Methods For Forecasting

Nonlinear Modeling of Economic and Financial TimeSeries
Downloaded by Huazhong University of Science and Technology At 01:13 20 December 2016 (PT)
Chapter 8 Alternative Methods for Forecasting GDP

Dominique Gugan, Patrick Rakotomarolahy
Article information:
To cite this document: Dominique Gugan, Patrick Rakotomarolahy. "Chapter 8
Alternative Methods for Forecasting GDP" In Nonlinear Modeling of Economic and
Financial Time-Series. Published online: 08 Mar 2015; 161-185.
Permanent link to this document:
http://dx.doi.org/10.1108/S1571-0386(2010)0000020013
Downloaded on: 20 December 2016, At: 01:13 (PT)
References: this document contains references to 51 other documents.
To copy this document: permissions@emeraldinsight.com
The fulltext of this document has been downloaded 262 times since 2015*
Users who downloaded this article also downloaded:

(2009),"Energy and GDP growth", International Journal of Energy Sector Management,
Vol. 3 Iss 2 pp. 157-170 http://dx.doi.org/10.1108/17506220910970560
(2006),"GDP per capita and its challengers as measures of happiness",
International Journal of Social Economics, Vol. 33 Iss 10 pp. 698-709 http://
dx.doi.org/10.1108/03068290610689732
Access to this document was granted through an Emerald subscription provided by

emerald-srm:313548 []
For Authors
If you would like to write for this, or any other Emerald publication, then please
use our Emerald for Authors service information about how to choose which
publication to write for and submission guidelines are available for all. Please visit
www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com

Emerald is a global publisher linking research and practice to the benefit of society.
The company manages a portfolio of more than 290 journals and over 2,350 books and
book series volumes, as well as providing an extensive range of online products and
additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner
of the Committee on Publication Ethics (COPE) and also works with Portico and the
LOCKSS initiative for digital archive preservation.
*Related content and download information correct at time of download.
Chapter 8
Alternative Methods for Forecasting GDP

Dominique Guegana and Patrick Rakotomarolahyb
a
Paris School of Economics, Universite Paris 1 Pantheon-Sorbonne,

CES-MSE, 106 boulevard de lHopital 75647 Paris Cedex 13, France,
e-mail: dguegan@univ-paris1.fr
b
Paris School of Economics, CES-MSE, Universite Paris 1 PantheonSorbonne, 106 boulevard de lHopital 75647 Paris Cedex 13, France,
e-mail: rakotopapa@yahoo.fr
Abstract
Purpose The purpose of this chapter is twofold: to forecast gross domestic
product (GDP) using nonparametric method, known as multivariate k-nearest
neighbors method, and to provide asymptotic properties for this method.
Methodology/approach We consider monthly and quarterly macroeconomic variables, and to match the quarterly GDP, we estimate the missing
monthly economic variables using multivariate k-nearest neighbors method
and parametric vector autoregressive (VAR) modeling. Then linking these
monthly macroeconomic variables through the use of bridge equations, we
can produce nowcasting and forecasting of GDP.
Findings Using multivariate k-nearest neighbors method, we provide a forecast
of the euro area monthly economic indicator and quarterly GDP, which is better
than that obtained with a competitive linear VAR modeling. We also provide
the asymptotic normality of this k-nearest neighbors regression estimator for
dependent time series, as a condence interval for point forecast in time series.
Originality/value of chapter We provide a new theoretical result for
nonparametric method and propose a novel methodology for forecasting
using macroeconomic data.
Keywords: economic indicators, euro area GDP, VAR, multivariate
k-nearest neighbors regression, asymptotic normality, forecast
JEL Classication: C22, C53, E32
International Symposia in Economic Theory and Econometrics, Vol. 20
F. Jawadi and W.A. Barnett (Editors)
Copyright r 2010 by Emerald Group Publishing Limited. All rights reserved
ISSN: 1571-0386/DOI: 10.1108/S1571-0386(2010)0000020013
162
Dominique Guegan and Patrick Rakotomarolahy
1. Introduction
Forecasting macroeconomic variables such as gross domestic product
(GDP) and ination plays an important role for monetary policy decisions
and for assessment of future state of the economy. Policymakers and
economic analysts either adapt their theoretical analysis of economic
conditions according to the macroeconomic variable forecasts or even
probably use them as a support and a justication of their theoretical
analysis. Better forecast performance for macroeconomic variables will lead
to better decisions.
The objective of this chapter is to propose a new way to forecast GDP,
reducing the forecasting errors for this macroeconomic variable. We work
with a nonparametric technique, called the nearest neighbors method, and we
compare our methodology with classical well-known parametric methods.
Considering forecasting issues, we can identify two kinds of methodologies in the literature: methods based on parametric modeling and methods
based on nonparametric techniques. The former methods include among
others the linear autoregressive models (Box and Jenkins, 1970) and the
nonlinear SETAR-STAR and Markov switching models (Tong (1990) and
Pena et al. (2003)). The latter ones include, for instance, kernel method,
nearest neighbors method, neural network and wavelet methods (Silverman,
1986; Hardle et al., 2004). The methods based on parametric modeling
have had great consideration in economic forecasting due to the huge
development of theoretical results concerning consistent, asymptotic
properties and robustness of the parameters modeling. These methods
have also had different problems concerning the strong hypotheses of model
specication, estimation method, and asymptotic properties of the estimated
parameters among others. The methods based on nonparametric techniques
have overcome some of these problems by avoiding an a priori specication
of the modeling and the distribution of residuals. Indeed, this methodology
is based on the fact that it lets the data speak by themselves. Hence, it avoids
the subjectivity of choosing a specic parametric modeling before looking at
the data. However, there is the cost of more complicated mathematical
arguments such as the selection of smoothing parameters. Nevertheless,
recent studies help to avoid these problems and also the speed of computers
that can develop search algorithms from appropriate selection criteria
(Devroye and Gyor, 1985; Becker et al., 1988).
The forecasting of GDP has been studied for a long time starting from the
growing use of linear autoregressive models. Indeed, in 1980, Sims forecasted
American GDP using linear vector autoregressive (VAR) model, then
Litterman (1986) extended this work using the Bayesian VAR aiming at
reducing estimation by VAR model. Engle and Granger (1987) pointed out
possible cointegration between US GDP and monetary aggregate M2 using
163
vector error correction modeling; this approach has been recently used by
Gupta (2006) to forecast South African GDP. Besides the linear modeling for
forecasting GDP, we observe the development of nonlinear methods for
forecasting GDP using Markov switching models (Hamilton, 1989) or
SETAR models (Clements and Krolzig, 1998). Another approach combines
linearity and aggregation, for instance, through the bridge equations method
(Bafgi et al., 2004; Diron, 2008). In this approach, the number of economic
indicators is also diminished. Alternatively, factor models based on a great
number of indicators have been proposed by Stock and Watson (2002), with a
dynamic extension developed by Bernanke and Boivin (2003), Forni et al.
(2005), and Kapetanios and Marcellino (2006). Recently, forecasting GDP
based on microeconomic foundation appears with the so-called dynamic
stochastic general equilibrium models (Smets and Wouters, 2004). Nevertheless, the linear univariate ARIMA or multivariate VAR models remain the
benchmarks in the literature.
Alternatively, the nonparametric techniques have not been considered a
lot in economic literature for forecasting: an interesting review is by
Yatchew (1998). Concerning the use of these techniques to forecast GDP, in
our knowledge very few works exist. We can cite Tkacz and Hu (1999) who
use neural networks to forecast Canadian GDP or Ferrara et al. (2010) and
Guegan and Rakotomarolahy (2010) who use nearest neighbors method and
radial basis function methods to forecast euro area GDP. The kernel
method, which is one of the well-known nonparametric methods, is rarely
used in forecasting because it is based on important restrictions that we
recall in Section 2. In this chapter, we focus on the nearest neighbors method
because we obtain robust theoretical results for the nearest neighbors
estimates, which permit to build condence intervals: these results do not
exist for radial basis function in a simple way, nor with the neural networks
method, and the kernel approach. Some other references on these
nonparametric techniques are Prakasa Rao (1983), Donoho and Johnstone
(1992), Kuan and White (1994), Friedman (1988), and Mack (1981).
Predicting a time series requires estimating the conditional mean and
variance of this time series in some period, given an information set based on
the past. Thus, our work based on nearest neighbors method rst concerns
the estimate of the conditional mean, which allows one to get the point
forecast. To build the condence intervals, we will need to estimate the
conditional variance.
In order to estimate the conditional mean associated with a time series, we
proceed in the following way. Given a time series (Xn)n, we consider the
following representation for the regression function m( ) associated with
this time series:
mx EX n1 jX n x.
(1)
164
Model (1) is a nonlinear regression problem that can be estimated using

