Journal of Mathematics and Statistics

Original Research Paper

Advanced Multivariate Time Series Forecasting Models

Miss Lei Wang

Department of Mathematics, Statistics and Physics, Wichita State University, USA

Article history Abstract: In this study, the focus is to collect and summarize various
Received: 27-10-2018 advanced forecasting models for multivariate time series dataset. We have
Revised: 05-12-2018 discussed about the inherent forecasting strengths and weaknesses related
Accepted: 19-12-2018 to these time series modelings. Also, the main section deal with the
experience of using such data in econometric analysis. Besides, the
Email: Fiona901587@yahoo.com
implementation of SAS and R softwares improve the parameter estimation
and forecasting accuracy. Eventually, we evaluated these forecasting
models by different criterions and select the best one for the future
tendency of land market value.

Keywords: Advanced Forecasting Model, Macroeconomic Variables,

Multivariate Time Series Analysis

Introduction mathematical studies within the field of biological

mathematics, physical and chemical mathematics and
Nowadays, there are a lot of methods and techniques others. Perhaps the Delphi Method is the most formal
to analyze and forecasting time series dataset. A lot of and widely known qualitative forecasting method. This
researchers have been studying time series forecasting technique was developed by the RAND Corporation. It
for approximately one century in order to get better employs a panel of experts who are assumed to be
forecasts for the future. To achieve best forecast knowledgeable about the problem. The panel members
accuracy level, various time series forecasting are physically separated to avoid their deliberations
approaches have been proposed in the literature. After being impacted either by social pressures or by a single
1980s, more sophisticated algorithms could be improved dominant individual.
since properties of computers were enhanced. Quantitative forecasting techniques is used to
This research presents a time series estimation and quantify the problem by way of generating numerical
prediction methods with the use of classic and advanced data or data that can be transformed into usable statistics.
forecasting tools. Our discussion about different time Quantitative methods are much more structured and
series models is supported by giving the experimental
reliable results than Qualitative methods. In subsequent
forecast results, performed on several macroeconomic
sections, as Fig. 1 shown, we will discuss all several
variables. Also, this paper gives an introduction to the
different types of quantitative forecasting models.
basic concepts of time series modeling, together with
some associated ideas such as stationarity and
parsimony. Finally, the summary of various existing Advanced Forecasting Models
forecasting models can provide information to develop Smoothing Model
an appropriate forecasting model which describes the
inherent feature of the series. For this method, we will present mechanics of the
The forecasting techniques can be broadly most important exponential smoothing methods and
categorized as consisting of qualitative and quantitative their application in forecasting time series with
methods. various characteristics. This helps us develop an
Qualitative forecasting techniques are mainly intuition to how these methods work. We apply the
exploratory research. Qualitative forecasts are often simple exponential smoothing method, Holt's linear
used in providing the insights into the problems. method, exponential smoothing method and additive
However, although some data analysis may be damped method and multiplicative damped method.
performed, the expectations are based on the For Exponential Smoothing Methods (ETS) model, we

© 2018 Miss Lei Wang. This open access article is distributed under a Creative Commons Attribution (CC-BY) 3.0
license.
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

take into innovation by considering multiplicative yt − ( lt −1 + bt −1 )

error equations: εt = (4)
lt −1 + bt −1

yt = ( lt −1 + bt −1 )(1 + ε t ) (1)
where, the ε ∼ NID(0, σ2), lt denotes an estimate of the
level of the series at time t, bt denotes an estimate of the
lt = ( lt −1 + bt −1 )(1 − αε t ) (2)
trend (slope) of the series at time t, α denotes the
smoothing parameter for level. β denotes the smoothing
bt = bt −1 + β ( lt −1 + bt −1 ) ε t (3) parameter for the trend.

