International Journal of Forecasting 20 (2004) 169 – 183 www.elsevier.


Forecasting economic and financial time-series with non-linear models
Michael P. Clements a, Philip Hans Franses b, Norman R. Swanson c,*
a Department of Economics, University of Warwick, Warwick, RI, USA Econometric Institute, Erasmus University Rotterdam, Rotterdam, NY, USA c Department of Economics, Rutgers University, 75 Hamilton Street, New Brunswick, NJ 08901-1248, USA b

Abstract In this paper we discuss the current state-of-the-art in estimating, evaluating, and selecting among non-linear forecasting models for economic and financial time series. We review theoretical and empirical issues, including predictive density, interval and point evaluation and model selection, loss functions, data-mining, and aggregation. In addition, we argue that although the evidence in favor of constructing forecasts using non-linear models is rather sparse, there is reason to be optimistic. However, much remains to be done. Finally, we outline a variety of topics for future research, and discuss a number of areas which have received considerable attention in the recent literature, but where many questions remain. D 2004 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
Keywords: Economic; Financial; Non-linear models

1. Introduction Whilst non-linear models are often used for a variety of purposes, one of their prime uses is for forecasting, and it is in terms of their forecasting performance that they are most often judged. However, a casual review of the literature suggests that often the forecasting performance of such models is not particularly good. Some studies find in favor, but equally there are studies in which their added complexity relative to rival linear models does not result in the expected gains in terms of forecast accuracy. Just over a decade ago, in their review of non-linear time series models, De Gooijer and Kumar (1992) concluded that there was no clear evidence in favor of
* Corresponding author. E-mail address: (N.R. Swanson).

non-linear over linear models in terms of forecast performance, and we suspect that the situation has not changed very much since then. It seems that we have not come very far in the area of non-linear forecast model construction. We argue that the relatively poor forecasting performance of non-linear models calls for substantive further research in this area, given that one might feel uncomfortable asserting that non-linearities are unimportant in describing economic and financial phenomena. The problem may simply be that our non-linear models are not mimicing reality any better than simpler linear approximations, and in the next section we discuss this and related reasons why a good forecast performance ‘across the board’ may constitute something of a ‘holy grail’ for non-linear models. We discuss the current state-of-the-art in non-linear modeling and forecasting, with particular emphasis

0169-2070/$ - see front matter D 2004 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved. doi:10.1016/j.ijforecast.2003.10.004

especially if it is forecasts in those states which are most valuable to the user of the forecasts. Clements and Smith (1999) compare the forecasting performance of empirical self-exciting threshold autoregressive (SETAR) models and AR models using simulation techniques which ensure that past nonlinearities are present in the forecast period. / International Journal of Forecasting 20 (2004) 169–183 placed on outlining a number of open issues. In Section 3 we discuss recent theoretical and methodological issues to do with forecasting with non-linear models. and almost all other continuous time finance models are non-linear. Hamilton. selection and testing approaches become more sophisticated. Sichel. become useful for forecasting output growth). only when oil prices increase by a large amount do they have a significant effect on output growth. many of which go beyond the traditional preoccupation with point forecasts to consider the whole predictive density. Granger ¨ and Terasvirta (1993b). ch. loss functions. Why consider non-linear models? Many of us believe that linear models ought to be a relatively poor way of capturing certain types of economic behavior.g. ranging from regime-switching models. On a more cautionary note. but references are innumerable). In Section 2 we discuss why one might want to consider non-linear models. and Boero and Marrocu (2004) for an application of regime-specific evaluation to exchange rate forecasts. As such. Similarly. and data-mining. to neural networks and genetic algorithms. argue that the superior in-sample performance of such models will only be matched out-ofsample if that period contains ‘non-linear features’. Shortly after De Gooijer and Kumar. As our non-linear model estimation. bond pricing models. 1994. Granger and Terasvirta (1993a). at certain times. recessions and expansions). 9 (see also Terasvirta & ¨ ¨ Anderson. 1992) in their review of smooth transition autoregressive (STAR) models of US industrial production. there is ample reason to continue to look at non-linear models. It is thus not surprising that nonlinear models. as such models can be viewed as reasonable approximations to the non-linear phenomenon of interest. estimation and specification.g. Thus.g. this paper complements the rest of the papers in this special issue of the International Journal of Forecasting. 2. In macroeconomics and finance theory a host of non-linear models are already in vogue. Tong (1995). Clements et al. Other types of non-linearity might include the possibility that the effects of shocks accumulate until a process ‘‘explodes’’ (self-exciting or catastrophic behavior). pp. as can financial variables (periods of high and low volatility). and how these are dealt with in the papers collected in this issue. 1989. The predominance of non-linear models in economics and finance is not inconsistent with the use of linear models by the applied practitioner. as well as the notion that some variables are relevant for forecasting only once in a while (e. or economic performance. The topics we focus on include joint and conditional predictive density evaluation. and therefore. and a number of reasons why their forecasting ability relative to linear models may not be as good as expected. Box –Jenkins ARMA) model of output growth in a Western economy subject to the business cycle. 409 – 410 argues strongly that for non-linear models ‘how well we can predict depends on where we are’ and that there are ‘windows of opportunity for substantial reduction in prediction errors’. See Tiao and Tsay (1994) for a four-regime TAR model applied to US GNP. . are receiving a great deal of attention in the literature.g. The rest of the paper is organized as follows. For example. Concluding remarks are gathered in Section 5. As well as which. one might expect to see their forecast performance improve commensurately. almost all-real business cycle (RBC) models are highly non-linear.170 M. Output growth non-linearities can be characterized by the presence of two or more regimes (e. Section 4 highlights a number of empirical issues. This suggests that an important aspect of an evaluation of the forecasts from nonlinear models relative to the linear AR models is to make the comparison in a way which highlights the favorable performance of the former for certain states. where the properties of output growth in recessions are in some ways quite different from expansions (e. which might count against the aforementioned improvement in the relative performance of non-linear models.P. from the perspective of forecasting. diffusion processes describing yield curves. amongst others. we review a number of factors. The obvious example would be a linear (e.

