McLeod I (1977) Derivation of the theoretical autocovariance function of statistics, 461–464.
of autoregressive-moving average time series. J Roy Stat Soc: Ser Shahwan T, Odening M (2007) Computational Intelligence in Eco-
C (appl Stat) 26(2):194–194 nomics and Finance. Springer, Berlin Heidelberg, New York,
Mikosch T, Gadrich T, Kluppelberg C, Adler RJ (1995) Parameter pp 63–74
estimation for ARMA models with infinite variance innova- Shibata R (1976) Selection of the order of an autoregressive model by
tions. Ann Stat 305–326 Akaike’s information criterion. Biometrika AC-63(1):117–126
Mirzavand M, Sadatinejad S J, Ghasemieh H, Imani R, Mot- Siew LY, Chin LY, Wee PMJ (2008) ARIMA and integrated ARFIMA
lagh M S (2014) Prediction of ground water level in arid models for forecasting air pollution index in Shah Alam Selan-
environment using a non-deterministic model. J Water Res gor. Malaysian J Analytical Sci 12(1):257–263
Protect 2014 Silvey SD (1959) The Lagrangian multiplier test. Ann Math Stat
Monahan JF (1983) Fully Bayesian analysis of ARMA time series mod- 30(2):389–407
els. J Econ 21(3):307–331 Singh DP, Kumar P, Prabakaran K (2013) Application of ARIMA
Mondal P, Shit L, Goswami S (2014) Study of effectiveness of time model for forecasting Paddy production in Bastar division of
series modeling (ARIMA) in forecasting stock prices. Int J Comp Chhattisgarh. Amer Int J Res Sci Technol Engg Math 5:82–87
Sci Engg Appl 4(2):13 Singh S, Parmar KS, Kaur J, Kumar J, Makkhan SJS (2021) Prediction
Morana C (2001) A semiparametric approach to short-term oil price of COVID-19 pervasiveness in six major affected states of India
forecasting. Energ Eco 23(3):325–338 and two-stage variation with temperature. Air Qual Atmos Health
Narayanan P, Basistha A, Sarkar S, Kamna S (2013) Trend analysis 14(12):2079–2090
and ARIMA modelling of pre-monsoon rainfall data for western Singh S, Parmar KS, Kumar J, Makkhan SJS (2020) Development
India. CR Geosci 345(1):22–27 of new hybrid model of discrete wavelet decomposition and