Forecasting techniques

Qualitative forecasting Quantitative forecasting

Casual model Time series data

Delph method Brainstorming
Cause and effect

Intuition Unit root test ADF

Expert’s idea

Without trend With trend

ARIMA Non-stationary

Cointegration linear regression ETS

VECM

Fig. 1: Summary of advanced forecasting models

254
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

Autoregressive Integrated Moving Average Model  LLMVt −1 

 
The time series processes we have discussed so far  LCPI t −1 
are all stationary processes, but many applied time  LURt −1 
Yt = C + Φ1  
series, particularly those arising from economic and  LPPt −1 
business areas, are non-stationary. With respect to the  LCCI 
class of covariance stationary processes, non-stationary  t −1

time series can occur in many different ways. They could  LPMI t −1 
(7)
have non-constant means µt, time-varying second  LLMVt − 2   LLMVt − p 
   
moments such as non-constant variance σ t2 , or both of  LCPI t−2   LCPI t − p 
these properties. In this section, we will explain the  LURt − 2   LURt − p 
+Φ 2   + ⋯ + Φt − p   + at
construction of a very useful class of homogeneous non-  LPPt − 2   LPPt − p 
stationary time series models, the autoregressive  LCCI   LCCI 
integrated moving average models. Some useful  t −2
  t− p

 LPMI t − 2   LPMI t − p 
differencing and variance stabilizing transformations are
introduced to connect the stationary and non-stationary
time series models. where, at are white noise process. E(at) = 0 and:
Many models used in practice are of the simple
ARIMA type, which have a long history and were 0 when t = τ ,
E ( at at′ ) = 
formalized in Box and Jenkins (1970). ARIMA stands Ω when t ≠ τ ,.
for Autoregressive Integrated Moving Average and an
ARIMA(p, d, q) model for an observed series {yt}, t = In the reduced form, we will include a six variable
1 ⋅⋅⋅ T is a model where the dth difference zt = yt-yt-d is VAR with one lag in our forecasting model:
taken to induce stationarity of the series. The process
{zt} is then modeled as zt = µ + єt with:  LLMVt  φ11 φ12 φ13 φ14 φ15 φ16 
   
єt = φ1 єt −1 + φ2 єt − 2 + ⋯ + φ p єt − p + ut − η1 ut −1 − ⋯ − ηq ut − q (5)  LCPI t  φ21 φ22 φ23 φ24 φ25 φ 26 
 LUI t  φ φ32 φ33 φ34 φ35 φ36 
Yt =   Φ1 =  31 
or in terms of polynomials in the lag operator L (defined  LCCI t  φ41 φ42 φ43 φ44 φ45 φ 46 
 LPP  φ φ52 φ53 φ54 φ55 φ56 
through Lsxt = xt-s):  t
  51 
 LPMI t  φ61 φ62 φ63 φ64 φ65 φ66 
φ ( L ) єt = η ( L ) ut (6)
Coefficient φii,i indicates the influence of the iith lag
where, ut is white noise and usually Normally distributed of variable Yi on itself, while coefficient φij,i indicates the
as ut ∼ N(0, σ2). The stationarity and invertibility influence of the iith lag of variable Yj on Yi.
conditions are simply that the roots of φ(L) and η(L), A "VAR in levels" is known as the series modeled
respectively, are outside the unit circle circle (Mark, 1980). are stationary, we forecast them directly by fitting a
VAR to the data. A "VAR in differences" is known as
Vector Auto-Regression Model the series are non-stationary, we firstly take
differences to make them stationary and then we fit a
There are a variety of methods available for
VAR model. In both cases, the models and
forecasting economic variables. One common type of
coefficients are estimated equation by equation using
forecast is Vector Auto-regression modeling for
the principle of least squares.
multivariate Time Series approach. This type of forecast
We applied the VAR selection package for
is predominant in economics and financial analysis.
forecasting the raw data. The function returns
A Vector Auto-Regression (VAR) model is an useful
information criteria and final prediction error for
and flexible approach to describe the dynamic behavior
sequential increasing the lag order up to a VAR(p)-
of economic activity and financial time series dataset; process. which are based on the same sample size. For
that is, a vector of time series. In this system, we each equation, the parameters are estimated by
consider one equation for one variable as dependent minimizing the sum of squared ei,t values.
variable with constant and lags. Each variable is assumed Before we ran the R software, we take the log
to influence with each other in the system, which makes transformation of the raw data to stabilize the
direct interpretation of the estimated coefficients very variance. And then, we set the 80% of the data as
difficult (See Hyndman and Athanasopoulos, 2014). We training set and the remaining data as the test set
write a multi-dimensional VAR(p) as: (Brockwell and Davis, 2002).