capturing the distinct dynamics of the business cycle phases) does not necessarily translate in to a good outof-sample performance relative to a model such as an AR. feeling that the forecasts are in some sense ‘‘second’’. l2 < 0 if st ¼ 2ðW contractionW or W recessionWÞ. which take up the challenge of forecasting with nonlinear models. 2g: ð3Þ The model can be rewritten as the sum of two independent processes: Dyt À ly ¼ lt þ zt . .M. / International Journal of Forecasting 20 (2004) 169–183 171 In addition. if one believes the underlying phenomenon is non-linear. 2. which cannot be readily exploited to improve out-of-sample performance.1. Diebold and Nason (1990) give a number of reasons why non-linear models may fail to outperform linear models. and to seek alternatives. and may only be detected by careful analysis along the lines of Koop and Potter (2000). ru ]. but. it is worth considering a non-linear model. They also suggest that conditional –mean non-linearities may be a feature of the data generating process (DGP). because some aspects of the economy or financial markets do indeed display nonlinear behavior. We end this section with two short illustrations of some of the difficulties that can arise. This follows from the belief that a good model for the in-sample data should also be a good out-of-sample forecasting model. In the second. Clements et al. then neglecting these features in constructing forecasts would leave the end-user uneasy. The focus is on a two-regime MS model: Dyt À ðst Þ ¼ aðDytÀ1 À ðstÀ1 ÞÞ þ ut : ð1Þ 2 where ut f IN[0. As an example. and apply their analysis to post war US output growth. but may not be large enough to yield much of an improvement to forecasting. for example. knowing that the AR(2) model is incapable of capturing the distinct dynamics of expansions and contractions. defined by the transition probabilities: pij ¼ Prðstþ1 ¼ jAst ¼ iÞ. (2) The description of a MS– AR model is completed by the specification of a model for the stochastic and unobservable regimes on which the parameters of the conditional process depend. Once a law has been specified for the states st the evolution of regimes can be inferred from the data. will necessarily improve matters. but this does not contribute to an improved forecast performance. but that the wrong types of nonlinear models have been used to try and capture them. in the context of exchange rate prediction. The regime-generating process is assumed to be an ergodic Markov chain with a finite number of states st = 1. jaf1. there is distinct regime-switching behavior. is unlikely to engender confidence that one has a model that captures the true dynamics of the economy. But this is the crux of the matter—a good in-sample fit (here. 2 X j¼1 pij ¼ 1 bi.P. The subse- quent sections of this paper review some of the recent developments in model selection and empirical strategies. The sub-section below illustrates. as well as the explanation that they are present and important. Markov-switching models of US output growth Clements and Krolzig (1998) present some theoretical explanations for why Markov-switching (MS) models may not forecast much better than AR models. such as adding another regime. as well as the remaining papers in this issue. which are obviously deficient in some respect. we discuss the relationship between output growth and the oil price. (for a tworegime model). One is that apparent non-linearities detected by tests for linearity are due to outliers or structural breaks. but warn against the expectation that such models will always do well— there are too many unknowns and the economic system is too complex to support the belief that simply generalizing a linear model in one (simple!) direction. an AR(2) model for US GNP growth may yield lower average squared errors than an artificial neural network with one hidden unit. There is a view that. and see also Clements and Hendry (1999) for a more general discussion. 2. In the first. We take the view that. it is natural to be unhappy with models. That said. The conditional mean l(st) switches between two states: lðst Þ ¼ 8 < l1 > 0 : if st ¼ 1ðW expansionW or WboomWÞ.

Invoking the unrestricted VAR(1) representation of a Markov chain: ft ¼ ðp11 þ p22 À 1ÞftÀ1 þ tt . the contribution of the nonlinear part of (4) to the overall forecast depends on both the magnitude of the regime shifts. where ~ t = 1 À Pr(st = 2) if st = 2 and À Pr(st = 2) otherwise. the ˆ conditional mean of DyT + h is given by DyTþhAT À ly which equals: ˆ ˆ lT þhAT þ zTþhT ¼ ðl2 À l1 Þðp11 þ p22 À 1Þh fT AT þ ah ˆ ˆ  ½DyT À l À ðl À l ÞfTAT Š y 2 1 p11 + p22 À 1g a in (4). the conditional expectation collapses to a linear prediction rule. When the regimes are unpredictable. we can do no better than employing a simple linear model. More recently. With the advantage of several more years of data. To what extent did the OPEC oil price rises contribute to the recessions in the 1970s. Hamilton (2000) has used a new flexible non-linear approach (Dahl & ¼ ah ðDyT À ly Þ þ ðl2 À l1 Þ ˆ  ½ðp11 þ p22 À 1Þh À ah ŠfT AT : ð4Þ The first term in (4) is the optimal prediction rule for a linear model. e t fIN ½0. attempt to predict business cycle regimes. and otherwise takes on the value zero.172 M. Since the predictive power of detected regime shifts is extremely small. where fT AT contains the information about the most recent regime at the time the forecast is made. Artis. such that E[lt] = E[zt] = 0. Clements et al. and the transition probabilities in (4) can be made to depend on leading indicator variables. See also Krolzig (2003). 2. which simply reverse previous (within the preceding year) declines do not depress. such as Sensier. Thus.65. Hooker (1996) showed that the linear relationship proposed by Hamilton (1983) appears not to hold from 1973 onwards (the date of the first oil price hike!). Dacco and Satchell (1999) for an analysis of the effects of wrongly classifying the regime the process will be in. and might one expect reductions in prices to stimulate growth. output growth. represents the contribution of the Markov chain: lt ¼ ðl2 À l1 Þft . who suggested that the relationship was asymmetric. While the process zt is Gaussian: zt ¼ aztÀ1 þ e t . although perhaps to a lesser extent? Hamilton (1983) originally proposed a linear relationship between oil prices and output growth for the US. Pr(st = 2) = p12/( p12 + p21) is the unconditional probability of regime 2. Osborn and Birchenhall (2004). lt. r2 Š.64.2. and the largest root of the AR polynomial to be 0. and constructs a variable that is the percentage change in the oil price in the current quarter over the previous year’s high. Paap and Vroomen (2004) utilize information from extraneous variables to determine the regime in a novel approach to predicting the US unemployment rate. Clements and Krolzig estimate p11 + p22 À 1 = 0. given by a. he also cast doubt on the simple asymmetry hypothesis suggested by Mork (1989). e the other component. then predictions of the hidden Markov chain are given by: ˆ ˆ fTþhAT ¼ ðp11 þ p22 À 1Þh fT AT ˆ where fT AT = E[~ TAYT] = Pr(sT = 2AYT) À Pr(sT = 2) is the filtered probability Pr(sT = 2AYT) of being in regime 2 corrected for the unconditional probability. as a way of sharpening the forecasting ability of these models. but is unaffected by oil price declines. and the contribution of the Markov regime-switching structure is given by the ˆ ˆ term multiplied by fT AT . This was challenged by Mork (1989). / International Journal of Forecasting 20 (2004) 169–183 where ly is the unconditional mean of Dyt.P. However. A number of authors. Output growth and the oil price Of obvious interest in the literature is the relationship between oil prices and the macroeconomy. Hamilton (1996) proposes relating output growth to the net increase in oil prices over the previous year. and on the persistence of regime shifts p11 + p22 À 1 relative to the persistence of the Gaussian process. Al2 À l1A. Heuristically. Franses. so that the second reason explains the success of the linear AR model in forecasting. Thus. Thus. . when this is positive. the relative performance of non-linear regime-switching models would be expected to be better the more persistent the regimes. in that output growth responds negatively to oil price increases. increases in the price of oil. Recently.

