Building Neural Network Models For Time Series: A Statistical Approach

Journal of Forecasting
J. Forecast. 25, 49–75 (2006)

Published online in Wiley InterScience
(www.interscience.wiley.com). DOI: 10.1002/for.974
Building Neural Network Models for Time

Series: A Statistical Approach
MARCELO C. MEDEIROS,1 TIMO TERÄSVIRTA2* AND
GIANLUIGI RECH2
1
Department of Economics, Pontifical Catholic University of Rio
de Janeiro, Brazil
2
Department of Economic Statistics, Stockholm School of
Economics, Sweden
ABSTRACT
This paper is concerned with modelling time series by single hidden layer feed-
forward neural network models. A coherent modelling strategy based on sta-
tistical inference is presented. Variable selection is carried out using simple
existing techniques. The problem of selecting the number of hidden units is
solved by sequentially applying Lagrange multiplier type tests, with the aim of
avoiding the estimation of unidentified models. Misspecification tests are
derived for evaluating an estimated neural network model. All the tests are
entirely based on auxiliary regressions and are easily implemented. A small-
sample simulation experiment is carried out to show how the proposed mod-
elling strategy works and how the misspecification tests behave in small
samples. Two applications to real time series, one univariate and the other mul-
tivariate, are considered as well. Sets of one-step-ahead forecasts are con-
structed and forecast accuracy is compared with that of other nonlinear models
applied to the same series. Copyright © 2006 John Wiley & Sons, Ltd.
key words model misspecification; neural computing; nonlinear forecast-

ing; nonlinear time series; smooth transition autoregression
INTRODUCTION
Alternatives to linear models in econometric and time series modelling have increased in popular-
ity in recent years. Nonparametric models that do not make assumptions about the parametric form
of the functional relationship between the variables to be modelled have become more easily appli-
cable due to computational advances. Another class of models, the flexible functional forms, offers
an alternative that in fact also leaves the functional form of the relationship unspecified. While these
models do contain parameters, often a large number of them, the parameters are not globally iden-
tified or, using the statistical terminology, estimable. Identification or estimability, if achieved, is
local at best without additional parameter restrictions. The parameters are not interpretable either as
they often are in parametric models.
* Correspondence to: Timo Teräsvirta, Department of Economic Statistics, Stockholm School of Economics, Box 6501, SE-
113 83 Stockholm, Sweden. E-mail: timo.terasvirta@hhs.se
Copyright © 2006 John Wiley & Sons, Ltd.

50 M. C. Medeiros, T. Teräsvirta and G. Rech
The artificial neural network (NN) model is a prominent example of such a flexible functional
form. It has found applications in a number of fields, including economics. Kuan and White (1994)
surveyed the use of NN models in (macro)economics, and several financial applications appeared in
a recent special issue of IEEE Transactions on Neural Networks (Abu-Mostafa et al., 2001). The
use of the NN model in applied work is generally motivated by a mathematical result stating that
under mild regularity conditions, a relatively simple NN model is capable of approximating any
Borel-measurable function to any given degree of accuracy; see, for example, Fine (1999) and the
references therein. Such an approximator would still contain a finite number of parameters. How to
specify such a model, that is, how to find the right combination of parameters and variables, is a
central topic in the NN literature (Fine, 1999). Many popular specification techniques are ‘general-
to-specific’ or ‘top-down’ procedures: the investigator begins with a large model and applies appro-
priate algorithms to reduce the number of parameters using a predetermined stopping rule. Such
algorithms usually do not rely on statistical inference.
In this paper, we propose a coherent modelling strategy for single hidden-layer feedforward NN
time series models. The models discussed here are univariate, but adding exogenous regressors to
them does not pose problems. The difference between our strategy and the general-to-specific
approaches is that ours works in the opposite direction, from specific to general. We begin with a
small model and expand that according to a set of predetermined rules. The reason for this is that
we view our NN model as a statistical nonlinear model and apply statistical inference to the problem
of specifying the model or, as the NN experts express it, finding the network architecture. We shall
argue in the paper that proper statistical inference is not available if we choose to proceed from large
models to smaller ones, from general to specific. Our ‘bottom-up’ strategy builds partly on early
work by Teräsvirta and Lin (1993). More recently, Anders and Korn (1999) presented a strategy that
shares certain features with our procedure. Swanson and White (1995, 1997a,b) also developed and
applied a specific-to-general strategy that deserves mention here. Balkin and Ord (2000) proposed
an inference-based method for selecting the number of hidden units in the NN model. Refenes and
Zapranis (1999) developed a computer-intensive strategy based on statistics to select the variables
and the number of hidden units of the NN model. Our aim has been to develop a strategy that min-
imizes the amount of computation required to reach the final specification and, furthermore, con-
tains an in-sample evaluation of the estimated model. The proposed methodology is based on
auxiliary regressions and its implementation is straightforward. Moreover, selecting the right com-
bination of variables and parameters is of great importance when forecasting from NN models is
considered, as these models have a natural tendency to overfit. Usually, an over-parametrized NN
model will have a very good performance in-sample but will perform poorly in an out-of-sample
forecasting exercise. We shall consider the differences between our strategy and the others men-
tioned here in later sections of the paper.
The plan of the paper is as follows. The next section describes the model and a statistical inter-
pretation of it. A model specification strategy, consisting of specification, estimation and evaluation
of the model, is described in the third section. The results concerning a Monte-Carlo experiment are
reported in the fourth section. Two applications with real data sets, as well as a review of other appli-
cations, are presented in the fifth section. A final section contains concluding remarks.
THE AUTOREGRESSIVE NEURAL NETWORK MODEL
The autoregressive neural network (AR-NN) model is defined as
Copyright © 2006 John Wiley & Sons, Ltd. J. Forecast. 25, 49–75 (2006)
DOI: 10.1002/for
Neural Network Models for Time Series 51
h
yt = G(x t ; y ) + e t = a ¢ x˜ t + Â l i F(w˜ i¢x t - b i ) + e t (1)
i =1
where G(xt; y) is a nonlinear function of the variables xt with parameter vector y Œ ˙(q+2)h+q+1 defined
as y = [a ¢, l1, . . . , lh, w̃ 1¢, . . . , w̃ h¢, b1, . . . , bh]¢. The vector x̃t Œ ˙q+1 is defined as x̃t = [1, x¢t]¢,
where xt Œ ˙q is a vector of lagged values of yt and/or some exogenous variables. The function
F(w̃ i¢xt - bi), often called the activation function, is the logistic function
-1
F(w˜ i¢x t - b i ) = (1 + e - (w˜ i¢xt - bi ) ) (2)
where w̃ i = [w̃ 1i, . . . , w̃ qi]¢ Œ ˙q and bi Œ ˙, and the linear combination of these functions in (1)
forms the so-called hidden layer. Model (1) with (2) does not contain lags of et and is therefore called
a feedforward NN model. For other choices of the activation function, see Chen et al. (2001). Fur-
thermore, {et} is a sequence of independently normally distributed random variables with zero mean
and variance s2. The nonlinear function F(w̃i¢xt - bi) is usually called a hidden neuron or a hidden
unit. The normality assumption enables us to define the log-likelihood function, which is required
for the statistical inference we need, but it can be relaxed.
Certain special cases of (1) are of interest. When xt = yt-d in F, model (1) becomes a multiple
logistic smooth transition autoregressive (MLSTAR) model with h + 1 regimes in which only the
intercept changes according to the regime. The resulting model is expressed as
h
yt = a ¢ x̃ t + Â l i F(g i ( yt - d - ci )) + e t (3)
i =1
where gi = w̃ 1i and ci = bi/w̃ 1i. When h = 1, equation (3) defines a special case of an ordinary LSTAR
model considered in Teräsvirta (1994). When gi Æ •, i = 1, . . . , h, model (3) becomes a self-
exciting threshold autoregressive (SETAR) model with a switching intercept and h + 1 regimes.
An AR-NN model can thus be interpreted either as a semiparametric approximation to any Borel-
measurable function or as an extension of the MLSTAR model where the transition variable can be
a linear combination of stochastic variables. We should, however, stress the fact that model (1) is,
in principle, neither globally nor locally identified. Three characteristics of the model imply non-
identifiability. The first one is the exchangeability property of the AR-NN model. The value in the
likelihood function of the model remains unchanged if we permute the hidden units. This results in
h! different models that are indistinguishable from each other and in h! equal local maxima of the
log-likelihood function. The second characteristic is that in (2), F(x) = 1 - F(-x). This yields two
observationally equivalent parametrizations for each hidden unit. Finally, the presence of irrelevant
hidden units is a problem. If model (1) has hidden units such that li = 0 for at least one i, the param-
eters w̃ i and bi remain unidentified. Conversely, if w̃ i = 0 then li and bi can take any value without
the value of the likelihood function being affected.
The first problem is solved by imposing, say, the restrictions b1 £ . . . £ bh or l1 ≥ . . . ≥ lh. The
second source of under-identification can be circumvented, for example, by imposing the restrictions
w̃ 1i > 0, i = 1, . . . , h. To remedy the third problem, it is necessary to ensure that the model contains
no irrelevant hidden units. The difficulty is dealt with by applying statistical inference in the model
specification; see the next section. For further discussion of the identifiability of NN models see, for
example, Hwang and Ding (1997) and the references therein.
DOI: 10.1002/for
For estimation purposes it is often useful to reparametrize the logistic function (2) as
F(g i (w i¢x t - ci )) = (1 + e -g i ( w i¢xt -ci ) )

-1 (4)
where gi > 0, i = 1, . . . , h, and ||wi|| = 1 with

q
w i1 = 1 - Â w ij2 > 0, i = 1, . . . , h (5)
j =2
The parameter vector y of model (1) becomes y = [a¢, l1, . . . , lh, g1, . . . , gh, w12, . . . , w1q, . . . ,
wh2, . . . , whq, c1, . . . , ch]¢. In this case the first two identifying restrictions discussed above can be
defined as, first, c1 £ . . . £ ch or l1 ≥ . . . ≥ lh and, second, gi > 0, i = 1, . . ., h.
STRATEGY FOR BUILDING AR-NN MODELS
Three stages of model building