255
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

VAR model generate the forecasting in a recursive єt

∼ IID ( 0,1) . (16)
structure. The VAR system require each variable is ht
regressed on a constant and p las of its own lags as well
as on p lags of each of the other variables. To illustrate
where, ht is the conditional variance.
the process, assume that we have fitted the multi-
Especially, the order of the model needs to be
dimensional VAR (1) described in equations Equation
specified for each of the parametric models before fitting
(7) for all observations up to time T.
the model to the in-sample data.
Then the one-step-ahead forecasts are generated by:
The ability to model volatility clustering can be seen
in the definition of the conditional variance where it is
yˆ1,T +1|T = cˆ1 + φˆ11,1 y + φˆ12,1 y (8)
1,T 1,T evident that a large Z t2− i will give rise to a large ht2 :

yˆ 2,T +1|T = cˆ1 + φˆ21,1 y + φˆ22,1 y (9) ht = ζ 0 + ζ 1 Z t2−1 + ζ 2 Z t − 2 + ⋯ + ζ p Z t − p (17)

1,T 2,T

yˆ3,T +1|T = cˆ1 + φˆ31,1 y + φˆ32,1 y (10) Hence the conditional variance of Zt is:
1,T 3,T

Var ( Z t | Z t −1 ,⋯) = E ( Z t2 | Zt2−1 ,⋯) (18)

yˆ 4,T +1|T = cˆ1 + φˆ41,1 y + φˆ42,1 y (11)
1,T 4,T

= h (t ) (19)
yˆ5,T +1|T = cˆ1 + φˆ51,1 y + φˆ52,1 y (12)
1,T 5,T

= ζ 0 + ζ 1 Z t2−1 + ζ 2 Z t − 2 + ⋯ + ζ p Z t − p (20)
yˆ 6,T +1|T = cˆ1 + φˆ61,1 y + φˆ62,1 y (13)
1,T 6,T

We can also state that the current conditional

This is the same form as Equation (8) to Equation variance should also depend on the previous conditional
(13) except that the errors have been set to zero and variances as:
parameters have been replaced with their estimates.
ht = ζ 0 + ς 1 ht −1 + ς 2 ht − 2 + ⋯ + ς p ht − p
AR-GARCH Model (21)
+ζ 1 Z t2−1 + ζ 2 Z t2− 2 + ⋯ + ζ q Z t2− q
In 1982, Engle introduced the Auto-Regressive
Conditional Heteroscedasticity (ARCH) and explicitly In this notation, the error term Zt follow a generalized
recognizes this type of temporal dependence. Moreover, autoregressive conditional heteroskedastic process with
Crawford and Fratantoni (2003) applied a Markov- orders p and q, GARCH(p, q), which is proposed by
switching model to U.S. home price and compare the Bollerslev (2006).
performance with ARIMA and Generalized Although the error term, єt, can be autocorrelated in
Autoregressive Conditional Heteroscedasticity the regression model, it should be stationary. A non-
(GARCH) models. In 1991, Nelson put forward stationary error structure could produce a spurious
Exponential Garch (E-Garch) and Glosten et al. (1993), regression where a significant regression can be achieved
introduced GJR-GARCH model (See Engle, 2001). for a totally unrelated series as shown in Clive and Paul
Besides, Miles (2008) evaluate the Forecasting (1986) and Peter (1986). In such a case, one should
Performance of Linear and Non-linear Models of House properly difference the series before estimating the
Prices. To correlate changes in volatility with changes in regression (Brockwell and Davis, 2002).
returns, Engle et al. (1987) proposed the GARCH-M To fit the AR(m)- GARCH(p, q) model, we consider
model (Miles, 2008). the following formula:
The ARCH model assumes that the changes in
variance is a function of the realizations of squares of  xt = f ( t , xt −1 , xt − 2 ,⋯) + єt
preceding errors. 
To model a time series єt using an ARCH process, we є m
= ∑ k =1 βk єt − k + vt
 t
assume that: 
vt = ht et
 p q
ht = ω + ∑ i =1ηi ht − i + ∑ j =1 λ j vt2− j
ht Z t = єt (14) 