Pagan (1997a).M.g. Diebold. Hahn & Tay. requiring simulation or numerical methods in the latter case. For example. Giacomini and White (2003) As an aside. but not for a threshold autoregressive process. Christoffersen. 2001). Christoffersen. Hamilton. 2001) to characterize the appropriate non-linear transform of the oil price. in financial risk management interest often focuses on predicting a particular quantile (as the Value-at-Risk. For ¨ example. the recent surveys by Franses & van Dijk. An obvious starting point is which non-linear model to use. given the many possibilities that are available. 1998 (henceforth DGT). 2000.g.’ (p. Corradi and Swanson (2003a) draw on elements from both types of papers in order to provide a test for choosing among competing predictive density models which may be misspecified. 2001. 1999. Certain theoretical properties. 2004. Franses & Terasvirta. . the process of discovery of (an approximation to) the form of the non-linearity in the relationship between output growth and oil price changes has taken place over two decades. for example. where the ratio of the drift to the variance of the disturbance term is the crucial quantity.1 An approach to tackling these issues is to use predictive density or distributional testing. In this section we outline a number of these. Perhaps this warns against expecting too much in the near future using ‘canned’ routines and models. these papers have shown that the durations and amplitudes of expansions and contractions of the classical cycle can be reasonably well reproduced by simple random walk with drift models. Gunther & Tay. They are interested in whether the recurrent shifts between expansion and contraction identified by the MS model remain when oil prices have been included as an explanatory variable for the mean of output. Further. Raymond and Rich (1997) conclude that ‘while the behavior of oil prices has been a contributing factor to the mean of low growth phases of output. The different types of models often require different theoretical and empirical tools (see e.g. such as stability and stationarity. 2003. closed form solutions exist for the conditional mean forecast for an MS process. VaR) but alternatively the entire conditional distribution of a variable may be of interest. Clements & Smith. van Dijk. Hong. Corradi & Swanson. as a means of establishing which of a number of candidate forecasting models has distributional features that most closely match the historical record.P. Raymond and Rich (1997) have investigated the relationship between oil prices and the macroeconomy by including the net increase in the oil price in an MS model of US output for the period 1951 – 1995. Diebold. is far from simple. Harding and Pagan (2001) argue that non-linear models should be evaluated in terms of their ability to reproduce certain features of the classical cycle. At the same time. / International Journal of Forecasting 20 (2004) 169–183 173 Hylleberg. White. This could include. finding out which of the models yields the best distributional or interval predictions. a new strand of literature addressing the issue of predictive density evaluation has arisen (see e. are not always immediately evident. Over the last few years. 2002). 1 . Theoretical and methodological issues There are various theoretical issues involved in constructing non-linear models for forecasting. Non-linear models appear to add little over and above that which can be explained by the random walk with drift. 2000). the point forecast evaluation literature explicitly recognizes that all the candidate models may be misspecified (see e. Of course. the issue of which characteristics often arises. forecasting). For example. 2001. Giacomini & White. and often by the requirement that the model is capable of generating the key characteristics of the data at hand. The literature on the evaluation of predictive densities is largely concerned with testing the null of correct dynamic specification of an individual conditional distribution model. 1998. 2000. The choice of model might be suggested by economic theory. and the persistence of shocks. Clements and Krolzig (2002) investigate whether oil prices can account for the asymmetry in the business cycle using the tests proposed in Clements and Krolzig (2003). Clements et al. 3. rather than their ability to match the stationary moments of the detrended growth cycle. Bai. 2002. 196). Hahn & Inoue. . movements in oil prices generally have not been a principal determinant in the historical incidence of these phases . Pagan (1997b). 2002. Clearly. and has involved the application of state of the art econometric techniques. even once we have determined the purpose to which it is to be put (here.

. n models are estimated. hi aHi t j¼s and y hi ¼ arg min Eðqi ðyj . where s = max{s1. the squared (approximation) error associated with model i. h0 ÞÞ2 k¼2.P. .. Additionally. then R + 2. If model i is correctly specified. resulting in a sequence of P = T À R estimators. where fi (ÁAÁ. given some vector of variables. h1) is taken as the benchmark model. hi) = À ln fi ( ytAZ t À 1. hy) À F0(uAZ t. The hypotheses of interest are: Z y H0 : max EððF1 ðuAZ t . h0) than the benchmark. .. 2001). . is measured in terms of the average over U of E(( Fi (uAZ t.).. In the current discussion. where uaU. assume that qi ( yt. n. Clements et al. . Using the DGT terminology. h0 ÞÞ2 k¼2. hi aHi ð6Þ where qi denotes the objective function for model i. possibly taking into account that pt( ytAVt À 1) contains parameters that have been estimated (see also Bai. allowing for misspecification under both the null and the alternative hypotheses. The approach taken by Corradi and Swanson (2003a) differs from the DGT approach as they do not assume that any of the competing models are correctly specified. and define the true conditional distribution as F 0 (uAZ t . 1999. 1]. hn). . this estimator is first computed using R observations. . Diebold et al. i and U is a possibly unbounded set on the real line.t ¼ arg min qi ðyj . n ð5Þ . . F1(ÁAÁ. h1. so that hy i is the probability limit of a quasi maximum likelihood estimator (QMLE). . . yt À s1 + 1.n U y À ðFk ðuAZ t . then the probability of observing a value of yt no larger than that actually observed is a uniform random variable on [0.. By making use of the probability integral transform. we focus on 1-step ahead prediction. Hereafter. Xt À s2 + 1). Z jÀ1 . i = 1. Goodness-of-fit statistics can then be constructed that compare the empirical distribution function of the probabilities to the 45j degree line. . Now. Fn(uAZ t. Note that for a given u. hi).. Z jÀ1 . then the resulting sequence of probabilities should be identically independently distributed... . and the objective is to test whether some competitor model can provide a more accurate approximation of F0(ÁAÁ. U possibly unbounded. we compare R V t V T À 1. .. They proceed by ‘‘selecting’’ a conditional distribution model that provides the most accurate out-of-sample approximation of the true conditional distribution. hi ÞÞ. h1 Þ À F0 ðuAZ t . i =1. .. if pt( ytAVt À 1) is the ‘‘true’’ conditional distribution of ytAVt À 1.n U y À ðFk ðuAZ t . . DGT suggest a simple and effective means by which predictive densities can be evaluated. . Moreover. h0 ÞÞ2 Þ/ðuÞduV0 ð7Þ versus HA : max Z y EððF1 ðuAZ t . More precisely. Now. Many of the above papers seek to evaluate predictive densities by testing whether they have the property of correct (dynamic) specification. s2}. 2001. . . Z t=( yt.. hk Þ À F0 ðuAZ t . Assume that accuracy is measured using a distributional analog of mean square error. . . they posit that all models should be viewed as approximations to some unknown underlying DGP.. hk Þ À F0 ðuAZ t . Following standard practice (such as in the real-time forecasting literature). assume that the objective is to form parametric conditional distributions for a scalar random variable. hi) is the conditional density associated with Fi. ð8Þ where /(u) z 0 and mU /ðuÞ ¼ 1. . if we have a sequence of predictive densities. y then hiy = h0.. define the group of conditional distribution models from which we want to make a selection as y y F1(uAZ t. / International Journal of Forecasting 20 (2004) 169–183 tackle a similar problem.174 M. and X is a vector. . uaU oR. Z t À 1. Now. Xt. then R + 1 observations. Thus. h0))2). h1 Þ À F0 ðuAZ t . i ¼ 1.. hi Þ. h0 ) = Pr ( yt + 1 V uAZ t). .. t = 1.. so these estimators are then used to construct sequences of P 1-step ahead forecasts and associated forecast errors. . This is done using an extension of the conditional Kolmogorov test approach of Andrews (1997) due to Corradi and Swanson (2003b) that allows for the insample comparison of multiple misspecified models. h0 ÞÞ2 Þ/ðuÞdu > 0. . yt + 1. developing a framework that allows for the evaluation of both nested and nonnested models. n. define the recursive m-estimator for the parameter vector associated with model i as: t 1 X ˆ hi. T. assume that i = 1. and so on until the last estimator is constructed using T À 1 observations.. More specifically. Hong.

