J Syst Sci Complex

A Demand Forecasting Method Based on Stochastic

Frontier Analysis and Model Average: An Application in
Air Travel Demand Forecasting∗
ZHANG Xinyu · ZHENG Yafei · WANG Shouyang

DOI: 10.1007/s11424-018-7093-0
Received: 24 March 2017 / Revised: 10 January 2018
The
c Editorial Office of JSSC & Springer-Verlag GmbH Germany 2018

Abstract Demand forecasting is often difficult due to the unobservability of the applicable historical
demand series. In this study, the authors propose a demand forecasting method based on stochastic
frontier analysis (SFA) models and a model average technique. First, considering model uncertainty,
a set of alternative SFA models with various combinations of explanatory variables and distribution
assumptions are constructed to estimate demands. Second, an average estimate from the estimated
demand values is obtained using a model average technique. Finally, future demand forecasts are
achieved, with the average estimates used as historical observations. An empirical application of air
travel demand forecasting is implemented. The results of a forecasting performance comparison show
that in addition to its ability to estimate demand, the proposed method outperforms other common
methods in terms of forecasting passenger traffic.
Keywords Air travel demand, demand forecasting, model average, model uncertainty, stochastic
frontier analysis.

1 Introduction
Demand forecasting has long been investigated in the related literature, and it is still a
rapidly evolving research field. Researchers are now seeking to ‘exploit the availability of ever
Zhang Xinyu
Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China; Center
for Forecasting Science, Chinese Academy of Sciences, Beijing 100190, China.
ZHENG Yafei
Postdoctoral Working Station, Shenwan Hongyuan Securities Co., Ltd., Shanghai 200031, China; Academy of
Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China.
WANG Shouyang
Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing 100190, China; Center
for Forecasting Science, Chinese Academy of Sciences, Beijing 100190, China.
∗ This research was supported by the National Natural Science Foundation of China under Grant Nos. 71522004,

11471324 and 71631008 and a Grant from the Ministry of Education of China under Grant No. 17YJC910011.
 This paper was recommended for publication by Editor HUANG Haijun.
2 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

more plentiful data observations, greater computational power, and increasingly sophisticated
economic and statistical theory’[1] . Common methods applied in the demand forecasting in-
clude statistical methods such as linear multivariate regression models[2–6] , univariate time
series models[7–10] , and artificial intelligence methods such as neural network[11–13] , and sup-
port vector regression models[14–17] . Additionally, forecasting combination techniques are also
widely applied[18–24] .
In most of the existing literature, for the demand modeling and forecasting purposes, the
historical consumption quantities of a product or service are taken as proxies of the unob-
servable demands. However, the consumption quantity is mostly not equal to the quantity
that consumers actually demand. In addition to demand, consumers’ behavior can be con-
strained by other factors, such as the supply shortages and poor logistics services. We name
the effects caused by these factors as inefficiency of the sellers in this study. Unconstrained
demand estimating is an important issue, and has been discussed mostly in hotel and airline
revenue management[25] by using either reservation denial data[26] or various mathematical
models[27, 28] . The current study addresses the unconstrained demand forecasting problem,
and in contrast to previous studies, we focus on demand at a more macro level, i.e., the na-
tional demand. Although the existing forecasting methods have been more sophisticated and
their forecasting accuracy has been markedly improved, the forecasting issue for this type of
unconstrained demand is still ignored by most researchers.
The main purpose of this study is to develop a method that can estimate and forecast
the unconstrained demand, i.e., the ‘true demand’, properly and scientifically. To achieve
this goal, we propose a demand forecasting method based on statistic frontier analysis (SFA)
models and a model average technique. The SFA is a parametric method that was developed
to estimate efficient frontier and efficiency scores. It has been widely applied in efficiency
assessment problems[29–31] . Thus, the application of SFA in demand forecasting is one of the
contributions of the current study. Additionally, considering uncertainty in the selection of SFA
models, a cross-validation type model averaging is applied to improve the forecasting accuracy.
Model averaging, i.e., setting weights (which have values between 0 and 1, and sum to 1)
to candidate models in some way, can prevent us from putting all of our inferential eggs in
one unevenly woven basket[32] . Model averaging often reduces the risk in regression estimation,
as ‘betting’ on multiple models provides a type of insurance against a poor singly selected
model[33] . Bayesian model averaging has long been a popular statistical technique; see [34]
for a comprehensive review. Recently, frequentist model averaging has also been garnering
interest; see, for example, [35–38]. In the current paper, we follow the idea in [38] and use
a cross-validation type model averaging to combine demand forecasts based on various SFA
models.
In the empirical application section, we aim to forecast Chinese national air travel demand.
For the nationwide air travel demand, linear multivariate regression models are widely used
in [39–41]. Other than regression models, Holt–Winters forecasting models were also applied in
several studies; see, for example, [42] and [43]. These two studies were both based on a long time
series, which could demonstrate the whole historical trend of the demand. Other studies on
 A DEMAND FORECASTING METHOD 3

nationwide air travel demand forecasting include [13] where a hybrid approach based on a back-
propagation neural network was utilized, and [44] where a gray combination forecasting model
based on GM (1, 1) for the long-term forecasting of nationwide civil air traffic was adopted.
However, to the best of our knowledge, all those studies used historical passenger traffic as the
proxy of the air travel demand and actually forecasted passenger traffic. In contrast, by using
the proposed method in this paper, we can estimate and forecast the unconstrained air travel
demand based on the historical passenger traffic and some explanatory variables.
The plan of this study is as follows. Section 2 introduces our proposed demand forecasting
methodology. Section 3 presents an empirical application on air travel demand forecasting and
compares the proposed method with other widely used methods. Section 4 concludes this study.