where, f(t, xt-1, xt-2,⋅⋅⋅) is the regression function of xt, et ∼

Var ( єt | past data ) = ht . (15) N(0.1).

256
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

GARCH model is a generalization of ARCH model. The Φijs gives the coefficient relating yit to yj,t-s
In GARCH (p,q), when p = 0, it is ARCH(q) model.
Litterman (1980) assumed that Φ1ii = 1 and all the other
To fit a GARCH model for єt, we must follow.
Φ(ij ) = 0. These 0 and 1 values characterize the mean of
s
A portmanteau Q test is used for autocorrelation in
errors: the prior distribution for the coefficients. Moreover,
Litterman (1980) assumed that:
H0: There is no evidence show that there are
autocorrelation in residuals for some lag p.  γ2  s
H1: There are some evidence show that there are
1 s

 s 
(
Φii( ) ∼ N (1, γ 2 ) Φii( ) ∼ N 1, 2  Φij( ) ∼ N 0,  S ( i, j , l ) 
2
) (25)
autocorrelation in residuals for some lag p.

The investigation of conditional variance models has Although each equation i = 1, 2,..., n of the VAR is
been one of the main areas of study in time series estimated separately, the same value γ is used for each
analysis of financial markets. Towards these ends, the i. Smaller values of γ mean greater confidence in the
GARCH model and its variations have been applied to prior information.
many risk and volatility studies. We use above simple According to Litterman (1986), the prior for the
examples to illustrate the procedures (William, 2006). variance is the only other prior to be set, the standard error
on the coefficient estimate of lag l of variable j in equation
Bayesian VAR Model i is denoted by a standard deviation of the form S(i, j, l):
The Bayesian approach was introduced to a reevaluation
of the VAR approach based on the Bayesian principles. γ g ( l ) f ( i, j )  (σ i )
Thus, the VAR approach was characterized by several S ( i, j , l ) =  (26)
σi
deficiencies, especially due to the over-parameterization
problems. The Bayesian approach proposes a solution to
and:
this problem due to the fact that it does not ponder too much
any of the parameters of the model. However the emphasis
falls on the use of prior distributions for the parameters, the 1 when i = j ,
f ( i, j ) = 
prior distributions being a key factor in the BVAR  wij when i ≠ j ,.
approach. Therefore, the advantages and disadvantages of
VAR model and BVAR were listed in Table 1. σi
When developing the BVAR model, Litterman has where, is correction for the scale for the series i
σj
made some assumptions on the unrestricted VAR model
given by the following equations (Volkan and Gu, 2009): compared with j and 0 < γ < 1. In Equation (26), the
model requires choosing specific values for g(l) (the lag
Yt = µ + Π1 yt −1 + Π 2 yt − 2 + ⋯ + Π p yt − p + єt (22) decay) and γ, the tightness parameter and the standard
deviation on the first own lag, will improve forecasting
performance. Thus, the parameter g(l) measures the
yt − yt −1 = c + єt (23) tightness on lag l with respect to lag 1 and is assumed
to have a harmonic shape with a decay factor of λ. (See
As for example, writing the nth equation in a BVAR Gupta and Sichei, 2006).
model: Litterman (1984a) a found that tight priors around
zero on coefficients of other variables provide better
Yi ,t = ci + Φ1i1 y1,t −1 + Φ1i 2 y2,t −1 + Φ1in yn ,t −1 ⋯ forecast. Choon-Shan and Roy (2004) recommended a
value of λ = 0.7 in concert with γ = 0.9. Kinal and Ratner
+Φi21 y1,t − 2 + Φi22 y1,t − 2 + ⋯ + Φin2 yn ,t − 2 (24)
(1986) used λ = 0.40 and γ = 0.90 (See Choon-Shan and
p p p
+⋯ + Φ y n1 1, t − p + Φ yn , t − p + Φ yn , t − p + є n , t .
i2 in Roy, 2004).