. hk.P. where uaU. the approach can be used for the comparison of the joint CDF of current (and lagged) output and hours worked. ð9Þ k¼2. m to be a sample of length S drawn (simulated) from model j and evaluated at the parameters estimated under model j. T denote actual historical output (growth rates).n U where T À1 1 X ZP. the rule is to choose Model 1 over Model 2 if Z y ðF1 ðu.. . S.u ð1. j = 1. where j j ˆ hy is the probability limit of hj. and let Dlog Xj. h2 Þ À F0 ðu. j = 1.. Note that within our context. Clearly... In general. .. As above. In order to obtain valid asymptotic critical values. hjy) À F0(u. . each model is estimated via QMLE. say a technology and a preference shock. Clements et al. h1 Þ À F0 ðu. . h0) denote the distribution of Yt evaluated at u and Fj (u.n(hj.u ð1. they suggest two block bootstrap procedures.t is defined as hi.. etc. t jÀ1 Z . For further details.n(hj. let Yt = (Dlog Xt. and j where uaUoR2 . the hypotheses of interest are: Z H0 : max j¼2. .T ). h1. their testing framework can be used to address questions of the following sort: (i) For a given RBC model. denote the output series simulated under model j. R V t V T À 1. Thus. and let F0(u. their approach can be used to evaluate the model’s joint CDF of current and lagged output. m and n = 1. Now. consider m RBC models. kÞ ¼ pffiffiffi ðð1fytþ1 Vug P t¼R ˆ À F1 ðuAZ t . . . h0 Þ À F1 ðu. Corradi and Swanson (2003c) develop a statistic based on comparison of historical and simulated distributions. ... say. including autocorrelation and second moment structures. n = 1. For any evaluation point. The squared (approximation) error associated with model j.t ¼ arg max hi aHi 1 j¼s ln fi ðyj ...t ÞÞ2 Þ: ð10Þ Here.n À 1(hj. where S denotes the length of the ˆ simulated sample. . say in the form of a RBC model. h1 ÞÞ2 y À ðF0 ðuÞ À Fj ðu. kÞ/ðuÞdu. each of which depends on the relative rate of growth of the actual and simulated sample size. accuracy is measured in terms of squared error. so that in terms of the above notation. hj ÞÞ2 Þ/ðuÞduV0 . please refer to Corradi and Swanson (2003a). hy) denote the distribution of Yj. the above approach can be used to evaluate multiple non-linear forecasting models. .n(hy). in the case of RBC models driven by only one shock.. qi = À ln fi. For example. j = 1. what is the relative usefulness of different sets of calibrated parameters for mimicing different dynamic features of output growth? (ii) Given a fixed set of calibrated parameters. .. possibly unbounded.T ) = (Dlog Xj. h0))2). n. and assume that the variables of interest are output and lagged output. . taken as T ! l. say a technology shock. . is measured in terms of the (weighted) average over U of (( Fj(u. this measure defines a norm and is a typical goodness-of-fit measure. U where mU /ðuÞdu ¼ 1 and /(u) z 0 for all uaU oR2. As the RBC data are simulated using estimated parameters (as well as previously calibrated parameters). . i = 1. .M.T ). Further. . assume that focus centers on evaluating the joint dynamics of a non-linear prediction model. . .m U y ððF0 ðu.. and U is a possibly unbounded set on R2.n(hj. Denote Dlog Xj. and t ˆ ˆ hi. m.n.T . h0 ÞÞ2 Þ/ðuÞdu. This limiting distribution is thus not nuisance parameter free. where parameter estimation is done using the T available historical observations. . Dlog Xt À 1). S. h0 ÞÞ2 Þ/ðuÞdu U Z y < ððF2 ðu. the standard issue of singularity that arises when testing RBC models is circumvented by considering a subset of variables (and their lagged values) for which a non-singular distribution exists. . Dlog Xj. where fi is the conditional density associated with P model i. ˆ ˆ ˆ Yj. t = 1. In addition. / International Journal of Forecasting 20 (2004) 169–183 175 conditional distributions in terms of their (mean square) distance from the true distribution. and critical values cannot be tabulated. .t ÞÞ2 À ð1fytþ1 Vug ˆ À Fk ðuAZ t . Now. The statistic is: Z ZP ¼ max ZP. hi Þ . the limiting distribution of their test statistic is a Gaussian process with a covariance kernel that reflects the contribution of parameter estimation error.. In the case of models driven by two shocks.T )). what is the relative performance of RBC models driven by shocks with a different marginal distribution? More specifically. set model 1 as the benchmark model. Let Dlog Xt.

h1 ÞÞ À ðF0 ð¯ .. Also. y À ðF0 ðuÞ À Fj ðu. T t¼1 S n¼1 ˆ with hj. h1 ÞÞ2 U !2 T S 1 X 1 X ˆ1. hj ÞÞ Þ/ðuÞdu > 0: Thus. h2 ÞÞ > 0: The KLIC is a sensible measure of accuracy. h0 ÞÞÞ2 À ððFj ð¯ . h0 ÞÞÞ2 À ððFj ð¯ .m U 2 V V V H 0 versus HAVand H 0Vversus HAVcan be tested in a similar manner. h1 ÞÞ2 Þ À ððF0 ðuÞ A U y À Fj ðu. h0 Þ j¼2.... or in measuring accuracy for a given confidence interval.5%.m U and y ððF0 ðuÞ À F1 ðu. White.5% and 1. 1989).S ðuÞ À ðF0 ðuÞ À Fj ðu. h0 Þ j¼2..T . Giacomini (2002) proposes an extension. hj ÞÞ2 Þ/ðuÞdu p 0. but not using the KLIC. no model can provide a better approximation (in a squared error sense) to the distribution of Yt than the approximation provided by model 1. h1 ÞÞ À ðF0 ð¯ .m ¯ y y u ÀF0 ð u. / International Journal of Forecasting 20 (2004) 169–183 and HA : max Z j¼2. hj ÞÞ2 Þ/ðuÞdu ¼ 0 versus Z y HW : ððF0 ðuÞ À F1 ðu. h0 ÞÞÞ2 Þ > 0: u ¯ If interest focuses on testing the null of equal accuracy of two distribution models (analogous to the pairwise conditional mean comparison setup of Diebold & Mariano. Clements et al.... Fernandez-Villaverde and Rubio-Ramirez (2001) have shown that the best model under the KLIC is also the model with the highest posterior probability. and Kitamura (2002) suggests a generalization for choosing among models that satisfy some conditional moment restrictions. If interest focuses on confidence intervals. h0 ÞÞÞ2 ÞV0: u ¯ versus y y u u HAV: max ðððF1 ð¯ . we can simply state the hypotheses as: Z y V H 0V: ððF0 ðu.n ðhj.g. and preferred in others. In pffiffiffiffi to test H0 versus HA.S ¼ max Zj.176 M.. We can do so quite easily using the squared error criterion. h1 Þ À F1 ð u. it leads to simple Likelihood Ratio tests. this cannot be done in an obvious manner using the KLIC. h0 Þ À F0 ð u..T an estimator of hy that satisfies Assumption 2 j below. See Corradi and Swanson (2003c) for further details... The above approach is an alternative to the KLIC that should be viewed as complementary in some cases. hj ÞÞ u ¯ ¯ ÀðF0 ð¯ .. which uses a weighted (over Yt) version of the KLIC. hj Þ À Fj ð u. ð11Þ j¼2.. As an illustration. For example. hj Þ À Fj ð u.P.m ¯ y y ÀF0 ð u. hj ÞÞ ¯ ¯ ÀðF0 ð¯ . Furthermore. so that ¯ the objective is to ‘‘approximate’’ Pr(¯ V Yt V uÞ then u the null and alternative hypotheses can be stated as: y y u u H0V : max ðððF1 ð¯ . according to which we should choose Model 1 over Model 2 if:3 y y Eðlog f1 ðYt . h2 Þ À log f2 ðYt . 1995). 1982. we often do not have an explicit form for the density implied by the various models we are comparing. Of course.T .T ÞVug ¼ 1fYt Vug À 1fY1. while it can easily be done using our measure. if we are interested in measuring accuracy over a specific region.n ðh T t¼1 S n¼1 !2 T S 1 X 1 X ˆ À 1fYt Vug À 1fYj..S where:2 Z ZT . is the KLIC. model comparison can be done using kernel density estima3 Recently. h0 Þ À F1 ðu. which have actually occurred. . h0 Þ À F0 ð u. Another measure of distributional accuracy available in the literature (see e.S ðuÞ/ðuÞdu. under H0. Interestingly. assume that we wish to evaluate the accuracy of different models in approximating the probability that the rate of growth of output is say between 0. Vuong. the relevant test statistic order is T ZT . which on average gives higher probability to events.. h1 ÞÞ2 y 2 Zj.T ÞVug . h1 Þ À F1 ð u. as it chooses the model.