many smoothing methods (Hart, 1994).
In this chapter, we reconstruct the function m( ) using multivariate knearest neighbors method. Among the nonparametric techniques, we focus
on this method because apart from its advantage away from risk of model
misspecication or some strong hypotheses of parametric method, it handles
the nonlinearity of the data sets by the embedding concept. Indeed, working
in a multivariate environment allows us to discover and take into account
the structural behavior that cannot always be discerned on a path. The other
advantages of the method are for practicians because it allows one to use a
small set of parameters, it is the easiest method to implement, and it is not
time consuming. Finally, it is the only nonparametric method for which
robust theoretical results are available. These results build condence
intervals that help compare its accuracy with classical parametric modeling.
The use of such nonlinear method for forecasting GDP is fundamental as
soon as nonlinearity behavior Inside GDP is detected, for example for US
GDP in Hamilton (1989). On the other hand, the interest of the nearest
neighbors methods takes into account among others the nonlinear features
of the nancial data sets (Mizrach, 1992; Nowman and Saltoglu, 2003;
Guegan and Huck, 2005). Finally, working in a multivariate setting is not
too constrained as recent results have made available a method for selecting
the number of neighbors within a given space (Ouyang et al., 2006).
Our theoretical result is a contribution to the general problem concerning
the nonparametric estimate of the regression function m( ) with k-NN
method, extending well-known results obtained for independent and
identically distributed random variables (Stone, 1977; Mack, 1981; Devroye,
1982; Stute, 1984). In case of dependent variables, Collomb (1984) provides
piecewise convergence for univariate variables, and Yakowitz (1987) gets the
quadratic mean error for uniformly weighted k-NN estimates for univariate
samples. Here, working with multivariate time series, we control the bias of
a general multivariate k-NN estimate, using several weights, and we
establish the asymptotic normality of this estimate from which we can
construct condence intervals.
In this chapter, we use our theoretical result to propose a new way to
estimate and forecast the macroeconomic indicators used to build GDP. For
that, we follow the work developed by Diron (2008), which is used a lot in
central banks. Her method is based on a limited number of economic
indicators, which are plugged in eight linear equations from which an
estimate of GDP is obtained. Her method associates bridge equations and
forecasts combinations incorporating a large number of economic activities,
including different single forecasts based on production sectors, survey data,
nancial variables, and leading index constructed from a large number of
economic indicators. Her method is competitive with methods allowing a
165
huge number of indicators and we do not compare here these different

modelings. Considering one specic parametric modeling (bridge equations), we compare our forecasting errors based on nonparametric
techniques with the same methodology based on linear modeling (Runstler
and Sedillot, 2003; Darne, 2008). Our estimation procedure is different, in
the sense that we estimate the economic indicators with multivariate
k-nearest neighbors method.
The chapter is organized as follows. In Section 2, we establish our
theoretical result: the asymptotic normality of the multivariate k-NN
regression estimate for mixing time series. In Section 3, an empirical
forecasting exercise is provided. It allows one to compare nonparametric
and parametric approaches for monthly indicators and their impact in the
nal GDP estimate. Section 4 concludes and the appendix is devoted to the
proofs.
2. Theoretical Result
We consider a real time series (Xn)n, and we transform the original
data set by embedding it in a space of dimension d, building
X n X nd1 ; . . . ; X n 2 Rd . The embedding concept is important because
it takes into account some characteristics of the series that are not always
observed on the trajectory in R.
Considering expression (1), we are interested in getting an estimate of
mx, x 2 Rd , using the k closest vectors to X n x inside the training set
denoted by S, that is, S fX t X td1 ; ; X t jt d; :::; n 1g Rd .
We dene a neighborhood around x 2 Rd such that Nx fiji 1; . . . ; kn
whose X i represents the ith nearest neighbor of x in the sense of a given
distance measure}. Then the k-NN regression estimate of mx, x 2 Rd is
given by
X
mn x
wx X i X i1 ,
(2)
X i 2S; i2Nx
where w( ) is a weighting function associated with neighbors and it is

noteworthy that the parameter k has to be estimated. A general form for the
weights is
wx X i
1=nRdn Kx X i =Rn
,
P
1=nRdn ni1 Kx X i =Rn
(3)
where Rn corresponds to the distance between x and the further neighbors,

and K( ) is a given weighting function. Two weighting functions have been
mainly used, the exponential function Kx X i =Rn expjj x X i jj2
and the uniform function Kx X i =Rn 1=k.
166
The result established in Theorem 1 proves the asymptotic convergence of

the k-NN regression estimate belonging to Rd for dependent variables: this
result extends the study by Yakowitz (1987). This result is interesting
because, in practice, it is not necessary to lter the observed data to make
them independently work using this approach. On the other hand, this result
guarantees the consistency of the estimate mn( ), and therefore the
conditional mean forecast will asymptotically coincide with the expected
true value. Indeed, the knowledge of the bias and speed rate of the variance
of the estimates provides consistent estimates, and their asymptotic
normality provides condence intervals. These condence intervals can be
used to compare the quality of point forecasts obtained from different
methods, which enhances comparison of several methods (parametric and
nonparametric methods) beyond point forecast. Since no rigorous test is
available to discuss the choice between the parametric and the nonparametric approaches, predictive methodology can be used for this objective.
To establish our main result, we assume that the time series (Xn)n is
strictly stationary and is characterized by an invariant measure with density
f. Moreover, the random variable X n1 jX n x has a conditional density
f yj x, and the invariant measure associated with the embedded time series
X n n is h.
Theorem 1. Assuming that (Xn)n is a stationary time series, and that the
following assumptions are veried:
(Xn)n is f-mixing.
mx; f yj x and hx are p-continuously differentiable.
f yj x is bounded.
Pkn
wi 1, then k-NN regresThe sequence knon is such that i1
sion function mn x dened in Equation (2) veries:
p
(4)
nQ mn x Emn x!D N0; s2 ,
(i)
(ii)
(iii)
(iv)
with
Emn x mx2 OnQ ;
(5)
where 0rQr1, Q 2p=2p d, and s2 g2 VarX n1 jX n x B2 ;