Table 1: Comparison of Bayesian VAR and VAR models

Advantages Disadvantages
Bayesian VAR model Reducing root mean square imposing some Overcome problems with prior distributions
prior restrictions on parameters percent error
Easier and more accurate assessment of uncertainty.
VAR model Fairer assumptions about data interaction of different Over-parameterization the loss of degrees of
related variables in forecasting macroeconomic freedom which exponentially decrease for the
variable number of lags included over-fitting phenomenon

257
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

VECM Model the short-run behavior of the model. It is pointed out

that residual misspecification can arise as a
VECMs are a theoretically-driven approach useful
for estimating both short-term and long-term effects consequence of omitting these important conditioning
of one time series on another. Inherent in the variables and increasing the lag-length is often not the
distinction is the notion of equilibrium. Failure to solution (as it usually is, for example, when
establish co-integration often leads to spurious autocorrelation is present). The procedures for testing
regression problems. the properties of the residuals are discussed and
The VECM(P) form with the cointegration rank r(≤ k) illustrated through examples. We then consider the
is written as: method of testing for "reduced rank", that is, testing
how many co-integration vectors are present in the
∆LLMVt = δ + π LLMVt −1 + Φ∆LLMVt −1 + єt (27) model. At this stage a major issue is confronted which
presents considerable difficulty in applied work,
where, єt is Gaussian random variable and are namely that the reduced rank regression procedure
matrices of parameters estimated using OLS. The provides information on how many unique co-
component π produces different linear combinations integration vectors span the co-integration space, while
of levels of the time series Xt as such the matrix any linear combination of the stationary vectors is itself
LLMVt contains information about the long run also a stationary vector (See Richard, 1995).
properties of the system describe by the model. For
instance, if the rank of the matrix π is 0, then no series Forecasting Accuracy
of the variables can be expressed as a linear
In this paper, the author focus on statistical
combination of the remaining series. This indicates
methodology and forecast macroeconomic variable on
that there does not exists a long run relationship
among the series of the VAR model as a test of co- time series datasets regarding real estate scenario. The
integration a rank of 0 means integration is rejected. Table 2 showed all the potential univariate forecasting
On the other hand, if the rank of the coefficient matrix π is models. However, the Log transformation reduced the
1, or greater than 1 then there exists 1 or more co- Root Mean Square Error (RMSE) significantly, which is
integrating vectors. This indicates a long run relationship almost near to 0.1.
or that the series exhibits significant evidence or behaving We mainly evaluate forecasting models based on
as a co-integrated system (See Kamal, 2014). the two performance measures of RMSE and Akaike
Since the Johansen approach requires a correctly Information Criterion (AIC) for univariate forecasting
specified VECM, it is necessary to ensure that the model. As was the case with the forecast in Table 4, land
residuals in the model are "white noise". This involves market value is projected to continue increase in the
setting the appropriate lag-length in the model and following years. It shows the stable increase in the future.
including (usually dummy) variables that only affect This number will significantly rise to 7.3 in 2018.