Hansen (2001). as well as measures and tests based on other loss functions. from smooth transition models to projection pursuit and wavelet models. / International Journal of Forecasting 20 (2004) 169–183 177 tors. Timmerman and White (1999). although progress has been made. including smooth transition models (see e. Swanson and White (1997a) survey a number of the important contributions. and can perhaps be viewed as a leading indicator for the predictive density literature. Linton. In sum. Timmerman and White (2003). Diebold and Chen (1996).M. In some contexts it is even difficult to establish the consistency of some econometric parameter estimates. as discussed in Christoffersen and Diebold (1996). Much remains to be done in this area. Zheng. Nevertheless. For example.b). West (1996). 1993a. Pesaran and Timmerman (2000). Christoffersen and Diebold (1997). Sullivan. The ‘datasnooping procedures’ developed by White (2000). 1995. 2000). with much work remaining to be done.g. there are many recent papers that propose novel approaches to estimation. this leads to tests with non-parametric rates (see e. Leybourne and Newbold (1997). whilst others are less clear. comparison via our squared error measure of accuracy is carried out using empirical distributions. such as cointegrating vectors in certain non-linear cointegration models. Nevertheless.g. it is often difficult to come up with asymptotically valid inferential strategies using standard estimation procedures (that essentially minimize one-step ahead errors) for many varieties of non-linear models. or that one model forecast encompasses another.P. McCracken (1999). it is only recently that much attention has been given to the notion that the same loss function used in-sample for parameter estimation is often that which should be used out-of-sample for forecast evaluation. A number of these papers emphasise the role of parameter estimation uncertainty in testing for equal forecast accuracy. but this would appear to be an open area. That said. and threshold parameters in some types of regimeswitching models. although not germane to forecasting. Hansen and Jeganathan (1997). with much room for advance. The study by De Gooijer . Stekler (1991). which include Chao. Clements and Hendry (1996). where the sheer magnitude of the problem can quickly grow out of hand. Timmerman and White (2001). For example. Harvey. Chatfield (1993). Finally. Diebold and Mariano (1995). there is also the issue of whether the in-sample model estimation criterion and out-ofsample forecast accuracy criterion should be matched. Clements et al.b) and neural ¨ network models (see e. estimation and in-sample inference of non-linear forecasting models remains a potentially difficult task. Granger & Terasvirta. Pesaran and Timmerman (1994). but in addition tests based on directional forecast accuracy and sign tests. A counter to the fear that such models may be overfitting in-sample—in the sense of picking up transient. within the KLIC framework. Heaton and Luttmer (1995). and constructing tests using non-linear and/or non-differentiable loss functions. Leybourne and Newbold (1998). There are now a host of tests based on traditional squared error loss criteria. the extension of the datasnooping methodology to multivariate models awaits attention. can also be used. such as the variety of new cross-validation related techniques in the area of neural nets. On the other hand. Point forecast production and evaluation continues to receive considerable attention. Swanson & White. Granger (1993). Corradi and Swanson (2001). 1997a. Sullivan. However. or on a ‘hold-out’ sample. so that resulting test statistics converge at parametric rates. However. as well as allowing for parameter estimation error and misspecification when the comparisons involve non-linear models. among others. accidental connections between variables— is of course to compare the models on out-of-sample performance. much remains to be done with regard to making these tests applicable to non-linear models. tests of forecast encompassing. Hansen. West (2001). Stekler (1994). Clark and McCracken (2001). Weiss (1996). Clements and Hendry (1993). and used in Sullivan. Maasoumi and Whang (2003). as well as the consequences of the models being nested. The above paragraph notes the role of the loss function in determining how accuracy is to be assessed.g. Granger (1999). Pesaran and Timmerman (1992). A general problem with non-linear models is the ‘curse of dimensionality’ and the fact that such models tend to have a large number of parameters (at least relative to the available number of macroeconomic data points)—how to keep the number of parameters at a tractable level? This sort of issue is relevant to the specification of many varieties of nonlinear models. Harvey. some models have parameters that are readily interpretable. West and McCracken (2002).

or should models be estimated and/or evaluated against specific (dynamic) features of the historical record? Should we split the sample into in. More specifically. Taylor (2004) uses adaptive exponential smoothing methods that allow smoothing parameters to change over time. he presents a new adaptive method for predicting the volatility in financial returns. and to estimate the size of the shock that is required to cause the nonlinear behavior. Difficult issues arise when cointegration is no longer ‘global’.P.178 M. Paap and Vroomen (2004). Traditionally this literature has focused on the conditional mean. They analyze the outof-sample performance of SETAR models relative to a linear AR and a GARCH model using daily data for the Euro effective exchange rate. These non-linear models allow the dynamics of the conditional variance model to be influenced by the sign and size of past shocks. Bradley and Jansen (2004) propose a model in which the dynamics that characterize stock returns are allowed to differ in periods following a large swing in stock returns—that is. in order to adapt to changes in the characteristics of the time series.and out-of-sample periods. where the smoothing parameter varies as a logistic function of user-specified variables. Kunst (1992). Developments in these areas look set to continue apace with the more parametric approaches considered in this issue. Their approach allows them to test for the existence of non-linearities in returns. Their evaluation is conducted on point. Using a variety of forecast evaluation methods. Their results show that the GARCH model is better able to capture the 4 There is a growing literature on nonparametric and semiparametric forecasting methods and models. Gannoun and Zerom (2002). A number of possible approaches are available to account for the possibility that the parameters in forecasting models are changing over time. the design of non-linear models for actual data and the estimation of parameters are less straightforward. Clements and Galvao (2004) review specification and estimation procedures in systems when cointegration is ‘global’. Gannoun and De Gooijer (1998). Matzner-Løber. De Gooijer. Clements et al. inference and forecasting. Empirical issues In this section we discuss various practical issues. The authors discuss issues relating to estimation. and to provide competitive forecasts compared to alternative models. and conditional on specific regimes. and the references therein. How should we select a model?4 Should all the observations be used. and evaluate the forecasting ability of non-linear systems in the context of interest rate prediction. who propose a model in which a key autoregressive parameter depends on a leading indicator variable. but a number of recent papers have looked at nonparametric estimation of other aspects of conditional ditributions. De Gooijer and Vidiellai-Anguera establish the superiority of their model from a forecasting perspective. The autoregressive parameter is constant unless a linear function of the leading indicator plus a disturbance term exceeds a certain threshold level. Samanta (1989). unconditionally. Parameters are estimated for the method by minimizing the sum of squared deviations between realized and forecast volatility. building on earlier contributions by Anderson (1997). Another new model is developed by Franses. The model is applied to forecasting unemployment. and is shown to be capable of capturing the sharp increases in unemployment in recessions. intervals and density regions: see e. and it mimics some of the ideas put forward in Franses and Paap (2002). such as quartiles. where should the split occur? Boero and Marrocu (2004) provide evidence related to some of these questions. interval and density forecasts. In contrast to linear models. 4. the new method gives encouraging results when compared to fixed-parameter exponential smoothing and a variety of GARCH models. Using stock index data. and if so. / International Journal of Forecasting 20 (2004) 169–183 and Vidiella-i-Anguera (2004) extends ‘non-linear cointegration’ to allow the equilibrium. van Dijk and Franses (2000). inter alia. This model captures the notion that some variables are relevant for forecasting only once in a while. The approach is analogous to that used to model timevarying parameters in smooth transition GARCH models. cointegrating relationship to depend upon the regime. and the relationship of their model to existing non-linear models. . These factors can also be used as transition variables in the new smooth transition exponential smoothing approach. and the theory-justification for the long-run relation˜ ship is less clear. over the whole forecast period. a non-linear state-dependent model.g.