2 Methodology
2.1 Stochastic Frontier Analysis Models
The standard parametric SFA was proposed independently by [45] and [46]. The general
setup is
yi = xT
i β − ui + vi , i = 1, 2, · · · , n, (1)

where xi is the exogenous explanatory variables vector, β is the coefficient vector, yi is the
response variable, ui captures inefficiency of the producer, and vi captures outside influences
beyond the control of the producer. It is assumed that ui and vi are independent of one other.
To estimate parameters in Equation (1), some distribution assumptions on both ui and vi
are often imposed. Then, all parameters in the model can be estimated using the maximum
likelihood estimation method. Typically, vi is assumed to be normally distributed with a
mean 0 and variance σv2 , and the most common distributions for ui are the truncated normal
distribution Normal+ (0, σu2 ) and exponential distribution Exp(σu ).
In the current paper, we propose a demand forecasting method based on SFA models, using
observable consumption quantities and demand explanatory variables. Following the general
setup of SFA model in Equation (1), we assume that

Ci = Di − ui , Di = f (xi ) + vi , (2)

where Ci denotes actual consumption quantity, Di denotes the unconstrained demand, f (xi )
is an unknown function, and xi contains exogenous explanatory variables that determine the
demand. Thus, our proposed demand estimation SFA model is as follows:

Ci = f (xi ) − ui + vi . (3)

2.2 Model Average

Assume that we have a total of S candidate SFA models and that the sth candidate model
is
Ci = xT
(s),i β(s) − u(s),i + v(s),i , (4)
4 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

where s ∈ {1, 2, · · · , S}. We assume S be unrelated to sample size n. We use a simple linear
function to approximate the unknown function f (xi ) in Equation (2). Because it is uncer-
tain which variables should be included in Equation (4), we allow different variable combina-
tions x(s) to be used in different candidate models. For the sth candidate model, we assume
2 + 2
v(s),i ∼ Normal(s) (0, σv,(s) ) and u(s),i ∼ D(s) (μ(s) , σu,(s) ) (note that these assumptions can be
misspecified), using which we can calculate E(u(s),i ), which is denoted by δ(s) (μ(s) , σu,(s) ). The
+
distribution D(s) is chosen between the truncated normal and exponential distributions. For
the truncated normal distribution D(s) + 2
(μ(s) , σu,(s) ) = Normal+ (μ, σu,(s)
2
), we have

φ(μ(s) /σu,(s) )
δ(s) (μ(s) , σu,(s) ) = μ(s) + σu,(s) . (5)
Φ(μ(s) /σu,(s) )
+ 2
For the exponential distribution D(s) (μ(s) , σu,(s) ) = Exp(σu,(s) ), we have
−1
δ(s) (μ(s) , σu,(s) ) = σu,(s) . (6)

Under the sth candidate model, the unknown parameters β(s) , μ(s) , σu,(s) and σv,(s) can be
estimated by maximum likelihood estimation. Let β(s) , μ (s) , σ
u,(s) and σ
v,(s) be the maximum
likelihood estimators of β(s) , μ(s) , σu,(s) and σv,(s) respectively. Then, for a new observation
xn+1 , the forecast of demand is
 n+1,(s) = xT
D 
n+1,(s) β(s) . (7)

The forecast of consumption quantity is

n+1,(s) = D
C  n+1,(s) − δ(s) (
μ(s) , σ
u,(s) ). (8)

Write w = (w1 , w2 , · · · , wS )T as the weight vector for model averaging, belonging to set W =
S
{w ∈ [0, 1]S : s=1 ws = 1}. Then, the model average forecasts of demand and consumption
quantity are as follows:
S

 n+1 (w)
D =  n+1,(s) ,
ws D (9)
s=1
S

n+1 (w)
C = n+1,(s) .
ws C (10)
s=1

To choose the weight vector w in (9) and (10), we propose a cross validation (CV) method.
 i,(s) = xT β(s) , C
Let D i,(s) = D
 i,(s) − δ(s) (
μ(s) , σ  i (w) = S ws D
u,(s) ), D i (w) =
 i,(s) , C
i,(s) s=1
S
ws C i,(s) and C  (−i)
(w) be the average estimator of Ci without using the ith observation
s=1 i
{Ci , xi,(1) , xi,(2) , · · · , xi,(S) }. Thus, our CV criterion is formulated as
n

CV(w) = 
{C
(−i)
(w) − Ci }2 . (11)
i
i=1

The resulting weight vector is

w
 = arg min CV(w). (12)
w∈W
 A DEMAND FORECASTING METHOD 5

 n+1 (w)
The resulting model average forecasts for demand and consumption quantity are D 

and Cn+1 (w),
 respectively. It is straightforward to obtain that
1 n+1 (w) − Cn+1 }2 .
CV(w) ≈ E{C (13)
n
n+1 (w) − Cn+1 }2 is also
Thus, by minimizing the CV criterion, the squared forecasting risk E{C
minimized approximately.
The following theorem shows that the squared forecasting risk is indeed minimized in large
sample sense. Let the squared forecasting risk functions of consumption and demand be, re-
spectively,
n+1 (w) − Cn+1 }2
riskC (w) = E{C  n+1 (w) − Dn+1 }2 .
and riskD (w) = E{D

Theorem 2.1 If Assumptions C.1–C.5 in Appendix A.1 are satisfied, then

riskC (w)

→1 (14)
inf w∈W riskC (w)
in probability, as n → ∞. If Assumption C.6 in Appendix A.1 is further satisfied, then
riskD (w)

→1 (15)
inf w∈W riskD (w)
in probability, as n → ∞.
Theorem 2.1 states that our model average consumption quantity and demand forecasts

Cn+1 (w)
 and D  n+1 (w)
 are asymptotically identical to those obtained by using the infeasible
best possible model average consumption quantity and demand forecasts, which means that our
method is asymptotically optimal. The detailed assumptions and the proof of Theorem 2.1 are
given in Appendix.