Table 2: Comparison Table for univariate forecasting model

Model ME RMSE MAE MPE MAPE AIC
Regression -1.103e-13 1127.403 918.1104 0.6624 21.1810 586.37193
Log Regression -2.7e-17 0.1661 0.1489 -0.0371 1.7665 -12.84073
Regression with -5.2326 590.5238 424.728 -1.1753 7.9002 545.45
AR (2) errors
ETS (M,A,N) -91.4374 617.2248 339.1251 -0.8840 408.4312 413.6146
ARIMA (1,1,0) 88.7736 713.7366 477.83 2.6121 7.973 532.83
Log ARIMA (2,0,1) 0.001 0.063 0.057 0.006 0.692 N/A

Table 3: Comparison Table for multivariate forecasting model

Model VAR (1) BVAR (2) VECM (2) AR(2)-GARCH(1,1)
RMSE 0.08125 0.00797 0.0059 0.139570
AICC -860.1579 -4.5159 -12.8474 -39.04328
HQC 1 -4.49395 -12.8065 -39.031194
AIC - 4.52431 -12.8951 -39.04328
SBC 2 -4.43271 -12.6122 -32.99622
FPEC 3 0.010844 2.514E-6
BIC 4 -805.7328

258
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

The author measured forecasting performance by Table 4: Forecasting VECM (2) Model for LLMV
RMSE and AIC, as the authors did and we will also Year Forecasting Error 95% L B 95% U B
compare performance in terms of Mean absolute Error 2013 8.61155 0.07584 8.46290 8.76019
(MAE), AICC, Bayesian Information Criterion (BIC) 2014 8.66936 0.15104 8.37334 8.96539
2015 8.76250 0.22786 8.31590 9.20910
and so on. The MAE criterion is most appropriate
2016 8.86550 0.30301 8.27161 9.45939
when the cost of a forecast error rises proportionally 2017 8.96023 0.37445 8.22633 9.69413
with respect to the absolute size of the error. With 2018 9.03590 0.44069 8.17217 9.89963
RMSE, the cost of the error rises as the square of the 2019 9.08805 0.50070 8.10670 10.06941
error and so large errors can be weighted far more 2020 9.11704 0.55394 8.03134 10.20275
than proportionally. 2021 9.12638 0.60040 7.94961 10.30315
2022 9.12117 0.64054 7.86573 10.37661
In addition, while RMSE and AIC are good
relative measures, both depend on the scale of the
forecast variable. Moreover, each could hypothetically Acknowledgement
be quite low and still contain systematic bias and do a I would like to thank Dr. Jenchi. Cheng for his
poor job of forecasting average value changes. A penetrating questions and insights during our
given forecasting model may have a systematic conversation and Dr.Gamal Weheba for his suggestions.
positive or negative bias and do a poor job of tracking I also want to thank Dr. Ziqi Sun and Jason Clemens for
the actual mean of value changes and measures such being a special voice during my research.
as RMSE and MAE could well miss this defect. We
will thus evaluate forecasts based on the four Ethics
performance measures of RMSE, AIC and MAE.
In Table 3, we provide the forecasting accuracy for This research paper follows the major requirement of
multivariate forecasting model. But some criterion are ethics, stressing accuracy influence on advanced
missing and not comparable. Eventually, we selected forecasting model. It also meet the required accurate
VECM (2) for forecasting land market value. Because definition of key concepts of ethical reflection such as
the VECM (2) minimize the RMSE and AICC (the principles and norms, values and virtues, rights and duties
small-sample-size corrected version of Akaike and the right perception of the implication of each one.
information criterion).
References
Conclusion Bollerslev, T., 2006. A conditionally heteroskedastic
Land market value, both directly and indirectly, is time series model for speculative prices and rates of
related to the housing market, commercial and return. Rev. Econom. Stat., 69: 542-547.
residential buildings, construction industry, job-hunting DOI: 10.2307/1925546
Brockwell, P.J. and R.A. Davis, 2002. Introduction to
market and home price. Therefore, the improved
Time Series and Forecasting. 1st Edn., Springer,
forecasting promise important benefits for any parties
New York, ISBN-10: 0387953515, pp: 434.
exposed to housing market (Miles, 2008). Litterman, R.B., 1980. A bayesian procedure for
The Table 4 shows the VECM (2) forecasting forecasting with vector autoregression, Working
model for the general tendency of the land market Paper-Massachusetts Institute of Technology,
value for about 10 years. The forecasting results Department of Economics.
shows the stable increase in the future. This number Choon-Shan, L. and A. Roy, 2004. Accuracy of
will increase to 9.12 in 2022. Bayesian VAR in forecasting the economy of
The forecasting in real estate market is more Indiana. Proceedings of the Midwest Business
important and necessary for the economy of Economics Association, (BEA’ 04), pp: 67-71.
American, because the tendency of Land Market Litterman, R.B., 1984a. Forecasting and Policy Analysis
Value would be helpful for government and investor with Bayesian Vector Autoregression Models Federal
to examine the problem in housing market, make the Reserve Bank of Minneapolis Quarterly Review.
appropriate policy and regulate the housing market. Crawford, G.W. and M.C. Fratantoni, 2003. Assessing
On the other hand, forecasting techniques are widely the forecasting performance of regime-switching,
used in the area of finance and housing market. Thus, ARIMA and GARCH models of house prices. Real
a given forecasting model did a good job of tracking Estate Econom., 31: 223-243.
the actual value of land market changes. Most DOI: 10.1111/1540-6229.00064
importantly, the advanced forecasting model can Engle, R., 2001. An introduction to the use of
improve the accuracy and provide a better accurate ARCH/GARCH models in applied econometrics.
guidance and more options for decision-maker. NYU Working Paper No. FIN-01-030.