who consider daily. Next. Hamilton’s flexible non-linear regression model (Hamilton. the daily return distributions have different moment properties in their right and left tails. Dahl and Hylleberg (2004) give a nice review of four non-linear models. 2001). while aggregation with time-varying weights (and the presence of common shocks) is more likely to generate a role for non-linear modeling of the resultant series. artificial neural networks and two versions of the projection pursuit regression model. Interestingly. In a related study. where . Of course the bottom line is: Are there clear-cut examples where actual forecast improvements are delivered by non-linear models? Provision of such examples might serve to allay the fears held by many sceptics. Corradi and Swanson (2004) focus on various approaches to assessing predictive accuracy. Therefore. (2004). and compares them with linear models. in a ‘real-time forecasting’ exercise. and the parameters re-estimated at each step. namely. Clements et al. In addition to wellknown modeling approaches such as the variance – covariance method and historical simulation. certain moments of the return distributions do not exist in some countries.P. In particular. may produce ‘smoother’ series better suited for linear models. their results also indicate that the performance of the SETAR models improves significantly conditional on being in specific regimes. whereby the specification of the model is chosen. He assesses the quality of these models in a real-time forecasting framework. (2004) examine the role of domestic and international variables for predicting classical business cycles regimes in four European countries. they employ extreme value theory (EVT) to generate VaR estimates and provide the tail forecasts of daily returns at the 0. Further. It is found that often non-linear models perform best.M. The other two papers consider macroeconomic variables. Three nice recent examples of this sort are ˜ Clements and Galvao (2004). risk and reward are not equally likely in these economies. weekly and monthly data. In addition. how can we reliably estimate the model parameters? Can we address the problem whereby we only find local optima? How can we select good starting values in non-linear optimization? Can we design methods to test the non-linear models. tests that are robust in this sense are obviously desirable. Because estimated models are approximations to the DGP and likely to be mis-specified in unknown ways. For example. Sensier et al. But. all of whom show the ß ß relevance of non-linear models for forecasting. One of the main conclusions of their study is that there are various easy to apply statistics that can be constructed using out-of-sample conditional-moment conditions. Sensier et al. Genc ay and ß Selcuk (2004) investigate the relative performance of ß VaR models with the daily stock market returns of nine different emerging markets. and find distinct models for different temporal aggregation levels. and demonstrate the relevance of the various testing methods. The forecasting performance of these four approaches is compared for US industrial production and an unemployment rate series. However. with constant weights over time. as well as the specific properties of the data under scrutiny. what happens if we aggregate over the cross-section dimension? Marcellino (2004) argues that cross-section aggregation of countries. Marcellino (2004) fits a variety of non-linear and time-varying models to aggregate EMU macroeconomic variables. which non-linear models should be entertained in any specific instance? This question can be answered by looking at the theoretical properties of the models. The latter paper suggests that the use of non-linearities is crucial for making sensible statements about the tail behavior of asset returns. and evaluates the resulting forecasts using standard MSE-related criteria as well as direction-of-change tests. they find evidence that some of the flexible non-linear regression models perform well relative to the non-linear benchmark. based on in-sample estimation and in-sample data? How should we aggregate and analyze our data? Some of these questions are examined in van Dijk and Franses (2000). The results indicate that EVT based VaR estimates are more accurate at higher quantiles.999 percentile along with 95% confidence intervals for stress testing purposes. Genc ay and Selc uk (2004). According to estimated Generalized Pareto Distribution parameters. / International Journal of Forecasting 20 (2004) 169–183 179 distributional features of the series and to predict higher-order moments than the SETAR models. as the forecast origin moves through the available sample. which are robust to the presence of dynamic misspecification. The paper provides some guidance for model selection. They provide an illustration of model selection via predictive ability testing involving the US money-income relation.

Concluding remarks In this paper we have summarized the state-of-theart in forecast construction and evaluation for nonlinear models. (1997). P. We conclude that the day is still long off when simple.. e. tion. reliable and easy to use non-linear model specification. F. Christoffersen. W. Clements et al. The performance of SETAR models: A regime conditional evaluation of point. Franses & Lucas. P. and new models and tests. Clark. as in periodic models of seasonality (see Franses & Paap. Testing and comparing Value-at-Risk measures. Calculating interval forecasts. F. (1997). Acknowledgements The authors wish to thank Valentina Corradi and Dick van Dijk for helpful conversations. It will be interesting to see to what extent developments in these areas give rise to tangible gains in terms of forecast performance—we remain hopeful that great strides will be made in the near future. B. then as computational capabilities increase. 121 – 135. 561 – 572. 11. F. D. P. Further results on forecasting and model selection under asymmetric loss. .. Boero. International Journal of Forecasting. interval and density forecasts. Testing parametric conditional distributions of dynamic models. 305 – 320. with important issues of non-differentiability. They assess the out-of-sample predictability of various models and find some evidence that non-linear models lead to more accurate short-horizon forecasts. International Economic Review. (1997). & Diebold. 325 – 342. In addition. Suppose one has data with trend components. P. 1097 – 1128. Much remains to be done in the areas of specification.P. Transaction costs and non-linear adjustment towards equilibrium in the US treasury bill market. Chatfield. 11. Forecasting with a nonlinear dynamic model of stock returns and industrial production. M. F. Review of Economics and Statistics. Christoffersen. F. seasonality. parameter-estimation error and data-mining remaining to be addressed. International Journal of Forecasting.. 321 – 342. 465 – 484. A. allowing the possibility that future generations of models will significantly outperform linear models. Journal of Econometrics.g. G. especially of the spread. A conditional Kolmogorov test. Tests of equal forecast accuracy and encompassing for nested models. & Diebold. A. ˜ Clements. 2004 for an up-to-date survey). 20. Christoffersen. P. Evaluating interval forecasts. K. 85 – 110. Swanson gratefully acknowledges financial support from Rutgers University in the form of a Research Council grant. in press. 65. (2004). a further area for research (both empirical and theoretical) is the following. 13. 5.180 M. D. Journal of Applied Econometrics. (2001). And. Optimal prediction under asymmetric loss. Econometric Theory. International Journal of Forecasting. & Marrocu. M. C. E. non-linearity and outliers. but developing coherent strategies for such data remains an important task. T. Journal of Empirical Finance. (2004). Oxford Bulletin of Economics and Statistics. 841 – 862. neglecting outliers may suggest nonlinearity (see van Dijk. The papers in this issue suggest that careful application of existing techniques. 39. W. Hahn. D. A comparison of tests of non-linear cointegration with an application to the predictability of US interest rates using the term structure. X. (2001). Bradley. 1999) and the trend can be intertwined with seasonality. M. (1993). Supposing that the world is inherently non-linear. & Jansen. One finding is that composite leading indicators and interest rates of Germany and the US have substantial ˜ predictive value. (2004). J. especially if such models become truly multivariate. Andrews. 105. 2004. X. / International Journal of Forecasting 20 (2004) 169–183 the regimes are classified as binary variables.. Econometrica. and are grateful to Jan De Gooijer for reading and commenting on the manuscript. and testing. estimation. & Galvao. 8. 808 – 817. 59. M. there are grounds for optimism... (2001). Journal of Business and Economic Statistics. W. Bai. F.. (1996). & Inoue. 219 – 236. as mentioned. due to the possible inter-reactions between these elements. & McCracken. Clements and Galvao (2004) test whether there is non-linearity in the response of short and long-term interest rates to the spread. more complex models become amenable to analysis. 20. and in selection among alternative nonlinear prediction models. How should one proceed? This is a complex issue. (1998). Much has been said about modeling each of these characteristics in isola- References Anderson. E. estimation and forecasting procedures will be readily available. Nevertheless. H. De Gooijer and Vidiella-i-Anguera (2004) report more marked gains to allowing for non-linearities. Christoffersen.. can result in significant advances in our understanding. J.