with B On1Qp=d , and g a positive constant which is equal to one
when we use uniform weights.
The proof of this theorem is provided in the appendix. Some points of this
proof are as follows:
1. As soon as the number of neighbors k is different from one, we remark
that 8u, 0owi uo1, whatever the weighting function used: uniform or
exponential function.
167
2. The main difference between k-NN method and kernel method (Silverman, 1986) lies on the information set that we use to estimate the
function m( ) at a given point x. In the latter case, the information set is
xed, and in the former case, it is exible with respect to the choice of
the number of neighbors k. In this case, such exibility has an impact on
the values of the weights. When the number of neighbors k increases, the
weights wi ki1 decrease, and the product k:wi ki1 turns around a
constant g that belongs to R. For uniform weights, wi 1=k and g 1.
This last property implies that the asymptotic variance of theR estimate
mn( ) depend neither on the true density nor on the quantity w2 udu.
This asymptotic property is not veried when we work with the kernel
method; details are provided in the appendix.
3. The mixing conditions characterize different behaviors of dependent
variables. Parametric processes like the bilinear models, including
ARMA models, the related GARCH processes, and the Markov
switching processes, are known to be mixing (Guegan, 1983; Carrasco
and Chen, 2002). Thus, in practice, this condition is not too restrictive.
4. The condition (iv) in Theorem 1 is veried in particular for the weights
introduced in Equation (3). The parameter g introduced before entails the
correlations between the vectors X n . Finally, Theorem 1 providing
asymptotic normality for the estimate mn x under regular conditions
permits to build condence interval whose expression is given in the
following corollary.
Corollary 1. Under the assumptions of Theorem 1, a general form for the
condence interval around mx, for a given risk level 0oao1, is
"
#
^ 1a
^ 1a
sz
sz
mx 2 mn x B p 2 ; mn x B p 2 .
k
k
(6)
where z1a=2 is the (1(a/2)) quantile of the Student law, s^ an estimate for
s, and B such that
(1) B is negligible, if kn=n ! 0, as n-N.
1=d
^
with
r kn=n dhxc
;
where
(2) If
not,
B O(rp),
d=2
^
is an estimate for the density hx.
c p =Gd 2=2, and hx
The proof of this corollary is provided in the appendix.
3. Forecasting Euro Area GDP

Information on the current state of economic activity is a crucial ingredient
for policy-making. Economic policymakers, international organizations,
168
and private sector forecasters commonly use short-term forecasts of real

GDP growth based on monthly indicators. There exist many studies
proposing real-time modeling in order to take into account some complexities inherent to the computation of GDP, which are the number of
economic indicators, the modeling for GDP, the impact of data revisions,
etc. (Koenig et al., 2003; Bafgi et al., 2004; Schumacher and Breitung,
2008). Koenig et al. (2003) focus on the choice of vintage data and found a
substantial improvement of GDP forecast when using real-time vintage
data; the paper by Bafgi et al. (2004) deals on the modeling for GDP by
comparing euro area GDP forecasts from linear ARIMA and VAR models
with bridge equations and concludes that the latter provides better forecasts.
The paper by Schumacher and Breitung (2008) suggests the use of large
factor models for mixed-frequency data supported by experiment on
monthly German GDP and concludes on a minor impact of the data
revision in forecasting performance. In the present exercise, we show that
beyond the model chosen for GDP forecasts, the estimates of the monthly
economic indicators are crucial and can lead to nonnegligible errors if they
are not properly estimated. Thus, our approach is slightly new and different
in comparison to most of the published works on this subject.
In order to illustrate our proposal, we restrict to the modeling of GDP
using the bridge equations modeling (Runstler and Sedillot, 2003). We
consider the eight equations introduced in the paper of Diron (2008), each
equation providing a model of GDP, denoted Y it ; i 1; . . . ; 8. All the eight
models are aggregated to provide a nal GDP, denoted Yt. Each equation is
calculated from 13 monthly economic indicators, denoted X it ; i 1; . . . ; 13.
We focus here on the forecasting of these 13 indicators. We estimate and
forecast these indicators from two models: the unrestricted VAR modeling
and the multivariate NN approach. For the latter method, we distinguish
forecasts obtained without embedding data sets (d 1) from forecasts
obtained when dW1. The 13 economic indicators that we consider are listed
in Table 1.
For this exercise, we use the real-time database provided by EABCN
through their Web site.1 The real-time information set starts in January 1990
when possible (exceptions are the condence indicator in services, which
starts in 1995, and EuroCoin (ERC), which starts in 1999) and ends in
November 2007. The vintage series for the OECD composite leading
indicator are available through the OECD real-time database.2 The ERC
index is taken as released by the Bank of Italy. The vintage database for a
given month takes the form of an unbalanced data set at the end of the
1
2
http://www.eabcn.org
http://stats.oecd.org/mei/
169
Table 1: Monthly Economic Indicators of Euro Area Used for Forecasting

GDP
Short
Notation
Notation
X1
X2
IPI
CTRP
X3
SER-CONF
X4
X5
X6
RS
CARS
MAN-CONF
X7
ESI
X8
CONS-CONF
X9
RT-CONF
X10
EER
X11
X12
PIR
OECD-CLI
X13
ERC
Indicator Names
Industrial Production Index

Industrial Production Index
in Construction
Condence Indicator in
Services
Retail Sales
New Passenger Registrations
Industry
European Economic
Sentiment Index
Consumers Condence
Indicator
Retail Trade
Effective Exchange Rate
Deated EuroStock Index
OECD Composite Leading
Indicator Trend Restored
ERC Indicator
Sources
Period
Eurostat
Eurostat
19902007
19902007
European
Commission
Eurostat
Eurostat
European
Commission
European
Commission
European
Commission
European
Commission
Banque de
France
Eurostat
OECD
19952007
19902007
19902007
Bank of Italy
19992007
19902007
19902007
19902007
19902007
19902007
19902007
19902007
Note: Summary of the 13 economic indicators of euro area used in the eight GDP bridge
equations.
sample. To solve this issue, we apply the two previous methodologies to

forecast the monthly variables in order to complete the values until the end
of the current quarter for GDP nowcast and until the end of the next quarter
for GDP forecasts, then we aggregate the monthly data to quarterly
frequencies.
Alternatively, we use a stationary VAR methodology for forecasting
economic indicators and combine later with bridge equations proposed by
Diron (2008) to get estimates for GDP, making all the data stationary using
the rst difference. Among the 13 indicators used in Diron equations, three
indicators (Economic Sentiment Indicator (ESI), Composite Leading
Indicator (CLI), and ERC) appear redundant, in the sense that they are
built from the 10 others. Thus, in order to avoid repetition of variables in the
model, which could produce extra contribution on the variance through
correlations, we consider only 10 indicators as endogenous variables in the
VAR modeling, building 10-variate VAR for these remaining 10 indicators.
Finally, using AIC and Schwartz criteria for order selection we retain a 10-
170
variate VAR(1) (Akaike, 1974; Schwartz, 1978). Nevertheless, we need