As mentioned in the Introduction, our aim is to construct a coherent strategy for building AR-NN
models using statistical inference. The structure or architecture of an AR-NN model has to be deter-
mined from the data. We call this stage specification of the model, and it involves two sets of deci-
sion problems. First, the lags or variables to be included in the model have to be selected. Second,
the number of hidden units has to be determined. Choosing the correct number of hidden units is
particularly important as selecting too many neurons yields an unidentified model. In this work, the
lag structure or the variables included in the model are determined using well-known variable selec-
tion techniques. The specification stage of NN modelling also requires estimation because we suggest
choosing the hidden units sequentially. After estimating a model with h hidden units we shall test it
against the one with h + 1 hidden units and continue until the first acceptance of a null hypothesis.
What follows thereafter is evaluation of the final estimated model to check if the final model is ade-
quate. NN models are typically only evaluated out-of-sample, but in this paper we also suggest the
use of in-sample misspecification tests for the purpose. Similar tests are routinely applied in evalu-
ating STAR models (Eitrheim and Teräsvirta, 1996), and in this work we adapt them to the AR-NN
models. All this requires consistency and asymptotic normality for the estimators of parameters of
the AR-NN model, conditions for which can be found in Trapletti et al. (2000).
Variable selection
The first step in our model specification is to choose the variables for the model from a set of poten-
tial variables (lags in the pure AR-NN case). Several nonparametric variable selection techniques
exist (Tschernig and Yang, 2000; Vieu, 1995; Tjøstheim and Auestad, 1994; Yao and Tong, 1994;
Auestad and Tjøstheim, 1990), but they are computationally very demanding, in particular when the
number of observations is not small. In this paper variable selection is carried out by linearizing the
model and applying well-known techniques of linear variable selection to this approximation. This
keeps computational cost to a minimum. For this purpose we adopt the simple procedure proposed
in Rech et al. (2001). Their idea is to approximate the stationary nonlinear model by a polynomial
of sufficiently high order. Adapted to the present situation, the first step is to approximate function
G(xt; y) in (1) by a general kth-order polynomial. By the Stone–Weierstrass theorem, the approxi-
DOI: 10.1002/for
mation can be made arbitrarily accurate if some mild conditions, such as the parameter space y being
compact, are imposed on function G(xt; y). Thus the AR-NN model, itself a universal approxima-
tor, is approximated by another function. This yields
q q q q
G(x t ; y ) = p ¢ x˜ t + Â Â q j1 j2 x j1,t x j 2 ,t + . . . + Â ... Âq j1 ... jk x j1 ,t . . . x jk ,t + R(x t ; y ) (6)
j1 =1 j2 = j1 j1 =1 jk = jk -1
where R(xt; y) is the approximation error that can be made negligible by choosing k sufficiently
high. The q’s are parameters, and p Œ ˙q+1 is a vector of parameters. The linear form of the approx-
imation is independent of the number of hidden units in (1).
In equation (6), every product of variables involving at least one redundant variable has the coef-
ficient zero. The idea is to sort out the redundant variables by using this property of (6). In order to
do that, we first regress yt on all variables on the right-hand side of equation (6) assuming R(xt; y)
and compute the value of a model selection criterion (MSC), AIC or SBIC for example. After doing
that, we remove one variable from the original model and regress yt on all the remaining terms in
the corresponding polynomial and again compute the value of the MSC. This procedure is repeated
by omitting each variable in turn. We continue by simultaneously omitting two regressors of the
original model and proceed in that way until the polynomial is a function of a single regressor and,
finally, just a constant. Having done that, we choose the combination of variables that yields the
lowest value of the MSC. This amounts to estimating Sqi=1 (qi) + 1 linear models by ordinary least
squares (OLS). Note that by following this procedure, the variables for the whole NN model are
selected at the same time. Rech et al. (2001) showed that the procedure works well already in small
samples when compared to well-known nonparametric techniques. Furthermore, it can be applied
successfully even in large samples when nonparametric model selection becomes computationally
infeasible. Another alternative, using equation (6) as a starting point, is to apply the PcGets proce-
dure; see Krolzig and Hendry (2001) for details.
Parameter estimation
As selecting the number of hidden units requires estimation of neural network models, we now turn
to this problem. A large number of algorithms for estimating the parameters of a NN model are avail-
able in the literature. In this paper we instead estimate the parameters of our AR-NN model by
maximum likelihood, making use of the assumptions made previously on et. The use of maximum
likelihood or quasi-maximum likelihood makes it possible to obtain an idea of the uncertainty in the
parameter estimates through (asymptotic) standard deviation estimates. This is not possible by using
the above-mentioned algorithms. It may be argued that maximum likelihood estimation of neural
network models is most likely to lead to convergence problems, and that penalizing the log-
likelihood function one way or the other is a necessary precondition for satisfactory results. Two
things can be said in favour of maximum likelihood here. First, in this paper model building pro-
ceeds from small to large models, so that estimation of unidentified or nearly unidentified models,
a major reason for penalizing the log-likelihood, is avoided. Second, the starting values of the param-
eter estimates are carefully chosen, and we discuss the details of this later.
The AR-NN model is similar to many linear or nonlinear time series models in that the informa-
tion matrix of the log-likelihood function is block diagonal such that we can concentrate the likeli-
hood and first estimate the parameters of the conditional mean. Conditional maximum likelihood is
thus equivalent to nonlinear least squares. The maximum likelihood estimator of the parameters of
the conditional mean equals
DOI: 10.1002/for
T
1
ŷ = argmax QT (y ) = - argmin Â qt (y )
y 2 y t =1
where qt(y) = (yt - G(xt; y)).2 Under standard regularity conditions, the maximum likelihood esti-
mator ỹ is almost surely consistent for y and
(
T 1 2 (ŷ - y ) Æ N 0, - plim A(y )
TÆ•
-1
) (7)
1 ∂ 2 QT (y )
where A(y ) = see Trapletti et al. (2000).
s 2T ∂ y ∂ y ¢ ;
In this paper, we apply the heteroskedasticity-robust large-sample estimator of the covariance
matrix of ŷ (White, 1980)
T -1 T T -1
Ê ˆ Ê ˆÊ ˆ
Bˆ (yˆ ) = Â hˆ t hˆ ¢t eˆ t2 hˆ t hˆ ¢t Â hˆ t hˆ ¢t
¯ ËÂ
(8)
Ë t =1 t =1
¯ Ë t =1
¯
∂ qt (y )
where hˆ t = , and êt is the residual. In the estimation, the use of algorithms such as
∂ y y =yˆ
the Broyden–Fletcher–Goldfarb–Shanno (BFGS) or the Levenberg–Marquardt algorithms is strongly
recommended. See, for example, Bertsekas (1995) for details about optimization algorithms or Fine
(1999, chapter 5) for ones especially applied to the estimation of NN models.
Concentrated maximum likelihood

In order to reduce the computational burden we can apply concentrated maximum likelihood to esti-
mate y as follows. Consider the ith iteration and rewrite model (1) as
y = Z(f )q + e (9)
where y¢ = [y1, y2, . . . , yT], e¢ = [e1, e2, . . . , eT], q¢ = [a¢, l1, . . . , lh] and
Ê x˜ 1¢ F(g 1 (w 1¢x1 - c1 )) L F(g h (w h¢ x1 - ch )) ˆ

Z(f ) = Á M M O M ˜
Á ˜
Ë x T¢ F(g 1 (w 1¢x T - c1 )) L F(g h (w h¢ x T - ch ))¯
with f = [g1, . . . , gh, w¢1, . . . , w¢h, c1, . . . , ch]¢. Assuming f fixed (the value is obtained from the pre-
vious iteration), the parameter vector q can be estimated analytically by
-1
(
q̂ = Z(f )¢ Z(f ) ) Z(f )¢ y (10)
The remaining parameters f are estimated conditionally on q by applying the Levenberg–Marquardt

algorithm, which completes the ith iteration. This form of concentrated maximum likelihood was
proposed by Leybourne et al. (1998) in the context of STAR models and it substantially reduces the
DOI: 10.1002/for
dimensionality of the iterative estimation problem. Numerical issues such as the choice of starting
values are discussed in the Appendix.
Determining the number of hidden units