617 – 637. Queen Mary. & Terasvirta. Granger. (1999). & White. R. J. International Journal of Forecasting. Forecasting 181 threshold cointegrated systems. A comparison of the forecast performance of Markov-switching and threshold autoregressive models of US GNP. The block bootstrap for recursive m-estimators with applications to predictive evaluation. J. W. H. & Tay. Journal of Forecasting. 657 – 684. M. H. University of Exeter and Rutgers University.. manuscript.. (1999). Extreme value theory and Valueß ß at-Risk: Relative performance in emerging markets. Franses. Clements. Journal of Econometrics (In press). W. X.. G. Hamilton. Econometrics Journal. 347 – 366.. 253 – 263. P. P. Working Paper.. S. 228 – 248. X. International Journal of Forecasting. De Gooijer. 185 – 199. T... & Krolzig. A. 18. J. (1989). 353 – 381. 58. J.M. 237 – 253. J. S. 20. P. & Paap. M. 1. Clements. Giacomini. R. J. G.. 13. Business cycle asymmetries: characterisation and testing based on Markov-switching autoregressions. 51 – 56. H. Oxford Bulletin of Economics and Statistics. (1996). K. & Smith. F. 201 – 217. Gunther. P. Dahl. 18. F. J. International Economic Review. New York: Oxford University Press. N. & Hendry. (2001). J. W. P. A test for comparing multiple misspecified conditional distribution models. X. (2004). R. J. (1999). N. V. & Smith. & Satchell. F. C47 – C75. Diebold. J. & Zerom. Journal of Econometrics. Queen Mary. & Selc uk. 185 – 204. G. C. & Swanson. 28. R. 221 – 241. 81. 20. Evaluating density forecasts with applications to finance and management. P. J. Modelling non-linear ¨ economic relationships. -M. Tests of conditional predictive ability. MA: MIT Press [The Zeuthen Lectures on Economic Forecasting]. 17. R. J. (1990). (2002). X. & Vidiella-i-Anguera. & Nason. 8.. Franses. M. 1. P. (1999). (2004). (1993a). 287 – 303. International Journal of Forecasting. C. Journal of Business and Economic Statistics. A. Franses. D. Multi-step estimation for forecasting. Clements. 110. Cambridge: Cambridge University Press. A new approach to the economic analysis of nonstationary time series and the business cycle. D.. R. Journal of International Economics. (2002). Gannoun. 12. T. 357 – 384. & Rubio-Ramirez. A consistent test for out of sample non-linear predictive ability. 27. C.. (2003c).. Diebold. 255 – 276. X. 20.. & Swanson. D. & Hylleberg.. Hamilton. V. Some recent developments in non-linear time series modelling. Corradi. (2002). 863 – 883. & Swanson.. Multivariate density forecast evaluation and calibration in financial risk management: High frequency returns on foreign exchange. testing. (2000). 161 – 173. 651 – 652. J. / International Journal of Forecasting 20 (2004) 169–183 Clements. S. Evaluation of dynamic stochastic general equilibrium models based on a distributional comparison of simulated and historical data. Oil and the macroeconomy since World War II. & Krolzig. M. (1996). & Paap. (2003). Testing structural stability with endogenous breakpoint: A size comparison of analytic and bootstrap procedures.. Evaluating the forecast densities of linear and non-linear models: Applications to output growth and unemployment. J. Some recent developments in predictive accuracy testing with nested models and (generic) non-linear alternatives. 14. Journal of Political Economy. Journal of Applied Econometrics. Dacco. D. Why do regime-switching models forecast so badly. & Chen. (1983). (1996). -M. Corradi. M.. (2000). with an application to US unemployment. This is what happened to the oil price- . M.. (1995). 661 – 673.. International Journal of Forecasting. A. F. P. International Journal of Forecasting.. (1993b). C. Econometrica. P. manuscript. & Smith. R. & van Dijk. (2004). On the limitations of comparing mean squared forecast errors: comment. H. B.. M. 12. Journal of Forecasting. C. Can oil shocks explain asymmetries in the US Business Cycle? Empirical Economics. D. Clements. Fernandez-Villaverde. V. Corradi. (1998). P. V.. Nonparametric exchange rate prediction.. 57. (1993). Statistics and Probability Letters. Clements. J. Outline of forecast theory using generalized cost functions. F. H. International Journal of Forecasting. Periodic time series models. N. (2003b). (2002).. Non-linear time series models in empirical finance. Journal of Applied Econometrics. M. Clements. Granger. S. Paap. 21. Journal of Forecasting. & Swanson. R. N. Clements et al. 39. (2003a). Review of Economics and Statistics. & Mariano. (2004). University of Pennsylvania. 19. 255 – 271. H. (1992). (2004). S. and forecasting. 20. (2003). Comparing density forecasts via weighted Likelihood Ratio tests: Asymptotic and bootstrap methods. (2004). Cambridge. University of London. (1999). 397 – 407. M.. M.. 1 – 16. San Diego: University of California. San Diego: University of California. Hahn. (1998).. 70. F. Giacomini. Journal of Econometrics. Forecasting unemployment using an autoregression with censored latent effects parameters. F. J. Comparing dynamic equilibrium models to data. International Journal of Forecasting. N. F. T. -M. J. P. A. H. Franses. manuscript. R. P. 91. & Tay. Forecasting non-stationary economic time series. & Swanson. Clements.. Censored latent effects autoregression. F. Genc ay. Corradi.. C. & Hendry. & Vroomen.. Working Paper.P. R. & Kumar. (2002). A Monte Carlo study of the forecasting performance of empirical SETAR models.. F. 56.. 196 – 211. Diebold. R. University of London and Rutgers University. Corradi. Journal of Business and Economic Statistics.. H. De Gooijer... 124 – 141.. 315 – 332. Comparing predictive accuracy. Journal of Forecasting. 135 – 156. Diebold. D. Diebold. & Terasvirta. Oxford: Oxford University Press. & Krolzig. 20. P. (2002). On the limitations of ¨ comparing mean squared forecast errors: Comment. D. & Hendry. De Gooijer. D. Spanish Economic Review. Mean squared error properties of the kernel-based multi-stage median predictor for time series. A. Clements. Granger. Hamilton. R.. Evaluating multivariate forecast densities: A comparison of two approaches. Flexible regression models and relative forecast performance. V. R.