estimates for the previous three indicators to nalize the computation of
GDP with the Diron equations. We adjust a specic ARIMA modeling for
each three variables using AIC criterion for determining the orders p and q.
Finally, we retain ARIMA(3,1,0), ARIMA(10,1,0), and ARIMA(1,1,0)
modelings, respectively, for ESI, CLI, and ERC indicators. In all parametric
models, the parameters are estimated by least-squares method. We use
recursive forecasts when computing the forecast beyond one step ahead.
Regarding the method of NN, d being given, we determine the number of
neighbors k by minimizing the root mean square error (RMSE) criterion:
v
u
n
X
u
1
i
t
(7)
jjX^ t1 X it1 jj2 i 1; . . . ; 13,
n k d tkd1
i
where n is the sample size, and X^ t1 the estimate of the ith economic
indicator X it1 calculated from expression (2). The number 1rkr5 of
nearest neighbors determined by this criterion at the horizon h 1 is used to
calculate the forecasting capabilities for hW1.
In the case of the multivariate approach (dW1), we describe below the
algorithm used to determine the embedding dimension d and the number of
neighbors k used to obtain the best predictor for X inh in the sense of the
previous RMSE. We present the method for all indicators, and thus for
simplicity we drop the index i in the algorithm. All the data sets have been
made stationary with the same transformation than the one used for VAR
modeling. Thus, now we assume that we observe a stationary data set
X 1 ; :::; X n in R.
(1) We embed this data set in a space of dimension d, 2rdr10, getting a

sequence of vectors in Rd : {X d ; X d1 ; . . . ; X n }, where X i X id1 ;
. . . ; X i .
(2) Ranking the vectors, we determine the k-nearest vectors of X n . Denoting
ri jjX n X i jj, i d; d 1; :::; n 1, the distance between these
vectors, we build the sequence rd ; rd1 ; :::; rn1 ordered in an increasing
way: rd ord1 o orn1 , which provides the k-nearest vectors X j
corresponding to these rj , j d, d1,y, (dk)1.
(3) The one-step-ahead forecast mn X n X^ n1 is obtained from
X^ n1
kd1
X
wjjX n X j jjX j1 .
(8)
jd
(4) Considering now the new information set X 1 ; :::; X n ; X^ n1 , redo step 1 to
step 3, we get the two-steps-ahead forecast. We obtain the forecast of
third step ahead in a similar way as for the two-steps-ahead forecast, and
so on. We limit the choice of the embedding dimension d to 10.
171
We consider exponential weighting function since it reects the local

behavior of nearest neighbors method that gives more weight to the closest
neighbors. We favor this kind of weights rather than the uniform weights
that give same importance to all neighbors. Now, for each indicator
X it ; i 1; . . . ; 13, the best pair (d, k) is determined by minimizing again the
RMSE criteria dened in Equation (7). Once the pair (d, k) is found, it is
used for all prediction horizons.
As soon as we get the estimates for the monthly indicators with the two
previous methods (VAR and k-NN methods), we compute the GDP ash
estimates that were released in real time by Eurostat from the rst quarter of
2003 to the third quarter of 2007 using the previous forecasts of the monthly
indicators. According to this scheme, the monthly series have to be
forecasted for a horizon h varying between three and six months in order to
complete the data set at the end of the sample. Recall that the h-step-ahead
predictor for h>1 is estimated recursively starting from the one-step-ahead
formula.
Using ve years of vintage data, from the rst quarter of 2003 to the third
quarter of 2007, we provide RMSEs for the euro area ash estimates of
GDP growth Y^ t in genuine real-time conditions. We have computed the
RMSEs for the quarterly GDP ash estimates, obtained with the forecasting
methods used to complete adequately in real time the monthly indicators,
that is, VAR modeling and k-NN methods (d 1 and d>1). More precisely,
we provide the RMSEs of the combined forecasts based on the arithmetic
mean of the eight Diron equations. Thus, for a given forecast horizon h,
j
we compute Y^ t h, which is the predictor stemming from the eight equations
j 1; . . . ; 8, in which we have plugged the forecasts of the monthly
economic indicators,
and we compute the nal estimate GDP at horizon h:
P
j
Y^ t h 1=8 8j1 Y^ t h. The RMSE criterion used for the nal GDP is
v
u T
u1 X
(9)
Y^ t h Y t 2 ,
RMSEh t
T t1
where T is the number of quarters between the rst quarter of 2003 and the
second quarter of 2007 (in our exercise, T 18) and Yt the euro area ash
estimate for quarter t. The RMSE errors for the nal GDP are provided in
Table 2 and comments follow.
For both methods, VAR modeling and k-NN method, the RMSE
becomes lower when the forecast horizon reduces from h 6 to h 1,
illustrating the accuracy of the nowcasting and forecasting, which increases
as soon as the information set becomes more and more efcient, thanks to
the released monthly data. This is the strength of GDP forecasting based on
monthly economic indicators, instead of considering only GDP itself, since
each month new true values of economic indicators are available.
172
Table 2: RMSE on GDP Growth from the Three Methods VAR, k-NN
with d 1, and k-NN with dW1
Horizon
VAR
k-NN(d 1)
k-NN(dW1)
6
5
4
3
2
1
0.225
0.224
0.214
0.192
0.181
0.173
0.198
0.203
0.202
0.186
0.176
0.174
0.214
0.192
0.196
0.177
0.177
0.171
Note: RMSE for the estimated mean quarterly GDP growth Yt computed from Equation (9),
using VAR(p) modeling (column 2) and k-NN predictions (d 1 (column 3), and dW1 (column
4)) for the monthly economic indicators X it ; i 1; . . . ; 13, h is the monthly forecast horizon.
Values in boldface correspond to the smallest error for a given forecast horizon.
We remark that few days before the publication of the ash estimate
(around 13 days with h 1), the lowest RMSE is obtained with the
multivariate k-NN method (RMSE 0.171).
Looking at forecast errors by comparing column 2 with columns 3 and 4
of Table 2, we nd that forecast errors are always lower with the method of
NN rather than with VAR modeling (except at horizon h 1 where VAR
modeling gives better forecast error than univariate k-NN). One source of
such gain comes from the use of nearest neighbors method that is adapted
even with small samples, but this is not the case when working with VAR
modeling that requires large samples to be robust.
Finally, if we focus on nearest neighbors method, we obtain smaller errors
when working with multivariate setting d>1 than with univariate d 1.
This result shows the gain of the method developed in a space of higher
dimension. We expect that in terms of predictions, any method developed in
higher dimension improves the forecast accuracy. This is conrmed when we
compare, for the same method, the forecast errors obtained only in R with
the error calculated from a treatment in Rd : in this latter case, the errors are
always smaller (e.g., comparing columns 3 and 4 of Table 2). This idea has
already been developed in other empirical works that consider multivariate
methods with factor models (Kapetanios and Marcellino, 2006), and
methods with multivariate nonparametric techniques (Guegan and Rakotomarolahy, 2010).
To see the evolution of the trajectory of forecasting both parametric and
nonparametric methods, we provide, in Figures 1 and 2, the graphs of the
observed and estimated GDP growth from k-NN methods and VAR
modeling for forecast horizons varying from one to six quarters.
These graphs show that the trajectories of GDP forecasts from both
methods are very close for horizons hr3 and are able to follow the true
173
Figure 1: Quarterly Observed (in Black) and Forecasted GDP Growth