The number of hidden units included in a NN model is usually determined from the data. A popular
method for doing that is pruning, in which a model with a large number of hidden units is estimated
first, and the size of the model is subsequently reduced by applying an appropriate technique
such as cross-validation. Another technique used in this connection is regularization, which may be
characterized as penalized maximum likelihood or least squares applied to the estimation of
neural network models. For discussion see, for example, Fine (1999, pp. 215–221). Bayesian regu-
larization, based on selecting a prior distribution for the parameters, may serve as an example.
As discussed in the Introduction, another possibility is to begin with a small model and sequen-
tially add hidden units to the model, for discussion see, for example, Fine (1999, pp. 232–233),
Anders and Korn (1999), or Swanson and White (1995, 1997a,b). The decision of adding another
hidden neuron is often based on the use of an MSC or cross-validation. This has the following draw-
back. Suppose the data have been generated by an AR-NN model with h hidden units. Applying an
MSC to decide whether or not another hidden unit should be added to the model requires estima-
tion of a model with h + 1 hidden neurons. In this situation, however, the larger model is not iden-
tified and its parameters cannot be estimated consistently. This is likely to cause numerical problems
in maximum likelihood estimation. Besides, even when convergence is achieved, lack of identifica-
tion causes a severe problem in interpreting the MSC. The NN model with h hidden units is nested
in the model with h + 1 units. A typical MSC comparison of the two models is then equivalent to a
likelihood ratio test of h units against h + 1 ones, see, for example, Teräsvirta and Mellin (1986) for
discussion. The choice of MSC determines the (asymptotic) significance level of the test. But then,
when the larger model is not identified under the null hypothesis, the likelihood ratio statistic does
not have its customary asymptotic c2 distribution when the null holds. For more discussion of the
general situation of a model only being identified under the alternative hypothesis, see, for example,
Davies (1977, 1987) and Hansen (1996).
We shall also select the hidden units sequentially starting from a small model, in fact from a linear
one, but circumvent the identification problem in a way that enables us to control the significance
level of the tests in the sequence and thus also the overall significance level of the procedure. In car-
rying out the sequence of tests we simply assume that the smaller (null) model and thus the corre-
sponding log-likelihood function are always correctly specified. This is a standard assumption when
the null model is a linear model, but we retain it even when our null model is already an AR-NN
model. Admittedly, from the point of view of application, neural network models are generally
regarded as approximations to a true relationship between the variables involved; for discussion see,
for example, Anders et al. (1998). Since the only aim of testing is to find a parsimonious but ade-
quate AR-NN model to serve as an approximation to the unknown true relationship for forecasting,
we do not consider the assumption of correct specification for this purpose to be harmful or mis-
leading. Following Teräasvirta and Lin (1993) we derive a test that is repeated until the first non-
rejection of the null hypothesis. Assume now that our AR-NN model (1) contains h + 1 hidden units
and write it as follows:
h
yt = a ¢ x̃ t + Â l i F(g i (w i¢x t - ci )) + l h +1 F(g h +1 (w h¢ +1x t - ch +1 )) + e t (11)
i =1
DOI: 10.1002/for
Assume further that we have not rejected the hypothesis of model (11) containing h hidden units
and want to test for the (h + 1)th hidden unit. The appropriate null hypothesis is
H 0 :g h+1 = 0 (12)
whereas the alternative is H1 : gh+1 > 0. Under (12), the (h + 1)th hidden unit is identically equal to a
constant and merges with the intercept in the linear unit.
We assume that under (12) the parameters of (11) can be estimated consistently. Model (11) is
only identified under the alternative, which means that the standard asymptotic inference is not avail-
able. This problem is circumvented as in Luukkonen et al. (1988) by expanding the (h + 1)th hidden
unit into a Taylor series around the null hypothesis (12). The order of expansion is a compromise
between a small approximation error (high order) and availability of data (short time series neces-
sarily imply a relatively low order). Using a third-order Taylor expansion, rearranging and merging
terms results in the following model:
h q q q q q
yt = p ¢ x˜ t + Â l i F(g i (w i¢x t - ci )) + Â Â q ij xi ,t x j ,t + Â Â Â q ijk xi ,t x j ,t x k ,t + e t* (13)
i =1 i =1 j = i i =1 j = i k = j
where e*t = et + lh+1R(xt); R(xt) is the remainder. It can be shown that qij = g 2h+1q̃ij , q̃ij π 0, i = 1,
. . . , q; j = i, . . . , q, and qijk = g 3h+1q̃ijk, q̃ijk π 0, i = 1, . . . , q; j = i, . . . , q, k = j, . . . , q. Thus the null
hypothesis is H¢0 : qij = 0, i = 1, . . . , q; j = i, . . . , q, qijk = 0, i = 1, . . . , q; j = i, . . . , q; k = j, . . . , q.
Note that under H0 : e*t = et, so that the properties of the error process remain unchanged under the
null hypothesis. It is important to stress that the use of a first-order Taylor expansion is not possible
because it is linear in xt. The second-order approximation is not useful either because the logistic
function is odd and thus its second derivative evaluated under the null is zero. Higher-order approx-
imations will introduce many regressors, reducing the degrees of freedom, and will not improve the
statistical properties of the test. Under H0 : gh+1 = 0 and standard regularity conditions, completed with
the assumption E|xt,i|d < •, i = 1, . . . , q, for some d > 6, the LM type statistic
-1 T -1
1 T
ÏT T
Ê
T
ˆ ¸ T
LM = 2 Â eˆ tn t¢ ÌÂ nˆ tnˆ t¢ - Â nˆ t hˆ ¢t Â hˆ t hˆ ¢t Â hˆ tnˆ t¢˝ Â nˆ eˆt t (14)
sˆ t =1 Ó t =1 t =1
Ë t =1
¯ t =1 ˛ t =1
where êt = yt - G(xt; ŷ ),

∂ G(x t ; y ) È ∂ Fˆ1 ∂ Fˆh ˆ ∂ Fˆ1 ∂ Fˆ1
hˆ t = = Íx˜ t¢, Fˆ1 , . . . , Fˆh , lˆ 1 , . . . , lˆ h , l1 , . . . , lˆ 1 , ...,
∂ y ¢ y =yˆ Î ∂g 1 ∂g h ∂ w˜ 12¢ ∂ w˜ 1¢q
∂ Fˆh ∂ Fˆh ˆ ∂ Fˆ1 ∂ Fˆh ˘ ¢

lˆ h , . . . , lˆ h , l1 , . . . , lˆ h
∂ w˜ h¢2 ¢
∂ w˜ hq ∂c1 ∂ch ˙˚
with F̂i ∫ F(gî(ŵi¢xt - ĉi)) and
∂ Fî -2
= (wˆ i¢x t - cî )[ 2 cosh (gˆ i (wˆ i¢x t - cî ))] , i = 1, . . . , h
∂g i
∂ Fî -2
= g i x˜ j ,t [ 2 cosh(gˆ i (wˆ i¢x t - cî ))] , i = 1, . . . , h; j = 2, . . . , q
∂w˜ ij
∂ Fî -2
= -g i [ 2 cosh (gˆ i (wˆ i¢x t - cî ))] , i = 1, . . . , h
∂ ci
DOI: 10.1002/for
2 3
and vt = [x1,t, x1,t x2,t, . . . , xi,t xj,t, . . . , x1,t, . . . , xi,t xj,t xk,t, . . ., x3h,t], has an asymptotic c2 distribution with
m = q(q + 1)/2 + q(q + 1)(q + 2)/6 degrees of freedom.
The test can also be carried out in stages as follows:
1. Estimate model (1) with h hidden units. If the sample size is small and the model is thus diffi-
cult to estimate, numerical problems in applying the maximum likelihood algorithm may lead to
a solution such that the residual vector is not precisely orthogonal to the gradient matrix of G(xt;
ŷ). This has an adverse effect on the empirical size of the test. To circumvent this problem, we
T
regress the residuals êt on ĥt and compute the sum of squared residuals SSR0 = St=1 ẽ 2t . The new
residuals ẽt are orthogonal to ĥt.
T
2. Regress ẽt on ĥt and n̂t. Compute the sum of squared residuals SSR1 = St=1 n̂ 2t.
2
3. Compute the c statistic.
SSR0 - SSR1
LM chn2 = T
SSR0 (15)
or the F version of the test
( SSR0 - SSR1 ) m
LM Fhn =
SSR1 (T - n - m) (16)
where n = (q + 2)h + p + 1. Under H0, LMchn2 has an asymptotic c2 distribution with m degrees of
freedom and LMFhn is approximately F-distributed with m and T - n - m degrees of freedom.
The following cautionary remark is in order. If any gî, i = 1, . . . , h, is very large, the gradient
matrix becomes near-singular and the test statistic numerically unstable, which distorts the size of
the test. The reason is that the vectors corresponding to the partial derivatives with respect to gi, wi
and ci, respectively, tend to be almost perfectly linearly correlated. This is due to the fact that the
time series of those elements of the gradient resemble dummy variables being constant most of the
time and nonconstant simultaneously. The problem may be remedied by omitting these elements
from the regression in step 2. This can be done without significantly affecting the value of the test
statistic; see Eitrheim and Teräsvirta (1996) for discussion.
Testing zero hidden units against at least one is a special case of the above test. This amounts to
testing linearity, and the test statistic is in this case identical to the one derived for testing linearity
against the AR-NN model in Teräsvirta et al. (1993). A natural alternative to our procedure is the
one first suggested in White (1989) and investigated later in Lee et al. (1993). In order to test
the null hypothesis of h hidden units, one adds q hidden units to model (1) by randomly selecting
the parameters w̃ h+j, bh+j, j = 1, . . . , q. This solves the identification problem as the extra neurons
are observable, and the null hypothesis lh+1 = . . . = lh+q = 0 can be tested using standard inference.
When h = 0, this technique also collapses into a linearity test; see Lee et al. (1993). Simulation
results in Teräsvirta et al. (1993) and Anders and Korn (1999) indicate that the polynomial approx-
imation method presented here compares well with White’s approach, and it is applied in the rest of
this work.
In financial applications, at least in ones to high-frequency data, such as intradaily, daily or even
weekly series, the series typically contain conditional heteroskedasticity. This possibility can be
DOI: 10.1002/for
accounted for by robustifying the tests against heteroskedasticity following Wooldridge (1990). A
heteroskedasticity-robust version of the LM type test, based on the notion of robustifying statistic
(18), can be carried out as follows:
1. As before.
2. Regress n̂t on ĥt and compute the residuals rt.
3. Regress 1 on ẽtrt and compute the sum of squared residuals SSR1.
4. Compute the value of the test statistic
LM cr 2 = T - SSR1 (17)
The test statistic has the same asymptotic c2 null distribution as before.
It should be noticed that in the case of conditional heteroskedasticity, the maximum likelihood
estimates discussed previously are just quasi-maximum likelihood estimates. Under mild regularity
conditions they are still consistent and asymptotically normal.
Evaluation of the estimated model

After a model has been estimated it has to be evaluated. The specification test for determining the
number of hidden units is also a misspecification test of the estimated model and thus an in-sample
evaluation tool. We shall discuss another such test but before doing that we would like to empha-
size the need for checking stationarity. The modelling strategy we propose applies to stationary
variables. The estimated AR-NN model is not stationary when the lag polynomial of the linear com-
ponent contains nonstationary roots. (Note that neural network models without a linear unit trivially
satisfy the stationary condition.) This has an adverse effect on forecasting several periods ahead.
Thus, the roots of the lag polynomial have to be computed, and nonstationary roots must lead to
respecification of the model.
Test of no serial correlation