2.3 Demand Forecasting Process

Figure 1 demonstrates our proposed demand forecasting framework based on the SFA models
and the model averaging technique described in the previous section. There are four main steps
in this demand forecasting framework.
Step 1 Raw data collection. The main determinants and drivers that have significant influ-
ences on demand are selected according to the experience of previous studies. Hence, historical
records of both consumption quantities and demand explanatory variables are collected in this
step.
Step 2 Model construction. Considering the model uncertainty, a set of alternative SFA
models with different combinations of explanatory variables and different distribution assump-
tions about ui are constructed.
Step 3 Demand estimation. With the model averaging technique, the average demand
estimates are calculated based on the constructed SFA models from Step 2 and Equation (9).
Step 4 Demand forecasting. With the average demand estimates as historical observations,
the future demand forecasts can be achieved with standard forecasting models. In this study,
the univariate time series models are applied to implement the future demand extrapolation.
6 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

Figure 1 Our proposed demand forecasting framework

3 Empirical Application
In this section, we provide an empirical application of our proposed demand forecasting
method to an air travel demand forecasting problem. The main objective is to forecast the
national air travel demand of mainland China in the medium-term future (3–5 years). Hence,
according to the framework shown in Figure 1, we first select the main explanatory variables
for air travel demand modeling. Second, we construct all alternative SFA models with differ-
ent explanatory variable combinations and two different distribution assumptions, including
the truncated normal distribution and the exponential distribution, of the inefficiency term in
Equation (4). Then, the unconstrained demand series is generated by the CV-type model aver-
aging technique. Finally, with the estimated demand series, future air travel demand forecasts
 A DEMAND FORECASTING METHOD 7

are achieved by the standard univariate time series models.

3.1 Determinants of Air Travel Demand

Before implementing the demand forecasting, we provide a comprehensive review for the
determinants of air travel demand.
Many studies have summarized the determinants of air travel demand for their specific re-
search purposes. For example, [47] pointed out that the demand of airline transportation was
generally dependent on two main groups of drivers. The first group contained geo-economic
factors, which included variables about the economic activities and geographical characteristics
of the location where the air transportation occur. The second group contained service-related
factors, which mainly included the quality and price of airline products. Based on previous
studies, [48] classified the determinants of airline transportation demand into five groups, in-
cluding the economic activity factors (e.g., population, GDP per capita), the locational factors
(e.g., distance, tourist market), the price factors (e.g., airfare), the quality-of-service factors
(e.g., aircraft size, flight frequency) and the market factors (e.g., year, hubbing strategy).
However, [49] proposed a different division of air travel demand determinants, which were
economic factors, geographic and demographic factors, market structure, and social factors and
maturity. He claimed that each group affected air travel demand in a different way. The eco-
nomic factors, such as GDP, expenditure, airfares and inflation, were the most influential drivers.
Geographic and demographic factors mainly included population, country size and urbanization
level. Market structure meant that whether the market was regulated or deregulated and what
the airlines business model was. Social factors include variables that represented the public
psychological perception, such as the level of education and urbanization. These social factors
have rarely been discussed in air travel demand forecasting literature, but they are believed to
be significant promoters for people’s desire to travel[50] and to directly influence the national air
travel demand. The final factor group, maturity, was a phase of the air transportation market
where the air travel demand became less sensitive to the income changes.
We summarize and reorganize the determinants for nationwide air travel demand in Figure 2.
In this empirical application, we focus mainly on the macroscopic demand, which represents
the air travel demand of a country, a region or a city. The macroscopic demand can also be
considered the aggregate of city-pair air travel demands.

3.2 Data
According to the determinants summarized in Figure 2, we use yearly passenger traffic of
mainland China as the dependent variable as well as several explanatory variables. All variables
are defined and summarized in Table 1.
8 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

Economic Factors Geographic/Demographic Factors

Social Factors

Others
Market Structure

Figure 2 Determinants of air travel demand

Table 1 Summary of variables

Variable Definition Category Data Source Related Studies

Dependent variable
Yearly number of air passengers in Civil Aviation
Passenger Traffic (P) mainland China (Million person) Administration
of China
Explanatory variables
Yearly gross domestic product of Economic World Bank E.g., [44, 51, 52]
GDP China (Billions of US dollars) factor
Trade value (TV) Yearly total trade value of Economic UNCTAD STAT E.g., [53, 54]
China (Billions of US dollars) factor
Oil price (OP) The price of crude oil Economic WIND Database E.g., [49]
(Ten dollars per barrel) factor
Urban population (UP) Percentage of Chinese urban Demographic/ World Bank E.g., [50]
population to total population (%) Social factor

The yearly historical passenger traffic collected in our application is during the period of
year 1978 to 2014. Figure 3 shows a general tendency of sustained growth in the Chinese
air transportation market. Over the past several decades, most of the yearly growth rate was
 A DEMAND FORECASTING METHOD 9

between 0% and 30%, and the average growth rate of the last 5 years was approximately 11.2%.
In addition, Figure 4 shows the historical trends of all explanatory variables. In the subsequent
modeling processes, all value–related variables in Table 1 are recalculated in US dollars at the
constant prices (2005) to eliminate the effects of inflation. All aggregative variables, i.e., gross
domestic product and trade value, are used on a per capita basis and marked as GDPpc and
TVpc , respectively.
500 50
Yearly passenger traffic (Million person)
Growth rate (%)
450
40