Rate Computed from k-NN with d 1 (in Green), k-NN with dW1 (in
Blue), and VAR (in Red) Models between the rst quarter of 2003 and the
second quarter of 2007 for Different Forecast Horizons. Panel (a): Graph of
GDP Growth Rate Forecast at Horizon h 1 Using VAR and k-NN
Methods. Panel (b): Graph of GDP Growth Rate Forecast at Horizon h 2
Using VAR and k-NN methods. Panel (c): Graph of GDP Growth Rate
Forecast at Horizon h 3 Using VAR and k-NN Methods.
174
Figure 2: Quarterly Observed (in Black) and Forecasted GDP Growth

Rate Computed from k-NN with d 1 (in Green), k-NN with dW1 (in
Blue), and VAR (in Red) Models between the rst quarter of 2003 and the
second quarter of 2007 for Different Forecast Horizons. Panel (a): Graph of
GDP Growth Rate Forecast at Horizon h 4 Using VAR and k-NN
methods. Panel (b): Graph of GDP Growth Rate Forecast at Horizon h 5
Using VAR and k-NN Methods. Panel (c): Graph of GDP Growth Rate
Forecast at Horizon h 6 Using VAR and k-NN methods.
175
trajectory. In addition, they permit to detect also some declines of euro area
GDP, for example, in the second quarter of 2003. On the other hand, for
forecast horizons h>3, the VAR modeling provides forecasts for GDP that
converge to the sample mean when the forecast horizon increases. Conversely,
the k-NN modeling provides forecasts that follow the observed GDP
trajectory.
4. Conclusion
Knowing the importance of the nowcast and the forecast of macroeconomic
variables (such as GDP or ination) when analyzing the current state
of the economics and setting policy for the future economic conditions, we
suggest in this chapter alternative methods based on nonparametric
multivariate k-nearest neighbors method to improve the accuracy of GDP
forecasts.
We focus on detecting the best predictor for economic indicators using an
RMSE criterion, working in an embedded space of dimension d, and
focusing on the relevant set of data that helps solve this specic criterion
(Han et al., 1997; Hoover and Perez, 1999).
Our application used a new theoretical result that extends, for the
multivariate k-nearest neighbors estimation method, the L2-consistent result
obtained with uniform weighting (Yakowitz, 1987).
Some factors are questionable: the use of the aggregated monthly economic
indicators to match quarterly GDP; a specic test to decide a good strategy
between parametric and nonparametric modeling; the trade-off between
stationarity and nonlinearity when we work with nonparametric techniques.
Acknowledgments
The authors are extremely grateful to the Guest Editor Fredj Jawadi for
helpful comments and discussions. We have also beneted from the comments
of participants of the 3rd International Conference on Computation and
Financial Econometrics (CFE) in October 2009 in Limassol, Cyprus; of the
1st International Symposium on Computational Economics and Finance
(ISCEF) in February 2010 in Sousse, Tunisia; of the 9th German Open
Conference on Probability and Statistics (GOCPS) in March 2010 in Leipzig,
Germany; and of the 16th Annual Conference on Computing in Economics
and Finance (CEF) in July 2010 in London, UK.
Appendix. Euro Area Monthly Indicators

We provide in Table 1 the list of the monthly economic indicators used in
this study for the computation of GDP using bridge equations.
176
The Bridge Equations
We specify the bridge equations we use; details can be found in Diron

(2008). Let us dene Yt as Y t log GDPt log GDPt1 100, where
GDPt is GDP at time t. The nal GDP Yt used in the chapter is the mean of
the eight values computed below:
(1) EQ1.
Y 1t a10 a11 log X 1t log X 1t1 a12 log X 2t log X 2t1
1 3
a3 X t1 t .
(2) EQ2.
Y 2t a20 a21 log X 1t log X 1t1 a22 log X 2t log X 2t1 a23
4
log X t log X 4t1 a24 log X 5t log X 5t1 t .
(3) EQ3. Y 3t a30 a31 X 7t a32 X 7t1 t .
(4) EQ4. Y 4t a40 a41 X 6t X 6t1 a42 X 3t t .
(5) EQ5. Y 5t a50 a51 X 6t X 6t1 a52 X 9t a53 X 8t t .
10
11
11
6
(6) EQ6. Y 6t a60 a61 log X 10
t2 log X t3 a2 log X t1 log X t2 t .
7
12
12
12
7
7
7
(7) EQ7.
Y t a0 a1 log X t log X t1 a2 log X t2 log X 12
t3
7 7
a3 Y t1 t .
(8) EQ8. Y 8t a80 a81 X 13
t t .
Proofs of Theorem 1 and Corollary 1
We start giving the proof of Theorem 1. We rst establish a preliminary
lemma.
Lemma 1. Under the hypotheses of Theorem 1, either the estimate mn x is
asymptotically unbiased or
Emn x mx Onb ,
(A.1)
with b 1 Qp=d.
Proof of Lemma 1
We denote Bx; r0 fz 2 Rd ; jj x zjj r0 g, with the ball centered at x
with radius r0>0. We characterize the radius r ensuring that k(n) observations
fall in the ball Bx; r; since the function h( ) is p-continuously differentiable,
for a given i the probability qi of an observation xi to fall in Bx; r is
qi Pxi 2 Bx; r
Z
Z
hxi dxi hx:
Bx;r
hxcrd ord ,
Z
dxi
Bx;r
hxi hxdxi
Bx;r
A:2
where c is the volume of the unit ball and d x dx1 dx2 dxd . Thus, qi
qj ord for all i6j. We consider now the k-NN vectors xk and we denote q
the probability that they are in the ball Bx; r, that is, q Pxk 2 Bx; r,
177
then
qi q ord .
(A.3)
If N(r, n), the number of observations falling in the ball Bx; r, is provided for
a given r>0, we characterize r such that k(n) observations fall in Bx; r. We
proceed as follows. We denote Sni all nonordered combinations of the i tuple
indices from (nd) indices, then
ENr; n
nd
X
iPNr; n i
i0
nd
X
i0
nd
X
j 1 ;...;j i 2Sni
nd
X
nd
i
ji
Y
j 1 ;...;j i 2Sni jj 1
qj
nY
d
1 q
efj 1 ;...;j i g
qi 1 qndi
i0
i0
!
qi 1 qndi
qn d1 q qnd ,
A:4

where q and q are, respectively, the smallest and largest probabilities,
qi i 1; . . . ; n d. Thus, we obtain a lower bound for ENr; n. If
ENr; n kn, using Equations (A.2)(A.4), we obtain

r
kn
n d
1=d
(A.5)
Dx,
with Dx 1=hxc1=d .
Now, using Equation (2), we get
X
Ewx X i Y i ,
Emn x
(A.6)
i2Nx
R R
where Y i X i1 . We can remark that Ewx X i Y i Rd R wx xi
R i ; xi dxi dyi . Since f yi ; xi f yi jxi hxi , we obtain Ewx X i Y i
Ryi f y
d
R R wx xi yi f yi jxi hxi dxi dyi . Thus, as soon as the weighting function
w( ) vanishes outside the ball Bx; r,
Z
Z
Ewx X i Y i
wx xi yi f yi jxi dyi hxi dxi
Z
Bx;r
wx xi mxi hxi dxi :
Bx;r
A:7
178
To compute the bias we need to evaluate Emn x mx. We begin to

evaluate:
X
X Z
wx xi mxhxi dxi mxE
wx X i mx.
i2Nx
Bx;r
i2Nx
(A.8)
Then,
Emn x mx
X Z
i2Nx
wx xi mxi mxhxi dxi .
Equation (A.9) holds because