259
Miss Lei Wang / Journal of Mathematics and Statistics 2018, Volume 14: 253.260
DOI: 10.3844/jmssp.2018.253.260

Hyndman, R.J. and G. Athanasopoulos, 2014. Peter, P., 1986. Understanding spurious regressionin
Forecasting: Principles and Practice. 1st Edn., econometrics. J Econometrics, 33: 311-340.
OTexts, Obtext, ISBN-10: 0987507109, pp: 291. Engle, R., M. Lilien and R.P. Robins, 1987. Estimating
Kamal, R.D., 2014. Estimation of short and long run
time varying risk premia in the term structure: The
equilibrium coefficients in error correction model:
An empirical evidence from Nepal. Int. J. Arch-M model. Econometrica, 55: 391-407.
Econometr. Finan. Manage., 2: 214-219. Gupta, R. and M.M. Sichei, 2006. A BVAR model for
DOI: 10.12691/ijefm-2-6-1 the South African economy. South African J.
Mark, S., 1980. Bayesian time series analysis. University Economics, 74: 391-409.
of Warwick. Kinal, T. and J. Ratner, 1986. A VAR forecasting model
Miles, W., 2008. Boom-bust cycles and the forecasting of a regional economy: Its construction and
performance of linear and non-linear models of
comparative accuracy. Int. Regional Sci. Rev., 10:
house prices. J. Real Estate Finance Econom., 36:
249-264. DOI: 10.1007/s11146-007-9067-1 113-126.
Litterman, R.B., 1986. Forecasting with bayesian vector Richard, H., 1995. Using Cointegration Analysis in
autoregressions-five years of experience. J. Bus. Econometric Modelling. 1st Edn., Harvester
Economic Statistics American Statistical Wheatsheaf, Prentice Hall, ISBN-10: 0133558924,
Association, 4: 25-38. pp: 176.
Box, G. and G. Jenkins, 1970. Time Series Analysis: Volkan, S. and E. Gul, 2009. Usage of different prior
Forecasting. Time Series Analysis: Forecasting and distributions in Bayesian vector autoregressive
Control, Holden-Day, San Francisco.
Clive, G. and N. Paul, 1986. Forecasting economic time models. J. Math. Stat., 38: 85-93.
series. Harcourt Brace Jovanovich, New York. William, W., 2006. Time Series Analysis: Univariate
Glosten, L.R., R. Jagannathan and D.E. Runkle, 1993. and Multivariate Methods. 2nd Edn., Pearson
On the relation between the expected value and the Addison Wesley, Boston, ISBN-10: 0321322169,
volatility of the nominal excess return on stocks. J. pp: 614.
Finance, 48: 1779-1801.