400

30
350
Passenger traffic (Million person)

300

Growth rate (%)

20

250

10
200

150
0

100

−10
50

0 −20
1980 1985 1990 1995 2000 2005 2010 2015

Figure 3 Historical passenger traffic and its growth rate

In this study, we include the explanatory variable ‘urban population’, to replace the widely
used ‘total population’[54] . In addition to reflecting the tendency of population size, the ‘urban
population’ can also reflect the national urbanization degree, which has been proven to be
a significant determinant of air travel demand. Table 2 lists the estimation results for these
explanatory variables from a linear regression model.
Table 2 shows that all of the explanatory variables selected in this study are significant for
explaining the passenger traffic, where the ‘GDP per capita’ and ‘trade value per capita’, as
expected, have positive influences on passenger traffic. These variables represent the level of
the economic activities and people’s living standards. In the present stage of China’s economic
development, people will fly relatively more when their living standards are getting higher,
and the variable ‘urban population’, which represents the nation’s urbanization degree, is also
proven to be a positive determinant for the passenger traffic. In contrast, ‘oil price’, which is a
proxy for the travelling cost, is a significantly negative determinant, as expected.
10 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

G ros s domes tic product per ca pita (Thousand dollars ) T rade value per capita (Hundred dollars)
7 30

6 25

5
20
4
15
3
10
2
5
1

0 0
1980 1985 1990 1995 2000 2005 2010 1980 1985 1990 1995 2000 2005 2010

Urban population (%) Real oil price (Ten dollars )

60 10

50 8

40 6

30 4

20 2

10 0
1980 1985 1990 1995 2000 2005 2010 1980 1985 1990 1995 2000 2005 2010

Figure 4 Historical trends of explanatory variables

Table 2 The estimation results of a linear regression model

Dependent Variable: P
Sample period: 1978–2014
Variable Coefficient Standard Error Prob.
GDPpc 33.152 4.2799 0
TVpc 7.1258 0.8736 0
UP 0.3155 0.0964 0.0025
OP −3.826 0.5788 0
R-squared 0.9949 Adjusted R-squared 0.9945

3.3 Method Evaluation and Comparison

In this empirical application, the forecasting performance evaluation and the performance
comparisons of various methods are based primarily on the following two frequently used error
measures, i.e., root mean squared error (RMSE) and mean absolute percentage error (MAPE),
which are presented as follows,

 N  
1  N
1   Pi − Pi 
  2
(Pi − Pi ) , MAPE =
RMSE =
N N  Pi  ∗ 100,
i=1 i=1
 A DEMAND FORECASTING METHOD 11

where Pi is the forecast of the passenger traffic and N is the number of observations in the
testing dataset.
To assess the forecasting performance of our proposed method, named as MA-SFA below by
combining the abbreviations of model averaging (MA) and statistical frontier analysis (SFA),
we compare it with several other common forecasting methods in terms of their passenger traffic
forecasting performances. According to the literature review, three widely applied methods,
i.e., linear multivariate regression models (Regression), autoregressive moving average (ARMA)
models and back-propagation neural network (BPNN) algorithms, are constructed for the per-
formance comparisons. The ARMA model and BPNN algorithm are applied on the single
passenger traffic series, and the Regression model is built with all explanatory variables listed
in Table 2. In this study, the regression model and ARMA model are constructed with the
Eviews software, and the BPNN and MA-SFA models are built by the Matlab software. The
original sample data are separated into two parts; the first 32 observations are used for the
model construction, and the remaining 5 observations are used for the model evaluation.

3.4 Results and Discussions

The estimation results of the MA-SFA method are summarized in Table 3, where we list
the top four SFA models according to the weights for each candidate SFA model. Note that
with the four explanatory variables and the two distribution assumptions for ui , we have (C41 +
C42 + C43 + C44 ) ∗ 2 = 30 candidate SFA models to calculate the average value, although most
of them have lower weights than the SFA(4) model in Table 3. Thus, we only list the top four
for reference, and the remaining models are available upon request.

Table 3 The main SFA models and the corresponding estimated coefficients
SFA(1) SFA(2) SFA(3) SFA(4) Averaged
Weights 0.5561 0.2481 0.1897 0.0058
GDPpc 77.7218 36.3557 78.0373 26.6449
TVp c 12.0937 6.8652 8.0281
UP 0.6362 0.338 0.3416 0.0308 0.5027
OP −3.2051 −3.1626 −2.5721 −3.0543
δ(s) (
μ(s) , σ
u,(s) ) 0.9006 16.3616 9.0356 20.3535 6.3936

The distributions of ui in the four SFA models in Table 3 all happen to be the truncated
normal one, and the final averaged SFA model for air travel demand estimation, where the
coefficients of each explanatory variables represent an average effect on air travel demand, can
be described as follows,

P = 26.6449 ∗ GDPpc + 8.0281 ∗ TVpc + 0.5027 ∗ UP − 3.0543 ∗ OP − 6.3936, (16)

where 6.3936 is the average uncovered air travel demand in mainland China, i.e., the inefficiency
of the air travel market.
12 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