(iv) in Theorem 1). Then,
X Z
jEmn x mxj
i2Nx
(A.9)
Bx;r
i2Nx Bx;r
wx xi hxi dxi 1 (assumption
wx xi ajjxi x jjp hxi dxi .
(A.10)
Bx;r
We get this last expression since the constant a is known and m( ) is

p-continuously differentiable. The inequality (A.10) implies that
X
wx X i .
(A.11)
jEmn x mxj arp E
i2Nx
The relationship in Equation (A.11) holds because jjxi x jjp orp , as

soon as xi 2 Bx; r. Now, both cases can be considered:
When r is very small, then the bias is negligible and Emn x mx.
If the bias is not negligible, using Equations (A.5) and (A.11),
we get
kn
jEmn x mxj a
n d
p=d
Dx p .
(A.12)
If we choose k(n) as in integer part of nQ, and knowing that

k=n d k=n, then jEmn x mxj Onb with b 1 Qp=d.
Proof of Theorem 1
(1) We begin to establish the relationship Equation (5). In the following,
we denote Y i X i1 . We rewrite the left part of Equation (5) as
follows:
Emn x mx2 Varmn x Emn x mx2 .
(A.13)
We rst compute the variance of mn x, considering two cases:

(a) First case: The weights wi, i 1; :::; k, are independent of (Xn). In
that case the variance of mn(x) is equal to
(A.14)
Varmn x A B,
179
kn 2
kn
wi VarY i and B Si1
Sjai wi wj covY i ; Y j . Using assumpwhere Ai1
kn
Sjai jcovY i ; Y j j. This last term is
tion (ii) of Theorem 1, we get jBj Si1
negligible due to the result obtained by Yakowitz (1987) on the sum
kn
knwi 2 vx EY i mx2 .
of covariances. Now, A 1=kn2 Si1
Using the fact that the weights are decreasing with respect to the chosen
distance, wk w1 , we get
1 X
knwk 2 vx EY i mx2 A
kn2 i1
kn
1 X
knw1 2 vx EY i mx2 .

kn2 i1
kn
(A.15)
As soon as kn ! 1; the product knwi converges to 1 in case of uniform

weights, and can be bounded for exponential weights for all i and for all n;
thus, there exist two positive constants c0 and c1 such that Equation (A.15)
becomes
c21 X
c20 X
2
vx
EY

mx

A

vx EY i mx2 .
i
kn2 i1
kn2 i1
kn
kn
(A.16)
where vx VarX n1 jX n x. Using assumption (iv) of Theorem
1, we remark that EY i Emn x. If kn nQ where [ ] corresponds
to the integer part of a real number, then A OnQ thanks to Lemma
1 when n ! 1, and it follows that the relationship (Equation A.14)
becomes
Varmn x OnQ ,
(A.17)
and
Emn x mx2 On2b .
(A.18)
Substituting Equations (A.17) and (A.18) in Equation (A.13), we get

2b Q or Q 2p=2p d and the proof is complete.
(b) Second case: the weights wi, i 1,y,k, depend on (Xn). We use
kn
Varwx X i
again the relationship
(Equation A.14) with A Si1
Pkn
Y i and B i1 Sjai covwx X i Y i ; wx X j Y j . Remarking
that wx X j Y j are f-mixing since (Xj) and (Yj) are f-mixing
(Pagan and Ullah, 1999), then B is negligible from the result
obtained by Yakowitz (1987). We also remark that
180

kn
A Si1
Ewx X i Y i 2 Ewx X i Y i 2 , then
Z
kn Z
X
wx xi 2 y2i f yi ; xi dxi dyi
A
i1
Rd
Z
wx xi yi f yi ; xi dxi dyi

Rd
2 #
(A.19)
.
When k increases, the weights wi decrease, and knwi g where g is a real

constant, then

Z
Z Z
kn Z
g2 X
2
2
y
f
y
;
x
dx
dy
y
f
y
;
x
dx
dy
A
i i
i
i
i
i i
i
i
i
kn2 i1 Rd R
Rd R
g2 X
EY 2i EY i 2 .
kn2 i1
kn
A:20
Under stationary conditions for (Xn) and recalling that Y i X i1 ,

thus,
Equation
(A.20)
becomes
A g2 =knEX 21 EX 1 2 ;
2
A g =knVarX 1 . Finally, expression (A.14) becomes
Varmn x
g2
VarX 1 .
kn
(A.21)
Moreover, when we take kn nQ , Equation (A.21) becomes

Varmn x OnQ .
(A.22)
Using Equations (A.22) and (A.18) in Equation (A.13) gives 2b Q, and

Q 2p=2p d, and the proof is complete.
(2) We prove now the asymptotic normality of mn x. We assume that
the variance sn varmn x exists and is not null, thus
mn x Emn x X wi Y i Ewi Y i
.
sn
sn
i1
kn
(A.23)
To establish the asymptotic normality of mn x, we distinguish three cases

corresponding to the choice of the weighting functions.
(i) The weights are uniform, wi 1=kn, then Equation (A.23)
becomes
mn x Emn x X 1
Zi ,
sn
kn
i1
kn
(A.24)
where Z i Y i EY i =sn . The asymptotic normality of Equation (A.24) is

obtained using Theorem 2.2 given in Peligrad and Utev (1997). To compute
181
the variance,Pwe follow the work by Yakowitz (1987): varmn x

kn
Y i 1=knvarYj X x On21Qp=d ,
then
1=kn2 var i1
Equation (A.24) becomes
X wi Y i Ewi Y i
mn x Emn x p
,
nQ
sn
s
i1
kn
(A.25)
when kn nQ and s2 varYj X x, and the proof is complete.

(ii) The weights wi are real numbers and do not depend on (Xn)n, then
mn x Emn x X
wi Z i ,
sn
i1
kn
(A.26)
where Z i Y i EY i =sn . Now, applying again Theorem 2.2 given in

Peligrad
the asymptotic normality remarking that
Pget
Pkn and Utev (1997), we
kn
wi Z i 0 and Var i1
wi Z i 1. To compute s2n Varmn x,
E i1
we use the stationary condition of the time series (Xn)n, thus
Varmn x
kn
X
w2i VarY i
i1
kn
X
w2i VarY n1 jX n x B2 ,
i1
wherePB is given in Lemma 1. Remarking that 1=kn2

kn 2
wi o1 and
then i1
Varmn x VarY i jX i x B2
kn
X
w2i .
Pkn
(A.27)
i1 knwi
o1,
(A.28)
i1
Pkn
g2 =kn, and kn nQ , we get the result.