As already mentioned, the adequacy of the final model is tested by testing the hypothesis of no addi-
tional hidden units. In this subsection we discuss another misspecification test, one for testing the
null hypothesis of no serial correlation in the errors. Rejecting this hypothesis suggests that the errors
contain forecastable structure, which should not be the case in the AR-NN model intended for fore-
casting. The test is an application of the results in Eitrheim and Teräsvirta (1996) and Godfrey (1988,
pp. 112–121) and may be performed after finding the number of hidden units. We assume that the
errors in equation (1) follow an rth-order autoregressive process defined as
e t = p ¢n t + ut (18)
where p¢ = [p1, . . . , pr] is a parameter vector, v¢t = [et-1, . . . , et-r], and ut~NID(0, s2). Consider the
null hypothesis H0 : p = 0 whereas H1 : p π 0. The conditional normal log-likelihood of (1) with (18)
for observation t, given the fixed starting values, has the form
r r 2
1 1 1 Ï ¸
lt = - ln(2p ) - ln s 2 - Ìyt - Â p j yt - j - G(x t ; y ) + Â p j G(x t - j ; y )˝ (19)
2 2 2s 2 Ó j =1 j =1 ˛
DOI: 10.1002/for
The first partial derivatives of the normal log-likelihood for observation t with respect to p and y
are
∂ lt Ê ut ˆ
= { yt - j - G(x t - j ; y )}, j = 1, . . . , r
∂p j Ë s 2 ¯
(20)
∂ lt ut Ï ∂ G(x t , y ) r ∂ G( x t - j ; y ) ¸
= -Ê 2 ˆ Ì - Âp j ˝
∂y Ës ¯Ó ∂y j =1 ∂y ˛
Under the null hypothesis, the consistent estimators of the score are
T
∂ lˆt 1 T T
∂ lˆt 1 T
Â
t =1 ∂ p
=
sˆ 2
Â eˆtnˆ t
t =1
and Â
t =1 ∂ y
=-
sˆ 2
Â eˆ hˆ ,
t =1
t t
H0 H0
∂ G(x t ; yˆ )
where n̂t¢ = [êt-1, . . . , êt-r], êt-j = yt-j - G(xt-j; ŷ), j = 1, . . . , r, hˆ t = , and ŝ 2 = (1/T) St=1
T 2
ê t .
∂y
The LM statistic is (14) with ĥt and n̂ t defined as above, and it has an asymptotic c2 distribution with
r degrees of freedom under the null hypothesis. For details, see Godfrey (1988, pp. 112–121). The
test can be performed in three stages as shown before. It may be pointed out that the Ljung–Box
test or its asymptotically equivalent counterpart, the Box–Pierce test, both recommended for use in
connection with NN models by Zapranis and Refenes (1999), are not available. Their asymptotic
null distribution is unknown when the estimated model is an AR-NN model.
Modelling strategy
At this point we are ready to combine the above statistical ingredients into a coherent modelling
strategy. We first define the potential variables (lags) and select a subset of them, applying the vari-
able selection technique considered earlier. After selecting the variables we select the number of
hidden units sequentially. We begin testing linearity against a single hidden unit as described above
at significance level a. The model under the null hypothesis is simply a linear AR model. If the null
hypothesis is not rejected, the AR model is accepted. In case of a rejection, an AR-NN model with
a single unit is estimated and tested against a model with two hidden units at the significance level
ar, 0 < r < 1. Another rejection leads to estimating a model with two hidden units and testing it
against a model with three hidden neurons at the significance level ar2. The sequence is terminated
at the first nonrejection of the null hypothesis. The significance level is reduced at each step of the
sequence and converges to zero. In our applications later, we use r = 1/2. This way we avoid exces-
sively large models and control the overall significance level of the procedure. An upper bound for
the overall significance level a* may be obtained using the Bonferroni bound. For example, if a =
0.1 and r = 1/2 then a* £ 0.187. Note that if, instead of our LM type test, we apply a model selec-
tion criterion such as AIC or SBIC to this sequence, we in fact use the same significance
level at each step. Besides, the upper bound that can be worked out in the linear case, see, for
example, Teräsvirta and Mellin (1986), remains unknown due to the identification problem men-
tioned above.
In following the above path we have indeed assumed that all hidden neurons contain the variables
that are originally selected to the AR-NN model. Another variant of the strategy is the one in which
the variables in each hidden unit are chosen individually from the set of originally selected vari-
DOI: 10.1002/for
ables. In the present context this may be done, for example, by considering the estimated parame-
ter vector ŵ h of the most recently added hidden neuron, removing the variables whose coefficients
have the lowest t-values and re-estimating the model. Anders and Korn (1999) recommended this
alternative. It has the drawback that the computational burden may become extremely high. Because
of this we suggest another technique that combines sequential testing for hidden units and variable
selection. Consider equation (11). Instead of just testing a single null hypothesis as is done within
(11), we can do the following. First test the null hypothesis involving all variables. Then remove
one variable from the extra unit under test and test the model with h hidden units against this reduced
alternative. Remove each variable in turn and carry out the test. Continue by removing two vari-
ables at a time. Finally, test the model with h neurons against the alternatives in which the (h + 1)th
unit only contains a single variable and an intercept. Find the combination of variables for which
the p-value of the test is minimized. If this p-value is lower than a prescribed value, ‘significance
level’, add the (h + 1)th unit with the corresponding variables to the model. Otherwise accept the
AR-NN model with h hidden units and stop. This way of selecting the variables for each hidden unit
is analogous to the variable selection technique discussed earlier.
Compared to our first strategy, this one adds to the flexibility and on average leads to more par-
simonious models than the other one. On the other hand, as every choice of hidden unit involves a
possibly large number of tests, we do not control the significance level of the overall hidden unit
test. We do that, albeit conditionally on the variables selected, if the set of input variables is deter-
mined once and for all before choosing the number of hidden units.
Evaluation following the estimation of the final model is carried out by subjecting the model to
misspecification tests and controlling stationarity. If the model does not pass the checks, the model
builder has to reconsider the specification. Another way of evaluating the model is out-of-sample
forecasting. As AR-NN models are most often constructed for forecasting purposes, this is impor-
tant. This part of the model evaluation is carried out by saving the last observations in the series for
forecasting and comparing the forecast results with those from at least one benchmark model. The
results are of course conditional on the structure of the model remaining unchanged over the fore-
casting period, which may not necessarily be the case. For more discussion about this situation, see
Clements and Hendry (1999, chapter 2). In our view, out-of-sample and in-sample evaluations of
the estimated model are complementary rather than competing model evaluation techniques.
Discussion and comparisons

It is useful to compare our modelling strategy with other bottom-up approaches available in the lit-
erature. Swanson and White (1995, 1997a,b) apply a model selection criterion as follows. They start
with a linear model, adding potential variables to it until SBIC indicates that the model cannot be
further improved. Then they estimate models with a single hidden unit and select regressors sequen-
tially to it one by one unless SBIC shows no further improvement. Next, Swanson and White add
another hidden unit and proceed by adding variables to it. The selection process is terminated when
SBIC indicates that no more hidden units or variables should be added or when a predetermined
maximum number of hidden units has been reached. This modelling strategy can be termed fully
sequential.
Anders and Korn (1999) essentially adopt the procedure of Teräsvirta and Lin (1993) described
above for selecting the number of hidden units. After estimating the largest model they suggest pro-
ceeding from general-to-specific by sequentially removing those variables from hidden units whose
parameter estimates have the lowest (t-test) p-values. Note that this presupposes parameterizing the
hidden units as in (2), not as in (4) and (5).
DOI: 10.1002/for
Balkin and Ord (2000) select the ordered variables (lags) sequentially using a linear model and a
forward stepwise regression procedure. If the F-test statistic of adding another lag obtains a value
exceeding 2, this lag is added to the set of input variables. The number of variables selected also
serves as a maximum number of hidden units. The authors suggest estimating all models from the
one with a single hidden unit up to the one with the maximum number of neurons. The final choice
is made using the generalized cross-validation criterion of Golub et al. (1979). The model for which
the value of this model selection criterion is minimized is selected.
Refenes and Zapranis (1999) propose adding hidden units into the model sequentially. The number
of units, however, is selected only after adding all units up to a predetermined maximum number,
so the procedure is not genuinely sequential. The choice is made by applying the network informa-
tion criterion (Murata et al., 1994). The model is then pruned by removing redundant variables from
the neurons and re-estimating the model. Unlike the others, Refenes and Zapranis (1999) underline
the importance of misspecification testing which also forms an integral part of our modelling pro-
cedure. They suggest, for example, that the hypothesis of no error autocorrelation should be tested
by the Ljung–Box or the asymptotically equivalent Box–Pierce test. Unfortunately, these tests do
not have their customary asymptotic null distribution when the estimated model is an AR-NN model
instead of a linear autoregressive one.
Of these strategies, the Swanson and White one is computationally the most intensive, as the
number of steps involving an estimation of a NN model is large. Our procedure is in this respect the
least demanding. The difference between our scheme and the Anders and Korn one is that in our
strategy, variable selection does not require estimation of NN models because it is wholly based on
LM type tests (the model is only estimated under the null hypothesis). Furthermore, there is a pos-
sibility of omitting certain potential variables before even estimating neural network models.
Like ours, the Swanson and White strategy is truly sequential: the modeller proceeds by consid-
ering nested models. The difference lies in how to compare two nested models in the sequence.
Swanson and White apply SBIC whereas Anders and Korn and we use LM type tests. The problems
with the former technique have been discussed above. The problem of estimating unidentified models
is still more acute in the approaches of Balkin and Ord and Refenes and Zapranis. Because these
procedures require the estimation of NN models up to one containing a predetermined maximum
number of hidden units, several estimated models may thus be unidentified. The problem is even
more serious if statistical inference is applied in subsequent pruning as the selected model may also
be unidentified. The probability of this happening is smaller in the Anders and Korn case, in par-
ticular when the sequence of hidden unit tests has gradually decreasing significance levels.
MONTE-CARLO STUDY
In this section we report results from two Monte Carlo experiments. The purpose of the first one is
to illustrate some features of the NN model selection strategy described earlier and compare it with
the alternative in which model selection is carried out using an appropriate model selection crite-
rion. In the second experiment, the performance of our misspecification test is considered. In both
experiments we make use of the following model:
yt = 0.10 + 0.75 yt -1 - 0.05 yt - 4
+ 0.80 F(2.24(0.45 yt -1 - 0.89 yt - 4 + 0.09))
(21)
- 0.70 F(1.12(0.44 yt -1 + 0.89 yt - 4 + 0.35)) + e t
e t = ke t -1 + ut , ut ~ NID(0, s 2 )
DOI: 10.1002/for
In the first experiment, k = 0. The number of observations in the first experiment is either 200 or
1000, in the second one we report results for 100 observations. In every replication, the first 500
observations are discarded to eliminate the initialization effects. The number of replications equals
500.
Architecture selection
Results from simulating the modelling strategy can be found in Table I. The table also contains results
on choosing the number of hidden units using SBIC. This model selection criterion was chosen for
the experiment because Swanson and White (1995, 1997a,b) applied it to this problem. In this case
it is assumed that the model contains the correct variables. This is done in order to obtain an idea
of the behaviour of SBIC free from the effects of an incorrectly selected model.
Different results can be obtained by varying the error variance, the size of the coefficients of
hidden units and, in the case of our strategy, the significance levels. In this experiment, the signifi-
cance level is halved at every step, but other choices are of course possible. It seems, at least in the
present experiment, that selecting the variables is easier than choosing the right number of hidden
units. In small samples, there is a strong tendency to choose a linear model but, as can be expected,
nonlinearity becomes more apparent with an increasing sample size. The larger initial significance
level (a = 0.10) naturally leads to larger models on average than the smaller one (a = 0.05). Over-
fitting is relatively rare but the results suggest, again not unexpectedly, that the initial significance
level should be lowered when the number of observations increases. Finally, improving the signal-
to-noise ratio improves the performance of our strategy.
The results of the hidden unit selection by SBIC show that the empirical significance level implied
by it is, at least in this experiment, very low for both T = 200 and T = 1000, although it changes
with the sample size. Compared to our approach, the linear model is still chosen relatively often for
T = 1000 whereas the correct model with two hidden units is not selected at all.
Serial correlation test