Equation (16) shows the average effect of the four key determinants on air travel demand.
As expected, the GDP per capita and the trade value per capita both have positive influences
on the air travel demand. The estimated coefficient indicates that when the GDP per capita of
China increases by one unit, the air travel demand will increase by approximately 26 units, and
when the percentage of urban population increases by 1 unit, the air travel demand will increase
by approximately 0.5 unit. As expected, oil price, which is generally considered a proxy for the
flying costs of the passengers, has a negative impact on the air travel demand. Our method
does not provide standard errors for the estimated coefficients because it is very challenging, if
not impossible, to study the distribution theory of our model average estimate.
Figure 5 shows the actual passenger traffic, the air travel demand estimated by our proposed
MA-SFA method with sample data from the years 1978–2014, and the passenger traffic forecasts
for years 2010–2014 obtained from all the four methods. The MAPE and RMSE of the various
methods are also listed in Table 4.
500
Actual passenger traffic
Estimated passenger demand
450
Forecasts by Regression
Forecasts by BPNN
400 Forecasts by ARMA
Forecasts by MA−SFA

350

300
Million person

250

200

150

100

50

0
1980 1985 1990 1995 2000 2005 2010 2011 2012 2013 2014 2015
Year

Figure 5 The actual passenger traffic and estimated demand series

Table 4 Performance comparison

Performance
Methods
MAPE (%) RMSE

N =1 N =3 N =5 N=1 N =3 N =5
ARMA 0.1962 5.7741 10.8953 0.5251 23.473 48.3275
Regression 1.4963 5.5916 4.4099 4.0055 19.1907 16.9211
BPNN 3.877 5.4613 12.0241 10.3783 18.7301 54.6355
MA-SFA 1.4484 2.9005 2.8175 3.8773 9.4983 11.8885

Note: N is the number of observations used for model evaluation, and also denotes the forecasting
time horizons; The calculation of MAPE and RMSE is introduced in Subsection 3.3; And the smallest
 A DEMAND FORECASTING METHOD 13

value of each column is bold.

From the performance comparison results given in Table 4 and Figure 5, we can conclude
that the forecasting performance of our proposed method is very promising. Several conclusions
can be drawn from the above comparisons as follows.
1) In terms of both MAPE and RMSE, the MA-SFA method performs best in the most
forecasting time horizons, except for the 1-year forecast. These results suggest that although
ARMA models are suitable for the short-term forecasting, they tend to generate larger forecast-
ing errors as the forecasting time horizon becomes longer. Thus, we could get a conclusion that
the proposed MA–SFA method is appropriate for a medium–term demand forecasting, which
is exactly the main forecasting objective in this empirical application.
2) Compared to the ARMA and BPNN models, the multivariate models constructed in this
study, i.e., the regression model and MA-SFA method, perform better in the longer forecasting
time horizons, such as a 5-year forecasting. The main reason for this performance improvement
is that the other explanatory variables can usually provide additional information for demand
forecasting, especially in a longer forecasting time horizon.
3) By comparing the forecasting performance of the MA-SFA method and the linear regres-
sion model, there is a significant improvement in both MAPE and RMSE when applying the
MA-SFA method. This result confirms the performance improvement that is obtained when
adopting the model averaging technique and SFA models.
To sum up, on one hand, the proposed method, i.e., MA-SFA, can estimate the uncon-
strained air travel demand (the red dashed line shown in Figure 5), which is the main contri-
bution of this study. On the other hand, the proposed method also outperforms other common
demand forecasting methods in forecasting passenger traffic.
Finally, adopting a standard ARIMA models for the estimated demand series, we forecast
the Chinese air travel demand in year 2017, 2018 and 2019 as 486.58, 529.50 and 578.11 million
passengers, respectively.

4 Conclusions
In the existing demand forecasting literature, the conventional way for quantifying the
unobservable demand is to take the actual consumption quantity as a proxy. This will usually
lead to a substantial forecasting error. In this study, to deal with this problem, we propose a
new method that models the demand with the SFA model. Simultaneously, a model averaging
technique is utilized to consider the model uncertainty and improve the forecasting accuracy.
We further provide an empirical application for the Chinese national air travel demand fore-
casting. With our proposed method, the unconstrained demand series is estimated by utilizing
the historical passenger traffic and several explanatory variables of the air travel demand. An
ARIMA model is then constructed for modeling the estimated demand series to achieve the de-
mand forecasts over the next 3–5 years. By comparing its forecasting performance with those
of other common forecasting methods, the performance of our method is obviously promising.
14 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

References

[1] Clements M P and Hendry D F, The Oxford Handbook of Economic Forecasting, Oxford University
Press, Oxford, 2011.
[2] Kulendran N and Wilson K, Is there a relationship between international trade and international
travel? Applied Economics, 2000, 32(8): 1001–1009.
[3] Lim C and McAleer M, Cointegration analysis of quarterly tourism demand by Hong Kong and
Singapore for Australia, Applied Economics, 2001, 33(12): 1599–1619.
[4] Song H and Witt S F, Forecasting international tourist flows to Macau, Tourism Management,
2006, 27(2): 214–224.
[5] Koen R and Holloway J, Application of multiple regression analysis to forecasting South Africa’s
electricity demand, Journal of Energy in Southern Africa, 2014, 25(4): 48–58.
[6] McNown R, Aburizaizah O S, Howe C, et al., Forecasting annual water demands dominated
by seasonal variations: The case of water demands in Mecca, Applied Economics, 2005, 47(6):
544–552.
[7] Cho V, Tourism forecasting and its relationship with leading economic indicators, Journal of
Hospitality and Tourism Research, 2001, 25(4): 399–420.
[8] Goh C and Law R, Modeling and forecasting tourism demand for arrivals with stochastic non-
stationary seasonality and intervention, Tourism Management, 2002, 23(5): 499–510.
[9] Gil-Alana L A, Modelling international monthly arrivals using seasonal univariate long-memory
processes, Tourism Management, 2005, 26(6): 867–878.
[10] Ediger V Ş and Akar S, ARIMA forecasting of primary energy demand by fuel in Turkey, Energy
Policy, 2007, 35(3): 1701–1708.
[11] Cho V, A comparison of three different approaches to tourist arrival forecasting, Tourism Man-
agement, 2003, 24(3): 323–330.
[12] Kon S C and Turner L W, Neural network forecasting of tourism demand, Tourism Economics,
2005, 11(3): 301–328.
[13] Alekseev K P G and Seixas J M, A multivariate neural forecasting modeling for air transport —
Preprocessed by decomposition: A Brazilian application, Journal of Air Transport Management,
2009, 15(5): 212–216.
[14] Tay F E and Cao L, Application of support vector machines in financial time series forecasting,
Omega, 2001, 29(4): 309–317.
[15] Pai P F and Hong W C, Forecasting regional electricity load based on recurrent support vector
machines with genetic algorithms, Electric Power Systems Research, 2005, 74(3): 417–425.
[16] Chen K Y and Wang C H, Support vector regression with genetic algorithms in forecasting
tourism demand, Tourism Management, 2007, 28(1): 215–226.
[17] Xie G, Wang S Y, and Lai K K, Short–term forecasting of air passenger by using hybrid seasonal
decomposition and least squares support vector regression approaches, Journal of Air Transport
Management, 2014, 37: 20–26.
[18] Armstrong J S, Principles of Forecasting: A Handbook for Researchers and Practitioners, Springer
Science and Business Media, New York, 2001.
[19] Oh C O and Morzuch B J, Evaluating time-series models to forecast the demand for tourism in
 A DEMAND FORECASTING METHOD 15