P
(iii) Finally, we assume that wi wx X i = K
i1 wx X i ; where
w( ) is a given function. In that latter case, the weights depend on
the process (Xn)n. In the following, we denote by N(i) the order of
the ith neighbor. We rewrite the neighbor indices in an increasing
order such that M1 minfNi; 1 i Kg and Mk minfNi
efMj; 8jokg; 1 i Kg for 2 k K, and K k(n) is the
number of neighbors. We introduce a real triangular sequence
faKi ; 1 i K and aKi a08ig such that
As soon as
Sup
K
K
X
i1
a2Ki o1
2
i1 wi
and
max jaKi j ! 0.
1iK
n!1
(A.29)
Now using the sequences Mj; j 1; . . . ; K and aKi ; 1 i K, we can

rewrite expression (A.23) as
K
mn x Emn x X
aKi Si ,
sn
i1
(A.30)
182
with Si wMi X Mi1 EwMi X Mi1 =aKi sn . The sequence S 2i is uniformly integrable and Si is function only of X j ; j Mi 1; thus, if we
denote Fi , Gi , Fji , and Gji , the sigma algebras generated are fX r gri , fSr gri ,
fX r gjri , and fSr gjri , respectively, then Si 2 FMi1 and Gi FMi1 . For a
1
given integer , we have also G1
n
FnM1 ; since M1oM1
1 M2o Mn oMn 1 Mn 1. Then
sup
Sup
A2G1 ;B2G1
n ;PAa0
sup
jPBjA PBj
Sup
jPBjA PBj.
(A.31)
A2FM1
;B2F1
nM1 ;PAa0
1
Under the f-mixing assumption on (Xn)n, the right-hand part of the

expression (A.31) tends to zero as n-N and then the left-hand part of
Equation (A.31) converges to zero, hence the sequence (Si)i is f-f-mixing.
Moreover, for all i:
!

K
X
mn x
1.
(A.32)
aKi S i var
S i is centered and var
sn
i1
Then, using expressions (A.29)(A.32), and the Theorem 2.2 given in
Peligrad and Utev (1997), we get
mn x Emn x
!D N0; 1
sn
(A.33)
The variance of mx(x) is given by Equation (A.21). The proof of Theorem 1

is complete. We provide now the proof of Corollary 1.
Proof of Corollary 1
From Theorem 1, a condence interval for a given a can be computed,
and has the expression:
z1a=2
mn x Emn x
z1a=2 ,
s^ n
(A.34)
where z1a=2 is the (1a/2) quantile of Student law. Previously, we have

established that the estimate mn x can be biased; thus, the relationship
(Equation A.34) becomes:
mn x B s^ n z1a=2 mx mn x B s^ n z1a=2 .
(A.35)
When the bias is negligible, the corollary is established. If the bias is not
negligible, we can bound it. The bound is obtained using expressions (A.5)
183
and (A.36):
BO
kn
^
n dhxc
!!p=d
,
(A.36)
^
with c pd=2 =Gd 2=2, hx
being an estimate of the density hx.
Introducing this bound in expression (A.35) completes the proof.
References
Akaike, H. (1974). A new look at the statistical model identication. IEEE
Transaction on Automatic Control 19, 716723.
Bafgi, A.R., Golinelli, R. and Parigi, G. (2004). Bridge models to forecast the euro
area GDP. International Journal of Forecasting 20, 447460.
Becker, R.A., Chambers, J.M. and Wilks, A.R. (1988). The New S Language.
New York: Chapman and Hall.
Bernanke, B.S. and Boivin, J. (2003). Monetary policy in a data-rich environment.
Journal of Monetary Economics 50, 525546.
Box, G.E.P. and Jenkins, G.M. (1970). Time Series Analysis: Forecasting and
Control. New York: Holden Day.
Carrasco, M. and Chen, X. (2002). Mixing and moment properties of various
GARCH and stochastic volatility models. Econometric Theory 18, 1739.
Clements, M.P. and Krolzig, H.M. (1998). A comparison of the forecast
performance of Markov-switching and threshold autoregressive models of US
GNP. Econometrics Journal 1, C47C75.
Collomb, G. (1984). Nonparametric time series analysis and prediction: Uniform
almost sure convergence of the window and k-NN autoregression estimates.
Mathematic Operational Statistics, Series Statistics 16, 297307.
Darne, O. (2008). Using business survey in industrial and services sector to nowcast
GDP growth: The French case. Economics Bulletin 3 (32), 18.
Devroye, L.P. (1982). Necessary and sufcient conditions for the pointwise
convergence of nearest neighbor regression function estimates. Z. Wahrscheinlichkeneitstheorie Verw Gebiete 61, 467481.
Devroye, L.P. and Gyor, L. (1985). Nonparametric Density Estimation: The L1
View. New York: Wiley.
Diron, M. (2008). Short-term forecasts of euro area real GDP growth: An assessment
of real-time performance based on vintage data. Journal of Forecasting 27, 371390.
Donoho, D.L. and Johnstone, I.M. (1992). Minimax estimation via wavelet
shrinkage. Department of Statistics, Stanford University, Technical Report 402.
Engle, R.F. and Granger, C.W.J. (1987). Co-integration and error correction:
Representation, estimation, and testing. Econometrica 55, 251276.
Ferrara, L., Guegan, D. and Rakotomarolahy, P. (2010). GDP nowcasting with
ragged-edge data: A semi-parametric modeling. Journal of Forecasting 29, 186199.
184
Forni, M., Giannone, D., Lippi, M. and Reichlin, L. (2005). Opening the black box:
Structural factor models with large cross-sections. European Central Bank
Working Paper No. 571.
Friedman, J.H. (1988). Multivariate adaptive regression splines (with discussion).
Annals of Statistics 19, 1141.
Guegan, D. (1983). Une condition dergodicite pour des mode`les ilineaires a` temps
discret. CRAS Serie 1, 297301.
Guegan, D. and Huck, N. (2005). On the use of nearest neighbors in nance. Revue
de Finance 26, 6786.
Guegan, D. and Rakotomarolahy, P. (2010). A short note on the nowcasting and the
forecasting of euro-area GDP using non-parametric techniques. Economics
Bulletin 30 (1), 508518.
Gupta, R. (2006). Forecasting the South African economy with VARs and VECMs.
South African Journal of Economics 74, 611628.
Hamilton, J.D. (1989). A new approach to the economic analysis of nonstationary
time series and the business cycle. Econometrica 57, 357384.
Han, J., Fu, Y., Wang, W., Koperski, K. and Zaine, O. (1997). DMQL: A Data
Mining Query Language for Relational Databases. Vancouver: WP Simon Fraser
University.
Hardle, W., Muller, M., Sperlich, S. and Werwatz, A. (2004). Non-Parametric and
Semi-Parametric Models. New York: Springer Verlag.
Hart, J.D. (1994). Nonparametric Smoothing and Lack-of-Fit Tests. New York:
Springer-Verlag.
Hoover, K.D. and Perez, S.J. (1999). Data mining reconsidered: Encompassing and
the general-to-specic approach to specication search. Econometrics Journal 2,
167191.
Kapetanios, G. and Marcellino, M. (2006). A parametric estimation method for
dynamic factors models of large dimensions. IGIER Working Paper No. 305.
Koenig, E.F., Dolmas, S. and Piger, J. (2003). The use and abuse of real-time data in
economic forecasting. The Review of Economics and Statistics 85, 618628.
Kuan, C.M. and White, H. (1994). Articial neural networks: An econometric
perspective. Econometric Reviews 13, 191.
Litterman, R.B. (1986). Forecasting with Bayesian vector autoregressions ve years
of experience. Journal of Business Economic Statistics 4, 2538.
Mack, Y.P. (1981). Local properties of k-NN regression estimates. SIAM Journal on
Algebraic and Discrete Methods 2, 311323.
Mizrach, B. (1992). Multivariate nearest-neighbor forecasts of EMS exchange rates.
Journal of Applied Econometrics 7, S151S163. Supplement: Special Issue on
Nonlinear Dynamics and Econometrics.
Ouyang, D., Li, D. and Li, Q. (2006). Cross-validation and nonparametric k nearest
neighbor estimation. Econometrics Journal 9, 448471.
Nowman, B. and Saltoglu, B. (2003). Continuous time and nonparametric modeling
of U.S. interest rate models. International Review of Financial Analysis 12, 2534.
Pagan, A. and Ullah, A. (1999). Nonparametric Econometrics. Cambridge, UK:
Cambridge University Press.
Peligrad, M. and Utev, S.A. (1997). Central limit theorem for linear processes. The
Annals of Probability 25, 443456.
185
Pena, D., Tiao, G.C. and Tsay, R.S. (2003). A course in time series analysis. New
York: Wiley Series in Probability and Statistics.
Prakasa Rao, B.L.S. (1983). Nonparametric Functional Estimation. Orlando, FL:
Academic press.
Runstler, G. and Sedillot, F. (2003). Short-term estimates of euro area real GDP by
means of monthly data. European Central Bank Working Paper No. 276.
Schumacher, C. and Breitung, J. (2008). Real-time forecasting of German GDP
based on a large factor model with monthly and quarterly data. International
Journal of Forecasting 24, 368398.
Schwartz, G. (1978). Estimating the dimension of ARIMA models. Annals of
Statistics 6, 461464.
Silverman, B.W. (1986). Density Estimation for Statistics and Data Analysis. London:
Chapman and Hall.
Sims, C.A. (1980). Macroeconomics and reality. Econometrica 48, 148.
Smets, F. and Wouters, R. (2004). Forecasting with a Bayesian DSGE model: An
application to the euro area. Journal of Common Market Studies 42, 841867.
Stock, J.H. and Watson, M.W. (2002). Macroeconomic forecasting using diffusion
indexes. Journal of Business & Economic Statistics 20, 147162.
Stone, C. (1977). Consistent nonparametric regression. Annals of Statistics 5,
595645.
Stute, W. (1984). Asymptotic normality of nearest neighbor regression function
estimates. Annals of Statistics 12, 917926.
Tkacz, G. and Hu, S. (1999). Forecasting GDP growth using articial neural
networks. Bank of Canada Working Paper No. 99.
Tong, H. (1990). Non-Linear Time Series: A Dynamical System Approach. Oxford,
UK: Oxford University Press.
Yakowitz, S. (1987). Nearest neighbors method for time series analysis. Journal of
Time Series Analysis 8, 235247.
Yatchew, A.J. (1998). Nonparametric regression techniques in economics. Journal of
Economic Literature 36, 669721.
This article has been cited by:
1. Volkan lke, Afsin Sahin, Abdulhamit Subasi. 2016. A comparison of time series
and machine learning models for inflation forecasting: empirical evidence from the
USA. Neural Computing and Applications . [CrossRef]