The test of no error autocorrelation is simulated using model (21) with p = 0, 1 and s = 0.125, 0.25.
The maximum lags in the alternative equal 1, 2 and 4. The size discrepancy plots appear in Figure
1. Again, the test is somewhat conservative for s = 0.25 and less so for s = 0.125. The results of
power simulations do not offer any surprise: the power increases with parameter k. They are thus
omitted to save space but are available upon request.
CASE STUDIES
Example 1: annual sunspot numbers, 1700–2000

In this section we illustrate our modelling strategy by two empirical examples. In the first example
we build an AR-NN model for the annual sunspot numbers over the period 1700–1979 and forecast
with the estimated model up until the year 2001. The series, consisting of the years 1700–2001, was
obtained from the National Geophysical Data Center web page.1 The sunspot numbers are a heavily
modelled nonlinear time series: for a neural network example see Weigend et al. (1992). In this work
we adopt the square-root transformation of Ghaddar and Tong (1981) and Tong (1990, p. 420). The
1
http://www.ngdc.noaa.gov/stp/SOLAR/SSN/ssn.html.
DOI: 10.1002/for
Table I. Outcomes of the experiments of selecting the number of hidden units and the set of input variables
s=1
a = 0.05, r = 1/2
200 observations 1000 observations
Correct Too many Too few SBIC Correct Too many Too few SBIC
variables variables variables variables variables variables
ĥ = 0 405 6 10 483 114 0 0 394
ĥ = 1 73 2 1 17 363 0 0 106
ĥ = 2 3 0 0 0 3 0 0 0
ĥ > 2 0 0 0 0 0 0 0 0
a = 0.10, r = 1/2
ĥ = 0 335 7 0 483 68 0 0 394
ĥ = 1 122 5 0 17 387 0 0 106
ĥ = 2 4 0 0 0 33 0 0 0
ĥ > 2 1 0 0 0 4 0 0 0
s = 0.5
a = 0.05, r = 1/2
ĥ = 0 365 2 0 475 18 0 0 256
ĥ = 1 127 1 0 25 440 0 0 244
ĥ = 2 5 0 0 0 38 0 0 0
ĥ > 2 0 0 0 0 4 0 0 0
a = 0.10, r = 1/2
ĥ = 0 282 0 0 475 4 0 0 256
ĥ = 1 205 1 0 25 438 0 0 244
ĥ = 2 11 0 0 0 54 0 0 0
ĥ > 2 1 0 0 0 4 0 0 0
DOI: 10.1002/for
Table I. Continued
s = 0.125
a = 0.05, r = 1/2
ĥ = 0 116 0 0 423 0 0 0 4
ĥ = 1 360 0 0 77 304 0 0 495
ĥ = 2 23 0 0 0 177 0 0 1
ĥ > 2 1 0 0 0 19 0 0 0
a = 0.10, r = 1/2
ĥ = 0 86 0 0 423 0 0 0 4
ĥ = 1 382 0 0 77 262 0 0 495
ĥ = 2 30 0 0 0 205 0 0 1
ĥ > 2 2 0 0 0 33 0 0 0
Notes: (a) The test sequence starts at significance levels a = 0.05 and 0.10 and sample sizes 200 and 1000 based on 500
replications of model (21) for k = 0 and three different values for s and the same using SBIC. (b) Table entries represent
the number of times a given model is selected as the final specification. (c) ĥ is the estimated number of hidden units. (d)
The cases where the number of variables is correct but the combination is not the correct one appear under the heading ‘Too
few variables’. (e) The results concerning model selection using SBIC do not depend on the value of a.
0.05
0.025
0.045
0.04
0.02
0.035
Nominal Size - Empirical Size
Nominal Size - Empirical Size
0.03 0.015
0.025
0.01
0.02
0.015
0.005
0.01
o o
0 line 0 line
r=1 r=1
0.005 r=2 0 r=2
r=4 r=4
0
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Nominal Size Nominal Size
(a) (b)
Figure 1. Size discrepancy curves of the no error autocorrelation test. Panel (a): k = 0 and s = 0.25. Panel (b):
k = 0 and s = 0.125
DOI: 10.1002/for
20
t
y
10
0
50 100 150 200 250
t
0.6
0.4
F1
0.2
50 100 150 200 250

t
1
F2
0.5
0
50 100 150 200 250
t
Figure 2. Panel (a): transformed sunspot time series, 1700–1979. Panel (b): output of the first hidden unit in
(22). Panel (c): output of the second hidden unit in (22)
Table II. Test of no additional hidden units: minimum p-value of the set
of tests against each null model
Number of hidden units under the null hypothesis

0 1 2
-14 -9
p-value 3 ¥ 10 2 ¥ 10 0.019
transformed observations have the form yt = 2[ (1 + Nt ) - 1], t = 1, . . . , T, where Nt is the original

number of sunspots in the year t. The graph of the transformed series appears in Figure 2.
We use the observations for the period 1700–1979 to estimate the model and the remaining ones
for a forecast evaluation. We begin the AR-NN modelling of the series by selecting the relevant lags
using the variable selection procedure described earlier. We use a third-order polynomial approxi-
mation to the true model. Applying SBIC, lags 1, 2 and 7 are selected whereas AIC yields the lags
1, 2, 4, 5, 6, 7, 8, 9 and 10. We proceed with the lags selected by the SBIC. However, the residuals
of the estimated linear AR model are strongly autocorrelated. The serial correlation is removed by
also including yt-3 in the set of selected variables. When building the AR-NN model we select the
input variables for each hidden unit separately using the specification test described previously.
Linearity is rejected at any reasonable significance level and the p-value of the linearity test mini-
mized with lags 1, 2 and 7 as input variables. The sequence of including hidden units is discontin-
ued after adding the second hidden unit, see Table II, and the final estimated model is
yt = -0.17+ 0.85 yt -1 + 0.14 yt -2 - 0.31 yt -3 + 0.08 yt -7

( 0.83 ) ( 0.09 ) ( 0.12 ) ( 0.06 ) ( 0.05 )
[ (
+ 12.80 ¥ F 0.46 0.29 yt -1 - 0.87 yt -2 + 0.40 yt -7 - 6.68
( 7.18 ) ( 0.23 ) ( -) ( 0.83 ) ( 0.09 ) ( 0.05 )
)]
È ˘ (22)
( 0.48 ) Î ( 8.45 ¥ 10 )
3 ( - ) ( 0. 12
(
+ 2.44 ¥ F Í1.17 ¥10 3 0.83 yt -1 - 0.53 yt -2 - 0.18 yt -7 + 0.38 ˙ + eˆ t
) ( 0. 08 ) ( 7. 18 ) ˚
)
sˆ = 1.89, sˆ sˆ L = 0.70, R 2 = 0.89, pLJB = 1.8 ¥ 10 -7
pARCH (1) = 0.94, pARCH (2) = 0.75, pARCH (3) = 0.90, pARCH (4) = 0.44
DOI: 10.1002/for
where the figures in parentheses below the estimates are standard deviation estimates, ŝ is the resid-
ual standard deviation, ŝ L is the residual standard deviation of the linear AR model, R2 is the deter-
mination coefficient, pLJB is the p-value of the Lomnicki–Jarque–Bera test of normality, and
pARCH( j), j = 1, . . . , 4, is the p-value of the LM test of no ARCH against ARCH of order j. The
estimated correlation matrix of the linear term and the output of the hidden units is
Ê 1 -0.30 0.74 ˆ
Sˆ = Á -0.30 1 -0.19˜ (23)
Á ˜
Ë 0.74 -0.19 1 ¯
It is seen from (23) that there are no redundant hidden units in the model as none of the correlations
is close to unity in absolute value. Figure 2 illuminates the contributions of the two hidden units to
the explanation of yt. The linear unit can only represent a symmetric cycle, so that the hidden units
must handle the nonlinear part of the cyclical variation in the series. It is seen from Figure 2 that
the first hidden unit is activated at the beginning of every upswing, and its values return to zero
before the peak. The unit thus helps explain the very rapid recovery of the series following each
trough. The second hidden unit is activated roughly when the series is obtaining values higher than
its mean. It contributes to characterizing another asymmetry in the sunspot cycle: the peaks and the
troughs have distinctly different shapes, peaks being rounder than troughs. The switches in the value
of the hidden unit from zero to unity and back again are quite rapid (g2 large), which is the cause of
the large standard deviation of the estimate of g2, see the discussion above.
Table III shows the results of the test of no serial correlation described earlier. The results of the
misspecification tests of model (22) indicate no model misspecification. Furthermore, checking the
roots of the AR polynomial we can see that the estimated AR-NN model is stationary.
In order to assess the out-of-sample performance of the estimated model we compare our fore-
casting results with the ones obtained from the two SETAR models, the one reported in Tong (1990,
p. 420) and the other in Chen (1995), a NN model with 10 hidden neurons and the first 9 lags as
input variables, estimated with Bayesian regularization (MacKay, 1992a,b), and a linear autoegres-
sive model with lags selected using SBIC. The SETAR model estimated by Chen (1995) is one in
which the threshold variable is a nonlinear function of lagged values of the time series whereas it is
a single lag in Tong’s model.
Table IV shows the results of the one-step-ahead forecasting for the period 1980–2001. The results,
summarized by the root mean squared error (RMSE) and mean absolute error (MAE) measures, are
quite favourable for our AR-NN model. Turning away from the neural network models, the less than
impressive performance of the SETAR models may raise questions about their feasibility. However,
as Tong (1990, p. 421) has pointed out, these models are at their best in forecasting several years
Table III. Tests of no error autocorrelation in (22)
LM test for qth-order serial correlation