Singapore: Comparing within-sample and post-sample results, Journal of Travel Research, 2005,
43(4): 404–413.
[20] Karlaftis M G, Demand forecasting in regional airports: Dynamic Tobit models with GARCH
errors, Sitraer, 2008, 7: 100–111.
[21] Yu L A, Wang S Y, and Lai K K, Forecasting China’s foreign trade volume with a kernel-based
hybrid econometric-AI ensemble learning approach, Journal of Systems Science and Complexity,
2008, 21(1): 1–19.
[22] Xiao Y, Xiao J, Liu J, et al., A multiscale modeling approach incorporating ARIMA and anns for
financial market volatility forecasting, Journal of Systems Science and Complexity, 2014, 27(1):
225–236.
[23] Xiao Y, Liu J J, Hu Y, et al., A neuro-fuzzy combination model based on singular spectrum
analysis for air transport demand forecasting, Journal of Air Transport Management, 2014, 39:
1–11.
[24] Huang A Q, Lai K K, Li Y H, et al., Forecasting container throughput of Qingdao port with a
hybrid model, Journal of Systems Science and Complexity, 2015, 28(1): 105–121.
[25] Weatherford L R and Kimes S E, A comparison of forecasting methods for hotel revenue man-
agement, International Journal of Forecasting, 2003, 19(3): 401–415.
[26] Orkin E B, Wishful thinking and rocket science: The essential matter of calculating unconstrained
demand for revenue management, The Cornell Hotel and Restaurant Administration Quarterly,
1998, 39(4): 15–19.
[27] Wickham R R, Evaluation of forecasting techniques for short-term demand of air transportation,
Department of Aeronautics and Astronautics, Flight Transportation Laboratory, Massachusetts
Institute of Technology, Cambridge, 1995.
[28] Weatherford L R and Pölt S, Better unconstraining of airline demand data in revenue management
systems for improved forecast accuracy and greater revenues, Journal of Revenue and Pricing
Management, 2002, 1(3): 234–254.
[29] Battese G E and Coelli T J, A model for technical inefficiency effects in a stochastic frontier
production function for panel data, Empirical Economics, 1995, 20(2): 325–332.
[30] Jacobs R, Alternative methods to examine hospital efficiency: Data envelopment analysis and
stochastic frontier analysis, Health Care Management Science, 2001, 4(2): 103–115.
[31] Cullinane K, Wang T F, Song D W, et al., The technical efficiency of container ports: Compar-
ing data envelopment analysis and stochastic frontier analysis, Transportation Research Part A:
Policy and Practice, 2006, 40(4): 354–374.
[32] Longford N T, Editorial: Model selection and efficiency is ‘Which model· · · ?’ the right question?
Journal of the Royal Statistical Society A, 2005, 168: 469–472.
[33] Leung G and Barron A R, Infromation theory and mixing least squares regressions, IEEE Trans-
actions on Information Theory, 2006, 52: 3396–3410.
[34] Hoeting J A, Madigan D, Raftery A E, et al., Bayesian model averaging: A tutorial, Statistical
Science, 1999, 14: 382–417.
[35] Buckland S T, Burnham K P, and Augustin N H, Model selection: An integral part of inference,
Biometrics, 1997, 53: 603–618.
[36] Hansen B E, Least squares model averaging, Econometrica, 2007, 75: 1175–1189.
[37] Zhang X and Liang H, Focused information criterion and model averaging for generalized additive
partial linear models, Annals of Statistics, 2011, 39: 174–200.
16 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