Methods For Forecasting

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Methods For Forecasting

Transféré par

Droits d'auteur :

Formats disponibles

Nonlinear Modeling of Economic and Financial TimeSeries

Chapter 8 Alternative Methods for Forecasting GDP

Users who downloaded this article also downloaded:

Access to this document was granted through an Emerald subscription provided by

About Emerald www.emeraldinsight.com

*Related content and download information correct at time of download.

Alternative Methods for Forecasting GDP

Paris School of Economics, Universite Paris 1 Pantheon-Sorbonne,

Dominique Guegan and Patrick Rakotomarolahy

Alternative Methods for Forecasting GDP

Dominique Guegan and Patrick Rakotomarolahy

Model (1) is a nonlinear regression problem that can be estimated using

Alternative Methods for Forecasting GDP

huge number of indicators and we do not compare here these different

where w( ) is a weighting function associated with neighbors and it is

where Rn corresponds to the distance between x and the further neighbors,

Dominique Guegan and Patrick Rakotomarolahy

The result established in Theorem 1 proves the asymptotic convergence of

where 0rQr1, Q 2p=2p d, and s2 g2 VarX n1 jX n x B2 ;

Alternative Methods for Forecasting GDP

3. Forecasting Euro Area GDP

Dominique Guegan and Patrick Rakotomarolahy

and private sector forecasters commonly use short-term forecasts of real

Alternative Methods for Forecasting GDP

Table 1: Monthly Economic Indicators of Euro Area Used for Forecasting

Industrial Production Index

sample. To solve this issue, we apply the two previous methodologies to

Dominique Guegan and Patrick Rakotomarolahy

variate VAR(1) (Akaike, 1974; Schwartz, 1978). Nevertheless, we need

(1) We embed this data set in a space of dimension d, 2rdr10, getting a

Alternative Methods for Forecasting GDP

We consider exponential weighting function since it reects the local

Dominique Guegan and Patrick Rakotomarolahy

Alternative Methods for Forecasting GDP

Figure 1: Quarterly Observed (in Black) and Forecasted GDP Growth

Dominique Guegan and Patrick Rakotomarolahy

Figure 2: Quarterly Observed (in Black) and Forecasted GDP Growth

Alternative Methods for Forecasting GDP

Appendix. Euro Area Monthly Indicators

Dominique Guegan and Patrick Rakotomarolahy

The Bridge Equations

We specify the bridge equations we use; details can be found in Diron

Alternative Methods for Forecasting GDP

wx xi mxi hxi dxi :

Dominique Guegan and Patrick Rakotomarolahy

To compute the bias we need to evaluate Emn x mx. We begin to

wx xi mxi mxhxi dxi .

Equation (A.9) holds because

wx xi hxi dxi 1 (assumption

wx xi ajjxi x jjp hxi dxi .

We get this last expression since the constant a is known and m( ) is

The relationship in Equation (A.11) holds because jjxi x jjp orp , as

If we choose k(n) as in integer part of nQ, and knowing that

We rst compute the variance of mn x, considering two cases:

Alternative Methods for Forecasting GDP

As soon as kn ! 1; the product knwi converges to 1 in case of uniform

Substituting Equations (A.17) and (A.18) in Equation (A.13), we get

Dominique Guegan and Patrick Rakotomarolahy

wx xi yi f yi ; xi dxi dyi

When k increases, the weights wi decrease, and knwi g where g is a real

Under stationary conditions for (Xn) and recalling that Y i X i1 ,

Moreover, when we take kn nQ , Equation (A.21) becomes

Using Equations (A.22) and (A.18) in Equation (A.13) gives 2b Q, and

Now using the sequences Mj; j 1; . . . ; K and aKi ; 1 i K, we can