Lag
1 2 3 4 8 12
p-value 0.55 0.61 0.34 0.49 0.47 0.22
DOI: 10.1002/for
Table IV. One-step-ahead forecasts, their root mean square errors and mean absolute errors for the annual
number of sunspots from a set of time series models, for the period 1980–2001
Year Observation AR-NN NN model SETAR model SETAR model AR model

(Tong, 1990) (Chen, 1995)
Forecast Error Forecast Error Forecast Error Forecast Error Forecast Error
1980 154.6 153.4 1.2 136.9 17.7 161.0 -6.4 134.3 20.3 159.8 -5.2
1981 140.4 128.4 12.0 130.5 9.9 135.7 4.7 125.4 15.0 123.3 17.1
1982 115.9 95.8 20.1 101.1 14.8 98.2 17.7 99.3 16.6 99.6 16.3
1983 66.6 76.7 -10.1 88.6 -22.0 76.1 -9.5 85.0 -18.4 78.9 -12.3
1984 45.9 29.8 16.1 45.8 0.1 35.7 10.2 41.3 4.7 33.9 12.0
1985 17.9 21.9 -4.0 29.5 -11.6 24.3 -6.4 29.8 -11.9 29.3 -11.4
1986 13.4 13.5 -0.1 9.5 3.9 10.7 2.7 9.8 3.6 10.7 2.7
1987 29.4 23.7 5.7 25.2 4.2 20.1 9.3 16.5 12.9 23.0 6.4
1988 100.2 86.7 13.5 76.8 23.4 54.5 45.7 66.4 33.8 61.2 38.9
1989 157.6 161.6 -3.9 152.9 4.6 155.8 1.8 121.8 35.8 159.2 -1.6
1990 142.6 159.7 -17.1 147.3 -4.7 156.4 -13.8 152.5 -9.9 175.5 -32.9
1991 145.7 118.2 27.5 121.2 24.5 93.3 52.4 123.7 22.0 119.1 26.6
1992 94.3 98.1 -3.8 114.3 -20.0 110.5 -16.2 115.9 -21.7 118.9 -24.6
1993 54.6 64.8 -10.2 71.0 -16.4 67.9 -13.3 69.2 -14.6 57.9 -3.3
1994 29.9 21.0 8.9 32.9 -3.0 27.0 2.9 35.7 -5.8 29.9 -0.1
1995 17.5 14.9 2.6 19.2 -1.7 18.4 -0.9 18.9 -1.4 17.6 -0.1
1996 8.6 19.2 -10.6 10.2 -1.6 18.1 -9.5 11.6 -3.0 15.7 -7.1
1997 21.5 17.6 3.9 21.3 0.2 12.3 9.2 11.8 9.7 16.0 5.5
1998 64.3 64.6 -0.3 67.6 -3.3 46.7 17.6 58.5 5.8 52.5 11.8
1999 93.3 113.0 -19.7 105.2 -11.9 105.7 -12.5 122.7 -29.4 109.2 -15.9
2000 119.6 102.4 17.2 101.8 17.8 99.5 20.1 102.7 16.8 115.1 4.4
2001 111.0 102.9 8.1 112.5 -1.5 110.2 0.8 112.5 -1.5 121.0 -10.0
RMSE 12.2 12.8 18.1 17.3 15.9
MAE 9.9 9.9 12.9 14.3 12.1
ahead because they are able to reproduce the distinct nonlinear structure of the sunspot series clearly
better than the linear autoregressive models.
In order to find out whether or not model (22) generates more accurate one-step-ahead forecasts
than the other models we have applied the modified Diebold–Mariano test (Diebold and Mariano,
1995) of Harvey et al. (1997) to these series of forecasts. Table V shows the values of the statistic
and the corresponding p-values. The null hypothesis of no difference in the theoretical MAE or
RMSE between the AR-NN model and a competitor can be rejected only when the competitor is
any of the SETAR models. The AR-NN model thus appears somewhat better than the SETAR alter-
natives but not better than the linear AR model and the NN one obtained by Bayesian
regularization.
We also compared multi-step forecasts made by our model and the alternative models described
above. The forecasts were made according to the following procedure:
(1) For t = 1980, . . . , 1988, compute the out-of-sample forecasts of one to eight-step-ahead of each
model, ŷt(k), and the forecast errors denoted by êt(k) where k is the forecasting horizon.
(2) For each forecasting horizon, compute the RMSE and the MAE statistics.
DOI: 10.1002/for
Table V. Modified Diebold–Mariano test of the null of no difference

between the forecast errors of the different models
Comparison MDM statistic p-value

Squared errors
AR-NN vs. NN 0.41 0.34
AR-NN vs. SETAR (Tong, 1990) 1.42 0.08
AR-NN vs. SETAR (Chen, 1995) 1.89 0.04
AR-NN vs. AR 1.29 0.10
Absolute errors
AR-NN vs. NN 0.21 0.42
AR-NN vs. SETAR (Tong, 1990) 1.52 0.07
AR-NN vs. SETAR (Chen, 1995) 2.10 0.02
AR-NN vs. AR 1.10 0.15
Table VI. Multi-step-ahead forecasts, their root mean square errors and mean absolute errors for the annual
number of sunspots from a set of time series models, for the period 1981–2001
Horizon AR-NN NN model SETAR model SETAR model AR

(Tong, 1990) (Chen, 1995)
RMSE MAE RMSE MAE RMSE MAE
RMSE MAE RMSE MAE
2 18.4 14.6 20.7 16.7 31.6 21.0 27.2 21.6 26.5 18.8
3 21.6 14.6 24.3 19.3 38.4 25.2 33.6 24.8 28.2 19.9
4 22.2 15.6 27.3 21.6 42.2 26.4 31.8 23.6 27.8 20.2
5 22.4 14.0 32.4 23.2 42.2 27.0 30.6 21.6 26.9 19.1
6 20.6 14.0 36.5 25.3 41.6 26.4 31.9 23.0 26.8 19.7
7 27.5 18.4 42.2 30.2 43.3 30.3 34.0 25.0 27.5 19.8
8 25.1 20.0 39.6 30.1 45.2 35.0 33.8 26.0 26.7 19.6
Table VI shows the root mean squared error and the mean absolute errors for the annual number
of sunspots from a total of forecasts each made by each model for forecast horizons from 2 to 8
years. Several interesting facts emerge from the results. The forecastability of sunspots using the AR
model deteriorates very slowly with the forecast horizon. This is clearly due to the extraordinarily
persistent cycle in the series. As to the AR-NN model that also contains a linear unit, the advantage
in forecast accuracy compared to the AR model is clear at short horizons but vanishes at the 7-year
horizon. The large NN model obtained by Bayesian regularization does not contain a linear unit and
fares less well in this comparison. For the two SETAR models, forecastability deteriorates quite
slowly with the forecast horizon after a quick initial decay. The accuracy of forecasts, however,
measured by the root mean squared error or the mean absolute error, is somewhat inferior to that of
the linear AR model.
Example 2: financial prediction

Our second example has to do with forecasting stock returns. We have chosen it because our results
can be compared with ones from previous studies and because this is a multivariate example. Pesaran
and Timmermann (1995) provided evidence in favour of monthly US stock returns being predictable.
They constructed a linear model containing nine economic variables and showed that using the model
DOI: 10.1002/for
for managing a portfolio consisting of either the S&P500 index or bonds gave results superior to the
ones obtained from a simple random walk model. The variables are: excess returns on the S&P500
index, rt; dividend yield, DYt; earnings–price ratio; first lag of the one-month Treasury bill rate,
I1t-1; second lag of the one-month Treasury bill rate, I1t-2; first lag of the 12-month Treasury bond
rate, I12t-1; second lag of the 12-month Treasury bond rate, I12t-2; second lag of the year-on-year
rate of inflation, pt-2; second lag of the year-on-year rate of change in industrial production, DIPt-2;
and second lag of the year-on-year growth rate of narrow money stock, DMt-2. The choice between
stocks and bonds was reconsidered every month, and profits were reinvested. The time period
extended from January 1954 to December 1992. Later, Qi (1999) applied a NN model based on
Bayesian regularization to the same data set and obtained results vastly superior to the ones Pesaran
and Timmermann (1995) had reported. Recently, however, Maasoumi and Racine (2001) found that
with no model could one come close to the level of accumulated wealth Qi’s model generated, even
though some models had a similar in-sample performance. When Racine (2001) reproduced the
experiment, he was indeed unable to demonstrate similar results for the NN model.
Following the others, we respecify our model for each observation period. Thus, our modelling
strategy is applied as follows:
1. For t = 1, . . . , T; with T = 1960.1 to 1992.12.