[38] Hansen B E and Racine J S, Jackknife model averaging, Journal of Econometrics, 2012, 167(1):
38–46.
[39] Ba-Fail A O, Abed S Y, Jasimuddin S M, et al., The determinants of domestic air travel demand
in the Kingdom of Saudi Arabia, Journal of Air Transportation World Wide, 2000, 5(2): 72–86.
[40] Carson R T, Cenesizoglu T, and Parker R, Forecasting (aggregate) demand for US commercial
air travel, International Journal of Forecasting, 2011, 27(3): 923–941.
[41] Teyssier N, How the consumer confidence index could increase air travel demand forecast accu-
racy? PhD’s degree thesis, Cranfield University, Bedfordshire, 2012.
[42] Grubb H and Mason A, Long lead-time forecasting of UK air passengers by Holt-Winters methods
with damped trend, International Journal of Forecasting, 2001, 17(1): 71–82.
[43] Bermúdez J D, Segura J V, and Vercher E, Holt-Winters forecasting: An alternative formulation
applied to UK air passenger data, Journal of Applied Statistics, 2007, 34(9): 1075–1090.
[44] Wang S J, Sui D, and Hu B, Forecasting technology of national–wide civil aviation traffic, Journal
of Transportation Systems Engineering and Information Technology, 2010, 10(6): 95–102.
[45] Aigner D, Lovell C K, and Schmidt P, Formulation and estimation of stochastic frontier produc-
tion function models, Journal of Econometrics, 1977, 6(1): 21–37.
[46] Meeusen W and Van den Broeck J, Efficiency estimation from Cobb-Douglas production functions
with composed error, International Economic Review, 1977, 18(2): 435–444.
[47] Jorge–Calderón J D, A demand model for scheduled airline services on international European
routes, Journal of Air Transport Management, 1997, 3(1): 23–35.
[48] Castelli L, Pesenti R, and Ukovich W, An airline-based multilevel analysis of airfare elasticity
for passenger demand, Proceeding of the 7th Air Transport Research Society World Conference,
France, 2003.
[49] Demirsoy C, Analysis of stimulated domestic air transport demand in Turkey, Master’s degree
thesis, Erasmus University, Rotterdam, 2012.
[50] Lim C, Review of international tourism demand models, Annals of Tourism Rresearch, 1997,
24(4): 835–849.
[51] Marazzo M, Scherre R, and Fernandes E, Air transport demand and economic growth in Brazil:
A time series analysis, Transportation Research Part E: Logistics and Transportation Review,
2010, 46(2): 261–269.
[52] Fernandes E and Pacheco R R, The causal relationship between GDP and domestic air passenger
traffic in Brazil, Transportation Planning and Technology, 2010, 33(7): 569–581.
[53] Dargay J and Hanly M, The determinants of the demand for international air travel to and from
the UK, Proceeding of the 9th World Conference on Transport Research, Edinburgh, Scotland,
2001.
[54] Abed S Y, Ba-Fail A O, and Jasimuddin S M, An econometric analysis of international air travel
demand in Saudi Arabia, Journal of Air Transport Management, 2001, 7(3): 143–148.
[55] Zhang X, Wan A T K, and Zou G, Model averaging by jackknife criterion in models with depen-
dent data, Journal of Econometrics, 2012, 174(2): 82–94.
[56] Zhang X, Zou G, and Liang H, Model averaging and weight choice in linear mixed effects models,
Biometrika, 2014, 101(1): 205–218.
[57] Ando T and Li K C, A model-averaging approach for high-dimensional regression, Journal of the
American Statistical Association, 2014, 109(505): 254–265.
[58] White H, Maximum likelihood estimation of misspecified models, Econometrica: Journal of the
 A DEMAND FORECASTING METHOD 17

Econometric Society, 1982, 50(1): 1–25.

[59] Zhang X, Model averaging and its applications, Ph. D. dissertation, Chinese Academy of Sciences.

Appendix
A.1 Notations and Assumptions
T
Let θ(s) = (βT , μ
(s) , σ
(s) μ,(s) , σ
ν,(s) ) .
∗
Assumption C.1 For any candidate model s, there exists a constant vector θ(s) such that
as n → ∞, θ(s) − θ(s) = Op (n
∗ −1/2
).
From [58], we know that even for a misspecified model, Assumption C.1 is satisfied un-
T T
der some regularity conditions. Let θ = (θ(1) T T
, θ(2) , · · · , θ(S) ) and θ(−i) be a cross validation
version of θ,  i.e., without using the ith observation {Ci , x(1),i , x(2),i , · · · , x(S),i }. Let R(w) =
n  2 n (−i) (w) − Ci }2 . Furthermore, we rewrite C i (w) =
i=1 E{Ci (w) − Ci } and R(w) = i=1 E{Ci
 xi ) and C 
Ci (w, θ,  (−i)
(w) = C i (w, θ(−i) , xi ). Define L∗ (w) = n  ∗ 2
i i=1 {Ci (w, θ , xi ) − Ci } .
  ∗
Assumption C.2 ∂ Ci (w, θ, xi )/∂θ = Op (1), Ci (w, θ , xi ) = Op (1) and Ci = Op (1) uni-
formly for i = 1, 2, · · · , n + 1, w ∈ W , any θ between θ and its limit, and any θ between θ(−j)
and its limit for j = 1, 2, · · · , n.
Assumption C.3 The left hands of the first and third parts in (A.6) are uniformly inte-
grable.

Assumption C.4 supw∈W |R(w) − R(w)|/R(w) = o(1).
−1 1/2
Assumption C.5 {inf w∈W R(w)} n = o(1).
Assumption C.6 supw∈W {|riskC (w) − riskD (w)|/riskC (w)} = o(1).
Assumptions C.2 and C.3 are about the derivative and integrals and are very regular. As-
sumption C.4 is standard in cross validation analysis and is similar to the condition (10) of [55]
and the condition (A.5) of [38]. Assumption C.5 requires that inf w∈W R(w) grows at a rate
no slower than n1/2 and is very similar to the condition (A7) of [56] and the conditions (7)
of [57]. Assumption C.6 requires the squared forecasting risk of consumption dominates that of
demand.