13
Environmental Science and Pollution Research (2023) 30:19617–19641 19641

autoregressive integrated moving average (ARIMA) models in Wang JJ, Wang JZ, Zhang ZG, Guo SP (2012) Stock index forecast-
application to one month forecast the casualties cases of COVID- ing based on a hybrid model. Omega 40(6):758–766
19. Chaos Solitons Fractals 135:109866 Wang P, Zhang H, Qin Z, Zhang G (2017) A novel hybrid-Garch
Singh S, Parmar KS, Kumar J (2021) Soft computing model coupled model based on ARIMA and SVM for PM2. 5 concentrations
with statistical models to estimate future of stock market. Neural forecasting. Atmos Poll Res 8(5):850–860
Comput Appl 33(13):7629–7647 Wang YW, Shen ZZ, Jiang Y (2018) Comparison of ARIMA and
Singh S, Parmar KS, Makkhan SJS, Kaur J, Peshoria S, Kumar J (2020) GM (1, 1) models for prediction of hepatitis B in China. PLoS
Study of ARIMA and least square support vector machine (LS- ONE 13(9):e0201987
SVM) models for the prediction of SARS-CoV-2 confirmed Wei W, Jiang J, Liang H, Gao L, Liang B, Huang J, ... Chen H (2016)
cases in the most affected countries. Chaos Solitons Fractals Application of a combined model with autoregressive inte-
139:110086 grated moving average (ARIMA) and generalized regression
Siregar B, Nababan EB, Yap A, Andayani U (2017) Forecasting of raw neural network (GRNN) in forecasting hepatitis incidence in
material needed for plastic products based in income data using Heng County, China. PloS one, 11(6):e0156768
ARIMA method. In: 5th Int Conf Electl, Electron Infor Engg Weigend AS, Gershenfeld NA (1994) Time Series Prediction: Fore-
(ICEEIE) 135–139 casting the Future and Understanding the Past Addison-Wesley.
Slini T, Karatzas K, Moussiopoulos N (2002) Statistical analysis of Reading
environmental data as the basis of forecasting: an air quality Weiss E (2000) Forecasting commodity prices using ARIMA. Techl
application. Sci Total Environ 288(3):227–237 Anal Stocks Commod 18(1):18–19
Somvanshi VK, Pandey OP, Agrawal PK, Kalanker NV, Prakash MR, Williams BM, Hoel LA (2003) Modeling and forecasting vehicular
Chand R (2006) Modeling and prediction of rainfall using arti- traffic flow as a seasonal ARIMA process: Theoretical basis
ficial neural network and ARIMA techniques. J Ind Geophys and empirical results. J Transpo Engg 129(6):664–672
Union 10(2):141–151 Wilson G (1969) Factorization of the covariance generating func-
Sreenivasan M, Sumathi K (1997) Generalised parameters technique tion of a pure moving average process. SIAM J Numer Anal
for identification of seasonal ARMA (SARMA) and non-sea- 6(1):1–7
sonal ARMA (NSARMA) models. Korean J Comput Appl Math Woodward WA, Gray HL (1981) On the relationship between the S
4(1):135–146 array and the Box-Jenkins method of ARMA model identifica-
Suresh KK, Priya SK (2011) Forecasting sugarcane yield of Tamilnadu tion. J Am Stat Assoc 76(375):579–587
using ARIMA models. Sugar Tech 13(1):23–26 Yaziz SR, Azizan NA, Ahmad MH, Zakaria R (2016) Model-
Taheri Tizro A, Ghashghaie M, Georgiou P, Voudouris K (2014) Time ling gold price using ARIMA-TGARCH. Appl Mathl Sci
series analysis of water quality parameters. J Appl Res Water 10(28):1391–1402
Wastewater 1(1):40–50 Yule GU (1921) On the time-correlation problem, with especial ref-
Tiao GC, Tsay RS (1994) Some advances in non-linear and adaptive erence to the variate-difference correlation method. J Roy Stat
modelling in time-series. J Forecast 13(2):109–131 Soc 84(4):497–537
Tsay RS, Tiao GC (1984) Consistent estimates of autoregressive Yule GU (1927) On a method of investigating periodicities dis-
parameters and extended sample autocorrelation function for turbed series, with special reference to Wolfer’s sunspot
stationary and nonstationary ARMA models. J Am Stat Assoc numbers. Philosophical Transactions of the Royal Society of
79(385):84–96 London. Series A, Containing Papers of a Mathematical or
Tsay RS, Tiao GC (1985) Use of canonical analysis in time series Physical Character, 226(636–646), 267–298
model identification. Biometrika 72(2):299–315 Zakaria S, Al-Ansari N, Knutsson S, Al-Badrany T (2012) ARIMA
Tseng FM, Yu H, Tzeng G (2002) Combining neural network model Models for weekly rainfall in the semi-arid Sinjar District at
with seasonal time series ARMA model. Technol Forecast Soc Iraq. J Earth Sci Geotech Eng 2(3)
Change 69:71–87 Zhang GP (2003) Time series forecasting using a hybrid ARIMA and
Tuan PD (1984) The estimation of parameters for autoregressive mov- neural network model. Neurocomputing 50:159–175
ing average models. J Time Ser Anal 5(1):53–68 Zhang L, Lin J, Qiu R, Hu X, Zhang H, Chen Q, Wang J (2018) Trend
Tucker WT (1982) On the pade tablé and its relationship to the r and analysis and forecast of PM2. 5 in Fuzhou, China using the
s arrays and arm a modeling. Commun Stat-Theory Methods ARIMA model. Ecol Indicat 95:702–710
11(12):1335–1379 Zhang XD, Zhang YS (1993) Singular value decomposition-based
Unbehauen H, Göhring B (1974) Tests for determining model order in MA order determination of non-Gaussian ARMA models. IEEE
parameter estimation. Automatica 10(3):233–244 Trans Signal Process 41(8):2657–2664
Van den Boom AJW, Van Den Enden AWM (1974) The determina- Zhu B, Wei Y (2013) Carbon price forecasting with a novel hybrid
tion of the orders of process-and noise dynamics. Automatica ARIMA and least squares support vector machines methodology.
10(3):245–256 Omega 41(3):517–524
Wabomba MS, Mutwiri MP, Fredrick M (2016) Modeling and fore- Zhu S, Lian X, Liu H, Hu J, Wang Y, Che J (2017) Daily air quality
casting Kenyan GDP using autoregressive integrated moving index forecasting with hybrid models: a case in China. Environtl
average (ARIMA) models. Sci J Appl Math Stat 4(2):64–73 Poll 231:1232–1244
Wadi SAL, Almasarweh M, Alsaraireh AA, Aqaba J (2018) Predict-
ing closed price time series data using ARIMA Model. Modern Publisher's note Springer Nature remains neutral with regard to
Appl Sci 12(11):181–185 jurisdictional claims in published maps and institutional affiliations.
Wang CC (2011) A comparison study between fuzzy time series
model and ARIMA model for forecasting Taiwan export. Springer Nature or its licensor (e.g. a society or other partner) holds
Expert Syst Appl 38(8):9296–9304 exclusive rights to this article under a publishing agreement with the
Wang JH, Leu JY (1996) Stock market trend prediction using author(s) or other rightsholder(s); author self-archiving of the accepted
ARIMA-based neural networks. IEEE Int Conf Neural Net- manuscript version of this article is solely governed by the terms of
works 4(6):2160–2165 such publishing agreement and applicable law.

13