(a) Select the variables with the procedure described earlier using a third-order Taylor
expansion.
(b) Test linearity with all the selected variables in the transition function using the het-
eroskedasticity-robust version of the linearity test.
(c) If linearity is rejected, estimate an AR-NN model. Otherwise, estimate a linear regression
including all the covariates. The number of hidden units is determined using the het-
eroskedasticity-robust version of the LM test. The initial significance level of the tests equals
0.05.
The first model is estimated with the data extending to the end of 1959, and the whole modelling
procedure is repeated after adding another month to the sample. It is seen from Figure 3 that the
composition of variables varies quite considerably, although there are periods of stability, such as
the years 1978–1984, for example. Not a single one of the nine variables appears in every model,
however. Perhaps quite predictably when the sample is small, linearity is not rejected at the 5% level,
see Figure 4. There is a single period between 1984 and 1988 when the model selection strategy
yields a NN model and another one at the end of the period. This already leads one to expect only
minor differences in wealth at the end of the period between the strategies based on the linear and
the NN model. In fact, the linear model containing all variables and the NN strategy (either a linear
model with a subset of variables or a NN model) lead to a different investment decision in only 10
cases out of 396. Out of these 10, our technique yielded a correct direction forecast in four cases
and the linear model in the remaining six.
The accumulated wealth is shown in Table VII. The linear model gives the best results. Our NN
model (Panel E) is slightly better than the Bayesian regularization NN model of Racine (2001) (Panel
D) for no or low transaction costs. For high transaction costs, the relationship is the opposite. Thus,
our NN modelling strategy compares well with the Qi–Racine approach but is not any better than a
linear model with a constant composition of variables. The main reason for the linear model doing
well is that there is not much structure to be modelled in the relationship between the returns and
the explanatory variables. A nonlinear model cannot therefore be expected to do better than a linear
DOI: 10.1002/for
Figure 3. Variables selected using the AIC
Figure 4. p-value of the linearity test (heteroskedasticity-robust version). The dashed lines are the 0.1, 0.05
and 0.01 bounds
one. Furthermore, NN models most often require a large sample to perform well, and in this example
a clear majority of samples must be considered small.
It has been pointed out, see Fama (1998), that the accumulated wealth comparisons may be mis-
leading in assessing the forecasting performance of different models because the cumulative effect
of a single pair of different direction forecasts and thus investment decisions early on may grow
quite large. In our case, the different decisions are few and appear relatively late in the sample. As
a result, repeating the same exercise without reinvesting the profits leads to the conclusion that there
is no difference in performance between the AR model and the models, either linear or NN, obtained
by our technique.
Other examples
To complete the presentation, we briefly review other studies that have relevance in the present
context. Anders and Korn (1999) contains a simulation study in which neural network models con-
structed using various strategies are evaluated out-of-sample. One of the strategies is based on
Teräsvirta et al. (1993) and is therefore relevant for this paper. The results indicate that it is the one
that on average generates the most accurate forecasts. In a related paper, Anders et al. (1998) apply
neural network models to pricing call options. Their conclusion is that models specified using sta-
tistical inference have the best out-of-sample performance.
DOI: 10.1002/for
Table VII. Risks and profits of market, bond and switching portfolios based on the out-of-sample forecasts of
alternative models, 1960.1 to 1992.12
Transaction costs Mean return (%) Std. of return Sharpe ratio Final wealth ($)
Panel A: market portfolio
Zero 11.15 14.90 0.35 2503
Low 11.13 14.90 0.43 2463
High 11.11 14.89 0.43 2424
Panel B: bond portfolio
Zero 5.93 2.74 — 700
Low 4.72 2.74 — 471
High 4.72 2.74 — 471
Panel C: switching portfolio based on linear forecasts
Zero 13.66 10.08 0.77 7458
Low 12.21 10.18 0.74 4631
High 11.23 10.34 0.63 3346
Panel D: switching portfolio based on NN forecasts of Racine (2001)
Zero 13.23 10.89 0.67 6624
Low 11.98 10.88 0.67 4204
High 11.23 10.93 0.60 3292
Panel E: switching portfolio based on AR-NN forecasts
Zero 13.50 10.12 0.75 7054
Low 12.00 10.20 0.71 4319
High 10.99 10.34 0.61 3089
Rech (2002) presents a forecasting exercise in which 30 time series from different fields such as
macroeconomics, climatology and biology are being forecast using linear autoregressive models as
well as various neural network models. The linear autoregressive model turns out to be the best per-
former overall. The modelling strategy described in the present paper performs equally well as the
other, computationally more involved, strategies that include early stopping and various forms of
pruning.
In a recent study, Teräsvirta et al. (2005) apply the modelling strategy described in this paper to
47 monthly macroeconomic time series from G7 countries. Another neural network modelling strat-
egy they use is Bayesian regularization. It turns out that on average the latter strategy generates more
accurate forecasts than the AR-NN approach, without dominating it. The authors report, however,
that a number of the estimated AR-NN models are nonstationary, which has a negative effect on
forecast accuracy at long forecast horizons. The models specified and estimated using Bayesian reg-
ularization do not have this drawback because either they do not have a linear unit or the estimates
of parameters of the linear unit are shrunk towards zero. This outcome rather stresses the impor-
tance of model evaluation, which the authors, due to the vastness of the study (in total they esti-
mated more than 900 models and generated forecasts for four forecast horizons from each of them),
had to ignore. As a whole, the forecasting performance of the neural network models was somewhat
inferior to that of linear autoregressive and logistic smooth transition autoregressive models also
considered in that study. Combining all the information it appears that the AR-NN modelling strat-
egy can be successfully used to produce effective forecasting models, but the role of model evalu-
ation should not be forgotten.
DOI: 10.1002/for
CONCLUSIONS
In this paper we have demonstrated how statistical methods can be applied in building neural network
models. The idea is to specify parsimonious models and keep the computational cost down. An
advantage of our modelling strategy is that the modelling procedure is transparent rather than a black
box. Every step in model building is clearly documented and motivated. On the other hand, using
this strategy requires active participation of the model builder and willingness to make decisions.
Choosing the model selection criterion for variable selection and determining significance levels for
the test sequence for selecting the number of hidden units are not automated, and different choices
may often produce different models. However, after these decisions have been made, the modelling
strategy can easily be automated. As a whole, the method shows promise, and research is being
carried out in order to learn more about its properties in modelling and forecasting stationary time
series. The Matlab code for carrying out the modelling cycle exists and is downloadable at
www.hhs.se/stat/research/nonlinear.htm.
APPENDIX: STARTING VALUES AND COMPUTATIONAL ISSUES
Many iterative optimization algorithms are sensitive to the choice of starting values, and this is cer-
tainly so in the estimation of AR-NN models. Besides, an AR-NN model with h hidden units con-
tains h parameters, gi; i = 1, . . . , h; that are not scale-free. Our first task is thus to rescale the input
variables in such a way that they have the standard deviation equal to unity. This, together with the
fact that ||wh|| = 1, gives us a basis for discussing the choice of starting values of gi, i = 1, . . . , h. A
natural choice of initial values for the estimation of parameters in the model with h neurons is to
use the previous estimates for the parameters in the first h - 1 hidden units and the linear unit. The
starting values for the parameters gh, q, wh and ch are obtained in three steps as follows.
(1) For k = 1, . . . , K:
(a) Construct a vector v(k) (k) (k) (k) (k)
h = [v1h , . . . , v qh]¢ such that v 1h Œ (0, 1] and vjh Œ [-1, 1], j = 2, . . . ,
q. The values for v1h are drawn from a uniform (0, 1] distribution and the ones for v(k)
(k)
jh ,
j = 2, . . . , q, from a uniform [-1, 1] distribution.
(k) -1
(b) Define w(k) (k) (k)
h = vh || v h || , which guarantees || w h || = 1.
(k) (k)¢
(c) Let ch = median(w h x), where x = [x1, . . . , xT].
(2) Define a grid of N positive values g (n) h , n = 1, . . . , N, for the slope parameter.
(3) For k = 1, . . . , K and n = 1, . . . , N, estimate q using (10) and compute the value of QT(y) for
each combination of starting values. Choose those values of the parameters that minimize QT(y).
After selecting the initial values of the hth hidden unit we have to reorder the units if necessary
in order to ensure that the identifying restrictions discussed are satisfied.
Typically, choosing K = 1000 and N = 20 ensures good initial estimates. We should stress, however,
that K is a nondecreasing function of the number of input variables. If the latter is large we have to
select a large K as well.
Even from appropriate starting values it may sometimes be difficult to obtain reasonably accurate
estimates for those slope parameters gi, i = 1, . . . , h, that are very large. This is the case unless the
sample size is also large. To obtain an accurate estimate of a large gi it is necessary to have a large
number of observations such that w¢i xt lies in a small neighbourhood of ci. When the sample size is
DOI: 10.1002/for
not very large, there are generally few observations of this kind in the sample, which results in impre-
cise estimates of the slope parameter. This manifests itself in low absolute t-values for the estimates
of gi. In such cases, the model builder cannot take a low absolute value of the t-statistic of the param-
eters of the transition function as evidence for omitting the hidden unit in question. Another reason
for not doing so is that the t-value does not have its customary interpretation as a value of an asymp-
totic t-distributed statistic. This is due to an identification problem; see, for example, Bates and Watts
(1988, p. 87) or Teräsvirta (1994).
ACKNOWLEDGEMENTS
This research has been supported by the Tore Browaldh’s Foundation. The research of the first author
has been partially supported by CNPq. The paper is partly based on chapter 2 of the PhD thesis of
the third author. A part of the work was carried out during the visits of the first author to the Depart-
ment of Economic Statistics, Stockholm School of Economics and the second author to the Depart-
ment of Economics, PUC-Rio. The hospitality of these departments is gratefully acknowledged.
Material from this paper has been presented at the 5th Brazilian Conference on Neural Networks,
Rio de Janeiro, April 2001, the ESF-EMM Network First Annual Meeting, Arona, September 2001,
the 20th International Symposium on Forecasting, Dublin, June 2002, and seminars at CORE
(Louvain-la-Neuve), Monash University (Clayton, VIC), Swedish School of Economics (Helsinki),
University of California, San Diego, Cornell University (Ithaca, NY), Federal University of Rio de
Janeiro, and PUC-Rio. We wish to thank the participants of these occasions, Hal White in particu-
lar, for helpful comments. Our thanks also go to Chris Chatfield, Dick van Dijk and Marcelo
Fernandes for useful remarks, and Allan Timmermann for the data used in the second empirical
example of the paper and for fruitful suggestions. Comments from two anonymous referees have
been helpful. The responsibility for any errors or shortcomings in the paper remains ours.
REFERENCES
Abu-Mostafa YS, Atiya AF, Magdon-Ismail M, White H. 2001. Introduction to the special issue on neural net-
works in financial engineering. IEEE Transactions on Neural Networks 12: 653–655.
Anders U, Korn O. 1999. Model selection in neural networks. Neural Networks 12: 309–323.
Anders U, Korn O, Schmitt C. 1998. Improving the pricing of options: a neural network approach. Journal of
Forecasting 17: 369–388.
Auestad B, Tjøstheim D. 1990. Identification of nonlinear time series: first order characterization and order deter-
mination. Biometrika 77: 669–687.
Balkin SD, Ord JK. 2000. Automatic neural network modeling for univariate time series. International Journal of
Forecasting 16: 509–515.
Bates DM, Watts DG. 1988. Nonlinear Regression Analysis and its Applications. John Wiley & Sons: New York.
Bertsekas DP. 1995. Nonlinear Programming. Athena Scientific: Belmont, MA.
Chen R. 1995. Threshold variable selection in open-loop threshold autoregressive models. Journal of Time Series
Analysis 16: 461–481.
Chen X, Racine J, Swanson NR. 2001. Semiparametric ARX neural-network models with an application to fore-
casting inflation. IEEE Transactions on Neural Networks 12: 674–683.
Clements MP, Hendry DF. 1999. Forecasting Non-stationary Economic Time Series. MIT Press: Cambridge, MA.
Davies RB. 1977. Hypothesis testing when the nuisance parameter is present only under the alternative. Biometrika
64: 247–254.
DOI: 10.1002/for
Daviese RB. 1987. Hypothesis testing when the nuisance parameter is present only under the alternative. Bio-
metrika 73: 33–44.
Diebold FX, Mariano RS. 1995. Comparing predictive accuracy. Journal of Business and Economic Statistics 13:
253–263.
Eitrheim Ø, Teräsvirta T. 1996. Testing the adequacy of smooth transition autoregressive models. Journal of
Econometrics 74: 59–75.
Fama EF. 1998. Market efficiency, long-term returns, and behavioral finance. Journal of Financial Economics 49:
283–306.
Fine TL. 1999. Feedforward Neural Network Methodology. Springer: New York.
Ghaddar DK, Tong H. 1981. Data transformations and self-exciting threshold autoregression. Journal of the Royal
Statistical Society C30: 238–248.
Godfrey LG. 1988. Misspecification Tests in Econometrics. Econometric Society Monographs, Vol. 16, 2nd edn.
Cambridge University Press: New York.
Golub G, Heath M, Wahba G. 1979. Generalized cross-validation as a method for choosing a good ridge param-
eter. Technometrics 21: 215–223.
Hansen BE. 1996. Inference when a nuisance parameter is not identified under the null hypothesis. Econometrica
64: 413–430.
Harvey D, Leybourne S, Newbold P. 1997. Testing the equality of prediction mean squared errors. International
Journal of Forecasting 13: 281–291.
Hwang JTG, Ding AA. 1997. Prediction intervals for artificial neural networks. Journal of the American Statisti-
cal Association 92: 109–125.
Krolzig H-M, Hendry DF. 2001. Computer automation of general-to-specific model selection procedures. Journal
of Economic Dynamics Control 25: 831–866.
Kuan CM, White H. 1994. Artificial neural networks: an econometric perspective. Econometric Reviews 13: 1–91.
Lee T-H, White H, Granger CWJ. 1993. Testing for negleted nonlinearity in time series models. A comparison of
neural network methods alternative tests. Journal of Econometrics 56: 269–290.
Leybourne S, Newbold P, Vougas D. 1998. Unit roots and smooth transitions. Journal of Time Series Analysis 19:
83–97.
Luukkonen R, Saikkonen P, Teräsvirta T. 1988. Testing linearity in univariate time series models, Scandinavian
Journal of Statistics 15: 161–175.
Maasoumi E, Racine J. 2001. Entropy and predictability of stock market returns. Journal of Econometrics 107:
291–312.
MacKay DJC. 1992a. Bayesian interpolation. Neural Computation 4: 415–447.
MacKay DJC. 1992b. A practical Bayesian framework for backpropagation networks. Neural Computation 4:
448– 472.
Murata N, Yoshizawa S, Amari S-I. 1994. Network information criterion—determining the number of hidden units
for an artificial neural network model. IEEE Transactions on Neural Networks 5: 865–872.
Pesaran MH, Timmermann A. 1995. Predictability of stock returns: robustness and economic significance. Journal
of Finance 50: 1201–1228.
Qi M. 1999. Nonlinear predictability of stock returns using financial and economic variables. Journal of Business
and Economic and Statistics 17: 419– 429.
Racine J. 2001. On the nonlinear predictability of stock returns using financial and economic variables. Journal
of Business and Economic Statistics 19: 380–382.
Rech G. 2002. Forecasting with artificial neural network models. Working Paper Series in Economics and Finance
491, Stockholm School of Economics.
Rech G, Teräsvirta T, Tschernig R. 2001. A simple variable selection technique for nonlinear models. Communi-
cations in Statistics, Theory and Methods 30: 1227–1241.
Refenes APN, Zapranis AD. 1999. Neural model identification, variable selection, and model adequacy. Journal
of Forecasting 18: 299–332.
Swanson NR, White H. 1995. A model selection approach to assessing the information in the term structure using
linear models and artificial neural networks. Journal of Business and Economic Statistics 13: 265–275.
Swanson NR, White H. 1997a. Forecasting economic time series using flexible versus fixed specification and linear
versus nonlinear econometric models. International Journal of Forecasting 13: 439–461.
Swanson NR, White H. 1997b. A model selection approach to real-time macroeconomic forecasting using linear
models and artificial neural networks. Review of Economic and Statistics 79: 540–550.
DOI: 10.1002/for
Teräsvirta T. 1994. Specification, estimation, and evaluation of smooth transition autoregressive models. Journal
of the American Statistical Association 89: 208–218.
Teräsvirta T, Lin, C-FJ. 1993. Determining the number of hidden units in a single hidden-layer neural network
model. Research Report 1993/7, Bank of Norway.
Teräsvirta T, Mellin I. 1986. Model selection criteria and model selection tests in regression models. Scandina-
vian Journal of Statistics 13: 159–171.
Teräsvirta T, Lin CF, Granger CWJ. 1993. Power of the neural network linearity test. Journal of Time Series Analy-
sis 14: 309–323.
Teräsvirta T, van Dijk D, Medeiros M. 2005. Linear models, smooth transition autoregressions, and neural
networks for forecasting macroeconomic time series: a re-examination. International Journal of Forecasting,
forthcoming.
Tjøstheim D, Auestad B. 1994. Nonparametric identification of nonlinear time series—selecting significant lags.
Journal of the American Statistical Association 89: 1410–1419.
Tong H. 1990. Non-linear Time Series: A Dynamical Systems Approach. Oxford University Press: Oxford.
Trapletti A, Leisch F, Hornik K. 2000. Stationary and integrated autoregressive neural network processes. Neural
Computation 12: 2427–2450.
Tschernig R, Yang L. 2000. Nonparametric lag selection for time series. Journal of Time Series Analysis 21:
457– 487.
Vieu P. 1995. Order choice in nonlinear autoregressive models. Statistics 26: 307–328.
Weigend A, Huberman B, Rumelhart D. 1992. Predicting sunspots and exchange rates with connectionist net-
works. In Nonlinear Modeling and Forecasting, Casdagli M, Eubank S (eds). Addison-Wesley: Reading, MA.
White H. 1980. A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedastic-
ity. Econometrica 48: 817–838.
White H. 1989. An additional hidden unit test for neglected nonlinearity in multilayer feedforward networks. In
Proceedings of the International Joint Conference on Neural Networks. IEEE Press: New York; 451– 455.
Wooldridge JM. 1990. A unified approach to robust, regression-based specification tests. Econometric Theory 6:
17– 43.
Yao Q, Tong H. 1994. On subset selection in non-parametric stochastic regression. Statistica Sinica 4: 51–70.
Zapranis A, Refenes A-P. 1999. Principles of Neural Model Identification, Selection and Adequacy: With Appli-
cations to Financial Econometrics. Springer-Verlag: Berlin.
Authors’ biographies:
Marcelo C. Medeiros is Assistant Professor at the Department of Economics of the Pontifical Catholic Univer-
sity of Rio de Janeiro (PUC-Rio). His main research interest is nonlinear time series econometrics and the link
between machine learning and econometric theory.
Timo Teräsvirta is Professor of Econometrics at the Stockholm School of Economics, Sweden. His main research
interest is time series econometrics, nonlinear models and modelling in particular. He is a co-author of a book on
nonlinear econometrics and has published a number of articles in international journals.
Gianluigi Rech obtained his PhD from the Stockholm School of Economics, Sweden in 2002. His main research
interest is nonlinear time series econometrics.
Authors’ addresses:
Marcelo C. Medeiros, Department of Economics, Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro,
Brazil.
Timo Teräsvirta and Gianluigi Rech, Department of Economic Statistics, Stockholm School of Economics, Box
6501, SE-113 83, Stockholm, Sweden.
DOI: 10.1002/for