A.2 Proof of Theorem 1

0
Let wm be a weight vector belonging to W , in which the mth component is one. It is seen
that

riskC (w)
= E{C n+1 (w) − Cn+1 }2
 xn+1 ) − Cn+1 }2
n+1 (w, θ,
= E{C
n+1 (w, θ(−n) , xn+1 ) − Cn+1 + C
= E{C  xn+1 ) − C
n+1 (w, θ, n+1 (w, θ(−n) , xn+1 )}2
n+1 (w, θ(−n) , xn+1 ) − Cn+1 }2 + E{C
= E{C  xn+1 ) − C
n+1 (w, θ, n+1 (w, θ(−n) , xn+1 )}2
n+1 (w, θ(−n) , xn+1 ) − Cn+1 }{C
+E2{C  xn+1 ) − C
n+1 (w, θ, n+1 (w, θ(−n) , xn+1 )}

= n−1 R(w)  xn+1 ) − C
n+1 (w, θ,
+ E{C n+1 (w, θ(−n) , xn+1 )}2
18 ZHANG XINYU · ZHENG YAFEI · WANG SHOUYANG

+E2{C n+1 (w, θ(−n) , xn+1 ) − Cn+1 }{C  xn+1 ) − C

n+1 (w, θ, n+1 (w, θ(−n) , xn+1 )}

≡ n−1 R(w) + EΛ1 (w), (17)

where the fifth equality uses the fact that {Ci , xi }n+1
i=1 are i.i.d.,

n

CV(w) = 
{C
(−i)
(w) − Ci }2
i
i=1
n

= i (w, θ(−i) , xi ) − Ci }2
{C
i=1
n

= i (w, θ∗ , xi ) − Ci + C
{C i (w, θ(−i) , xi ) − C
i (w, θ∗ , xi )}2
i=1
n

= L∗ (w) + i (w, θ(−i) , xi ) − C
{C i (w, θ∗ , xi )}2
i=1
n

+2 i (w, θ∗ , xi ) − Ci }{C
{C i (w, θ(−i) , xi ) − C
i (w, θ∗ , xi )}
i=1
≡ L∗ (w) + Λ2 (w), (18)

n


R(w) = 
E{C
(−i)
(w) − Ci }2
i
i=1
n

=E i (w, θ(−i) , xi ) − Ci }2
{C
i=1
n

=E i (w, θ∗ , xi ) − Ci + C
{C i (w, θ(−i) , xi ) − C
i (w, θ∗ , xi )}2
i=1
n

= EL∗ (w) + E i (w, θ(−i) , xi ) − C
{C i (w, θ∗ , xi )}2
i=1
n

+E2 i (w, θ∗ , xi ) − Ci }{C
{C i (w, θ(−i) , xi ) − C
i (w, θ∗ , xi )}
i=1
≡ EL∗ (w) + EΛ3 (w), (19)

and
EL∗ (w) − L∗ (w)
n
 n

=E i (w, θ∗ , xi ) − Ci }2 −
{C i (w, θ∗ , xi ) − Ci }2
{C
i=1 i=1
n

 
= i (w, θ∗ , xi ) − Ci }2 − {C
E{C i (w, θ∗ , xi ) − Ci }2
i=1
n

 S 2  
S 2 
= E  0 ∗
ws (Ci (ws , θ , xi ) − Ci ) − 0  0 ∗
ws (Ci (ws , θ , xi ) − Ci )
i=1 s=1 s=1
 S
S 
n 
= i (ws0 , θ∗ , xi )
ws wm {E(C i (wm
− Ci )(C 0
, θ∗ , xi ) − Ci )
i=1 s=1 m=1
i (w0 , θ∗ , xi ) − Ci )(C
−(C i (w0 , θ∗ , xi ) − Ci )}
s m
 A DEMAND FORECASTING METHOD 19

S 
S 
  n
≤ ws wm  {E(C i (w0 , θ∗ , xi ) − Ci )(Ci (w0 , θ∗ , xi ) − Ci )
s m
s=1 m=1 i=1 

−(Ci (ws0 , θ∗ , xi ) − Ci )(C
i (wm0
, θ∗ , xi ) − Ci )}

= Op (n1/2 ). (20)

From Assumptions C.4 and C.5, we have

 −1

inf R(w) n1/2 = op (1), (21)
w∈W

which, along with Assumptions C.1 and C.2, implies,

n|Λ1 (w)| n|Λ2 (w)| n|Λ3 (w)|

sup = op (1), sup = op (1), sup = op (1). (22)
w∈W
R(w) w∈W
R(w) w∈W
R(w)

From this first part of (A.6) and Assumption C.3, we have

nE|Λ1 (w)| n|Λ1 (w)|

sup ≤ E sup = o(1). (23)
w∈W
R(w) w∈W
R(w)

Similarly, we have

E|Λ3 (w)|
sup = o(1). (24)
w∈W
R(w)

Now, from (A.1)–(A.8), we have

|CV(w) − n ∗ riskC (w)|

|n ∗ riskC (w) − R(w)|
sup = op (1), sup = o(1). (25)
w∈W
R(w) w∈W
R(w)

From (A.9) and Lemma 1 in [59], we can get (14).

Next, we prove (15). From (A.9) and Assumption C.6, we know that

n|riskC (w) − riskD (w)|

sup
w∈W
R(w)
|riskC (w) − riskD (w)| n ∗ riskC (w)
= sup
w∈W riskC (w)
R(w)
|riskC (w) − riskD (w)| n ∗ riskC (w)
≤ sup sup
w∈W risk C (w) w∈W
R(w)

|riskC (w) − riskD (w)| |n ∗ riskC (w) − R(w)|
≤ sup 1 + sup
w∈W riskC (w) w∈W
R(w)
= o(1), (26)

which, along with (A.9) and Lemma 1 in [59], implies (15).