& Tsay. Clements et al. & Timmerman. Journal of Business and Economic Statistics. & Anderson. In W. (2001). H. Raymond. D. E. Artis.. S. (2003). P. A recursive modelling approach to predicting UK stock returns. L. (1998). 365 – 381. & White. (1997b). Nonparametric forecasting: A comparison of three kernel-based methods. H. P. M. (1997a). Journal of Monetary Economics. manuscript.. M. 7. (2004). Mork. Cambridge: Cambridge University Press. Harding. Evaluation of out of sample probability density forecasts with applications to S&P 500 stock prices. 52. Asymptotics for out of sample tests of causality. mimeo. UC San Diego. Review of Financial Studies. Leybourne. J. & A. Department of Economics. What is an oil shock? Mimeo. Brown University. 38. 61 – 78). A. Timmerman. T. & White. E. A. A simple nonparametric test of predictive performance. 7. Stekler.. M.182 M. (2001). 740 – 744. (1999). Whatever happened to the oil price-macroeconomy relationship? Journal of Monetary Economics. D. R. 19.. D. Journal of Applied Econometrics. I.. D. Non-linearity. Policy. Southern Methodist University and Ewha University. A model selection approach to assessing the information in the term structure using linear models and artificial neural networks. M. 20. Y. Australian Economic Review. & White.. O. Taylor. G. Louisiana State University. R. Forecasting economic time series using adaptive versus nonadaptive and linear versus nonlinear econometric models. R. & Rich. J. R. F. & White. International Journal of Forecasting. Hamilton. Towards an understanding of some Business Cycle characteristics. H. (2002).. (1989). International Journal of Forecasting. (2001). Hendry. A. W. Statistics and Probability Letters. Credit and Banking. (1997). 281 – 291. H. S. Cornell University. (2000). Kitamura. Pesaran. Threshold cointegration in interest rates. International Journal of Forecasting. Oxford Review of Economic Policy. M.. 249 – 286. 495 – 505. D. 10. P. Maasoumi. 215 – 220. N. Hylleberg. T. (2003). H. Matzner-Løber. & De Gooijer. J. Harvey. 540 – 550. A. 537 – 573. K. Marcellino. A. Sichel. (1994). R. Communications in Statistical Theory Methods. Tests for forecast encompassing. (1997b). & Whang. Koop. Tiao. Wurtz ¨ (Eds. Linton. 195 – 213. A. H. Hansen. M. Oil and the macroeconomy: A Markov state-switching approach. Review of Economics and Statistics. M. (1998).. 13. (1994). 69. 359 – 372. International Journal of Forecasting. Gannoun. H. R. (2001). (1995). & Luttmer. Working Paper. Review of Economics and Statistics. 20. Econometrica. 79. (2004). A model selection approach to real-time macroeconomic forecasting using linear models and artificial neural networks. R. (2001). Assessing specification errors in stochastic discount factor models. G. Kunst. D. A. Krolzig. A. & White. Hansen. 20. (1994). (2000). 439 – 461. Journal of Econometrics. M. G. D. Journal of Monetary Economics. (2000). Discussion Paper 92-26. P. Y. 159 – 191. Volatility forecasting with smooth transition exponential smoothing. An unbiased and powerful test for superior predictive ability. Terasvirta. W. 557 – 590. Pagan. Leybourne. A. structural breaks. H.. G. H. O. R. University of Pennsylvania. Journal of Finance.. (2004). N. R.. Forecasting EMU macroeconomic variables.. UCSD. S. A generalization of the non-parametric Henriksson – Merton test of market timing. 54. G. Macroeconomic forecast evaluation techniques. Heaton. 30. Timmerman. Some advances in non-linear and macroeconomy relationship. 29. A. R. 217 – 227. & Newbold. M. 13. E. & Timmerman. Sullivan. Journal of Business and Economic Statistics.).. Hooker. LSE.. 461 – 465... H. -M. Journal of Finance. C. Tjostheim. Sensier. & Potter. Characterizing non¨ linearities in business cycles using smooth transition autoregressive models.P. Journal of Money. Swanson. Nonparametric estimation of conditional quantiles.. 265 – 279. (1989). S. Forecast evaluation with shared data sets. Barnett. 19 – 33. Journal of Political Economy. Hansen. 237 – 279. 375 – 384. (1992). (1999). Hong. Timmerman. R. Stekler. (1997a). Testing the equality of prediction mean squared errors. Samanta. P. 273 – 286. O. E. N. A parametric approach to flexible nonlinear inference. Hamilton. . Harvey. McCracken. R. 1 – 15.. (1992). manuscript. Economics Letters. International Journal of Forecasting. C. 254 – 259. / International Journal of Forecasting 20 (2004) 169–183 Pagan. (1992). A. Economic Journal. G. 12. 13. D. 27. Pesaran. & Pagan. (1996). 38. E. 1 – 7. Are economic forecasts valuable? Journal of Forecasting.. Osborn. 13. 16. or outliers in economic time series. (1997). 44.. Oil and the macroeconomy when prices go up and down: An extension of Hamilton’s results. A. Terasvirta..... Domestic and international influences on business cycle regimes in Europe. S. (1991). International Journal of Forecasting. J. 407 – 412.. M. 269 – 277. theory and the cycle. D. Inventories and the three phases of the business cycle. G. Y. Non-linear econometric modelling in time series analysis ( pp. Econometric evaluation of asset pricing models. 13. Swanson. Sullivan. & White. G. 49.... H. M. International Journal of Forecasting. (1995). 1647 – 1692. 193 – 213. Consistent testing for stochastic dominance under general sampling schemes. W. A. & Jeganathan. Sullivan. Predicting markov-switching vector autoregressive processes. Dangers of data-driven inference: The case of calendar effects in stock returns. Swanson. & Birchenhall. G. Econometric comparisons of conditional models. G. Data-snooping. & Timmerman. R. technical trading rules and the bootstrap. 7. R. J. & Newbold. (1994). L. Mimeo. 119 – 139. CA: Department of Economics. La Jolla. 1593 – 1617. (1997). J. 97. 8.. H. Journal of Business and Economic Statistics. Journal of Forecasting (In press). Dissecting the cycle: A methodological investigation. J. Pesaran. 343 – 357. (2003).. J. H.

H. Journal of Applied Econometrics. Scandinavian Journal of Statistics. D. He has recently been awarded a National Science Foundation research grant entitled the Award for Young Researchers. 22. A personal overview of non-linear time series analysis from a chaos perspective. (1996). H. Swanson has recent publications in Journal of Development Economics. including the Econometric Society. P. K. Clements. Zheng. Clements et al. 68. (1995). A. and the American Economic Association.. 307 – 333. Franses. a 1994 University of California. He has published in a variety of international journals. 667 – 691. Journal of Forecasting. Blackwells. D. / International Journal of Forecasting 20 (2004) 169–183 adaptive modelling in time-series. Philip Hans FRANSES is Professor of Applied Econometrics and Professor of Marketing Research. M. In W. Econometrica. & Franses.. F. H. 17. 203 – 227). Barnett. & McCracken. W. 64. C. He publishes on his research interests. Journal of Business and Economic Statistics. Journal of Empirical Finance and International Journal of Forecasting. C.. among others. (2002). San Diego Ph. Likelihood ratio tests for model selection and non-nested hypotheses. C. Inference about predictive ability. which are applied econometrics. (2000). (2002). Non-linear econometric modelling in time ¨ series analysis ( pp. CLEMENTS is a Reader in the Department of Economics at the University of Warwick. Hylleberg. A consistent test of conditional parametric distributions. Journal of Business and Economic Statistics. T. (2000). In M. both at the Erasmus University Rotterdam. (1996). Weiss.). Hendry (Eds. (2000). 109 – 131. Hendry. is currently Associate Professor at Rutgers University. West.. Tong.. 183 Biographies: Michael P. Cambridge: Cambridge University Press. Asymptotic inference about predictive ability. and Studies in Non-linear Dynamics and Econometrics. Franses. J. J. 539 – 560. D. and is a member of various professional organizations. J. marketing research and empirical finance.P. Journal of Time Series Analysis. D. 1 – 47. Estimating time series models using the relevant cost function. His research interests include forecasting. Econometrica. financial. 1097 – 1126. (1989). F. A reality check for data snooping. D. 21. W. forecasting.D. Further details about Swanson. 13. S. 399 – 445. A companion to economic forecasting ( pp. are available at his website: http:// econweb. 11. Journal of Econometrics. & Terasvirta. H. He is currently an associate editor of the Journal of Business and Economic Statistics. Vuong. Review of Economics and Statistics. He is also an editor of the International Journal of Forecasting. Hendry) of A Companion to Economic Forecasting 2002. Oxford: Blackwells. Smooth ¨ transition autoregressive models—a survey of recent developments. & D. Journal of the American Statistical Association. Testing for smooth transition non-linearity in the presence of outliers. Econometric Theory. K. & Lucas. the American Statistical Association. the International Journal of Forecasting. Terasvirta. D. . van Dijk. P. West.rutgers. H. Econometric Reviews.). His research interests include time-series modeling and forecasting.and macro-econometrics. Non-linear error correction models for interest rates in The Netherlands. P. J. Wurtz (Eds. van Dijk. 16. T. and co-edited (with David F. 57. P. A. Q. Econometrica. including copies of all papers cited above. 1067 – 1084. & ¨ A. time series. Tjøstheim. has co-authored two books on forecasting. van Dijk. 299 – 321 ).M. X. (1999). 217 – 235. Norman SWANSON.