Building Neural Network Models For Time Series: A Statistical Approach

Uploaded by

Copyright:

Available Formats

Building Neural Network Models For Time Series: A Statistical Approach

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Building Neural Network Models For Time Series: A Statistical Approach

Uploaded by

Copyright:

Available Formats

Journal of Forecasting

J. Forecast. 25, 49–75 (2006)

Building Neural Network Models for Time

key words model misspecification; neural computing; nonlinear forecast-

Copyright © 2006 John Wiley & Sons, Ltd.

THE AUTOREGRESSIVE NEURAL NETWORK MODEL

The autoregressive neural network (AR-NN) model is defined as

F(g i (w i¢x t - ci )) = (1 + e -g i ( w i¢xt -ci ) )

where gi > 0, i = 1, . . . , h, and ||wi|| = 1 with

STRATEGY FOR BUILDING AR-NN MODELS

Three stages of model building

Concentrated maximum likelihood

Ê x˜ 1¢ F(g 1 (w 1¢x1 - c1 )) L F(g h (w h¢ x1 - ch )) ˆ

The remaining parameters f are estimated conditionally on q by applying the Levenberg–Marquardt

Determining the number of hidden units

where êt = yt - G(xt; ŷ ),

∂ Fˆh ∂ Fˆh ˆ ∂ Fˆ1 ∂ Fˆh ˘ ¢

or the F version of the test

Evaluation of the estimated model

Test of no serial correlation

Discussion and comparisons

Serial correlation test

Example 1: annual sunspot numbers, 1700–2000

Nominal Size - Empirical Size

50 100 150 200 250

Number of hidden units under the null hypothesis

transformed observations have the form yt = 2[ (1 + Nt ) - 1], t = 1, . . . , T, where Nt is the original

yt = -0.17+ 0.85 yt -1 + 0.14 yt -2 - 0.31 yt -3 + 0.08 yt -7

Table III. Tests of no error autocorrelation in (22)

LM test for qth-order serial correlation

Year Observation AR-NN NN model SETAR model SETAR model AR model

Table V. Modified Diebold–Mariano test of the null of no difference

Comparison MDM statistic p-value

Horizon AR-NN NN model SETAR model SETAR model AR

Example 2: financial prediction

1. For t = 1, . . . , T; with T = 1960.1 to 1992.12.

Figure 3. Variables selected using the AIC

APPENDIX: STARTING VALUES AND COMPUTATIONAL ISSUES

You might also like