Appendixa: 1982-2015 Land Market Value Datasetsb

Year LMVc CPI GDPd IR UR CCI PP PMI Year LMV CPI GDP IR UR CCI PP PMI
1982 1274.88 96.5 6.49 6.2 9.7 43.4 231.66 42.8 2000 4509.19 172.2 12.68 3.4 4 75.9 282.16 43.9
1983 1232.25 99.6 7 3.2 9.6 44.70 233.79 69.9 2001 5428.39 177.1 12.71 2.8 4.7 79.7 284.97 45.3
1984 1387.16 103.9 7.4 4.3 7.5 46.7 235.82 50.6 2002 6123.09 179.9 12.96 1.6 5.8 81.7 287.63 51.6
1985 1546.45 107.6 7.71 3.6 7.2 47.9 237.92 50.7 2003 7208.82 184.13 13.53 2.3 6 85.9 290.11 60.1
1986 1879.09 109.6 7.94 1.9 7 50.4 240.13 50.5 2004 8646.18 188.9 13.95 2.7 5.5 93.1 292.81 57.2
1987 2297.13 113.6 8.29 3.6 6.2 2.7 242.29 61 2005 10708.93 195.3 14.37 3.4 5.1 100 295.52 55.1
1988 2678.79 118.3 8.61 4.1 5.5 54.5 244.50 56 2006 12547.31 201.6 14.72 3.2 4.6 106 298.38 51.4
1989 3097.56 124.8 8.85 4.8 5.3 56.4 246.82 47.4 2007 12290.28 207.3 14.99 2.8 4.6 107 301.23 49
1990 3257.63 130.7 8.91 5.4 5.6 58 249.62 40.8 2008 10464.64 215.3 14.58 3.82 5.8 103.3 304.09 33.1
1991 3050.34 136.2 9.02 4.2 6.8 58.2 252.98 46.8 2009 7537.82 214.5 14.54 -0.32 9.3 98.10 306.77 55.3
1992 3089.8 140.3 9.41 3 7.5 58.9 256.51 54.2 2010 7173.83 218.1 14.94 1.64 9.6 96.4 309.35 57.5
1993 2948.23 114.5 9.65 3 6.9 61.8 259.92 55.6 2011 6184.28 224.9 15.19 3.14 8.9 97.4 311.72 53.1
1994 2995.76 148.2 10.05 2.6 6.1 64.6 263.13 56.1 2012 5543.56 229.6 15.43 2.08 8.1 98.4 314.11 50.4
1995 2945.05 152.4 10.28 2.8 5.6 67.3 266.28 46.2 2013 6777.04 233 15.92 1.46 7.4 104.8 316.5 56.5
1996 3033.87 156.9 10.74 3 5.4 68.6 269.39 55.2 2014 8152 237.2 16.29 1.61 6.2 111.8 318.86 55.1
1997 3120.62 160.5 11.21 2.3 4.9 70.6 272.65 54.5 2015 8737.11 242.1 16.3 0.1 5.5 100.37 320.99 53.5
1998 3437.02 163.11 11.77 1.6 4.5 72.5 275.85 46.8
1999 3886.17 166.6 12.32 2.2 4.2 72.7 279.04 57.8
a
The data was based on the 34 years' national data on past and present real estate transaction from 1982 to 2015.
b
http://www.statista.com/statistics/188105/annual-gdp-of-the-united-states-since-1990/
Source: U.S. Bureau of Labor Statistics https://en.wikipedia.org/wiki/Main-Page.
c
The unit of land market value is million
d
The unit of GDP is trillion.

260